Scarred Consumption

K.7 Scarred Consumption Malmendier, Ulrike and Leslie Sheng Shen Please cite paper as: Malmendier, Ulrike and Leslie Sheng Shen (2019). Scarred Consumption. International Finance Discussion Papers 1259. https://doi.org/10.17016/IFDP.2019.1259 International Finance Discussion Papers Board of Governors of the Federal Reserve System Number 1259 October 2019

Board of Governors of the Federal Reserve System International Finance Discussion Papers Number 1259 October 2019 Scarred Consumption Ulrike Malmendier and Leslie Sheng Shen NOTE: International Finance Discussion Papers are preliminary materials circulated to stimulate discussion and critical comment. References to International Finance Discussion Papers (other than an acknowledgment that the writer has had access to unpublished material) should be cleared with the author or authors. Recent IFDPs are available on the Web at www.federalreserve.gov/pubs/ifdp/. This paper can be downloaded without charge from the Social Science Research Network electronic library at www.ssrn.com.

Scarred Consumption Ulrike Malmendiera,c,d and Leslie Sheng Shenb Abstract: We show that prior lifetime experiences can “scar” consumers. Consumers who have lived through times of high unemployment exhibit persistent pessimism about their future financial situation and spend significantly less, controlling for the standard life-cycle consumption factors, even though their actual future income is uncorrelated with past experiences. Due to their experience-induced frugality, scarred consumers build up more wealth. We use a stochastic lifecycle model to show that the negative relationship between past experiences and consumption cannot be generated by financial constraints, income scarring, and unemployment scarring, but is consistent with experience-based learning. Keywords: Household consumption, experience effects, expectation, lifecycle model JEL classifications: D12, D15, D83, D91, G51 * a University of California at Berkeley; b Federal Reserve Board; c NBER; d CEPR. We thank John Beshears, David Card, Alex Chinco, Ed Glaeser, Yuriy Gorodnichenko, Pierre-Olivier Gourinchas, Pat Kline, David Laibson, Jonathan Parker, Luigi Pistaferri, Joseph Vavra and seminar participants at Bocconi University, IZA/University of Bonn, MIT, Stanford University, UC Berkeley (Macro, Labor, Finance), University of Minnesota, University of Tilburg, University of Zurich, NBER SI (EFBEM, EFG-PD), the SITE workshop, Conference in Behavioral Finance and Decision Making, the CESifo Area Conference on Behavioral Economics, the Cleveland Fed Household Economics and Decision Making Conference, and 2016 EWEBE Conference in Cologne for helpful comments, and Clint Hamilton, Canyao Liu, and Ian Chin for excellent research assistance. The views in this paper are solely the responsibility of the author(s) and should not be interpreted as reflecting the views of the Board of Governors of the Federal Reserve System or of any other person associated with the Federal Reserve System.

The crisis has left deep scars, which will a↵ect both supply and demand for many years to come. — Blanchard (2012) I Introduction More than a decade after the Great Recession, consumers have been slow to return to prior consumption levels (Petev et al. 2011, De Nardi et al. 2012). As the quote above suggests, the crisis appears to have “scarred” consumers. Consumption has remained low not only in absolute levels, but also relative to the growth of income, net worth, and employment—a pattern that challenges standard life-cycle consumption explanations, such as time-varying financial constraints. For the same reason, low employment due to the loss of worker skills or low private investment, as put forward in the literature on “secular stagnation” and “hysteresis,” cannot account for the empirical pattern either.1 What, then, explains such long-term e↵ects of a macroeconomic crisis on consumption? Our hypothesis starts from the observation in Pistaferri (2016) that the long-lasting crisis e↵ects are accompanied by consumer confidence remaining low for longer periods than standard models imply. We relate this observation to the notion of experience-based learning. We show that consumers’ past lifetime experiences of economic conditions have a long-lasting e↵ect on beliefs and consumption, which is not explained by income, wealth, liquidity, and other life-cycle determinants. Prior research on experience e↵ects has shown that personally experienced stockmarket and inflation realizations receive extra weight when individuals form expectations about future realizations of the same variables.2 Here, we ask whether a similar mechanism is at work when individuals experience high unemployment rates. We apply the linearly declining weights estimated in prior work to both national and 1 TheliteratureonsecularstagnationconjecturedprotractedtimesoflowgrowthaftertheGreat Depression (Hansen 1939). Researchers have applied the concept to explain scarring e↵ects of the Great Recession (Delong and Summers 2012, Summers 2014a, 2014b). Blanchard and Summers (1986) introduce the term “hysteresis e↵ects” to characterize the high and rising unemployment in Europe. Cf.Cerra and Saxena (2008), Reinhart and Rogo↵ (2009), Ball (2014), Haltmaier (2012), and Reifschneider, Wascher, and Wilcox (2015). 2 Theoretical papers on the macro e↵ects of learning-from-experience in OLG models include Ehling, Graniero, and Heyerdahl-Larsen (2018), Malmendier, Pouzo, and Vanasco (2018), Collin- Dufresne, Johannes, and Lochstoer (2016), and Schraeder (2015). The empirical literature starts from Kaustia and Knu¨pfer (2008) and Malmendier and Nagel (2011, 2015). 1

local unemployment rates that individuals have experienced over their lives so far, and to their personal unemployment experiences. We show that past experiences predict both consumer pessimism (beliefs) and consumption scarring (expenditures) as well as several other empirical regularities, including generational di↵erences in consumption patterns, after controlling for wealth, income and other standard determinants. At the same time, past experiences do not predict future income, after including the same controls, and predict, if anything a positive wealth build up. We start by presenting four baseline findings on the relation between past experiences and consumption, beliefs, future income, and wealth build-up. We first document the long-lasting e↵ect of past experiences on consumption. Using the Panel Study of Income Dynamics (PSID) from 1999-2013, we find that past macro and personal unemployment experiences have significant predictive power for consumption, controlling for income, wealth, age fixed e↵ects, a broad range of other demographic controls (including current unemployment), as well as state, year, and even household fixed e↵ects. To the best of our knowledge, our analysis is the first to estimate experience e↵ects also within household, i.e., controlling for any unspecified household characteristics. Without household dummies, the identification comes both from cross-household di↵erences in consumption and unemployment histories, and from how these di↵erences vary over time. With household dummies, the estimation relies solely on within-household variation in consumption in response to lifetime experiences.3 In both cases, the e↵ects are sizable. A one standard deviation increase in the macro-level measure is associated with a 3.3% ($279) decline in annual food consumption, and a 1.6% ($713) decline in total consumption. A one standard deviation increase in personal unemployment experiences is associated with 3.7% ($314) and 2.1% ($937) decreases in annual spending on food and total consumption, respectively. The results are robust to variations in accounting for the spouse’s experience, excluding last year’s experience, or using di↵erent weights, from 3 We have also estimated a model with cohort fixed e↵ects. In that case, the identification controls for cohort-specific di↵erences in consumption. The results are very similar to estimations without cohort fixed e↵ects. Note that, di↵erently from most of the prior literature on experience e↵ects(MalmendierandNagel2011, 2015), theexperiencemeasureisnotabsorbedbycohortfixed e↵ectsastheconsumptiondatasetscontainsubstantialwithin-cohortvariationinexperiences. The unemploymentexperiencemeasureofagivencohortvariesovertimedependingonwherethecohort members have resided over their prior lifetimes. 2

equal to steeper-than-linearly declining.4 Second, we document that consumers’ past experiences significantly a↵ect beliefs. Using the Michigan Survey of Consumers (MSC) from 1953 to 2013, we show that people who have experienced higher unemployment rates over their lifetimes so far have more pessimistic beliefs about their financial situation in the future, and are more likely to believe that it is not a good time to purchase major household items in general. Importantly, these estimations control for income, age, time e↵ects, and a host of demographic and market controls. Third, werelatethesamemeasureoflifetimeunemploymentexperiencestoactual futureincome, uptothreePSIDwaves(sixyears)inthefuture. Again, wecontrolfor currentincome, wealth, demographics, aswell asage, state, year, and evenhousehold fixed e↵ects. We fail to identify any robust relation. In other words, while there is a strong reaction to prior lifetime experiences in terms of beliefs and consumption expenditures, actual future income does not appear to explain these adjustments. Our fourth baseline result captures the wealth implications of consumption scarring. If consumers become more frugal in their spending after negative past experiences, even though they do not earn a reduced income, we would expect their savings and ultimately their wealth to increase. Our fourth finding confirms this prediction in the data. Using a horizon of three to seven PSID waves (6 to 14 years), we find past lifetime experiences predict liquid and illiquid wealth build up, in particular for past personal unemployment experiences. Unobserved wealth e↵ects, the main alternative hypothesis, do not predict wealth build up, or even predict the opposite. These four baseline results—a lasting influence of economic experiences in the past on current expenditure decisions and on consumer optimism, but the lack of any e↵ect on actual future income, plus positive wealth build-up—are consistent with our hypothesis: Consumers over-weigh experiences they made during their prior lifetimes when forming beliefs about future realizations and making consumption choices, as predicted by models of experience-based learning (EBL). Considered jointly, and given the controls included in the econometric models, the results so far already distinguish our hypothesis from several alternative explanations: The inclusion of age controls rules out certain life-cycle e↵ects, such as an increase in 4 We also included lagged consumption in the estimation model to capture habit formation but do not find a significant e↵ect, while the experience proxy remains significant. 3

precautionary motives and risk aversion with age (cf. Caballero 1990, Carroll 1994), or declining income and tighter liquidity constraints during retirement (cf. Deaton 1991, Gourinchas and Parker 2002). The controls for labor market status and demographics account for intertemporal expenditure allocation as in Blundell, Browning, and Meghir (1994) or Attanasio and Browning (1995). The time fixed e↵ects control for common shocks and available information such as the current and past national unemployment rates. The PSID also has the advantage of containing information on wealth, a key variable in consumption models. Moreover, the panel structure of the PSID data allows for the inclusion of household fixed e↵ects and thus to control for time-invariant unobserved heterogeneity. To further distinguish EBL from other determinants that can be embedded in a life-cycle permanent-income model, we simulate the Low, Meghir, and Pistaferri (2010) model of consumption and labor supply and use estimations on the simulated data to illustrate directional di↵erences and other distinctive e↵ects. The Low et al. (2010) model accounts for various types of shocks, including productivity and job arrival, and allows for financial constraints as well as “income scarring”—the notion that job loss may have long-lasting e↵ects on future income because it takes time to obtain an o↵er of the same job-match quality as before unemployment. We further extend the Low et al. (2010) model to allow for “unemployment scarring”—the notion that job loss itself may induce a negative, permanent wage shock.5 We contrast these explanations with EBL by simulating the model for both Bayesian and experience-based learners. First, we utilize the simulations to show that, even with all of the life-cycle determinants and frictions built into the Low et al. model, it is hard to generate a negative correlation between past unemployment experiences and consumption when consumers are rational, after controlling for income and wealth. This holds both when we allow for financial constraints and income scarring, as in Low et al., and when we further add unemployment scarring. In fact, given the income control, the simulate-and-estimate exercise often predicts a positive relation between unemployment experiences and consumption. Intuitively, a consumer who has the same income as another consumer despite worse unemployment experiences likely has a higher permanent income component, and rationally consumes more. 5 WethanktheaudienceattheUniversityofMinnesotamacroseminarforthisusefulsuggestion. 4

We then turn to consumers who overweight their own past experiences when forming beliefs. Here, we find the opposite e↵ect: Higher life-time unemployment experiences predict lower consumption among EBL agents, controlling for income and wealth. Thus, the simulate-and-estimate exercise disentangles EBL from potential confounds such as financial constraints, income scarring, and unemployment scarring. There is a robust negative relation between past experiences and consumption under EBL, consistent with the empirical estimates, but not under Bayesian learning. Moreover, Bayesian learning is inconsistent with the estimated relation between past experiences and downward biased beliefs. The model also helps to alleviate concerns about imperfect wealth controls. We conduct both simulate-and-estimate exercises leaving out the wealth control in the estimation. In the case of rational consumers we continue to estimate a positive rather than negative relationship between past experiences and consumption; in the case of experience-base learners, we continue to estimate a negative relationship. Guided by these simulation results, we perform three more empirical steps: (1) a broad range of robustness checks and replications using variations in the wealth, liquidity, and income controls, and using di↵erent data sets; (2) a study of the implications of EBL for the quality of consumption and of the heterogeneity in consumption patterns across cohorts, and (3) a discussion of the potential aggregate e↵ects of EBL for consumption and savings. First, we replicate the PSID results using four variants of wealth controls: thirdand fourth-order liquid and illiquid wealth; decile dummies of liquid and illiquid wealth; separate controls for housing and other wealth; and controls for positive wealth and debt. Similarly, we check the robustness to four variants of the income controls: third- and fourth-order income and lagged income; quintile dummies of income and lagged income; decile dummies of income and lagged income; and five separatedummiesfortwo-percentilestepsinthebottomandinthetop10%ofincome and lagged income. All variants are included in addition to first- and second-order liquid and illiquid wealth and first- and second-order income and lagged income, and all estimations are replicated both with and without household fixed e↵ects. We also subsample households with low versus high liquid wealth (relative to the sample median in a given year), and find experience e↵ects in both subsamples.6 6 Our variants of wealth and income controls also address the concern that consumption may 5

As another, out-of-sample corroboration of our results, we replicate the PSID results in two additional data sets, the Consumer Expenditure Survey (CEX) and the Nielsen Homescan Data. The CEX contains a more comprehensive list of product categories, and sheds light on the impact of unemployment experience on durable and total consumption. The Nielsen data is a panel of consumption purchases by representative U.S. households. It contains detailed data on the products that households purchase at the Universal Product Code (UPC) level for each shopping trip, which allows us to control more finely for time (year-month) e↵ects. The estimated magnitudes in the Nielsen and CEX data are very similar to those in the PSID.7 Next, we exploit the richness of the Nielsen data to show that prior experiences a↵ect consumption also at the qualitative margin. We estimate a significant increase in several measures of frugality: (i) the use of coupons, (ii) the purchase of lowerquality items (as ranked by their unit price, within product module, market, and month), and (iii) the purchases of on-sale products. For example, households buy 9% more sale items at the 90th than at the 10th percentile of unemployment experiences. We then test a unique prediction of EBL: Since a given macroeconomic shock makes up a larger fraction of the lifetime experiences of younger than older people, macroeconomic shocks should have particularly strong e↵ects on younger cohorts. That is, the EBL model predicts that younger cohorts increase their consumption more than older cohorts during economic booms, and lower their spending more during busts. We confirm the prediction for both aggregate and personal unemployment experiences, and in both the positive and in the negative direction. Overall, our results on the lasting e↵ects of past experiences on consumption suggest that experience e↵ects constitute a novel micro-foundation of fluctuations in aggregate demand and long-run e↵ects of macro shocks. We provide suggestive evidence of this implication on the aggregate level by correlating aggregate lifetime experiences of past national unemployment among the U.S. population with real personal consumption expenditure (PCE) from the U.S. Bureau of Economic Analysis (BEA) from 1965 to 2013. The resulting plot shows that times of higher aggregate be a non-linear function of assets and earnings (Arellano, Blundell, and Bonhomme 2017). 7 We have also explored the Health and Retirement Survey (HRS), which contains information onconsumption(fromtheConsumptionandActivitiesMailSurvey)andwealthonabiennialbasis since 2001. However, given that cross-cohort variation is central to our identification, the lack of cohorts below 50 makes the HRS is not suitable for the analysis. 6

past-unemployment experience in the population coincide with lower aggregate consumer spending. This suggests that changes in aggregate consumption may reflect not only responses to recent labor-market adjustments, but also changes in belief formation due to personal lifetime experiences of economic shocks. Overall, our findings imply that the long-term consequences of macroeconomic fluctuations can be significant, thus calling for more discussion on optimal monetary and fiscal stabilization policy to control unemployment and inflation (Woodford 2003, 2010). Related Literature Our work connects several strands of literature. First and foremost, the paper contributes to a rich literature on consumption. Since the seminal work of Modigliani and Brumberg (1954) and Friedman (1957), the life-cycle permanent-income model has been the workhorse to study consumption behavior. Consumption decisions are an intertemporal allocation problem, and agents smooth marginal utility of wealth across predictable income changes over their life-cycle. Subsequent variants provide more rigorous treatments of uncertainty, time-separability, and the curvature of the utility function (see Deaton (1992) and Attanasio (1999) for overviews). A number of empirical findings, however, remain hard to reconcile with the model predictions. Campbell and Deaton (1989) point out that consumption does not react suciently to unanticipated innovation to the permanent component of income (excess smoothness). Instead, consumption responds to anticipated income increases, over and above what is implied by standard models of consumption smoothing (excess sensitivity; cf. West 1989, Flavin 1993). The empirical puzzles have given rise to a debate about additional determinants of consumption, ranging from traditional explanations such as liquidity constraints (Gourinchas and Parker 2002; see also Kaplan, Violante, and Weidner 2014; Deaton 1991; Aguiar and Hurst 2015) to behavioral approaches such as hyperbolic discounting (Harris and Laibson 2001), expectations-based reference dependence (Pagel 2017; Olafsson and Pagel 2018), and myopia (Gabaix and Laibson 2017).8 Experiencebased learning o↵ers a unifying explanation for both puzzles. The lasting impact of lifetime income histories can explain both consumers’ lack of response to permanent shocks and their overreaction to anticipated changes. Our approach is complementary to the existing life-cycle literature: Experience 8 See also Dynan (2000) and Fuhrer (2000) on habit formation. 7

e↵ects describe consumption after taking into account the established features of the life-cycle framework. EBL can explain why two individuals with similar income profiles, demographics, and household compositions make di↵erent consumption choices if they lived through di↵erent macroeconomic or personal employment histories. Our predictions and findings are somewhat reminiscent of consumption models with intertemporal non-separability, such as habit formation models (Meghir and Weber 1996, Dynan 2000, Fuhrer 2000). In both cases, current consumption predicts long-term e↵ects. However, the channel is distinct. Under habit formation, utility is directly linked to past consumption, and households su↵er a loss of utility if they do not attain their habitual consumption level. Under EBL, households adjust consumption patterns based on inferences they draw from their past experiences, without direct implications for utility gains or losses. Related research provides evidence on the quality margin of consumption. Nevo and Wong (2015) show that U.S. households lowered their expenditure during the Great Recession by increasing coupon usage, shopping at discount stores, and purchasing more goods on sale, larger sizes, and generic brands. While they explain this behavior with the decrease in households’ opportunity cost of time, we argue that experience e↵ects are also at work. The key element to identifying this additional, experience-based source of consumption adjustment are the inter-cohort di↵erences andthedi↵erencesinthosedi↵erencesovertime. Relatedly,Coibion,Gorodnichenko, and Hong (2015) show that consumers store-switch to reallocate expenditures toward lower-end retailers when economic conditions worsen. The second strand of literature is research on experience e↵ects. A growing literature in macro-finance, labor, and political economy documents that lifetime exposure to macroeconomic, cultural, or political environments strongly a↵ects their economic choices, attitudes, and beliefs. This line of work is motivated by the psychology literature on the availability heuristic and recency bias (Kahneman and Tversky 1974, Tversky and Kahneman 1974). The availability heuristic refers to peoples’ tendency to estimate event likelihoods by the ease with which past occurrences come to mind, with recency bias assigning particular weight to the most recent events. Taking these insights to the data, Malmendier and Nagel (2011) show that lifetime stock-market experiences predict subsequent risk taking in the stock market, and bond-market experiences explain risk taking in the bond market. Malmendier and Nagel (2015) 8

show that lifetime inflation experiences predict subjective inflation expectations. Evidence in line with experience e↵ects is also found in college students who graduate into recessions (Kahn 2010, Oreopoulos, von Wachter, and Heisz 2012), retail investors and mutual fund managers who experienced the stock-market boom of the 1990s (Vissing-Jorgensen 2003, Greenwood and Nagel 2009), and CEOs who grew up in the Great Depression (Malmendier and Tate 2005, Malmendier, Tate, and Yan 2011). In the political realm, Alesina and Fuchs-Schu¨ndeln (2007), Lichter, Lo¨✏er, and Siegloch (2016), Fuchs-Schu¨ndeln and Schu¨ndeln (2015), and Laudenbach et al. (2018) reveal the long-term consequences of living under communism, its surveillance system, and propaganda on preferences, norms, and financial risk-taking. Our findings on experience e↵ects in consumption point to the relevance of EBL in a new context and reveal a novel link between consumption, life-cycle, and the state of the economy. A novelty of our empirical analysis, compared to the existing literature, is that the detailed panel data allow us to identify e↵ects using withinhousehold variation, whereas earlier works such as Malmendier and Nagel (2011, 2015) rely solely on time variation in cross-sectional di↵erences between cohorts. In the rest of the paper, we first present the data (Section II), followed by the four baseline findings on consumption, beliefs, future income, and wealth build-up (Section III). The stochastic life-cycle model in Section IV illustrates the di↵erences between the consumption of rational and experience-based learners. Guided by the simulation results, we present additional wealth and income robustness tests in Section V, and replicate the results in the CEX and Nielsen data. Section VI shows further results on the quality margins of consumption and the cross-cohort heterogeneity in responses to shocks. Section VII discusses the aggregate implications of experience-based learning for consumer spending and concludes. II Data and Variable Construction II.A Measure of Consumption Scarring Our conjecture is that individuals who have lived through dicult economic times have more pessimistic beliefs about future job loss and income, and thus spend less than other consumers with the same income, wealth, employment situation, and 9

Figure 1: Monthly Consumption Expenditure by Age Group )$( erutidnepxE naeM morf noitaiveD 04 02 0 02- 04- 2004m1 2006m1 2008m1 2010m1 2012m1 Age<40 Age 40 to 60 Age>60 Notes. Six-monthmovingaveragesofmonthlyconsumptionexpendituresofyoung(below40),midaged(between40and60),andoldindividuals(above60)intheNielsenHomescanPanel,expressed asdeviationsfromthecross-sectionalmeanexpenditureintherespectivemonth,anddeflatedusing the personal consumption expenditure (PCE) price index of the U.S. Bureau of Economic Analysis (BEA). Observations are weighted with Nielsen sample weights. other demographics. The opposite holds for extended exposure to prosperous times. Consumers who have mostly lived through good times in the past will tend to spend more than others with the same income, wealth, and demographics. Moreover, the experience-based learning model has a second implication: Younger cohorts react more strongly to a shock than older cohorts since it makes up a larger fraction of their life histories so far. As a result, the cross-sectional di↵erences vary over time as households accumulate di↵erent histories of experiences. The raw time-series of household expenditures (from the Nielsen data) in Figure 1 illustrates the hypothesized e↵ects. Expenditures are expressed as deviations from the cross-sectional mean in each month. In general, the spending of younger cohorts (below40)ismorevolatilethanthatofoldercohorts, consistentwithyoungercohorts 10

exhibiting greater sensitivity. Zooming in on the Great Recession period, we also see that the spending of younger cohorts was significantly more negatively a↵ected than those of the other age groups. Such patterns are consistent with consumers being scarred by recession experiences, and more so the younger they are. To formally test the experience-e↵ect hypothesis, we construct measures of past experiences that apply the weighting function estimated in prior work to the experience of high and low unemployment rates. We focus on experiences of unemployment rates following Coibion, Gorodnichenko, and Hong (2015), who single out unemployment as the most spending-relevant variable. We construct measures of past experiences on both the macro (national and local) level and the personal level. The macro measure captures the experience of living through various spells of unemployment rates. The personal measure captures personal situations experienced so far. Specifically, unemployment experience accumulated by time t is measured as t 1 E = w(,t,k)W , (1) t t k k=0 X where W is the unemployment experience in year t k, and k denotes the time t k lag.9 Weights w are a function of t, k, and , where is a shape parameter for the weighting function. Following Malmendier and Nagel (2011), we parametrize w as (t k) w(,t,k) = . (2) t 1 (t k) k=0 P This specification of experience weights is parsimonious in that it introduces only one parameter, , to capture di↵erent possible weighting schemes for past experiences. It simultaneously accounts for all experiences accumulated during an individual’s lifetime and, for < 0, allows for experience e↵ects to decay over time, e.g., as memory fades or structural change renders old experiences less relevant. That is, for 9 In the empirical implementation, we utilize unemployment information from birth up to year t 1 while the theoretical p is constructed based on realizations of W for k = 0,...,t 1, i.e., t t k from the moment of birth to the realization at the beginning of the current period. It is somewhat ambiguouswhatcorrespondsbesttothetheoreticalset-up,especiallyas,inpractice,onlybackward looking (macro) information becomes available to every individual. However, since we do control for(macroeconomicandpersonal)contemporaneousunemploymentinallregressions, theinclusion orexclusionofmacroorpersonalunemploymentattimetintheexperiencemeasuredoesnotmake a di↵erence to the estimation results. 11

> 0, the weighting scheme emphasizes individuals’ recent experiences, letting them carryhigherweights, whilestillallowingforanimpactofearlierlifehistories. As ! , it converges towards the strongest form of recency bias. In our main empirical 1 analyses, we will apply linearly declining weights ( = 1), which approximate the weights estimated in Malmendier and Nagel (2011, 2015). For robustness, we also conduct the analysis using = 3. With this range of parameters, we capture that, say, in the early 1980s, when the national unemployment rate exceeded 10%, a then 30-year-old was still a↵ected by the experience of living through low unemployment in the early 1970s (around 5-6%) as a 20-year-old, but that this influence was likely smaller than more recent experiences. Empirically, we construct national, local, and individual measures of unemployment experiences, depending on the data set and individual information available. For national unemployment rates, we combine several historical time series: a) the data from Romer (1986) for the period 1890-1930; b) data from Coen (1973) for the period 1930-1939; c) the BLS series that counts persons aged 14 and over in the civilian labor force for the period 1940-1946; and d) the BLS series that counts persons aged 16 and over in the civilian labor force for the period 1947-present.10 For the more local, region-specific measure of unemployment experiences, we combine information on where a family has been living (since the birth year of the householdhead)withinformationaboutlocalhistoricalunemploymentrates. Ideally, both sets of information would be available since the birth year of the oldest generation in our data. However, the Bureau of Labor Statistics (BLS) provides state-level unemployment rates only since 1976, and there do not appear to be reliable sources of earlier historical unemployment data for all US states.11 These data limitations 10 Analternative,widelycitedsourceof1890-1940dataisLebergott(1957,1964). Laterresearch has identified multiple issues in Lebergott’s calculations and has sought to modify the estimates to better match the modern BLS series. Romer (1986) singles out two of Lebergott’s assumptions as invalid and generating an excessively volatile time series: (1) that employment and output move one-to-one in some sectors, and (2) that the labor force does not vary with the business cycle. Coen (1973) finds that both armed forces and cyclical variations in average hours/worker have been ignored in previous studies, and these variables appear to have significant e↵ects on measures of labor participation. 11 The state-level BLS rates are model-based estimates, controlled in “real time” to sum to national monthly (un)employment estimates from the Current Population Survey (CPS). While it is possible to construct estimates of state-level unemployment using the pre-1976 CPS, we do not do so to avoid inconsistencies and measurement errors. 12

imply that, if we were to work with “all available” data to construct region-specific measures, the values for family units from the later periods would be systematically more precise than those constructed for earlier periods, biasing the estimates. Hence, we have to trade o↵ restricting the sample such that all family units in a given data set have sucient location and employment-rate data, and ensuring sucient history to construct a reliable experience measure. We choose to use the five most recent years state-level unemployment rates, t 5 to t 1, either by themselves or combined with national unemployment rate data from birth to year t 6. In the former case, we weight past experiences as specified in (2) for k = 1,...,5, and then renormalized the weights to 1. In the latter case, we use weights exactly as delineated in (2). As we will see, the estimation results are very similar under all three macro measures, national, regional, and combined. We will show the combined measure in our main regressions whenever geographic information on the individual level is available. For the personal experience measure, we use the reported employment status of the respondent in the respective data set. We face the same data limitations as in the construction of the state-level macro measure regarding the earlier years in the lives of older cohorts. Mirroring our approach in constructing the local macro measure, we use the personal-experience indicator variables from year t 5 to t 1 and national unemployment rates from birth to t 6, with weights calculated as specified in (2). II.B Consumption Data Our main source of data is the PSID. It contains comprehensive household-level data on consumption and has long time-series coverage, which allows us to construct experience measures for each household. We will later replicate the results in Nielsen and CEX data. Compared to those data, the PSID has the advantage of containing rich information on household wealth, a key variable in consumption models. The PSID started its original survey in 1968 on a sample of 4,802 family units. Along with their split-o↵ families, these families were surveyed each year until 1997, when the PSID became biennial. We focus on data since 1999 when the PSID started to cover more consumption items (in addition to food), as well as information on household wealth. The additional consumption variables include spending on childcare, clothing, education, health care, transportation, and housing, and approx- 13

imately 70% of the items in the CEX survey (cf. Andreski et al. 2014). Regarding household wealth, the survey asks about checking and saving balances, home equity, and stock holdings. Those variables allow us to control for consumption responses to wealth shocks, and to tease out the e↵ects of experiences on consumption for different wealth groups. Indeed, compared to the Survey of Consumer Finances (SCF), which is often regarded as the gold standard for survey data on wealth, Pfe↵er et al. (2016) assess the quality of the wealth variables in the PSID to be quite similar. The exceptions are “business assets” and “other assets,” for which the PSID tends to have lower values. We construct separate controls for liquid and illiquid wealth, using the definitions of Kaplan, Violante, and Weidner (2014). Liquid wealth includes checking and savings accounts, money market funds, certificates of deposit, savings bonds, treasury bills, stock in public companies, mutual funds, and investment trusts. Illiquid wealth includes private annuities, IRAs, investments in trusts or estates, bond funds, and life insurance policies as well as the net values of home equity, other real estate, and vehicles. The PSID also records income and a range of other demographics, including years of education (ranging from 0 to 17), age, gender, race (White, African American, or Other), marital status, and family size. The information is significantly more complete for the head of household than other family members. Hence, while the family is our unit of analysis, our baseline estimations focus on the experiences and demographics of the heads, including our key explanatory variable of unemployment experiences. We then show the robustness to including the spouse’s experiences. Thekeyexplanatoryvariableisthepastexperienceofeachhouseholdheadateach point in time, calculated as the weighted average of past unemployment experiences as defined in (1) and (2). The PSID allows us to construct both macroeconomic and personal experience measures, and to further use both national and state-level rates for the macro measure. As discussed above, the more local measure has to account for several data limitations. The oldest heads of household in the survey waves we employ are born in the 1920s, but the PSID provides information about the region (state) where a family resides only since the start of the PSID in 1968, and the Bureau of Labor Statistics (BLS) provides state-level unemployment rates only since 1976. As specified above, we use the five most recent years state-level unemployment rates, t 5 to t 1, either by themselves or combined with national unemployment 14

rate data from birth to year t 6. We will show the combined measure in our main regressions; the results for (pure) national and regional measures are very similar. To measure personal experiences, we first create a set of dummy variables indicating whether the respondent is unemployed at the time of each survey.12 We employ the same approach as with the state-level data regarding the early-years data limitations. Figure 2: Unemployment Experience by Age Group and by Region Notes. Thegraphsshowtheunweightedmeansoflocalunemploymentexperiences,intheleftpanel for di↵erent age groups, and in the right panel for di↵erent regions. Figure 2 illustrates the heterogeneity in lifetime experiences, both in the crosssection and over time, for our PSID sample using the (combined) macro measure. The left panel plots the unweighted mean experiences of young (below 40), middleaged (between 40 and 60), and old individuals (above 60), while the right panel plots the measures for individuals in the Northeast, North Central, South, and West. The plots highlight the three margins of variation that are central to our identification 12 The PSID reports eight categories of employment status: working now, only temporarily laid o↵, looking for work, unemployed, retired, permanently disabled, housewife; keeping houses, student, and other. We treat other as missing, looking for work, unemployed as “unemployed,” and all other categories as “not unemployed.” One caveat is that the PSID is biennial during our sample period. For all gap years t, we assume that the families stay in the same state and have the same employment status as in year t 1. Alternatively, we average the values of t 1 and t+1, shown in Appendix A. 15

strategy: At a given point in time, people di↵er in their prior experiences depending on their cohort and location, and these di↵erences evolve over time. Table 1: Summary Statistics (PSID) Variable Mean SD p10 p50 p90 N Age 47.61 12.06 32 47 65 33,164 Experience (Macro) [in %] 6.00 0.28 5.67 5.97 6.36 33,164 Experience (Personal) [in %] 4.55 14.27 0.00 0.00 18.92 33,164 Household Size 2.75 1.45 1 2 5 33,164 Household Food Consumption [in $] 8,452 5,153 2,931 7,608 14,999 33,164 Household Total Consumption [in $] 44,692 31,786 16,626 39,608 76,823 33,164 Household Total Income [in $] 80k 51k 22k 70k 155k 33,164 Household Liquid Wealth [in $] 38k 320k -23k 0k 91k 33,164 Household Illiquid Wealth [in $] 222k 919k 1k 71k 513k 33,164 Household Total Wealth [in $] 260k 1,007k -3k 72k 636k 33,164 Notes. Summary statistics for the estimation sample, which covers the 1999-2013 PSID waves and exclude observations with a total income below the 5th or above the 95th percentile in each sample wave, as well as inthepre-sample1997wave(sincewecontrolforlaggedincome). Age,Experience(Macro),andExperience (Personal)arecalculatedfortheheadsofhouseholds. HouseholdTotalIncomeincludestransfersandtaxable incomeofallhouseholdmembersfromthelastyear. LiquidandilliquidwealtharedefinedfollowingKaplan, Violante and Weidner (2014). Values are in 2013 dollars (using the PCE), annual, and not weighted. Table 1 shows the summary statistics for our sample. We focus on household heads from age 25 to 75.13 In the main analysis, we run the regressions excluding observations with total family income below the 5th or above the 95th percentile in each wave. The sample truncation addresses known measurement errors in the income variable.14 After dropping the individuals for whom we cannot construct the experience measures (due to missing information about location or employment status in any year from t to t 5), and observations with missing demographic controls or that only appear once, we have 33,164 observations. The mean macro experience is 6.0%, and the mean personal experience is 4.6%. Average household 13 Controlling for lagged income, the actual minimum age becomes 27. We also conduct the analysis on a subsample that excludes retirees (households over age 65) since they likely earn a fixed income, which should not be a↵ected by beliefs about future economic fluctuations. The results are similar. 14 GouskovaandSchoeni(2007)evaluatethequalityofthefamilyincomevariableinthePSIDby comparingittofamilyincomereportedintheCPS.Theincomedistributionsfromthetwosurveys closely match between the 5th and 95th percentiles, but there is less consensus in the upper and lower five percentiles. As a robustness check, we use the full sample, cf.Appendix-Table A.1. 16

food consumption and average household total consumption in our sample are $8,452 and $44,692, respectively (in 2013 dollars). III Baseline Results Our analysis starts from the observation that macro shocks appear to have a longlasting impact on consumer behavior and that the puzzling persistence of reduced consumer expenditures correlates with consumer confidence remaining low for longer than standard models would suggest (Pistaferri 2016). We test whether we can better predict consumer confidence and consumer behavior if we allow for a role of consumers’ prior experiences of economic conditions. Prior lifetime experiences have been found to have long lasting e↵ects on individual beliefs and decision-making in the realms of stock returns, bond returns, inflation, and mortgage choices. Here, we ask whether a similar mechanism might help to explain patterns in consumption expenditures. Specifically, we measure past experiences of spending-relevant macro conditionsintermsofhigherorlowerunemploymentratesasinCoibionetal. (2015), both on the aggregate level (unemployment rates) and on the personal level. We then show that past experiences of unemployment have a measurable, lasting e↵ect on individual beliefs and consumption expenditures, but fail to predict (lower) future income or future wealth. III.A Past Experiences and Consumption We relate expenditures to prior experiences of economic conditions by estimating C = ↵+UE + UEP +0x +⌘ +& + +" , (3) it it it it t s i it where C is consumption, UE is i’s macroeconomic and UEP her personal unemit it it ployment experience over her prior life, x is a vector of controls including wealth it (first- and second-order logarithm of liquid and illiquid wealth), income (first- and second-order logarithm of income and lagged income), age dummies, household characteristics (dummy indicating if the household head is currently unemployed, family size, gender, years of education (ranging from 0 to 17), marital status, and race (White, African American, Other)); ⌘ are time (year) dummies, & state dummies, t s 17

and household dummies.15 We conduct our empirical analysis both with food i consumption, following the earlier literature, and with total consumption as dependent variables.16 Standard errors are clustered at the cohort level. All the regression results are quantitatively and qualitatively similar when clustered by household, household-time, and cohort-time or two-way clustered at the cohort and time level. Our main coecients of interest are and . The rational null hypothesis is that bothcoecientsarezero. Thealternativehypothesis, basedontheideaofexperience e↵ects, is that consumers who have experienced higher unemployment spend less on average and, hence, that both coecients are negative. Identification. Weestimatethemodelbothwithandwithouthouseholddummies. In the former case, we identify experience e↵ects solely from time variation in the within-household co-movement of consumption and unemployment histories. In the latter case, identification also comes from time variation in cross-sectional di↵erences in consumption and unemployment histories between households. We illustrate the sources of identification with a simple example of the unemployment experiences and household consumption of three individuals in our PSID data over the course of the Great Recession. Consider two individuals (A and B) who have the same age (born in 1948) but live in di↵erent states (Pennsylvania and Alabama) during the 2007-2013 period and a third (C) who lives in the same state as B (Alabama) but di↵ers in age (born in 1975). The two sets of bars in Figure 3 illustrate their lifetime experiences of unemployment at the beginning and at the end of the 2007-2013 period, based on the 15 We have also included region year fixed e↵ects, and the results remain very similar. One may ⇤ consider fully saturating the model with state year fixed e↵ects, to control for unspecified deter- ⇤ minants of consumption that a↵ect consumers di↵erently over time and by state. (Note that those alternative determinants would need a↵ect consumption exactly in the direction of the experiencee↵ecthypothesis,includingthedi↵erente↵ectsexperienceshaveonyoungerandolderpeople.) Since one of the key margins of variation in macroeconomic unemployment experience (UE ) is at the it state yearlevel,state yearfixede↵ectswouldabsorbmuchofthevariation,resultingininsucient ⇤ ⇤ statisticalpowertopreciselyestimatecoecients. Instead,wehaveestimatedthemodelcontrolling for current state-level unemployment rates as a sucient statistic. The results are similar. 16 Food consumption had been widely used in the consumption literature largely because food spending used to be the only available consumption variable in the PSID before 1999. We are separatingouttheresultsonfoodconsumptionpost-1999partlyforcomparison,butalsoincasethe data is more accurate as some researchers have argued. Food consumption and total consumption come directly from the PSID Consumption Data Package 1999-2013. 18

Figure 3: Examples of Experience Shocks from the Recession (PSID) 6.5 3000 6 2500 5.5 2000 5 1500 4.5 1000 4 500 2007 2013 Born 1948; 2007-2013 in Pennsylvania 1 2 Born 1948; 2007-2013 in Alabama Born 1975; 2007-2013 in Alabama )srab ,% ni( tnemyolpmenU fo ecneirepxEemitefiL )senil ,$ ni( rebmeM ylimaF reP erutidnepxE 6.20% 6.11% 6.06% 5.81% 5.70% 5.46% -13% $1877 $1677 $1632 -7% $1568 $777 -21% $613 A B C A B C Notes. Thered(dark)barsdepictthe2007and2013unemploymentexperiencesofpersonA,andthe red(dark)linethecorrespondingchangeoftotalconsumptionpermemberofA’sfamily. Similarly, theblue(mediumdark)barsandlineshowpersonB’sunemploymentexperiencesandconsumption, and the green (light) bars and line person C’s unemployment experiences and consumption. All consumption expenditures are measured in 2013 dollars, adjusted using PCE. Person A’s ID in the PSID is 45249; person B’s ID in the PSID is 53472; person C’s ID in the PSID is 54014. weighting scheme in (2) and their states of residence. Person A enters the crisis period with a higher macroeconomic unemployment experience than Person B (5.81% versus5.70%), butherlifetimeexperienceworsenslessoverthecourseofthefinancial crises and becomes relatively more favorable by 2013 (6.06% versus 6.11%) because unemployment rates were lower in Pennsylvania than in Alabama during the crisis period. Person C has even lower macroeconomic unemployment experiences before the crisis period than Person B (5.46%), but, being the younger person, C is more affected by the crisis which leads to a reversal of the lifetime unemployment experience between the old and the young by the end of the crisis (6.11% versus 6.20%). Figure 3 relates these di↵erences-in-di↵erences of lifetime experience over the crisis period to consumption behavior. The increase in unemployment experiences of Person A, 19

B, and C by 0.25%, 0.41%, and 0.74%, respectively, were accompanied by decreases in consumption in the same relative ordering, by 7%, 13%, and 21%, respectively. Results Table 2 shows the estimation results from model (3) with (log) food consumption as the dependent variable in the upper panel and with (log) total consumption in the lower panel. All regressions control for first- and second-order (logs of) income, lag income, liquid wealth, illiquid wealth, and for all other control variables listed above as well as the fixed e↵ects indicated at the bottom of the table. Columns (1)-(3) show results without household fixed e↵ects, and columns (4)-(6) with household fixed e↵ects. All estimated coecients on the control variables (not shown) have the expected sign, consistent with prior literature. The estimated negative coecients indicate that both macroeconomic and personal unemployment experiences predict reduced consumption expenditures in the long-run. In the estimations predicting food consumption, shown in the upper half of the table, we find a significantly negative e↵ect of both macroeconomic and personal experiences, controlling for the current unemployment status. The economic magnitudes remain the same whether we include the two types of experience measures separately or jointly, though the statistical significance of the macro measure diminishes somewhat in the specifications without household fixed e↵ects (columns 1-3) when we include both measures jointly (column 3). When we introduce household fixed e↵ects (in columns 4-6), the estimated coecient on macro experience becomes larger and more precisely estimated. Based on the column (6) estimates, a one standard-deviation increase in macroeconomic unemployment experience leads to a 3.3% decrease in food consumption, which translates to $279 less annual spending. Hence, the economic magnitude of the macro experience e↵ect alone is large, particularly considering that the estimates reflect behavioral change due to fluctuation in the macro-economy, not direct income shocks. As expected, the estimated personal experience e↵ects become slightly smaller when we include household fixed e↵ects. The decrease reflects that experience effects (also) predict cross-sectional di↵erences in consumption between households with “mostly good” versus “mostly bad” lifetime experiences, and this component of experience e↵ects is now di↵erenced out. Nevertheless, the e↵ect of personal experience is more than two times larger than macroeconomic experience in absolute value. 20

Table 2: Experience E↵ects and Annual Consumption (PSID) (1) (2) (3) (4) (5) (6) Dependent Variable: Food Consumption Experience (Macro) -0.097** -0.091* -0.120** -0.117** (0.047) (0.047) (0.054) (0.055) Experience (Personal) -0.322*** -0.320*** -0.263** -0.260** (0.097) (0.097) (0.119) (0.119) R-squared 0.192 0.193 0.193 0.541 0.542 0.542 Dependent Variable: Total Consumption Experience (Macro) -0.022 -0.018 -0.059*** -0.057*** (0.019) (0.019) (0.021) (0.021) Experience (Personal) -0.178*** -0.177*** -0.148*** -0.147*** (0.030) (0.030) (0.031) (0.031) R-squared 0.573 0.574 0.574 0.788 0.788 0.788 Demographic controls Yes Yes Yes Yes Yes Yes Income controls Yes Yes Yes Yes Yes Yes Wealth controls Yes Yes Yes Yes Yes Yes Age fixed e↵ects Yes Yes Yes Yes Yes Yes State fixed e↵ects Yes Yes Yes Yes Yes Yes Year fixed e↵ects Yes Yes Yes Yes Yes Yes Household fixed e↵ects No No No Yes Yes Yes Observations 33,164 33,164 33,164 33,164 33,164 33,164 Notes. Theconsumptionvariablescomefromthe1999-2013PSIDConsumptionExpenditureDatapackage. We take the logarithm of consumption, income, and wealth; non-positive values are adjusted by adding the absolute value of the minimum plus 0.1 before being logarithmized. “Experience (Macro)” is the macroeconomicexperiencemeasureofunemployment,and“Experience(Personal)”isthepersonalexperiencemeasure, as defined in the text. Demographic controls include family size, heads’ gender, race, marital status, education level, and a dummy variable indicating whether the respondent is unemployed at the time of the survey. Income controls include the first and second order of the logarithm of income and lagged income. Wealth controls include the first and second order of the logarithm of liquid and illiquid wealth. We exclude from the sample observations with total family income below the 5th or above the 95th percentile in each wave Robust standard errors (in parentheses) are clustered by cohort. *, **, *** denote 10%, 5%, and 1% significance, respectively. 21

The estimated e↵ect of a one standard-deviation increase in personal unemployment experiences is similar to that of macro experiences. It predicts a 3.7% decrease in food consumption, which is approximately $314 in annual spending. When we use total consumption as the dependent variable, in the lower half of Table2, theeconomicmagnitudeofthemacroexperiencee↵ectdecreasesinthespecification without household fixed e↵ects (columns 1-3) but is again as precise as in the case of food consumption when we include household fixed e↵ects (columns 4-6). In terms of economic magnitude, a one standard-deviation increase in macro experience lowers total consumption by 1.6%, or $713 annually, based on the estimated coecient from column (6). A one standard-deviation increase in personal lifetime unemployment experience lowers total consumption by 2.1%, or $937 annually. We also re-estimate the results on the entire sample, without excluding observations in the top and bottom 5 percentiles of income. As shown in Appendix-Table A.2, the coecients on macroeconomic and personal unemployment experiences become both larger (in absolute value) and more statistically significant. The results are also robust to several variations in the construction of the key explanatory variable. First, as discussed above, our baseline specification fills the gap years of the (biennial) PSID by assuming that families stay in the same state and have the same employment status as in the prior year. Alternatively, we average the values of the prior and the subsequent year, t 1 and t+1. This variation a↵ects both the experience proxy and several control variables. As shown in Appendix- Table A.3, the results are robust. Second, our results are robust to including both the head of the household and the spouse in the construction of the experience measure (Appendix-Table A.4), to excluding the experience of year t 1 from the measure (Appendix-Table A.5), and to using di↵erent weighting (Appendix-Table A.6). In terms of alternative approaches to calculating standard errors, we estimate regressions with standard errors clustered at di↵erent levels in Appendix-Table A.7. We also vary the weighting of observations by applying the PSID family weights, shown in Appendix-Table A.8. (We do not use PSID family weights in the main regression due to eciency concerns.) Overall, the results robustly show that consumers with more adverse macroeconomic and personal unemployment experience tend to spend less, controlling for wealth, income, employment, family structures, and demographics. 22

III.B Past Experiences and Beliefs Given the robust findings of a negative and significant relationship between people’s lifetime experiences of economic conditions and their consumption behavior, we turn to explore the channels through which past experiences a↵ect consumption. To what extent do personal lifetime experiences color beliefs about future outcomes? We relate past lifetime experiences of economic fluctuations to current beliefs about future economic prospects, using the Reuters/Michigan Survey of Consumers (MSC) microdata on expectations from 1953 to 2013. The MSC is conducted by the Survey Research Center at the University of Michigan, quarterly until Winter 1977 and monthly since 1978. The dataset is in repeated cross-section format and includes a total of 292,708 observation. On average, 605 individuals are surveyed each month. Among the multitude of belief elicitations, we identify two questions that capture expectations about economic conditions and consumption. The first question elicits beliefs about one’s future financial situation: “Now looking ahead – do you think that a year from now you will be better o↵ financially, or worse o↵, or just about the sameasnow?” Thesecond question is about expendituresfor (durable) consumption items and individuals’ current attitudes towards buying such items: “About the big things people buy for their homes – such as furniture, refrigerator, stove, television, and things like that. Generally speaking, do you think now is a good or bad time for people to buy major household items?” For the empirical analysis, we construct two binary dependent variables. The first indicator takes the value of 1 if the respondent expects better or the same personal financial conditions over the next 12 months, and 0 otherwise. The second indicator is 1 if the respondent assesses times to be good or the same for durable consumption purchases, and 0 otherwise. We also extract income and all other available demographic variables, including education, maritalstatus, gender, andageoftherespondent.17 Theexplanatoryvariable of interest is again our measure of lifetime unemployment experiences. Since the MSC does not reveal the geographic location of survey respondents, we apply equation (1) to the national unemployment rates to construct the “Experience (Macro)” 17 The MSC does not make information about race available anymore via their standard data access, the SDA system (Survey Documentation and Analysis), since it has been found to be unreliable. When we extract the variable from the full survey, all results are very similar with the additional control. 23

variable for each of individual i from birth until year t, and apply equation (2) to calculate the weighted average of past unemployment experiences. We construct the measure for each respondent at each point in time during the sample. We regress the indicators of a positive assessment of one’s future financial situation or a positive buying attitude on past unemployment experiences, controlling for current unemployment, income, demographics, age fixed e↵ects and year fixed effects. Year fixed e↵ects, in particular, absorb all current macroeconomic conditions as well as all historical information available at the given time. Table 3 shows the corresponding linear least-squares estimations. In columns (1) to (3), we present the estimates of the relation between prior unemploymentrate experiences and respondents’ forecasts of their own future situation. We find that people who have experienced times of greater unemployment during their lives so far are significantly more pessimistic about their future financial situation. The statistical and economic significance of the estimated e↵ect is robust to variations in the controls: Whether we include only (age and time) fixed e↵ects, control for income, or for all demographic variables, we always estimate a highly significant coecient between 0.015 and 0.011. The robustness of the estimates to the income control is reassuring since the controls for respondents’ financial situation are more limited in the MSC data. This renders the estimates in columns (1) to (3) open to alternative interpretations, especially unobserved wealth e↵ects. When we include income in columns (2) and (3), the estimation has the expected positive coecient, and the same holds for demographics that might proxy for unobserved wealth(e.g., education)incolumn(3). Thecoecientofpastexperiencesofnational unemployment rates remains highly significant and negative. In terms of the economic magnitude, consider the inter-decile range of lifetime experiences: Respondents at the 90th percentile are around 2 percentage points more likelytosayfinancialconditionswillbeworseinthenext12monthsthanrespondents at the 10th percentile. The estimations based on the second question, shown in columns (4) to (6), generate very similar results. We estimate a significantly negative e↵ect of lifetime experiences of unemployment on “buying attitude.” The coecient is again fairly stable across specifications, ranging from 0.061 to 0.047. Respondents who have experienced unemployment rates at the 90th percentile of the sample are around 24

snoitatcepxE dna stce↵E ecneirepxE :3 elbaT smeti dlohesuoh rojam yub ot emit dab/dooG raey gnimoc noitidnoc laicnanfi detcepxE )daB = 0 ,emaS ro dooG = 1( )esroW = 0 ,emaS ro retteB = 1( )6( )5( )4( )3( )2( )1( ***740.0- ***250.0- ***160.0- ***110.0- ***310.0- ***510.0- )orcaM( ecneirepxE )500.0( )500.0( )500.0( )400.0( )400.0( )400.0( ***440.0- ***440.0- ***240.0- ***510.0- ***510.0- ***510.0etar tnemyolpmenU )100.0( )100.0( )100.0( )100.0( )100.0( )100.0( ***440.0 ***050.0 ***020.0 ***710.0 emocnI )200.0( )100.0( )100.0( )100.0( seY oN oN seY oN oN slortnoc cihpargomeD seY seY seY seY seY seY stce↵e dexfi egA seY seY seY seY seY seY stce↵e dexfi raeY 734,002 909,102 596,412 612,602 247,702 511,122 snoitavresbO 660.0 260.0 550.0 740.0 740.0 540.0 derauqs-R ot esnopser eht si )3(-)1( snmuloc ni elbairav tnedneped ehT .)CSM( sremusnoC fo yevruS nagihciM eht morf era selbairav llA .setoN tuoba tsuj ro ,↵o esrow ro ,yllaicnanfi ↵o retteb eb lliw uoy won morf raey a taht kniht uoy od – daeha gnikool woN“ noitseuq eht fo yevruS nagihciM eht ni stnednopser laudividni yb detroper )↵o esroW = 0 ,emas eht tuoba ro ↵o retteB = 1( ”?won sa emas eht hcus – semoh rieht rof yub elpoep sgniht gib eht tuobA“ noitseuq eht ot esnopser si )6(-)4( snmuloc ni elbairav tnednepeD .sremusnoC elpoep rof emit dab ro doog a si won kniht uoy od ,gnikaeps yllareneG .taht ekil sgniht dna ,noisivelet ,evots ,rotaregirfer ,erutinruf sa tsael htiw enod si noitamitsE .stnednopser laudividni yb detroper )daB = 0 ,)emaS ro( dooG = 1( ”?smeti dlohesuoh rojam yub ot detcurtsnoc ,tnemyolpmenu fo erusaem ecneirepxe cimonoceorcam eht si ”)orcaM( ecneirepxE“ .sthgiew elpmas htiw dethgiew ,serauqs ,noitacudeedulcnislortnoccihpargomeD .sdlohesuohybdecneirepxeetartnemyolpmenulanoitandethgiewgninilced-ylraenilemitefilasa ,srorre dradnatS .2102 ot 3591 morf snur doirep elpmas ehT .ega hcae rof selbairav ymmud era slortnoc egA .redneg dna ,sutats latiram .ylevitcepser ,ecnacfiingis %1 dna ,%5 ,%01 etoned *** ,** ,* .yticitsadeksoreteh ot tsubor era ,sesehtnerap ni nwohs 25

7 percentage points more likely to say now is a bad time to buy major household items than those at the 10th percentile. This second analysis also addresses concerns unobserved wealth and other unobserved financial constraints even further, beyond the stability across specifications. Here, respondents are asked about “times in general,” and the confounds should not a↵ect their assessment of general economic conditions. Yet, they strongly rely on their personal experiences to draw conclusions about economic conditions more broadly. Our results suggest that the economic conditions individuals have experienced in the past have a lingering e↵ect on their beliefs about the future. Individuals who have lived through worse times consider their own financial future to be less rosy and times to be generally bad for spending on durables, controlling for all historical data, current unemployment, and other macro conditions. This evidence on the beliefs channel is consistent with prior literature on experience e↵ects, including Malmendier and Nagel (2011, 2015). III.C Past Experiences and Future Income Next we ask whether the long-term reduction in consumption after past unemployment experiences, as well as the ensuing consumer pessimism, might be the response to lower employment and earnings prospects. Might the consumer pessimism be explained by (unobserved) determinants of households’ future income that are correlated with past unemployment experiences? As we will show, the answer is no. To test whether past unemployment experiences are correlated with (unobserved) determinants of households’ future income, we re-estimate our baseline model from equation (3) with the dependent variable changed to future income either one or two or three survey waves in the future, i.e., two, four, and six years ahead. The estimation results are in Table 4. They suggest that unemployment experiences do not play a significant role in explaining future income. After controlling for income, wealth, employment status, the other demographics, and fixed e↵ects,18 the estimated coecients of personal unemployment experiences are all positive, small, and insignificant. For macroeconomic experiences, we estimate small negative coefficients, which are also insignificant with the exception of the estimation predicting 18 All results are similar if we do not include time fixed e↵ects in the regressions, which may more realistically capture how people form belief given information friction. 26

Table 4: Experience E↵ects and Future Income Income Income Income t+2 t+4 t+6 Experience (Macro) -0.030 -0.044* -0.050 (0.020) (0.023) (0.030) Experience (Personal) 0.010 0.021 0.017 (0.013) (0.013) (0.021) Income controls Yes Yes Yes Wealth controls Yes Yes Yes Demographic controls Yes Yes Yes Age fixed e↵ects Yes Yes Yes State fixed e↵ects Yes Yes Yes Year fixed e↵ects Yes Yes Yes Household fixed e↵ects Yes Yes Yes Observations 15,710 11,258 7,641 R-squared 0.865 0.884 0.903 Notes. The dependent variables are future income in two, four, and six years, respectively. ”Experience (Macro)” is the macroeconomic experience measure of unemployment, and ”Experience (Personal)” is the personal experience measure. Demographic controls include family size, heads’ gender, race, marital status, education level, and a dummy variable indicating whether the respondent is unemployed at the time of the survey. Income controls include the first and second order of the logarithm of income and lagged income. Wealth controls include the first and second order of the logarithm of liquid and illiquid wealth. We exclude from the sample observations with total familyincomebelowthe5th orabovethe95th percentileineachwavefrom1999to2013, aswellas thepre-sample1997wave(becausewecontrolforlaggedincome). Wetakethelogarithmofincome, and wealth; non-positive values are adjusted by adding the absolute value of the minimum plus 0.1 before being logarithmized. Robust standard errors (in parentheses) are clustered by cohort. *, **, *** denote 10%, 5%, and 1% significance, respectively. 27

income four years ahead, where it is marginally significant. In summary, our results imply that past experiences do not predict future earnings prospects. Relatedly, onemayaskwhetherpastunemploymentexperiencesa↵ectthevolatility of future income. Even if expected income is una↵ected by past experiences, a consumer might (correctly) perceive the variance of income to be a↵ected. If consumers feel greater uncertainty about the stability of their future employment, they will save more to mitigate risk and thus consume less as a result. To test if such a relationship between unemployment experience and income volatility exists, we re-estimate our baseline model (3) using income volatility as the dependent variable. Following Meghir and Pistaferri (2004) and Jensen and Shore (2015), we construct volatiltiy measures both for the transitory and the permanent income. The transitory income-variance measure is the squared two-year change in excess log income, where excess log income is defined as the residual from an OLS regression of log income on our full slate of control variables. The permanent-income variance measure is the product of two-year and six-year changes in excess log income (from year t 2 to t and t 4 to t + 2, respectively). Appendix-Table A.12 shows the results for either measure, two, four or six years ahead (i.e., t+2, t+4, or t+6). We do not find a strong correlation between unemployment experiences and income volatility, other than one marginally significant coecient on macroeconomic experience for the variance of permanent income in t+2. Hence, consumers’ long-term reduction in consumption after past unemployment experiences does not appear to be a rational response to future income uncertainty. III.D Past Experience and Wealth Build-up The significant e↵ect of past unemployment experiences on consumption, and the lack of a relation with future income, imply that household experiences could even a↵ect the build-up of wealth. In the case of negative lifetime experiences, consumers appear to restrain from consumption expenditures more than rationally “required” by their income and wealth positions. This experience-induced frugality, in turn, predicts more future wealth. Vice versa, consumers who have lived through mostly good times are predicted to be spenders and should thus end up with less wealth. In order to test whether experience e↵ects are detectable in long-run wealth 28

accumulation, we relate households’ lifetime experiences to their future wealth, using up to seven survey waves (14 years) in the future. We consider both liquid wealth and total wealth. This analysis also ameliorates potential concerns about the quality of the consumption data and alternative life-cycle interpretations of our findings. Figure 4 summarizes the coecients of interest graphically for 10 regressions, namely, the cases of wealth at t + 6, t + 8, t + 10, t + 12, and t + 14. The upper part shows the coecient estimates when studying the impact on liquid wealth, and the lower part shows the estimates for total wealth. All coecient estimates are positive. The impact of macro experiences is smaller and (marginally) significant only in a few cases, namely, for total wealth in the more recent years and for liquid wealthfurtherinthefuture. Theestimatesoftheroleofpersonallifetimeexperiences are much larger and typically significant, with coecients ranging from 0.02 to 0.03 for liquid wealth and from 0.08 to 0.10 for total wealth. The estimates imply that a one-standard deviation increase in personal lifetime experiences of unemployment will lead to additional precautionary savings and resulting wealth build-up of about 1.3% or $4,500 ten years later. In other word, households who have experienced high unemployment tend to accumulate more wealth down the road. Appendix-Table A.13 provides the details on the coecient estimates of both experience measures. In summary, individuals’ lifetime experiences strongly predict consumption expenditure, and beliefs about future economic conditions appear to play a role in explaining this result. However, such beliefs do not seem to be consistent with actual income and wealth changes. In fact, we see evidence of a positive relationship between past experience and future wealth build-up. IV Consumption with Experience-based Learning Our four baseline results suggest that past experiences can “scar” consumers. The combination of expenditure, belief, and future-wealth results are hard to fully explain in the traditional life-cycle consumption model. However, given the lack of exogenous, experimental variation in lifetime experiences, it is important to further explore potential confounds arising from unobserved determinants and frictions. We utilize the Low, Meghir, and Pistaferri (2010) model to account for several frictions and possible confounds, and to illustrate that, for a wide range of pa- 29

serusaeM ecneirepxE rof slavretnI ecnedfinoC dna stneiceoC detamitsE :pu-dliuB htlaeW :4 erugiF eht sa htlaew diuqil gnisu ledom laciripme ruo etamitse-er ew nehw setamitse eht tneserp )srab der( shparg evfi reppu ehT .setoN evfi ehT .elbairav tnedneped eht sa htlaew latot esu ew nehw setamitse eht wohs )srab eulb( shparg evfi rewol ehT .elbairav tnedneped dna deggal raey-21 ,deggal raey-01 ,deggal raey-8 ,deggal raey-6 esu ew nehw stneiceoc detamitse eht wohs redro latnoziroh ni shparg .level ecnedfinoc %09 ta era slavretni ecnedfinoc eht llA .ylevitcepser serusaem ecneirepxe deggal raey-41 30

rameterizations, we can distinguish experience e↵ects, even directionally. The Low et al. framework captures a broad array of standard life-cycle consumption factors, including financial constraints, social-insurance programs, and “income scarring,” i.e., the notion that job loss reduces income flows because of lower match quality in future jobs. Moreover, we extend the Low et al. model to include “unemployment scarring,” i.e., the notion that unemployment, once experienced, makes individuals inherently less employable. We distinguish both income scarring and unemployment scarring as well as other life-cycle features from scars due to longlasting experience e↵ects. The focus of Low et al. is on the interaction of di↵erent types of risk (productivity shocks, employment risk) with social insurance (unemployment insurance, food stamps, and disability insurance). While the social-insurance programs are not the focus of this paper, they add richness to our analysis and ensure that the experience-e↵ect estimates are not confounded. Towardsthatend,weintroducetwoclassesofconsumersintothemodel: standard rational agents and experience-based learners. Rational consumers use all available historicaldatatoupdatetheirbeliefsabouttheprobabilityofbeingunemployednext period. Experience-based consumers overweight their own experiences when forming beliefs. We simulate intertemporal consumption and labor decisions for both types of consumers and estimate the relation between experience measures and consumption in both settings, i.e., also for rational consumers, for whom they should not have a significantly negative relation. The simulate-and-estimate exercise illustrates the basic mechanism of experience-based learning, and distinguishes it from features of the standard consumption model, such as wealth or liquidity constraints. It provides guidance towards empirical robustness checks and additional tests. Low, Meghir, and Pistaferri (2010) Model Setup. Consumers can work for 40years, untilage62(startingatage23), thenhavemandatory10yearsofretirement where they receive social-security benefits, and die at the end of retirement. Periods are quarters, amounting to L = 200 periods of consumption and labor decisions in total. Their utility function is c e⌘P 1 U(c,P) = ⇥ , (4) 1 31

where c is consumption, and P an indicator equal to 1 if a person works. In each t, consumer i chooses consumption c and, when applicable, labor supply P to i,t i,t maximize lifetime expected utility L maxV = U(c ,P )+E s tU(c ,P ) . (5) i,t i,t i,t t i,s i,s ci,t " # Pi,t s X =t+1 We impose c < A , which rules out borrowing. As we will see below, by maxi,t t imizing the financial constraints of consumers, we are able to derive the sharpest distinction between the role of experience e↵ects and financial constraints.19 We assume that flow utility takes a near CRRA form which induces a precautionary savings motive. (A detailed description of the intertemporal budget constraint and the social-insurance programs is in Appendix B.) Income Process The wage in this model is determined by the following formula lnw = d +x +u +a , (6) i,t t 0i,t i,t i,j,t0 where d is the log-price of human capital at time t, x the component determined t 0i,t by i’s age at time t, u the stochastic component, and a the job-fit component i,t i,,j,t0 of i’s wage at firm j for a job o↵ered (and accepted) in period t . Gross quarterly 0 incomeisw h, wherehisthenumber ofhoursworkedin a quarter. Thethreesociali,t insurance programs Low et al. include in their model are detailed in Appendix B. Agents have the ability to make decisions about whether or not to work. For example, agents need not work if an o↵er is too low. They can also retire early. Note that this implies that experienced-based learners may make di↵erent labor supply choices depending on their concern about future employment and desire to save. The Deterministic Component of Wage. The deterministic component of wage d + x is the same for all individuals of a given age at time t. The size of t 0it 19 Thereasonisthat(unobserved)financialconstraintsareapotentialconfoundoftheempirical relation between prior experiences and consumption: Younger cohort tend to be more constrained in their borrowing ability and are predicted to react more strongly to a shock than older cohorts under the experience-e↵ect hypothesis. By eliminating borrowing altogether from the simulation, we maximize the impact of financial constraints. 32

this component is estimated via regression in Low et al. and of the form20 d +x = ↵+ age+ age2. (7) t 0i,t 1 · 2 · The Permanent Component of Wage. Thestochasticcomponentofthewage u is determined by a random walk. Consumers receive a shock to this component i,t on average once a year. If consumer i has an income shock in period t, then u is i,t u = u +⇣ , (8) i,t i,t 1 i,t where ⇣ is i.i.d. normal with mean 0 and variance 2. i,t ⇣ The Job-Match Component of Wage. Akey elementoftheLowet al. model is its job-match process. The consumer-firm job-match component, a , is drawn i,j,t0 from a normal distribution with mean 0 and variance 2. It is indexed by the period a t in which the consumer joined firm j, and not by t, since it is constant throughout 0 the duration of the consumer-firm interaction. Job Arrival. In each period, the probability of job destruction is , the probability of a job o↵er is (1 )e for an employed worker, and n for an unemployed worker. Agents receive job o↵ers with varying job matches. By construction, they accept all o↵ers with a higher job match and reject all o↵ers with a lower job match. The job match component, in combination with the processes of job destruction and job generation, is at the core of the “income scarring” result of Low et al. (2010). While employed, people successively trade up for jobs that are a better match. Theythusgainhigherincomesovertheirlife-cycle. Inturn, iftheyexperience job destruction, they lose their job match and must (re-)start getting better and better job o↵ers. Hence, agents typically earn a lower income after an unemployment spell, and job loss leads to a long-lasting reduction in earnings. By accounting for rational “income scarring,” we impose a high bar on our hypothesis. We test whether experienced-based learners reduce their consumption beyond this bar. Belief Formation. Both types of consumers, rational and experience-based learners, know the model, but di↵er in their beliefs about the probability of job loss . 20 While x includes a larger set of control variables in the empirical portion of Low et al., only 0i,t age and age squared are used to fit a general lifetime income profile to the model. 33

We denote consumer i’s believed probability of job destruction at time t as b . Rai,t tional consumers use all available data on unemployment to update their beliefs. If they have lived long enough, they know (or closely approximate) the true value of , b = t. Experience-based learners form their belief based on the history of i,t 8 realizations in their prior lives. Applying specification (1), with weighting scheme (2), we obtain t 1 b = w(,t,k)P D , (9) i,t i,t i,t k k=1 X where D is an indicator of i experiencing job destruction in t, and i,t (t k) w(,t,k) = . (10) t k = 1 0 P i,t (t k) P is the weight assigned to realizations D at k periods before period t. Model Estimates on Experience E↵ects in Consumption. We simulate the consumption-saving decisions for both rational and behavioral consumers using the parameters in Table 5.21 The values are identical to those in Low et al. (2010) whenever possible. Following Low et al., we distinguish between high- and loweducation individuals by varying the corresponding parameters. We show several plots of the resulting consumption paths for both rational and experience-based learners in Appendix B. In particular, we separate consumers who were “lucky” and “unlucky” early in life, in terms of their earnings in Figures B.2 and B.3. The graphs illustrate the corresponding over- and underconsumption of experience-based learners during their early lifetime, relative to rational consumers, as well as the need to then curtail consumption later in the first case (good experiences) and the excess wealth build-up in the second case (bad experiences). This corresponds to the empirical relationship we found in Section III.D. Using the simulated values, we estimate the relationship between consumers’ unemployment experience and consumption behavior, controlling for income and wealth. The corresponding OLS regressions are in Table 6, columns (1) and (2), for rational consumers, and in columns (3) and (4) for experience-based learners. In the case of rational agents, prior experiences do not actually enter their belief 21 The full list of parameters is in Appendix-Table B.1. 34

Table 5: Key Simulation Parameters Parameter Benchmark value(s) Preference parameters Relative risk aversion coecient ⇢ 1.5 Interest rate r 1.5% Discount factor 1/(1+r) Lifetime parameters Working years 40 Retirement years 10 Income process High education Low education Standard deviation of job matches 0.226 0.229 a Standard deviation of permanent shocks 0.095 0.106 ⇣ formation. The purpose of including the experience measure here is to identify possible confounds of the significantly negative e↵ect we have estimated in the PSID data. Specifically, as we are concerned about unobserved wealth e↵ects, we estimate one model where we do not include wealth as a control (column 1) and one where we include wealth (column 2), in both cases in addition to the experience-e↵ect proxy. Income scarring. We first conduct the simulation using linearly declining weights ( = 1) for the measure of prior experiences, as we did in our empirical analysis. As shown in the top panel of Table 6, income has the expected positive sign and significance across specifications, as does wealth when it is included. More noteworthy is that, when using the simulations with rational agents (columns 1 and 2), weestimateapositive coecientoftheexperiencemeasure, indicatingthathigher unemployment experiences predict higher consumption. This is the opposite of what we find empirically, and a first step towards ameliorating concerns about confounds: It appears to be hard to (falsely) estimate a negative experience e↵ect when agents are rational, whether or not we include perfect wealth controls. When we alter the belief-formation process to experience-based learning, instead, we estimate a significant negative coecient, both without and with wealth control (columns3and 4). Thatis, lifetimeexperiencesstronglypredictconsumption behavior of experience-based learners, after taking into account their income and wealth. Compared to the results obtained empirically, the coecients on unemployment experience in columns 3 and 4 are greater in magnitude, which may be attributed to the lack to other control variables in the simulation exercises. 35

Table 6: Estimations with Model-Simulated Data (1) (2) (3) (4) Rational Rational EBL EBL = 1: Income 0.576 0.383 0.606 0.398 (244.87) (67.63) (197.97) (55.72) Wealth 0.261 0.264 (52.30) (57.34) Unemployment Experience 0.0975 0.398 -0.392 -0.762 (2.08) (3.24) (-9.01) (-9.49) = 3: Income 0.567 0.379 0.619 0.401 (137.52) (50.58) (163.45) (52.05) Wealth 0.265 0.270 (40.46) (63.31) Unemployment Experience 0.280 0.286 -0.200 -0.576 (4.85) (3.39) (-7.48) (-9.97) = 0: Income 0.569 0.382 0.598 0.397 (209.47) (236.72) (198.14) (55.39) Wealth 0.260 0.259 (52.52) (50.97) Unemployment Experience -0.135 0.397 -0.496 -1.027 (-3.18) (3.20) (-8.37) (-8.39) Notes. Estimations with simulated consumption values as the dependent variable and simulated same-period income and wealth as regressors, for rational consumers in columns (1) and (2), and experienced-based learning (EBL) consumers in columns (3) and (4). Estimations are for with = 1 in the top panel, = 3 in the middle panel, and = 0 in the bottom panel. Consumption, income, and wealth are in log terms. All estimations include period and education fixed e↵ects and use period-clustered standard errors. Simulations are based on the working periods of 10,000 simulated consumers and thus 1,600,000 observations. t statistics in parentheses. 36

Theresultsarequalitativelyandquantitativelysimilarwhenweputhigherweights on consumers’ recent experiences ( = 3), as shown in the middle panel of Table 6. Note that the positive sign of the experience-e↵ect estimate in the data simulated for rational agents (columns 1 and 2) not only ameliorates concerns about wealth confounds, but also seems to contradict the basic intuition of “income scarring:” Unexpected job destruction lowers lifetime income, and thus consumption. Why do higher unemployment experiences predict higher consumption? To understand this result, consider two consumers, A and B, with the same income. A has experienced unexpected job loss in the past, while B has not. All else held equal, “income scarring” predicts that A earns less. However, by assumption, A and B have the same income, suggesting that A’s wage is driven by her permanent-income component rather than her job-match component. As a result, A is less worried about unexpected job destruction and rationally consumes more. In other words, if one introduces a proxy for experience e↵ects into a world with rational agents, it can act as a proxy for the permanent-income component and generate the opposite sign. Under this scenario, there is little concern about confounding experience e↵ects with traditional determinants of lower consumption, including (unobserved) wealth e↵ects and income scarring, as long as we control for current income. To recover the intuition of income scarring and generate a negative relationship between the experience measure and consumption under rational learning, we can employ an experience measure that is more backward looking. In our model, this amounts to lowering . In the bottom panel of Table 6, we repeat the estimation setting = 0, so all prior experiences get equal weights. In this case, the specification with rational learners and without wealth control (in column 1) shows a negative correlation between unemployment experience and consumption. That is, for = 0, we see “income scarring” and its possible confound at work: If two people have the same income today, but one person got fired more in the past, then that person likely has earned less in the past, thus has lower assets today, and consumes less. However, once we control for asset accumulation (in column 2), we re-estimate a positive coecient on unemployment experiences, with all coecients being similar to the ones estimated for = 1 (column 2 of the upper panel). At the same time, the estimated experience e↵ect is robustly negative for experience-based learners, both with and without wealth control, as shown in columns (3)-(4) of the lower panel. In other 37

words, under a more backward-looking proxy for experience e↵ects the potential wealth confound materializes: If we do not control for wealth, the experience-e↵ect proxy might pick up those e↵ects even though agents are not experience-based learners. If agents are experience-based learners, we expect to robustly identify their experience-induced consumption adjustment. Table7: Estimations with Model-Simulated Data, Unemployment Scarring (1) (2) (3) (4) = 1 = 1 = 3 = 3 Rational EBL Rational EBL Income 0.282 0.350 0.284 0.371 (141.89) (65.34) (123.91) (59.28) Wealth 0.314 0.260 0.311 0.262 (54.31) (32.30) (55.01) (35.44) Unemployment Experience 0.176 -1.757 0.138 -1.493 (2.88) (-25.31) (3.06) (-31.06) Notes. Estimationswithsimulatedconsumptionvaluesasthedependentvariableandthesimulated same-period income and wealth as regressors for rational consumers. The simulations account for unemployment scarring. Consumption, income, and wealth are in log terms. The experience proxy is calculated with =1 in columns (1) and (2) and =3 in columns (3) and (4). All estimations include period and education fixed e↵ects and use period-clustered standard errors. Simulations are based on the working periods of 10,000 simulated consumers and thus 1,600,000 observations. t statistics in parentheses. Unemployment scarring. As a last step, we consider an even higher hurdle to the identification of experience e↵ects, and introduce additional negative correlation between unemployment and future income as an alternative explanation. Our motivation for introducing such “unemployment scarring” comes from research in labor economics that has found a persistent negative e↵ect of being unemployed on future income, especially during a recession (Davis and Von Wachter 2011, Huckfeldt 2016, Jarosch2015). Whilethosefindingsmightactuallybeevidenceforexperiencee↵ects, the existing literature proposes more traditional explanations. The model of “unemployment scars” in Jarosch (2015), for example, features a job-security component that resembles the “job-match component” of wages in Low, Meghir, and Pistaferri (2010), albeit with the di↵erence is that wage gains lost due to “income scarring” can be regained by working for an extended period.22 22 See the ✓ component of the firm-type vector in Section 2.1 of Jarosch (2015). y 38

To add “unemployment scarring” to our simulation, we reduce a consumer’s permanent wage component by the average size of a permanent income shock, , every ⇣ time she experiences job destruction. We re-simulate the model with this additional, permanent e↵ect of job loss on income, and then re-estimate the specifications of Tables 6, i.e., analyze again which e↵ects the experience proxy might pick up in estimations using simulated data with rational agents, and whether it identifies experience e↵ects in estimations using simulated data with experience-based learners. Table 7 shows the results for rational and for EBL agents, both for = 1 and for = 3, in the specifications controlling for wealth. We find that the signs and significant levels of all coecients remain the same as in the simulation without unemployment scarring. For the simulations with rational agents, the coecient on the experience-e↵ect proxy remains significantly positive both for = 1 and for = 3,23 though the size of the coecients becomes (mechanically) lower. Intuitively, the experience measure still acts as an indirect proxy for a high permanent component, but now for a subgroup where the permanent component has been systematically reduced compared to the baseline model: Observing two people A and B with the same income today, where only A has experienced unemployment, still suggests that A has a higher permanent component. However, A’s distribution of the permanent component will be shifted down by one standard deviation (unemployment scarring). Overall, these results provide evidence that our predictions are robust even when future income and unemployment are strongly negatively correlated due to both “income scarring” and “unemployment scarring.” For empirically validated parameterizations of experience e↵ects (with linearly declining or steeper weighting functions), financial constraints, unobserved wealth factors, income scarring, and unemployment scarringfailtogenerateanegativerelationbetweenourproxyforpastunemployment experiences and consumption when agents are Bayesian learners. Instead, a negative coecient estimate likely indicates experience-based learning. The same holds when varying the unemployment-experience proxy further to overweight past experiences even more (e.g., equal weighting all life-time experiences) as long as we control for wealth e↵ects. However, if we fail to appropriately control for wealth e↵ects under those latter flat weighting functions, the confound might materialize. We do not im- 23 As before, we estimate a positive coecient even when not controlling for wealth, unless we alter the experience proxy to further overweight experiences far in the past, i.e., for very low ’s. 39

plement the latter specifications empirically; nevertheless, the simulate-and-estimate exercise with those parameters indicates that it is important to conduct exhaustive robustness checks with a variety of alternative wealth specifications—including varying proxies for liquid versus illiquid wealth, higher-order terms, decile dummies, separate dummies for housing wealth or for positive wealth versus debt, and for completeness a similar battery of variations of the income controls. We will also use the model to generate additional predictions of the experience-e↵ect model that are not generated by alternative interpretations.24 V Robustness using PSID, CEX, and Nielsen Guided by the theoretical model, we re-estimate the consumption model with a battery of alternative and additional wealth, income, and liquidity controls using the PSID data. We then turn to the CEX and the Nielsen data, both for replication and, in Section VI, to examine additional predictions of the experience-e↵ect model. V.A PSID: Wealth, Income, and Liquidity We start from concerns about imperfect measurement of individual wealth. Our simulate-and-estimate exercise in Section IV alleviates these concerns, as it appears hard to generate misattribution under our standard proxy for experience e↵ects and given the controls for unemployment status and income – even in the presence of such mismeasurement. Moreover, our prior results on future wealth build up, future income, and beliefs are also hard to reconcile with the unobserved-wealth interpretation. Nevertheless, we use a battery of alternative wealth measures, which we include in addition to the first- and second-order liquid- and illiquid-wealth controls that are already included in Table 2: (1) third and fourth order controls of (log) illiquid and illiquid wealth, (2) wealth decile dummies, separately for liquid and illiquid wealth, (3) log home equity value (home price minus mortgage) and log non-housing wealth, 24 One prediction we did not pursue regards the hours worked. In general, EBL implies a positiverelationbetweenpastunemploymentexperienceandthelikelihoodofworkingbecausework generates greater income bu↵er. (Note that “work” is a binary decision in the model.) However, this prediction does not hold if income or unemployment scarring is strong. In that case, the cost of working dominates the gain, and consumers are more likely to choose living o↵ social welfare programs instead of working. 40

and (4) log total debt and log positive wealth separately. We summarize all estimated coecients of the macro and personal experience measures in the left panel of Figure 5. (The detailed results are in Appendix-Table A.9.) All coecients of interest remain very similar, both in size and in statistical significance. A related concern is measurement error in the income variable. As with wealth, we re-estimate our empirical model using varying constructs of income: (1) third and fourth order of (log) income and lagged income, (2) quintile dummies of income and lagged income, (3) decile dummies of income and lagged income, and (4) controls the bottom 2, 2nd-4th, 4th-6th, 6th-8th, 8th-10th, 90th-92nd, 92nd-94th, 94th-96th, 96th- 98th, and top 2 percentile dummies of income and lagged income. The estimated coecients of interest, shown in the right panel of Figure 5, are again similar in both magnitude and significance. All estimates are in Appendix-Table A.10. A more specific concern is related to the role of liquidity. Even though the results are robust to variations in wealth measures, might the estimated experience e↵ect still be confounded with (unmeasured) liquidity constraints? Our separate controls for liquid and illiquid wealth in the baseline estimations in Table 2 and in columns (2) and (6) of Appendix-Table A.9, ameliorate these concerns. As a further step, we test whether the consumption of households that are disproportionately likely to be liquidity constrained, as proxied by their low liquid-assets position, are more a↵ected by their unemployment experience. Closely following the practice in the consumption literature, such as Johnson, Parker, and Souleles (2006) and Parker, Souleles, Johnson, and McClelland (2013), for each year we sort households into two groups based on whether their liquid wealth lies above or below the median liquid-wealth level in the sample. Expanding equation (3), we interact an indicator for being in the below-median group and the experience variables. As shown in Appendix-Table A.11, households in the bottom half of the liquid-wealth group tend to spend less relative to households in the top half on average. However, their consumption expenditure does not exhibit a significantly stronger reaction to unemployment experience. All coecients are either insignificant or point in the opposite direction. This suggests that the negative e↵ect of unemployment experiences on consumption is not explained by liquidity constraints. 41

serusaeM ecneirepxE rof slavretnI ecnedfinoC dna stneiceoC :slortnoC emocnI dna htlaeW :5 erugiF emocnI .B htlaeW .A noitpmusnoC dooF :elbairaV tnednepeD noitpmusnoC dooF :elbairaV tnednepeD noitpmusnoC latoT :elbairaV tnednepeD noitpmusnoC latoT :elbairaV tnednepeD 8erusaeM 7 erusaeM 6erusaeM 5erusaeM 4erusaeM 3erusaeM 2 erusaeM 1erusaeM dnaelbairavtnednepedehtsanoitpmusnocdoofesuewnehwsetamitseehttneserptnardauqtfelreppuehtnisrab)krad(derehT .setoN dna htlaew diuqil rof seimmud eliced )2( ,htlaew diuqilli dna diuqil redro-htruof dna -driht )1( :slortnoc htlaew evitanretla ruof edulcni htlaew llA .tbed dna htlaew evitisop )4( dna ,)htlaew gnisuoh sunim htlaew latot( htlaew rehto dna htlaew gnisuoh )3( ,htlaew diuqilli eht wohs tnardauq tfel rewol eht ni srab )thgil( eulb ehT .htlaew diuqilli dna diuqil redro-dnoces dna -tsrfi ot noitidda ni era slortnoc ni srab )krad( der ehT .slortnoc htlaew fo stes ruof emas eht dna elbairav tnedneped eht sa noitpmusnoc latot esu ew nehw setamitse evitanretla ruof edulcni dna elbairav tnedneped eht sa noitpmusnoc doof esu ew nehw setamitse eht tneserp tnardauq thgir reppu eht eliced )3( ,emocni deggal dna emocni rof seimmud elitniuq )2( ,emocni deggal dna emocni redro-htruof dna -driht )1( :slortnoc emocni ,dn29 ht09,ht01 ht8,ht8 ht6,ht6 ht4,ht4 dn2,2mottobehtrofseimmudetarapes)4(dna,emocnideggaldnaemocnirofseimmud -tsrfi ot noitidda ni era slortnoc emocni llA .emocni deggal dna emocni fo selitnecrep 2 pot dna ,ht89 ht69 ,ht69 ht49 ,ht49 dn29 latot esu ew nehw setamitse eht wohs tnardauq thgir rewol eht ni srab )thgil( eulb ehT .emocni deggal dna emocni redro-dnoces dna slavretni ecnedfinoC .stce↵e dexfi dlohesuoh edulcni snoisserger llA .slortnoc emocni ruof eht dna elbairav tnedneped eht sa noitpmusnoc .level ecnedfinoc %09 eht ta era 42

V.B CEX Next, we turn to a second source of consumption data, the Consumer Expenditure Survey (CEX). We now enlarge the set of consumption items to include durable goods as well as the CEX measure of total consumption, which is widely used in the literature. It encompasses further categories of expenditures, in addition to durables and non-durable items, including healthcare and education expenses.25 The CEX is a repeated cross-sectional survey of household spending across a comprehensive list of product categories at the quarterly frequency. It is considered the benchmark data in the consumption literature. Compared to the PSID, its two maindisadvantagesarethelackofwealthinformationandthelackofpanelstructure. As in the analysis of the PSID, we link measures of consumption to households’ lifetime unemployment experiences. As before, we construct lifetime experiences as the weighted average of experienced unemployment outcomes since birth, using linearly declining weights. In the CEX data, we are not able to construct the same type of macro and personal unemployment experience measures as in the PSID because the CEX does not provide information on where households resided prior to the sample period, nor on their prior employment status. We use the macro-level experience measure based on national unemployment rates at the quarterly frequency. The top panel of Table 8 provides the summary statistics. The average income, $47k, is in line with the average income at the national level. The sample period runs from 1980-2012. Note that durable and non-durable consumption do not add up to total consumption because of expenditures that are not considered durable or non-durable, such as healthcare and education expenses. The average non-durable and durable spending amount to 67.9% and 20.0% of the mean total expenditures, respectively. Non-durable spending and durable spending are weakly positively correlated, with durable spending being much more volatile than non-durable spending. We re-estimate the sensitivity of consumption to experienced unemployment conditions,usinganestimationmodelthatcloselymirrorsthePSIDmodelfromequation (3). Table 9 shows the results for total, durable, and non-durable consumption. 25 Estimationsinvolvingdurableconsumptioncanbea↵ectedbythetimingofdurablepurchases. PriorresearchsuchasBar-IlanandBlinder(1992)andBergerandVavra(2015)showsthatdurable purchasestendtobediscontinuousandgodownduringrecessions. However, theseconcernsdonot apply to our estimates of experience e↵ects on food and other non-durable consumption items. 43

Table 8: Summary Statistics (CEX and Nielsen) Variable Mean SD p10 p50 p90 N CEX (Quarterly) Age of male head of HH 51 17 29 49 75 417,607 Income 47,220 48,925 8,634 33,728 100,000 417,607 Household size 2.7 1.5 1 2 5 417,607 Total expenditure 6,116 6,145 1,902 4,490 11,479 417,607 Non-durable expenditure 4,152 3,189 1,537 3,452 7,364 417,607 Durable expenditure 1,226 4,082 0 170 2,085 417,607 Experience (Macro) 6.1 0.3 5.8 6.0 6.5 417,607 Nielsen (Monthly) Age of male head of HH 50 12 33 49 67 3,171,833 Income $50-$60k $20-$25k $50-$60k $100k+ 3,171,833 Household size 2.8 1.5 1 2 5 3,171,833 Total expenditure 714 537 205 586 1,366 3,171,833 Coupon use 0.03 0.05 0 0.01 0.09 3,171,833 Product ranking 0.47 0.11 0.34 0.47 0.61 3,171,833 Purchase of sale items 0.24 0.24 0 0.17 0.62 3,171,833 Experience (Macro) 6.0 0.2 5.8 5.9 6.3 3,171,833 Notes. ThetoppanelshowsquarterlyCEXdatafrom1980-2012. Thebottompanelshowsmonthly Nielsen data from 2004-2013. Nielsen reports income in 13 brackets. Coupon use is the value of coupons divided by total expenditures. Product ranking ranges from 0 to 1 based on the unit price of a good within its product module and market in a given month; lower-priced goods have lower values. Purchase of sale items is the number of sale items divided by the total number of items bought. Experience (Macro) is households’ lifetime experience of national unemployment rates. The results strongly confirm our prior findings, and reveal new quantitative implications for the di↵erent components of total consumption. All experience e↵ect coecients are negative and highly significant. Households who have experienced worse unemployment conditions during their lifetime spend significantly less in total, durable, and non-durable consumption. The economic magnitudes are large: A one standard-deviation increase in unemployment experience is associated with a decline in annual consumption of $432 for non-durables and $564 in total. The estimate on non-durable consumption is larger than in the PSID ($276 decline in annual food consumption), while the estimate on total consumption is smaller than the PSID estimate ($912 decline). This may reflect the fact that non-durable in the CEX include leisure, which tends to be elastic, while total expenditures in the CEX encompass healthcare and education, which tend to be more inelastic. The 44

Table 9: Experience E↵ects and Quarterly Consumption (CEX) Total Durables Non-durable Experience (Macro) -0.077*** -0.085*** -0.086*** (0.010) (0.027) (0.005) Income control Yes Yes Yes Household controls Yes Yes Yes Age fixed e↵ects Yes Yes Yes Time fixed e↵ects Yes Yes Yes Region fixed e↵ects Yes Yes Yes Observations 417,607 417,607 417,607 R-squared 0.390 0.126 0.409 Notes. Pooled regressions with (log) total consumption expenditure, durable consumption, and non-durable consumption as the dependent variables. “Experience (Macro)” is the macroeconomic experiencemeasureofunemployment,constructedasalifetimelinearly-decliningweightednational unemployment rate experienced by households. Household characteristics include unemployment status, household size, education, and race. Time fixed e↵ects include year-quarter fixed e↵ects. Region fixed e↵ects include dummies for the Northeast, Midwest, South, and West region. Regressions are weighted by household sampling weights from CEX. The sample period runs from 1980 to 2012. Robust standard errors (in parentheses) are clustered by cohort. *, **, *** denote 10%, 5%, and 1% significance, respectively. new estimate for durables indicates that a one standard-deviation increase in past unemployment experience predicts a $120 decline in annual durable consumption. V.C Nielsen As a final source of data on consumption choices, we turn to the Nielsen Homescan Dataset. The Nielsen data contains information on product purchases of a panel of more than 100,000 U.S. households from 54 geographically dispersed markets, each roughly corresponding to a Metropolitan Statistical Area (MSA), from 2004-2013. The households provide detailed information about their purchases, including price, quantity, date of purchase, identifier of the store, as well as product characteristics, including brand, size and packaging, at the Universal Product Code (UPC) level. Households record the dollar value of any coupons used and whether the purchase involved a deal from the retailer (sale item). The product categories are food and 45

non-food grocery, health and beauty aids, and general merchandise, summing to approximately 3.2 million unique UPCs covering 125 general product categories.26 Households also report information on their demographics, including age, sex, race, education, occupation, employment status, family composition, household income, and location of residency up to the zip code level. Note that the geographic informationismoreprecisethanthestate-levelidentificationinthePSID,asitallows us to control for the local (county-level) unemployment rate U . The information is mt updated annually, and the demographics of the households are representative of the population demographics at the national level. For our analysis, we drop households with heads below the age of 25 or above 75, as in the PSID sample.27 Our data sample consists of 3,171,833 observations of 105,061 households. The bottom panel of Table 8 provides the summary statistics. We note that the average consumption expenditure from Nielsen approximately corresponds to the food consumption expenditures in the PSID, which cross-validates the quality of the data sets as the Nielsen data covers mostly food products. We also conduct a robustness analysis that keeps the advantages of the Nielsen panel structure but exploits the comprehensiveness of the CEX by creating a synthetic Nielsen-CEX panel. Details and estimations are in Appendix-Section A.2. The high-frequency nature of the Nielsen data allows us to construct more precise experience measures than the PSID. However, we are not able to construct the same typeofmacroandpersonalunemploymentexperienceproxiesasinthePSIDbecause, like the CEX, Nielsen provides no information about households’ prior residence or employment status (pre-sample period). We thus construct the macro-level experience measure based on national unemployment rates. For the personal experience measure, we can, at best, measure unemployment experiences since the beginning of the Nielsen data. Such a measure is necessarily biased, as it is less precise at the beginning of the sample and for shorter household spells. We therefore report the estimations employing only the macro-experience measure in the main text.28 26 Severalstudieshaveexaminedthequalityofthedata. Forexample, Einav, Leibtag, andNevo (2010) compare the self-reported Nielsen data with data from cash registers. They conclude that the reporting error is of similar magnitude to that found in commonly used economic data sets. 27 As in the PSID data, we also conduct the analysis on a subsample that excludes households over the age of 65 (retirees) whose expectation of their future income should be immune to beliefs about future economic fluctuations. The results from both sets of regressions are similar. 28 We have re-estimated our model using a measure of personal unemployment experience that 46

Like the CEX, Nielsen lacks information about consumers’ wealth, which is an important component of consumption analyses. Our prior estimations alleviate concernsaboutunobservedwealthtosomeextent, giventhecomparableestimatesacross the PSID data (with wealth controls) and the CEX, and given the robustness of the estimates across a broad range of wealth, income, and liquidity proxies. To further address the issue of the missing wealth control in the Nielsen data, we follow recent advancements in the literature, such as Stroebel and Vavra (2017) and Dube, Hitsch, andRossi(2018), anduseZIP-codelevelhousepricesasameasureofhousingwealth. According to these studies, consumption dynamics respond strongly to house price movements and housing wealth (see also Mian, Rao, and Sufi (2013) and Berger and Vavra (2015)). Empirical analyses can exploit this insight since better measures of housing prices have become available. Specifically, we extract Zillow’s Home Value Index at the local ZIP code level as a proxy for local housing prices,29 and merge it with the Nielsen data. The match rate lies around 75%, and the resulting data set contains almost 3.2 million observations. We include the Home Value Index, an indicator for being a homeowner, and their interaction in all of our estimations.30 To re-estimate the sensitivity of consumption to experienced unemployment conditions in the Nielsen data, we use an estimation model that again closely mirrors the PSID model from equation (3), but accounts for the additional market-level information: C = ↵+UE +U +0x +⌘ +& + +" . (11) it it mt it t m i it The new variables are the current county-level unemployment rate U and localmt market dummies & , where local markets denote Nielsen’s designated market areas m (DMAs).31 As before UE denotes the lifetime (macro) experience of unemployment it takes the value 1 at time t if the head of household has ever been unemployed since the beginning of the sample period up to time t 1, and 0 otherwise. The coecient of interest remains similar. 29 ZillowInc.collectsdetaileddataonhomevaluesacrosstheU.S.andconstructsmonthlyindices using the median value for a ZIP code. Zillow’s estimates of home values (“Zestimates”) aim to providerealisticmarketvaluesgiventhesize,rooms,andotherknownattributesofthehouse,recent appraisals, geographic location, and general market conditions. (The exact formula is proprietary.) FordetailsaboutthedataandZillow’scoverageacrosstheU.S.seeDube,Hitsch,andRossi(2018). 30 We also conduct the analysis without including these wealth controls in the regressions, and the coecient on unemployment experience remains significant and of very similar magnitude. 31 DMAs are slightly bigger than a county but smaller than an MSA. We control for location at 47

rates. The vector of controls x includes income controls, wealth controls, household it characteristics (unemployment status, household size, education, race, and a dummy variable indicating whether the respondent is unemployed at the time of the survey), and age dummies, and the time dummies ⌘ are now year-month-specific. Standard t errors are clustered at the cohort level. All regression results are quantitatively and qualitatively similar when clustered by household, household-time, cohort-time, or two-way clustered at the cohort and time level. Table 10 present results from regression specification (11). Columns (1)-(2) show estimates from pooled OLS regressions, and columns (3)-(4) from regressions with household fixed e↵ects, thus controlling for time-invariant unobserved heterogeneity at the household level. We find that, exactly as in the PSID data, households who have experienced worse unemployment conditions during their lifetimes so far spend significantly less, controlling for contemporaneous macro conditions, local market conditions, and household controls. The economic magnitude is significant: A one standard deviation increase in unemployment experiences is associated with a $708 decline in annual consumption of non-durables, which amounts to around 8% of average spending for the households in our sample. When we introduce household fixed e↵ects, the estimated experience e↵ects become smaller – as expected given that we are di↵erencing out the cross-sectional di↵erences in consumption between householdswith“mostlygood”versus“mostlybad”lifetimeexperiences. Now, aone standard deviation increase in unemployment experiences predicts a $300 decline in annual non-durable consumption, comparable to the PSID estimates. In Figure 6, we illustrate the economic magnitude of the estimates in the context of unemployment conditions during the Great Recession, which falls in the Nielsen sample period. The average monthly unemployment rate from 2008-2012 was 8.1%, with the maximum during the period being 10%. Comparing these numbers with historical averages, the average unemployment rate during the 60 years prior to 2008, from 1947-2007, was 5.6%. Now consider two individuals, a 25-year-old and a 60year-oldasofDecember2007. Theirlifetimeunemploymentexperience, basedonour experience weighting scheme, was 5.3% and 5.8%, respectively, when they entered the crisis in 2008. By the end of 2012, their lifetime unemployment experience was the local market level instead of the county level because people may travel outside of counties to purchase goods. The results are similar if we use county fixed e↵ects instead. 48

Table 10: Experience E↵ects and Monthly Consumption (Nielsen) (1) (2) (3) (4) Experience (Macro) -0.415*** -0.415*** -0.178*** -0.177*** (0.044) (0.044) (0.034) (0.034) Unemployment rate (county) -0.002 -0.005*** (0.003) (0.001) Income control Yes Yes Yes Yes Wealth control Yes Yes Yes Yes Household characteristics Yes Yes Yes Yes Age fixed e↵ects Yes Yes Yes Yes Time fixed e↵ects Yes Yes Yes Yes Market-area fixed e↵ects Yes Yes Yes Yes Household fixed e↵ects No No Yes Yes Observations 3,171,833 3,171,833 3,171,833 3,171,833 R-squared 0.116 0.116 0.526 0.526 Notes. PooledOLSandfixede↵ectsregressionwith(log)totalconsumptionexpenditureasthedependentvariable. Experience(Macro)isthemacroeconomicexperiencemeasureofunemployment, constructed as a lifetime linearly-declining weighted national unemployment rate experienced by households. Wealth controls include the ZIP-code level house-price index from Zillow, an indicator variable for households that own at least one house, and an interaction term between the house price index and the homeowner dummy. Household characteristics include unemployment status, household size, education, race, and a dummy variable indicating whether the respondent is unemployed at the time of the survey. Time fixed e↵ects are year-month fixed e↵ects. Regressions are weighted using the household sampling weights from Nielsen. The sample period runs from 2004 to 2013. Robust standard errors (in parentheses) are clustered by cohort. *, **, *** denote 10%, 5%, and 1% significance, respectively. 49

Figure 6: Example of Unemployment Experience Shock from Recession, Nielsen 7 1000 6.3% 6.1% 5.8% 6 5.3% 800 5 $727.4 $691.0 600 4 $575.3 3 $477.5 400 2 200 1 0 0 2007 Birth year = 1982 Birth year = 1947 2012 )srab ,% ni( tnemyolpmenU fo ecneirepxEemitefiL Monthly Expenditure (in$, lines) -5% -17% Notes. Example of the impact of the Great Recession on weighted lifetime experiences of unemployment rates and monthly consumption expenditure of a 25- and a 60-year-old (as of 2007) from December 2007 to December 2012. The bars show the weighted lifetime experiences of unemploymentratesbasedonlinearly-decliningweights. Thelinesshowthemonthlyexpenditures: thevalues for 2007 are from actual data, and the values for 2012 are calculated based on model estimates. 6.3% vs. 6.1%, respectively. In other words, the unemployment experience for the 25-year-old increased by 1 pp, whereas that for the 60-year-old increased by 0.3pp. Relating these experiences to consumption behavior, our model estimates imply that the monthly consumption expenditure of the 25-year-old decreased by approximately 18% while that of the 60-year-old decreased by approximately 5%. VI Further Implications and Discussion Building on the robust results on the relation between past unemployment experiences and consumption in the PSID, CEX, and Nielsen data, we study two further implications of experience e↵ects and discuss alternative channels, besides belief. 50

VI.A Consumption Quality Motivated by the robust results on the quantity of consumption spending, we further test whether people’s lifetime unemployment experiences a↵ect also the quality of their consumption. To that end, we make use of the rich micro-level information on purchases in the Nielsen data, which also captures the qualitative margins. We constructthreemeasuresofconsumptionquality: (1)couponuse,normalizedbytotal expenditures, (2) the ranking of products based on their unit price (within module, market, andmonth), normalizedbetween0and1, wherelowervaluerepresentslowerpriced goods, and (3) number of on-sale products purchased, normalized by the total number of products purchased. The summary statistics are in Table 8. Theestimationmodelisexactlyasdelineatedinequation(11), onlywithswitched outcome variables. Table 11 displays the main coecients of interest. We find that households who have lived through worse employment conditions are more likely to use coupons, purchase lower-end products, and allocate more expenditures toward saleitems. Forexample, ourestimatessuggestthathouseholdswhohaveexperienced unemployment rates at the 90th percentile of the sample experiences use $13 more in coupon and purchase 8% more sale items monthly than respondents at the 10th percentile. In other words, people who have lived through periods of high unemploymentadjustthequalitymarginsoftheirconsumptionaccordingly. Hence, athorough study on the long-term impact of macroeconomics shocks on consumption calls for analyses not only of aggregate spending figures but also of product substitution and consumption reallocation—margins that entail important welfare implications. VI.B Heterogeneity Across Cohorts Experience-basedlearningnaturallygivesrisetoheterogeneityinconsumptionchoices across cohorts. While all consumers overweight their personal experiences, in particular their more recent experiences, the experience-e↵ect hypothesis also implies that younger cohorts do so more strongly than older cohorts. Experience-based beliefs, as defined in equations (1) and (2), assigns weights to lifetime realizations, and the shorter a consumer’s life is the more mass is assigned to the most recent realization. One implication of our findings, then, is that a given unemployment shock should have a stronger e↵ect on cohorts with shorter lifetime histories so far. We predict 51

Table 11: Experience E↵ects and Monthly Consumption Quality (Nielsen) (1) (2) (3) (4) A: Coupons Experience (Macro) 0.036*** 0.035*** 0.005* 0.005* (0.005) (0.005) (0.003) (0.003) Unemployment rate (county) 0.001*** 0.003*** (0.000) (0.000) (0.000) (0.000) R-squared 0.040 0.041 0.690 0.690 B: Product Ranking Experience (Macro) -0.104*** -0.104*** 0.004** 0.004** (0.0338) (0.0338) (0.002) (0.002) Unemployment rate (county) -0.001** -0.009*** (0.001) (0.002) R-squared 0.083 0.083 0.680 0.680 C: On-sale Items Experience (Macro) 0.159*** 0.156*** 0.009** 0.009* (0.018) (0.018) (0.004) (0.004) Unemployment rate (county) 0.003*** 0.005*** (0.000) (0.001) R-squared 0.073 0.074 0.830 0.830 Income control Yes Yes Yes Yes Household characteristics Yes Yes Yes Yes Age fixed e↵ects Yes Yes Yes Yes Time fixed e↵ects Yes Yes Yes Yes Market area fixed e↵ects Yes Yes Yes Yes Household fixed e↵ects No No Yes Yes Observations 3,171,833 3,171,833 3,171,833 3,171,833 Notes. OLS regressions with the ratio of coupons used over total expenditure as the dependent variable in Panel A; the (transformed) ranking of goods, based on their unit price in their specific product modules, markets, and months in Panel B (where we use the logit transformation ln(y/(1y)) to map the original ranking, which ranges from 0 to 1, to the real line); and with the ratio of on-saleitemspurchasedoverthetotalnumberofitemspurchasedasthedependentvariableinPanel C.Experience(Macro)isthemacroeconomicexperiencemeasureofunemployment,constructedasa lifetime linearly-declining weighted national unemployment rate experienced by households. Other controls are as in Table 10. The sample period runs from 2004 to 2013. Robust standard errors (in parentheses) are clustered by cohort. *, **, *** denote 10%, 5%, and 1% significance, respectively. 52

that the young lower their consumption expenditure to a greater degree than older cohorts during economic busts and, vice-versa, increase it more during booms. We test this implication directly, regressing the change in log consumption in the Nielsen data on the interaction of age with the change in log unemployment conditions from month t to t 1, controlling for the same battery of controls as in Table 10. We do so separately for positive and negative changes (in absolute value) in unemployment rates in order to identify possible asymmetries in the reaction to improving versus tightening economic conditions. Since we know where a household resided in t 1, we can use changes in either the national unemployment rate or the local (county-level) unemployment rate as our proxy for a recently experienced unemployment shock, controlling for the respective other rate change.32 The results are in Table 12. We interact age with the national-rate shock in columns (1)-(2), and with the local (county-level) rate shock in columns (3)-(4). We include all interactions in columns (5)-(6). The changes in log national unemployment rate are absorbed by the time (year-month) fixed e↵ects, and we include the positiveandnegativechangesinloglocalunemploymentrateacrossallspecifications. The estimated age-unemployment interaction e↵ects reveal that unemployment shocks, whether positive or negative, have a smaller e↵ect on expenditures as age increases. The coecients are always significantly negative. The e↵ects are a bit stronger for increases in national unemployment and for decreases in local unemployment. When we include all four interaction e↵ects, the coecient sizes remain similar, with the exception of the interaction of age with lower national employment, where the estimated coecient becomes smaller and insignificant. Overall, the results support our prediction of a significantly stronger response to recent experiences among the young than among the old. 32 Itwouldbemorediculttoestimatethee↵ectofrecentchangesinunemploymentexperience on changes in consumption in the PSID. The low (biannual rather than monthly) frequency of survey waves makes it harder to define the “most recent” experience in a uniform way, and reduces statistical power as we have only eight waves. Hence we use the Nielsen data for this analysis. 53

noitautculF tnemyolpmenU ot noitcaeR ni ytienegoreteH-egA :21 elbaT )6( )5( )4( )3( )2( )1( )C(nl )C(nl )C(nl )C(nl )C(nl )C(nl ***120.0- ***120.0- ***320.0- ***320.0- )nwod-pmenu lanoitaN(nl * egA )500.0( )500.0( )500.0( )500.0( 000.0- 100.0- ***700.0- ***600.0- )pu-pmenu lanoitaN(nl * egA )300.0( )200.0( )200.0( )200.0( 200.0- 100.0- **300.0- *200.0- )nwod-pmenu lacoL(nl * egA )100.0( )100.0( )100.0( )100.0( ***900.0- ***800.0- ***800.0- ***800.0- )pu-pmenu lacoL(nl * egA )100.0( )100.0( )100.0( )100.0( seY seY seY seY seY seY lortnoc tnemyolpmenu lacoL seY seY seY seY seY seY lortnoc emocnI seY seY seY seY seY seY scitsiretcarahc dlohesuoH seY seY seY seY seY seY stce↵e dexfi egA seY seY seY seY seY seY stce↵e dexfi emiT seY seY seY seY seY seY stce↵e dexfi aera-tekraM seY oN seY oN seY oN stce↵e dexfi dlohesuoH 338,171,3 338,171,3 338,171,3 338,171,3 338,171,3 338,171,3 snoitavresbO 410.0 010.0 410.0 010.0 410.0 010.0 derauqs-R srosserger niam eht dna erutidnepxe noitpmusnoc latot ylhtnom gol ni egnahc eht gnieb elbairav tnedneped htiw noisserger SLO .setoN gnidnepedselbairavowtotnidetarapesetartnemyolpmenulacolrolanoitangolniegnahcehtdnaeganeewtebmretnoitcaretniehtgnieb eht era slortnoc tnemyolpmenu lacoL .1 t ot t emit morf htob ,)eulav etulosba ni( evitagen ro evitisop si egnahc eht rehtehw no dlohesuoH .evitagen ro evitisop si egnahc eht rehtehw no gnidneped selbairav owt otni detarapes etar tnemyolpmenu lacol gol ni egnahc snur doirep elpmas ehT .stce↵e dexfi htnom-raey edulcni stce↵e dexfi emiT .ecar dna ,noitacude ,ezis dlohesuoh edulcni scitsiretcarahc deretsulc era sesehtnerap ni srorre dradnats tsuboR .sthgiew dlohesuoh nesleiN yb dethgiew era snoissergeR .3102 ot 4002 morf ylhtnom .ylevitcepser ,ecnacfiingis %1 dna ,%5 ,%01 etoned *** ,** ,* .emit dna trohoc yb 54

This finding also helps further distinguish the experience-e↵ect hypothesis from alternative theories such as liquidity constraints of the young (e.g. Zeldes (1989), Gourinchas and Parker (2002)). Models with liquidity constraints predict that the young react more strongly to negative unemployment shocks than the old, as they are more likely to hit liquidity constraints; but they do not easily predict a more positive reaction to positive shocks. To generate the latter prediction, these models need to rely on the argument that the young were previously constrained, and a positive shock allows them to adjust to their permanent-income optimum. However, our identification also exploits the di↵erences in consumption of the young at better and worse economic times. Here, an adjustment to the PIH optimum would predict the opposite outcome relative to the experience e↵ect hypothesis: the young with more negative prior experiences would exhibit a stronger reaction to recent good outcomes according to the PIH.33 Thus, our findings highlight experience e↵ects as a distinct force in a↵ecting people’s consumption behavior. VI.C Preference Channel In Section III.B, we showed that individuals’ past experiences significantly influence beliefs about their future financial situation. This evidence helped to distinguishexperience-basedlearningfromalternativeexplanationsofconsumers’response to past experiences. At the same time, lifetime experiences might influence not only consumers’ beliefs but also their preferences. In other words, the evidence on experience-based learning (beliefs channel) does not rule out that experience-based taste changes (preference channel) are also at work. Thereare manypossible specifications ofthe preference-based interpretation, and it is thus impossible to conclusively reject the instable-preferences explanation. As in the case of the beliefs-based channel, we can at best aim to provide evidence in favor of specific formalizations. We explore one preference specification that has garnered significant support in prior empirical literature: We study whether our findings on the significant relationship between consumption and lifetime experience 33 We estimated a set of regressions that augments the specifications from Table 12 with triple interactions of age, positive and negative national or local unemployment shocks, and a dummy variableindicatingabove-medianunemploymentexperiencefortherespondent’sage. Theestimated e↵ects of positive national and local unemployment shocks are weaker (given age) for respondents withworseunemploymentexperiences,aspredictedbyEBLbutnotbyastandardPIHframework. 55

may be correlated with habit persistence in consumption. To that end, we estimate an alternative version of the empirical model in equation (11) that includes a lagged consumption measure on the right hand side. This dynamic specification, with the lagged dependent variable included, requires a correction for the correlation between the lagged dependent variable and the fixed e↵ects in the error term, which gives rise to “dynamic panel bias” (Nickell (1981)). To obtain unbiased and consistent coecients, we estimate the specification using a dynamic GMM panel estimator, following Holtz-Eakin, Newey, and Rosen (1988), ArellanoandBond(1991),ArellanoandBover(1995),andBlundellandBond(1998). Accordingly, both level and di↵erenced equations are used, and the lagged dependent variable is instrumented using lagged di↵erences for the level equation and lagged levels for the di↵erenced equation.34 The goodness of fit statistics for the system GMM estimators are calculated as the square of the correlation coecients between the actual and the fitted values of the dependent variable. In Table 13, we present the results. The estimates show that the e↵ects of prior unemployment experience on consumption remain highly significant after taking into account possible habit persistence in consumption. The estimation results both confirm the robustness of experience e↵ects and indicate that they do not operate through the channel of habit formation. VII Aggregate Implications and Conclusion While it has been a decade since the start of the Great Recession, e↵ects of the crisis still linger, and a better understanding of the long-term e↵ects of economic shocks hasproven to be ofutmost importancefor both academicsand policy-makers. In this paper, wehaveputforwardtheideathatpastexperiencesofmacroeconomicandpersonal unemployment shocks play a significant role in shaping household consumption decisions and thereby the long-term consequences of macroeconomic shocks. Estimation results from three di↵erent data sources confirm this conclusion. Households who have experienced times of higher local and national unemployment 34 Note that we test for first- and second-order autocorrelation in the first-di↵erenced errors and find that they are first-order serially correlated, but not second-order serially correlated. This supports the validity of the moment conditions used by the system GMM estimators. 56

Table 13: Experience E↵ects and Consumption, GMM regressions PSID Nielsen CEX Experience (Macro) -0.181*** -0.266*** -0.045*** (0.063) (0.051) (0.006) Experience (Personal) -0.635** — — (0.120) Income control Yes Yes Yes Wealth control Yes Yes No Household characteristics Yes Yes Yes Age fixed e↵ects Yes Yes Yes Time fixed e↵ects Yes Yes Yes Location fixed e↵ects Yes Yes Yes Observations 29,813 3,016,952 235,834 R-squared 0.45 0.41 0.64 Notes. System GMM regressions with food consumption (in logarithm) as the dependent variable. “Experience (Macro)” is the macroeconomic experience measure, “Experience (Personal)” is the personal experience measure, specified as described above for the respective datasets. Time fixed e↵ects include year fixed e↵ects for the PSID sample, year and month fixed e↵ects for the Nielsen sample, and year and quarter fixed e↵ects for the CEX sample. Location fixed e↵ects include state fixed e↵ects for the PSID sample, market area fixed e↵ects for the Nielsen sample, and region fixed e↵ectsfortheCEXsample. Thesampleperiodrunsfrom1999-2013forthePSID,2004to2013for the Nielsen sample, and 1980 to 2012 for the CEX sample. Robust standard errors in parentheses are clustered on cohort. *, **, *** denote 10%, 5%, and 1% significance, respectively. 57

Figure 7: Aggregate Unemployment Experience and Consumer Spending )dehsad( % ,ecneirepxE tnemyolpmenU 5.6 6 5.5 5 53. 43. 33. 23. )dilos( dednerted ,PDG/ECP laeR 1965 1975 1985 1995 2005 2015 Notes. Aggregateunemploymentexperiencecalculatedasaweightedaverageofnationalunemploymentexperience,asdefinedinEquation1,withtheweightsbeingU.S.populationbyage(restricted to age 25 to 75) from the Census. Aggregate consumer spending is measured as real personal consumption expenditure (PCE) from the U.S. Bureau of Economic Analysis (BEA) normalized by real gross domestic product (GDP), detrended by removing a linear time trend from the series. and more personal unemployment spend significantly less, after controlling for income, wealth, and demographics, and tend to choose lower-quality items. We further showthatbeliefsaboutone’sfuturefinancialsituationbecomepessimistic, consistent with the consumption behavior, but that such beliefs do not seem to be consistent with actual income and wealth changes. In fact, we see evidence of a positive relationship between past experience and future wealth build-up. In light of our results on the lasting e↵ects of past experiences on consumption, experience e↵ects could potentially constitute a novel micro-foundation underlying fluctuations in aggregate demand and long-run e↵ects of macroeconomic shocks. While a thorough investigation of the macroeconomic implications of experience effects is beyond the scope of this paper, we now provide some suggestive evidence 58

on the aggregate level. Specifically, we relate an aggregate measure of lifetime experiences in the U.S. population to a measure of aggregate consumption expenditure in the U.S. from 1965 to 2013. For the former measure, we take a weighted average of national unemployment experience, as defined in Equation (1), using data on U.S. population broken down by age (age 25 to 75) from the Census as weights. For aggregate consumer spending, we use data on real personal consumption expenditure (PCE) from the U.S. Bureau of Economic Analysis (BEA) normalized by real gross domestic product (GDP). As shown in Figure 7, there exists a negative relationship between the two measures: times of higher aggregate unemployment experience coincide with times of lower aggregate consumer spending. The strong negative correlation pattern not only adds credibility to our micro-level estimates but also suggests the possibility that personally experienced labor market conditions may be a significant granular source of aggregate fluctuations. The evidence on experience e↵ects in consumption has potentially important policy implications. They appear to significantly dampen macroeconomic fluctuations, which in turn calls for considerations from policy-makers on optimal stabilization policy, monetary or fiscal. For future research, our empirical methodology could be applied to a larger crosssection of countries, particularly countries that have undergone more drastic and volatile macroeconomic events such as the emerging market countries and some European countries. Such exercises would help to determine the extent to which personal experiences a↵ect household consumption—the key ingredient in all macro and macro-finance frameworks. References Abadie, A. and G. Imbens (2011). Bias-Corrected Matching Estimators for Average Treatment E↵ects. Journal of Business and Economic Statistics 29(1), 1–11. Aguiar, M.andE.Hurst(2015).ConsumptionvsExpenditure.Journal of Political Economy 113(5), 919–948. Alesina, A. and N. Fuchs-Schu¨ndeln (2007). Goodbye Lenin (or Not?): The E↵ect of Communism on People’s Preferences. American Economic Review 97(4), 1507–1528. 59

Andreski, P., G. Li, M. Z. Samancioglu, and R. Schoeni (2014). Estimates of annual consumption expenditures and its major components in the PSID in comparison to the CE. American Economic Review 104(5), 132–135. Arellano, M., R. Blundell, and S. Bonhomme (2017). Earnings and consumption dynamics: a nonlinear panel data framework. Econometrica 85(3), 693–734. Arellano, M. and S. Bond (1991). Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations. Review of Economic Studies (58), 277–297. Arellano, M. and O. Bover (1995). Another Look at the Instrumental Variable Estimation of Error-Components Models. Journal of Econometrics (68), 29– 51. Attanasio, O. and M. Browning (1995). Consumption over the Life-cycle and over the Business-cycle. American Economic Review 85(5), 1118–1137. Attanasio, O.P.(1999).Consumption(Chapter11).Volume1, PartBofHandbook of Macroeconomics, pp. 741 – 812. Elsevier. Attanasio, O. P. and G. Weber (1995). Is Consumption Growth Consistent with Intertemporal Optimization? Evidence from the Consumer Expenditure Survey. Journal of political Economy 103(6), 1121–1157. Ball, L.(2014).Long-TermDamagefromtheGreatRecessioninOECDCountries. NBER Working Paper 20185. Bar-Ilan, A.andA.Blinder(1992).ConsumerDurables: EvidenceontheOptimality of Usually Doing Nothing. Journal of Money, Credit and Banking 24(2), 258–272. Berger,D.andJ.Vavra(2015).ConsumptionDynamicsDuringRecessions.Econometrica 83(1), 101–154. Blanchard, O. (2012). Sustaining a Global Recovery. Finance and Development 46(3), 9–12. Blanchard, O. and L. Summers (1986). Hysteresis and the European Unemployment Problem. NBER Macroeconomic Annual 1, 15–90. Blundell, R. and S. Bond (1998). Initial Conditions and Moment Restrictions in Dynamic Panel Data Models. Journal of Econometrics (87), 115–143. Blundell, R., M. Browning, and C. Meghir (1994). Consumer Demand and the Life-cycle Allocation of Household Expenditures. The Review of Economic Studies 61(1), 57–80. Caballero, R. J. (1990). Consumption Puzzles and Precautionary Savings. Journal of Monetary Economics 25(1), 113–136. 60

Campbell, J. and A. Deaton (1989). Why is Consumption So Smooth. Review of Economic Studies 56, 357–374. Carroll, C.D.(1994).HowDoesFutureIncomeA↵ectCurrentConsumption? The Quarterly Journal of Economics, 111–147. Cerra, V. and S. C. Saxena (2008). Growth Dynamics: The Myth of Economic Recovery. American Economic Review 98(1), 439–457. Coen, R. (1973). Labor Force and Unemployment in the 1920’s and 1930’s: A Reexamination Based on Postwar Experience. Review of Economics and Statistics 55(1), 46–55. Coibion, O., Y. Gorodnichenko, and G. H. Hong (2015). The Cyclicality of Sales, Regular and E↵ective Prices: Business Cycle and Policy Implications, Forthcoming. American Economic Review 105(3), 993–1029. Collin-Dufresne, P., M. Johannes, and L. A. Lochstoer (2016). Asset Pricing when’This Time is Di↵erent’. Swiss Finance Institute Research Paper (13-73), 14–8. Davis, S. J. and T. M. Von Wachter (2011). Recessions and the Cost of Job Loss. Working Paper, National Bureau of Economic Research. De Nardi, M., E. French, and D. Benson (2012). Consumption and the Great Recession. Federal Reserve Bank of Chicago Economic Perspectives (1), 1–16. Deaton, A. (1991). Saving and Liquidity Constraints. Econometrica 59(5), 1221– 1248. Deaton, A. (1992). Understanding Consumption. Oxford: Clarendon Press. Delong, B. and L. Summers (2012). Fiscal Policy in a Depressed Economy. Brookings Papers on Economic Activity 44(1), 233–297. Dube, J.-P., G. J. Hitsch, and P. E. Rossi (2018). Income and Wealth E↵ects on Private-Label Demand: Evidence From the Great Recession. Marketing Science. Dynan, K. E. (2000). Habit Formation in Consumer Preferences: Evidence from Panel Data. American Economic Review 90(3), 391–406. Ehling, P., A. Graniero, and C. Heyerdahl-Larsen (2018, July). Asset Prices and Portfolio Choice with Learning from Experience. Review of Economic Studies 85(1), 1752–1780. Einav, L., E. Leibtag, and A. Nevo (2010). Recording Discrepancies in Nielsen Homescan Data: Are They Present and Do They Matter? Quantitative Marketing and Economics 2(8), 207–239. 61

Flavin, M. (1993). The Excess Smoothness of Consumption: Identification and Interpretation. Review of Economic Studies 60(3), 651–666. Friedman, M. (1957). Theory of the Consumption Function. Prenceton University Pres. Fuchs-Schu¨ndeln, N. and M. Schu¨ndeln (2015). On the Endogeneity of Political Preferences: Evidence from Individual Experience with Democracy. Science 347(6226), 1145–1148. Fuhrer, J. (2000). Habit Formation in Consumption and Its Implications for Monetary-Policy Models. American Economic Review 90, 367–390. Gabaix, X. and D. Laibson (2017). Myopia and Discounting. Working paper. Gourinchas, P.-O. and J. Parker (2002). Consumption over the Life Cycle. Econometrica 70(1), 787–817. Gouskova, E. and R. Schoeni (2007). Comparing Estimates of Family Income in the PSID and the March Current Population Survey, 1968–2005. Panel Study of Income Dynamics Technical Paper Series. Michigan: Ann Arbor. Greenwood, R. and S. Nagel (2009). Inexperienced Investors and Bubbles. Journal of Financial Economics 93, 239–258. Haltmaier, J. (2012). Do Recessions A↵ect Potential Output? Board of Governors of the Federal Reserve System Finance and Economics Discussion Series 1066. Hansen, A. H. (1939). Economic Progress and Declining Population Growth. The American Economic Review 29(1), 1–15. Harris, C. and D. Laibson (2001). Dynamic Choices of Hyperbolic Consumers. Econometrica 64(4), 935–957. Holtz-Eakin, D., W. Newey, and H. Rosen (1988). Estimating Vector Autoregressions with Panel Data. Econometrica (56), 1371–1395. Huckfeldt, C. (2016). Understanding the Scarring E↵ect of Recessions. Jarosch, G. (2015). Searching for Job Security and the Consequences of Job Loss. Jensen, S. T. and S. H. Shore (2015). Changes in the Distribution of Earnings Volatility. Journal of Human Resources 50(3), 811–836. Johnson, D., J. A. Parker, and N. S. Souleles (2006). Household Expenditure and the Income Tax Rebates of 2001. American Economic Review 96(5), 1589– 1610. Kahn, L. (2010). The Long-term Labor Market Consequences of Graduating from College in a Bad Economy. Labor Economics (17), 303–316. 62

Kahneman,D.andA.Tversky(1974).SubjectiveProbability: AJudgmentofRepresentativeness. In The Concept of Probability in Psychological Experiments, pp. 25–48. Springer. Kaplan, G., G. L. Violante, and J. Weidner (2014). The Wealthy Hand-to-mouth. Working Paper, Brookings Papers on Economic Activity. Kaustia, M.andS.Knu¨pfer(2008).DoInvestorsOverweightPersonalExperience? Evidence from IPO Subscriptions. Journal of Finance 63, 2679–2702. Laudenbach, C., U. Malmendier, and A. Niessen-Ruenzi (2018). The Long-lasting E↵ectsofExperiencingCommunismonFinancialRisk-Taking.WorkingPaper. Lebergott, S. (1957). Annual Estimates of Unemployment in the United States, 1900-1954. In NBER (Ed.), The Measurement and Behavior of Unemployment. Princeton, NJ: Princeton University Press. Lebergott, S. (1964). Manpower in Economic Growth: The American Record since 1800. New York: McGraw-Hill. Lichter, A., M. Lo¨✏er, and S. Siegloch (2016). The Long-Term Costs of Government Surveillance: Insights from Stasi Spying in East Germany. Working Paper. Low, H., C. Meghir, and L. Pistaferri (2010). Wage Risk and Employment Risk over the Life Cycle. American Economic Review 100(4), 1432–67. Malmendier, U. and S. Nagel (2011). Depression Babies: Do Macroeconomic Experiences A↵ect Risk-Taking? Quarterly Journal of Economics 126, 373–416. Malmendier, U. and S. Nagel (2015). Learning from Inflation Experiences. Quarterly Journal of Economics 131(1), 53–87. Malmendier, U., D. Pouzo, and V. Vanasco (2018). Investor Experiences and Financial Market Dynamics. Working paper. Malmendier, U. and G. Tate (2005). CEO Overconfidence and Corporate Investment. Journal of Finance 60(6), 2661–2700. Malmendier, U., G. Tate, and J. Yan (2011). Overconfidence and Early-life Experiences: The E↵ect of Managerial Traits on Corporate Financial Policies. Journal of Finance 66(5), 1687–1733. Meghir, C.andL.Pistaferri(2004).IncomeVarianceDynamicsandHeterogeneity. Econometrica 72(1), 1–32. Meghir, C. and G. Weber (1996). Intertemporal Nonseparability or Borrowing Restrictions? A Disaggregate Analysis using a U.S. Consumption Panel. Econometrica 64(5), pp. 1151–1181. 63

Mian, A., K. Rao, and A. Sufi (2013). Household Balance Sheets, Consumption, and the Economic Slump. Quarterly Journal of Economics 128(4), 1687–1726. Modigliani, F. and R. Brumberg (1954). Utility analysis and the consumption function: An interpretation of cross-section data. In K. Kufihara (Ed.), Post- Keynesian Economics, pp. 128–197. Rutgers University Press. Nevo, A. and A. Wong (2015). The Elasticity of Substitution Between Time and Market Goods: Evidence from the Great Recession. Working paper. Nickell, S. (1981). Biases in Dynamic Models with Fixed E↵ects. Econometrica: Journal of the Econometric Society, 1417–1426. Olafsson, A. and M. Pagel (2018). The Retirement-Consumption Puzzle: New Evidence on Individual Spending and Financial Structure. Working paper. Oreopoulos, P., T. von Wachter, and A. Heisz (2012). The Short- and Long-Term Career E↵ects of Graduating in a Recession: Hysteresis and Heterogeneity in the Market for College Graduates. American Economic Journal: Applied Economics (4), 1–29. Oyer, P. (2008). The Making of an Investment Banker: Stock Market Shocks, Career Choice, and Lifetime Income. The Journal of Finance 63(6), 2601– 2628. Pagel, M. (2017). Expectations-Based Reference-Dependent Life-Cycle Consumption. Review of Economic Studies 84(2), 885–934. Parker, J. A., N. S. Souleles, D. S. Johnson, and R. McClelland (2013). Consumer Spending and the Economic Stimulus Payments of 2008. American Economic Review 103(6), 2530–2553. Petev, I., L. Pistaferri, and I. Saporta-Eksten (2011). An Analysis of Trends, Perceptions, andDistributionalE↵ectsinConsumption.InB.W.DavidB.Grusky and C. Wimer (Eds.), The Great Recession, Handbook of the Economics of Finance, pp. 161–195. Russell Sage Foundation. Pfe↵er, F. T., R. F. Schoeni, A. Kennickell, and P. Andreski (2016). Measuring Wealth and Wealth Inequality: Comparing Two US Surveys. Journal of economic and social measurement 41(2), 103–120. Pistaferri, L. (2016). Why has consumption grown only moderately after the Great Recession? Working paper. Reifschneider, D., W. L. Wascher, and D. Wilcox (2015). Aggregate Supply in the United States: Recent Developments and Implications for the Conduct of Monetary Policy. IMF Economic Review 63(1), 71–109. Reinhart, C. and K. Rogo↵ (2009). The Aftermath of Financial Crises. American Economic Review 99(2), 466–472. 64

Romer, C. (1986). Spurious Volatility in Historical Unemployment Data. Journal of Political Economy 94(1), 1–37. Rosenbaum, P. and D. Rubin (1983). The Central Role of the Propensity Score in Observational Studies for Causal E↵ects. Journal of Business and Economic Statistics 70, 41–55. Schraeder, S. (2015). Information Processing and Non-Bayesian Learning in Financial Markets. Review of Finance, 1–31. Stroebel, J. and J. Vavra (2017). House Prices, Local Demand, and Retail Prices. Journal of Political Economy. Forthcoming. Summers, L. H. (2014a). Reflections on the “New Secular Stagnation Hypothesis”. In C. Teulings and R. Baldwin (Eds.), Secular Stagnation: Facts, Causes and Cures, pp. 27–40. A VoxEU.org eBook. Summers, L. H. (2014b). US Economic Prospects: Secular Stagnation, Hysteresis, and the Zero Lower Bound. Business Economics 49(2), 65–73. Tversky, A. and D. Kahneman (1974). Judgment under Uncertainty: Heuristics and Biases. Science 185(4157), 1124–1131. Vissing-Jorgensen, A. (2003). Perspectives on Behavioral Finance: Does “Irrationality” Disappear with Wealth? Evidence from Expectations and Actions. In NBER Macroeconomics Annual. West, K. D. (1989). The Insensitivity of Consumption to News About Income. Journal of Monetary Economics 21, 17–23. Woodford, M. (2003). Interest and Prices: Foundations of a Theory of Monetary Policy. Princeton: Princeton University Press. Woodford, M. (2010). Optimal Monetary Stabilization Policy. Volume 3 of Handbook of Monetary Economics, pp. 723 – 828. Elsevier. Zeldes, S. P. (1989). Consumption and Liquidity Constraints: An Empirical Investigation. Journal of Political Economy 97(2), 305–346. 65

For Online Publication Scarred Consumption Ulrike Malmendier and Leslie Sheng Shen August 2019 Appendix A Empirical Analysis A.1 Robustness using PSID Data We present a series of robustness tests of the estimations relating unemployment experiences to consumption, as well as the estimations of the wealth build-up. The first twelve tables use the PSID data. Appendix-Table A.1 presents the summarystatisticsofthefullsample,i.e.,includingobservationswithtotalfamilyincome below the 5th or above the 95th percentile in each wave. Otherwise, we apply the same restrictions as in the construction of the main sample, namely, drop individuals for whom we cannot construct the experience measures (due to missing information about location or employment status in any year from t to t 5), and observations with missing demographic controls or that only appear once. The resulting sample has 37,156 observations, compared to 33,164 in the main sample. The sample statistics are very similar, with a mean macroeconomic experience measure of 6.0%, mean personal experience of 5.4%, average food consumption of $8,559, and average total consumption of $46,256 (both measured in 2013 dollars). In Appendix-Table A.2, we re-estimate the regression model of Table 2 on the full sample. The results become even stronger. The estimated macroeconomic experience and personal experience e↵ects are both larger and more significant those estimated in Table 2. In Appendix-Table A.3, we construct an alternative experience measures for the gap years (between the PSID biennial surveys). For the macroeconomic experience measure in the main text, we fill in the unemployment rate in a gap year t by 66

assuming that the family lived in the same state as in year t 1. Here, we assume that respondents spend half of year t in the state in which they lived in year t 1 and the other half in the state in which they lived in year t + 1. (This alternate construction does not change the value if respondents live in the same state in t 1 andt+1.) Similarlyforthepersonalexperiencemeasure, wereconstructrespondents’ employment status in year t as the average of their status in years t 1 and t+1, ratherthanapplyingthevaluefromyeart 1. Forexample, ifapersonisunemployed in t 1 and is employed in t+1, the personal experience in t will be denoted as 0.5. Re-estimating the model in (3), we find results very similar to those in Table 2. In Appendix-Table A.4, we present an alternative experience measure that incorporates the experiences of the spouses. For married households, we use the average of the household heads’ and spouses’ experiences, controlling for married-couples indicator. All variables other than the couple indicator and the experience measures are defined as in Table 2. The coecients of interest remain very stable, with some of the personal experience e↵ect estimates increasing in (absolute) magnitude. Appendix-Table A.5 presents yet another alternative experience measure, which excludes unemployment experiences from year t 1 to further rule out concurrent factors. All other variables are defined as in Table 2. The coecients of interest remain stable without households fixed e↵ects. When including households fixed e↵ects, the estimates are slightly smaller in magnitude but remain significant. In Appendix-Table A.6, we use weighting parameter = 3 instead of = 1 to construct experience measures, and re-estimate the fixed-e↵ect models of Table 2. Higher means individuals put more emphasis on their more recent experiences. As shown in Table A.6, the results remain similar. Hence, the significant relation between experience and consumption is robust to the variation in weighting parameter. Appendix-Table A.7 shows the results for di↵erent clustering units. Instead of clustering by cohort as in Table 2, we cluster the standard errors by cohort year, ⇤ household, household year, and we two-way cluster by cohort and year. The pooled ⇤ regressions in Appendix-Table A.7 correspond to the specification in column (3) of Table 2, and the specifications with household fixed-e↵ects correspond to column (6) in Table 2. The statistical significance of our results are not a↵ected in most cases. Once we included household fixed e↵ects, both experience variables are significant. 67

InAppendix-TableA.8,weapplythePSIDlongitudinalfamilyweights. Notethat some families are given zero weight and are thus dropped from the estimation, which explains the lower number of observations in the weighted regressions. As before the results remain very similar in the specifications with household fixed e↵ects. Appendix-Tables A.9, A.10, and A.11 address concerns about unobserved wealth, liquidity, or income components. Appendix-Table A.9 presents results from estimations using alternative wealth controls, in addition to the measures of liquid and illiquid wealth in Table 2: third- and fourth-order liquid and illiquid wealth (column 1); decile dummies of liquid and illiquid wealth (column 2); housing wealth and other wealth (column 3); positive wealth and debt (column 4). Columns (5)-(8) mirror columns (1)-(4) respectively but include household fixed e↵ects. The coecients of interest remain stable and (at least marginally) statistically significant. Appendix-Table A.10 uses alternative income controls, in addition to the controls of first and second order of income and lagged income: third- and fourth-order incomeandlaggedincome(column1); quintiledummiesofincomeandlaggedincome (column 2); decile dummies of income and lagged income (column 3); controls for bottom 2, 2nd 4th, 4th 6th, 6th 8th, 8th 10th, 90th 92nd, 92nd 94th, 94th 96th, 96th 98th, and top 2 percentile dummies of income and lagged income (column 4). Columns (5)-(8) have the same income controls as columns (1)-(4) respectively but include household fixed e↵ects. The coecients of interest remain stable. All of the estimates that were significantly negative before are still significant. In A.11, we test whether households that are more liquidity constrained are more a↵ected by their unemployment experience. Closely following the practice in the consumption literature such as Johnson, Parker, and Souleles (2006) and Parker, Souleles, Johnson, and McClelland (2013), we sort households into two groups based on whether their liquid wealth is above or below the sample median in the respective year. We then add an indicator for below-median liquid wealth as well as its interactions with the experience variables to the estimating equation (3). As Appendix- Table A.11 shows, households in the bottom half of liquid wealth tend to spend less, but do not exhibit stronger reactions to unemployment experience. This suggests households’ experience a↵ect consumption beyond potential liquidity constraints. In Appendix-Table A.13, we study the e↵ects of lifetime experiences on wealth 68

accumulation. This analysis tests whether, given the significant impact of unemployment experiences on consumption, we can also detect experience e↵ects in the build-up of wealth. The dependent variables are either liquid wealth or total wealth, and the main regressors are lagged experience measures. We lag the experience measures by six, eight, ten, twelve, and 14 years, instead of using the contemporary experience measures, recognizing that the e↵ects of experience on wealth may take time to realize. We include the same set of control variables as in our main analyses, including controls for income in years t 1 and t 2, and add a control for the av- erage family income between year t 2 and the year in which the lagged experience measures are based on (six, eight, ten, twelve, and 14 years ago, respectively). For example, when six-year lagged experience is the main regressor, we control for the average income between t 2 and t 6. This average-income control addresses the concern that previous experiences of economic boom or crisis may have implications for future income (Oyer (2008); Kahn (2010); Oreopoulos, von Wachter, and Heisz (2012)).35 In Appendix-Figure A.1, we plot the estimated coecients on the lagged experience measures. In Appendix-Table A.13, we show the estimates of the coefficients on the 10-year, 12-year, and 14-year lagged experience measures. We find a significant role of past experiences for the build-up of wealth and liquid wealth, especially in the context of personal experiences. 35 The results are similar if, instead of having an average-income control, we include the incomes for all years between year t 2 and the year in which the lagged experience measures are based on. 69

Table A.1: Summary Statistics (PSID), Full Sample Variable Mean SD p10 p50 p90 N Age 47.65 12.03 32 47 65 37,156 Experience (Macro) [in %] 6.00 0.28 5.67 5.97 6.37 37,156 Experience (Personal) [in %] 5.77 16.57 0.00 0.00 20.00 37,156 Household Size 2.73 1.45 1 2 5 37,156 Household Food Consumption [in $] 8,559 5,630 2,600 7,608 15,451 37,156 Household Total Consumption [in $] 46,256 36,497 14,733 39,559 82,765 37,156 Household Total Income [in $] 93k 133k 17k 69k 178k 37,156 Household Liquid Wealth [in $] 65k 718k -22k 0k 117k 37,156 Household Illiquid Wealth [in $] 282k 1,268k 0k 72k 606k 37,156 Household Total Wealth [in $] 346k 1,545k -3k 73k 762k 37,156 Notes. Summary statistics for the estimation sample, which covers the 1999-2013 PSID waves. Age, Experience(Macro), andExperience(Personal)arecalculatedfortheheadsofhouseholds. Householdtotal income includes transfers and taxable income of all household members from the last year. Liquid wealth andilliquidwealtharedefinedfollowingKaplan,ViolanteandWeidner(2014). Allvaluesarein2013dollars using the PCE. Observations are annual and not weighted. 70

Table A.2: Consumption (PSID), Full Sample (1) (2) (3) (4) (5) (6) Dependent Variable: Food Consumption Experience (Macro) -0.181*** -0.165*** -0.171** -0.163** (0.051) (0.050) (0.069) (0.069) Experience (Personal) -0.756*** -0.752*** -0.426*** -0.422*** (0.114) (0.114) (0.137) (0.137) R-squared 0.199 0.204 0.204 0.542 0.543 0.543 Dependent Variable: Total Consumption Experience (Macro) -0.059* -0.046 -0.079** -0.073** (0.031) (0.028) (0.031) (0.031) Experience (Personal) -0.603*** -0.602*** -0.328*** -0.326*** (0.073) (0.073) (0.082) (0.082) R-squared 0.496 0.507 0.507 0.755 0.757 0.757 Demographic controls Yes Yes Yes Yes Yes Yes Income controls Yes Yes Yes Yes Yes Yes Wealth controls Yes Yes Yes Yes Yes Yes Age fixed e↵ects Yes Yes Yes Yes Yes Yes State fixed e↵ects Yes Yes Yes Yes Yes Yes Year fixed e↵ects Yes Yes Yes Yes Yes Yes Household fixed e↵ects No No No Yes Yes Yes Observations 37,156 37,156 37,156 37,156 37,156 37,156 Notes. We include all observations i.e., also observations with total family income below the 5th or above the 95th percentile in each wave from 1999 to 2013, as well as the pre-sample 1997 wave (because we control for lagged income). All variables are defined as in Table 2. Robust standard errors (in parentheses) are clustered by cohort. *, **, *** denote 10%, 5%, and 1% significance, respectively. 71

Table A.3: Consumption (PSID), Alternative Experience Measure: Gap Years (1) (2) (3) (4) (5) (6) Dependent Variable: Food Consumption Experience (Macro) -0.099** -0.093* -0.124** -0.120** (0.047) (0.048) (0.055) (0.055) Experience (Personal) -0.337*** -0.335*** -0.267** -0.264** (0.104) (0.104) (0.127) (0.128) R-squared 0.192 0.193 0.193 0.541 0.542 0.542 Dependent Variable: Total Consumption Experience (Macro) -0.022 -0.018 -0.061*** -0.059*** (0.020) (0.019) (0.022) (0.022) Experience (Personal) -0.182*** -0.181*** -0.152*** -0.151*** (0.031) (0.031) (0.033) (0.033) R-squared 0.573 0.574 0.574 0.788 0.788 0.788 Demographic controls Yes Yes Yes Yes Yes Yes Income controls Yes Yes Yes Yes Yes Yes Wealth controls Yes Yes Yes Yes Yes Yes Age fixed e↵ects Yes Yes Yes Yes Yes Yes State fixed e↵ects Yes Yes Yes Yes Yes Yes Year fixed e↵ects Yes Yes Yes Yes Yes Yes Household fixed e↵ects No No No Yes Yes Yes Observations 33,164 33,164 33,164 33,164 33,164 33,164 Notes. All variables other than the experience measures are defined as in Table 2. The construction of the experience measures di↵ers as follows: For any gap year t (between PSID survey waves in t 1 and t+1), the baseline experience measures in the main text assume that families reside in the same state as in year t 1. ThealternativeconstructionusedinthisAppendix-Tableassumesthatfamiliesresidehalfofyeartin their(t-1)-stateofresidence,andhalfoftheyearintheir(t+1)-stateofresidence. (Thedi↵erentassumption does not matter when a family does not move between surveys.) Hence, the macro experience measure in this Appendix-Table uses the average of the year t unemployment rates of the (t-1)-state of residence and the(t+1)-stateresidenceasgapyeart’sunemploymentrate. Similarly, forthepersonalexperiencemeasure, we fill in the employment status of a household head in a gap year with the average of the years before and after. For example, if a person is unemployed in t 1 and is employed in t+1, then his personal experience in year t is denoted as 0.5. We exclude from the sample observations with total family income below the 5th or above the 95th percentile in each wave from 1999 to 2013, as well as the pre-sample 1997 wave (because we control for lagged income). Robust standard errors (in parentheses) are clustered by cohort. *, **, *** denote 10%, 5%, and 1% significance, respectively. 72

Table A.4: Consumption (PSID): Alternative Experience Measure: Spousal Experience (1) (2) (3) (4) (5) (6) Dependent Variable: Food Consumption Experience (Macro) -0.079* -0.071 -0.111** -0.106* (0.047) (0.047) (0.054) (0.055) Experience (Personal) -0.402*** -0.400*** -0.313** -0.309** (0.111) (0.111) (0.130) (0.130) R-squared 0.192 0.193 0.193 0.541 0.542 0.542 Dependent Variable: Total Consumption Experience (Macro) -0.021 -0.017 -0.059*** -0.056*** (0.019) (0.019) (0.021) (0.021) Experience (Personal) -0.213*** -0.212*** -0.161*** -0.159*** (0.035) (0.034) (0.033) (0.033) R-squared 0.573 0.574 0.574 0.788 0.788 0.788 Income controls Yes Yes Yes Yes Yes Yes Demographic controls Yes Yes Yes Yes Yes Yes Income controls Yes Yes Yes Yes Yes Yes Wealth controls Yes Yes Yes Yes Yes Yes Age fixed e↵ects Yes Yes Yes Yes Yes Yes State fixed e↵ects Yes Yes Yes Yes Yes Yes Year fixed e↵ects Yes Yes Yes Yes Yes Yes Household fixed e↵ects Yes Yes Yes Yes Yes Yes Observations 33,164 33,164 33,164 33,164 33,164 33,164 Notes. All variables other than the couple indicator, and experience measures are defined as in Table 2. Couple is an indicator equal to 1 for households who are married, and is now included as a demographic controle. The experience measures for the married households are constructed using an average of the household’sheadandthespouse. Weexcludefromthesampleobservationswithtotalfamilyincomebelow the 5th or above the 95th percentile in each wave from 1999 to 2013, as well as the pre-sample 1997 wave (because we control for lagged income). Robust standard errors (in parentheses) are clustered by cohort. *, **, *** denote 10%, 5%, and 1% significance, respectively. 73

Table A.5: Consumption (PSID), Alternative Experience Measure: Lagged Experience Dependent Varariable: Food Consumption Total Consumption (1) (2) (3) (4) (5) (6) Dependent Variable: Food Consumption Experience (Macro) -0.109* -0.103* -0.094** -0.093** (0.056) (0.057) (0.046) (0.046) Experience (Personal) -0.320** -0.318** -0.066 -0.064 (0.134) (0.134) (0.136) (0.136) R-squared 0.204 0.205 0.205 0.587 0.587 0.587 Dependent Variable: Total Consumption Experience (Macro) -0.013 -0.010 -0.047* -0.046* (0.024) (0.024) (0.026) (0.025) Experience (Personal) -0.179*** -0.178*** -0.120*** -0.119*** (0.038) (0.038) (0.038) (0.038) R-squared 0.572 0.573 0.573 0.806 0.807 0.807 Income controls Yes Yes Yes Yes Yes Yes Demographic controls Yes Yes Yes Yes Yes Yes Income controls Yes Yes Yes Yes Yes Yes Wealth controls Yes Yes Yes Yes Yes Yes Age fixed e↵ects Yes Yes Yes Yes Yes Yes State fixed e↵ects Yes Yes Yes Yes Yes Yes Year fixed e↵ects Yes Yes Yes Yes Yes Yes Household fixed e↵ects Yes Yes Yes Yes Yes Yes Observations 20,163 20,163 20,163 20,163 20,163 20,163 Notes. The experience measures (both macro and personal) does not contain unemployment experience from year t 1. All other variables are defined as in Table 2. We exclude from the sample observations with total family income below the 5th or above the 95th percentile in each wave from 1999 to 2013, as well as the pre-sample 1997 wave (because we control for lagged income). Robust standard errors (in parentheses) are clustered by cohort. *, **, *** denote 10%, 5%, and 1% significance, respectively. 74

Table A.6: Consumption (PSID), Alternative Experience Measure: Di↵erent Weights ( = 3) Dependent Varariable: Food Consumption Total Consumption (1) (2) (3) (4) (5) (6) Experience (Macro) -0.021** -0.019* -0.033*** -0.032** (0.010) (0.011) (0.012) (0.012) Experience (Personal) -0.175*** -0.174*** -0.150*** -0.149*** (0.029) (0.029) (0.031) (0.031) R-squared 0.573 0.574 0.574 0.788 0.788 0.788 Demographic controls Yes Yes Yes Yes Yes Yes Income controls Yes Yes Yes Yes Yes Yes Wealth controls Yes Yes Yes Yes Yes Yes Age fixed e↵ects Yes Yes Yes Yes Yes Yes State fixed e↵ects Yes Yes Yes Yes Yes Yes Year fixed e↵ects Yes Yes Yes Yes Yes Yes Household fixed e↵ects Yes Yes Yes Yes Yes Yes Observations 33,164 33,164 33,164 33,164 33,164 33,164 Notes. Allvariablesother than the experience measures are defined as inTable 2. The experience measures areconstructedusing=0intheupperpartofthetable,and=3inthelowerpart. Weexcludefromthe sample observations with total family income below the 5th or above the 95th percentile in each wave from 1999 to 2013, as well as the pre-sample 1997 wave (because we control for lagged income). Robust standard errors(inparentheses)areclusteredbycohort. *,**,***denote10%,5%,and1%significance,respectively. 75

stinU gniretsulC evitanretlA ,)DISP( noitpmusnoC :7.A elbaT )8( )7( )6( )5( )4( )3( )2( )1( noitpmusnoC dooF :elbairaV tnednepeD **711.0- **711.0- 711.0- **711.0- **190.0- **190.0- **190.0- **190.0- )orcaM( ecneirepxE )940.0( )050.0( )260.0( )350.0( )240.0( )640.0( )330.0( )140.0( **062.0- ***062.0- *062.0- **062.0- ***023.0- ***023.0- **023.0- ***023.0- )lanosreP( ecneirepxE )101.0( )990.0( )231.0( )901.0( )580.0( )590.0( )311.0( )680.0( 245.0 245.0 245.0 245.0 391.0 391.0 391.0 391.0 derauqs-R noitpmusnoC latoT :elbairaV tnednepeD ***750.0- ***750.0- **750.0- ***750.0- 810.0- 810.0- 810.0- 810.0- )orcaM( ecneirepxE )710.0( )910.0( )020.0( )710.0( )510.0( )710.0( )710.0( )510.0( ***741.0- ***741.0- **741.0- ***741.0- ***771.0- ***771.0- ***771.0- ***771.0- )lanosreP( ecneirepxE )720.0( )030.0( )340.0( )820.0( )320.0( )620.0( )640.0( )420.0( 887.0 887.0 887.0 887.0 475.0 475.0 475.0 475.0 derauqs-R seY seY seY seY seY seY seY seY slortnoc cihpargomeD seY seY seY seY seY seY seY seY slortnoc emocnI seY seY seY seY seY seY seY seY slortnoc htlaeW seY seY seY seY seY seY seY seY stce↵e dexfi egA seY seY seY seY seY seY seY seY stce↵e dexfi raeY seY seY seY seY seY seY seY seY stce↵e dexfi etatS seY seY seY seY oN oN oN oN stce↵e dexfi dlohesuoH 461,33 461,33 461,33 461,33 461,33 461,33 461,33 461,33 snoitavresbO raey dna trohoc ,raey trohoc yb deretsulc era )4( ot )1( snmuloc ni srorre dradnatS .2 elbaT ni sa denfied era selbairav llA .setoN ⇤ elpmas eht morf edulcxe eW .)8( ot )5( snmuloc rof emas eht dna ,ylevitcepser ,raey dlohesuoh dna dlohesuoh ,)gniretsulc yaw-owt( ⇤ eht sa llew sa ,3102 ot 9991 morf evaw hcae ni elitnecrep ht59 eht evoba ro ht5 eht woleb emocni ylimaf latot htiw snoitavresbo ,** ,* .trohoc yb deretsulc era )sesehtnerap ni( srorre dradnats tsuboR .)emocni deggal rof lortnoc ew esuaceb( evaw 7991 elpmas-erp .ylevitcepser ,ecnacfiingis %1 dna ,%5 ,%01 etoned *** 76

Table A.8: Consumption (PSID), Alternative Weights: PSID Weights (1) (2) (3) (4) (5) (6) Dependent Variable: Food Consumption Experience (Macro) 0.016 -0.093* -0.119** -0.115** (0.056) (0.047) (0.054) (0.054) Experience (Personal) -0.302*** -0.324*** -0.262** -0.260** (0.112) (0.098) (0.120) (0.120) R-squared 0.225 0.226 0.193 0.541 0.541 0.541 Dependent Variable: Total Consumption Experience (Macro) -0.003 -0.021 -0.058*** -0.056** (0.023) (0.019) (0.021) (0.022) Experience (Personal) -0.162*** -0.176*** -0.150*** -0.149*** (0.042) (0.030) (0.032) (0.032) R-squared 0.576 0.576 0.574 0.787 0.788 0.788 Demographic controls Yes Yes Yes Yes Yes Yes Income controls Yes Yes Yes Yes Yes Yes Wealth controls Yes Yes Yes Yes Yes Yes Age fixed e↵ects Yes Yes Yes Yes Yes Yes State fixed e↵ects Yes Yes Yes Yes Yes Yes Year fixed e↵ects Yes Yes Yes Yes Yes Yes Household fixed e↵ects No No No Yes Yes Yes Observations 32,834 32,834 32,834 32,834 32,834 32,834 Notes. All variables are defined as in Table 2, but observations are now weighted by the PSID family weights. The family with zero weights are dropped. Robust standard errors (in parentheses) are clustered by cohort. We exclude from the sample observations with total family income below the 5th or above the 95th percentileineachwavefrom1999to2013,aswellasthepre-sample1997wave(becausewecontrolfor lagged income). Robust standard errors (in parentheses) are clustered by cohort. *, **, *** denote 10%, 5%, and 1% significance, respectively. 77

slortnoC htlaeW lanoitiddA ,)DISP( noitpmusnoC :9.A elbaT )8( )7( )6( )5( )4( )3( )2( )1( noitpmusnoC dooF :elbairaV tnednepeD *690.0- *401.0- *101.0- **011.0- 280.0- *280.0- 970.0- *880.0- )orcaM( ecneirepxE )550.0( )450.0( )450.0( )450.0( )150.0( )840.0( )740.0( )740.0( *622.0- **162.0- **852.0- **852.0- **132.0- ***123.0- ***272.0- ***813.0- )lanosreP( ecneirepxE )021.0( )911.0( )021.0( )911.0( )890.0( )890.0( )990.0( )790.0( 155.0 245.0 345.0 245.0 302.0 491.0 991.0 491.0 derauqs-R noitpmusnoC latoT :elbairaV tnednepeD *040.0- **350.0- **640.0- **450.0- 300.0- 310.0- 500.0- 510.0- )orcaM( ecneirepxE )020.0( )220.0( )120.0( )120.0( )020.0( )910.0( )910.0( )910.0( ***031.0- ***841.0- ***441.0- ***641.0- ***960.0- ***871.0- ***131.0- ***571.0- )lanosreP( ecneirepxE )030.0( )130.0( )030.0( )130.0( )420.0( )920.0( )620.0( )920.0( 508.0 887.0 197.0 887.0 336.0 575.0 695.0 775.0 derauqs-R seY seY seY seY seY seY seY seY slortnoc cihpargomeD seY seY seY seY seY seY seY seY slortnoc htlaeW seY seY seY seY seY seY seY seY stce↵e dexfi egA seY seY seY seY seY seY seY seY stce↵e dexfi raeY seY seY seY seY seY seY seY seY stce↵e dexfi etatS seY seY seY seY oN oN oN oN stce↵e dexfi dlohesuoH 781,13 461,33 461,33 461,33 781,13 461,33 461,33 461,33 snoitavresbO redro-htruof dna -driht rof slortnoc )1( nmuloC .slortnoc htlaew eht fo smret ni ylno 2 elbaT ni esoht morf re↵id snoissergeR .setoN gnisuoh rof slortnoc )3( nmuloC .htlaew diuqilli dna htlaew diuqil fo seimmud eliced sedulcni )2( nmuloC .htlaew diuqilli dna diuqil slortnoc htlaew llA .tbed dna htlaew evitisop rof slortnoc )4( nmuloC .)htlaew gnisuoh sunim htlaew latot( htlaew rehto dna htlaew slortnoc htlaew emas eht evah )8( – )5( snmuloC .htlaew diuqilli dna diuqil fo redro dnoces dna tsrfi fo slortnoc eht ot noitidda ni era ,ecnacfiingis %1 dna ,%5 ,%01 etoned *** ,** ,* .trohoc yb deretsulc era srorre dradnats tsuboR .ylevitcepser )4( – )1( snmuloc sa .ylevitcepser 78

slortnoC emocnI lanoitiddA ,)DISP( noitpmusnoC :01.A elbaT )8( )7( )6( )5( )4( )3( )2( )1( noitpmusnoC dooF :elbairaV tnednepeD **311.0- **811.0- **611.0- **811.0- *980.0- **890.0- **590.0- **790.0- )orcaM( ecneirepxE )550.0( )650.0( )550.0( )550.0( )840.0( )840.0( )740.0( )740.0( *332.0- *232.0- **342.0- **042.0- **722.0- **832.0- ***472.0- ***762.0- )lanosreP( ecneirepxE )121.0( )121.0( )021.0( )911.0( )501.0( )101.0( )890.0( )990.0( 445.0 445.0 345.0 345.0 602.0 402.0 002.0 002.0 derauqs-R noitpmusnoC latoT :elbairaV tnednepeD **650.0- **650.0- **650.0- **750.0- 710.0- 710.0- 810.0- 020.0- )orcaM( ecneirepxE )120.0( )120.0( )220.0( )220.0( )910.0( )910.0( )910.0( )910.0( ***831.0- ***631.0- ***041.0- ***931.0- ***141.0- ***241.0- ***451.0- ***351.0- )lanosreP( ecneirepxE )130.0( )130.0( )130.0( )130.0( )920.0( )920.0( )920.0( )920.0( 987.0 987.0 887.0 987.0 185.0 185.0 975.0 975.0 derauqs-R seY seY seY seY seY seY seY seY slortnoc cihpargomeD seY seY seY seY seY seY seY seY slortnoc htlaeW seY seY seY seY seY seY seY seY stce↵e dexfi egA seY seY seY seY seY seY seY seY stce↵e dexfi raeY seY seY seY seY seY seY seY seY stce↵e dexfi etatS seY seY seY seY oN oN oN oN stce↵e dexfi dlohesuoH 461,33 461,33 461,33 461,33 461,33 461,33 461,33 461,33 snoitavresbO fo redro htruof dna driht rof slortnoc )1( nmuloC .slortnoc emocni eht fo smret ni ylno 2 elbaT ni esoht morf re↵id snoissergeR .setoN seimmud eliced sedulcni )3( nmuloC .emocni deggal dna emocni fo seimmud elitniuq sedulcni )2( nmuloC .emocni deggal dna emocni ,dn29 ht09 ,ht01 ht8 ,ht8 ht6 ,ht6 ht4 ,ht4 dn2 ,2 mottob eht rof yletarapes sedulcni )4( nmuloC .emocni deggal dna emocni fo noitidda ni era slortnoc emocni llA .emocni deggal dna emocni fo seimmud elitnecrep 2 pot dna ,ht89 ht69 ,ht69 ht49 ,ht49 dn29 )1( snmuloc sa slortnoc emocni emas eht evah )8( – )5( snmuloC .emocni deggal dna emocni fo redro dnoces dna tsrfi fo slortnoc eht ot .ylevitcepser ,ecnacfiingis %1 dna ,%5 ,%01 etoned *** ,** ,* .trohoc yb deretsulc era srorre dradnats tsuboR .ylevitcepser )4( – 79

Table A.11: Consumption (PSID), Additional Liquidity Controls (1) (2) (3) (4) (5) (6) Dependent Variable: Food Consumption Experience (Macro) -0.147** -0.144** -0.143** -0.143** (0.058) (0.058) (0.063) (0.063) Experience (Macro) * LLW 0.097 0.103* 0.048 0.055 (0.060) (0.059) (0.056) (0.055) Low Liquid Wealth -0.572 0.013 -0.602* -0.316 -0.023 -0.352 (0.358) (0.013) (0.355) (0.335) (0.014) (0.331) Experience (Personal) -0.302** -0.292** -0.241 -0.236 (0.142) (0.142) (0.149) (0.149) Experience (Personal) * LLW -0.037 -0.053 -0.038 -0.046 (0.177) (0.176) (0.156) (0.155) R-squared 0.192 0.193 0.193 0.542 0.542 0.542 Dependent Variable: Total Consumption Experience (Macro) -0.021 -0.023 -0.054** -0.055** (0.022) (0.022) (0.024) (0.023) Experience (Macro) * LLW 0.001 0.009 -0.012 -0.006 (0.015) (0.016) (0.015) (0.016) Low Liquid Wealth 0.034 0.047*** -0.006 0.080 0.013*** 0.046 (0.093) (0.006) (0.094) (0.093) (0.004) (0.095) Experience (Personal) -0.087** -0.086** -0.083** -0.082** (0.036) (0.036) (0.033) (0.033) Experience (Personal) * LLW -0.166*** -0.168*** -0.118*** -0.118*** (0.046) (0.046) (0.044) (0.044) R-squared 0.573 0.575 0.575 0.788 0.788 0.788 Demographic controls Yes Yes Yes Yes Yes Yes Income controls Yes Yes Yes Yes Yes Yes Wealth controls Yes Yes Yes Yes Yes Yes Age fixed e↵ects Yes Yes Yes Yes Yes Yes State fixed e↵ects Yes Yes Yes Yes Yes Yes Year fixed e↵ects Yes Yes Yes Yes Yes Yes Household fixed e↵ects No No No Yes Yes Yes Observations 33,164 33,164 33,164 33,164 33,164 33,164 Notes. Low Liquid Wealth (LLW) is an indicator variable equal to 1 for households with liquid wealth below the sample-year median. All other variables are defined as in Table 2. We exclude from the sample observationswithtotalfamilyincomebelowthe5th orabovethe95th percentileineachwavefrom1999to 2013, as well as the pre-sample 1997 wave (because we control for lagged income). Robust standard errors (in parentheses) are clustered by cohort. *, **, *** denote 10%, 5%, and 1% significance, respectively. 80

ytilitaloV emocnI erutuF dna stce↵E ecneirepxE :21.A elbaT emocnI fo ecnairaV :elbairaV tnednepeD )6( )5( )4( )3( )2( )1( yrotisnarT tnenamreP yrotisnarT tnenamreP yrotisnarT tnenamreP 6+t 6+t 4+t 4+t 2+t 2+t 161.0 070.0 041.0 930.0 520.0 *480.0 )orcaM( ecneirepxE )401.0( )860.0( )590.0( )150.0( )470.0( )940.0( 260.0- 831.0 571.0- 330.0 130.0- 211.0 )lanosreP( ecneirepxE )781.0( )011.0( )751.0( )431.0( )951.0( )461.0( 317,51 691,11 280,12 566,51 014,82 566,51 snoitavresbO 842.0 734.0 973.0 983.0 463.0 014.0 derauqs-R seY seY seY seY seY seY slortnoc cihpargomeD seY seY seY seY seY seY slortnoc emocnI seY seY seY seY seY seY slortnoc htlaeW seY seY seY seY seY seY stce↵e dexfi etatS seY seY seY seY seY seY stce↵e dexfi raeY seY seY seY seY seY seY stce↵e dexfi dlohesuoH 375,21 614,8 270,71 102,21 060,42 102,21 snoitavresbO 344.0 914.0 254.0 593.0 433.0 693.0 derauqs-R ecneirepxE” .ylevitcepser ,sraey xis dna ,ruof ,owt ni ytilitalov emocni yrotisnart dna tnenamrep era selbairav tnedneped ehT .setoN .erusaem ecneirepxe lanosrep eht si ”)lanosreP( ecneirepxE” dna ,tnemyolpmenu fo erusaem ecneirepxe cimonoceorcam eht si ”)orcaM( rehtehw gnitacidni elbairav ymmud a dna ,level noitacude ,sutats latiram ,ecar ,redneg ’sdaeh ,ezis ylimaf edulcni slortnoc cihpargomeD emocni fo mhtiragol eht fo redro dnoces dna tsrfi eht edulcni slortnoc emocnI .yevrus eht fo emit eht ta deyolpmenu si tnednopser eht morf edulcxe eW .htlaew diuqilli dna diuqil fo mhtiragol eht fo redro dnoces dna tsrfi eht edulcni slortnoc htlaeW .emocni deggal dna sa llew sa ,3102 ot 9991 morf evaw hcae ni elitnecrep ht59 eht evoba ro ht5 eht woleb emocni ylimaf latot htiw snoitavresbo elpmas eht era seulav evitisop-non ;htlaew dna ,emocni fo mhtiragol eht ekat eW .)emocni deggal rof lortnoc ew esuaceb( evaw 7991 elpmas-erp eht era )sesehtnerap ni( srorre dradnats tsuboR .dezimhtiragol gnieb erofeb 1.0 sulp muminim eht fo eulav etulosba eht gnidda yb detsujda .ylevitcepser ,ecnacfiingis %1 dna ,%5 ,%01 etoned *** ,** ,* .trohoc yb deretsulc 81

noitalumuccA htlaeW :31.A elbaT t emit ta htlaeW latoT t emit ta htlaeW diuqiL :.raV tnednepeD )21( )11( )01( )9( )8( )7( )6( )5( )4( )3( )2( )1( ***910.0 ***810.0 010.0 210.0 300.0 300.0 500.0 *600.0 )orcaM( .pxE 01 t )600.0( )600.0( )800.0( )800.0( )300.0( )300.0( )300.0( )300.0( 500.0- 300.0- ***380.0 ***380.0 100.0- 000.0- ***320.0 ***320.0 )lanosreP( .pxE 01 t )410.0( )410.0( )310.0( )310.0( )200.0( )200.0( )400.0( )400.0( 417.0 417.0 417.0 492.0 492.0 292.0 233.0 233.0 233.0 840.0 840.0 840.0 derauqs-R 196,12 196,12 196,12 196,12 196,12 196,12 196,12 196,12 196,12 196,12 196,12 196,12 snoitavresbO ***020.0 ***020.0 700.0 010.0 **700.0 **800.0 *600.0 **700.0 )orcaM( .pxE 21 t )700.0( )700.0( )900.0( )900.0( )300.0( )300.0( )300.0( )300.0( 100.0 300.0 ***190.0 ***290.0 100.0 200.0 ***620.0 ***620.0 )lanosreP( .pxE 21 t )410.0( )410.0( )410.0( )410.0( )300.0( )200.0( )500.0( )500.0( 037.0 037.0 037.0 692.0 692.0 492.0 333.0 333.0 333.0 050.0 050.0 940.0 derauqs-R 724,91 724,91 724,91 724,91 724,91 724,91 724,91 724,91 724,91 724,91 724,91 724,91 snoitavresbO 010.0 *110.0 100.0- 200.0 **800.0 **800.0 *700.0 **800.0 )orcaM( .pxE 41 t )600.0( )600.0( )010.0( )900.0( )300.0( )300.0( )300.0( )300.0( 900.0 010.0 ***590.0 ***590.0 200.0 300.0 ***820.0 ***820.0 )lanosreP( .pxE 41 t )900.0( )900.0( )310.0( )310.0( )300.0( )300.0( )500.0( )500.0( 728.0 728.0 728.0 083.0 083.0 873.0 133.0 133.0 133.0 250.0 250.0 250.0 derauqs-R 151,71 151,71 151,71 151,71 151,71 151,71 151,71 151,71 151,71 151,71 151,71 151,71 snoitavresbO seY seY seY seY seY seY seY seY seY seY seY seY slortnoc cihpargomeD seY seY seY seY seY seY seY seY seY seY seY seY slortnoc emocnI seY seY seY seY seY seY seY seY seY seY seY seY stce↵e dexfi egA seY seY seY seY seY seY seY seY seY seY seY seY stce↵e dexfi etatS seY seY seY seY seY seY seY seY seY seY seY seY stce↵e dexfi raeY seY seY seY oN oN oN seY seY seY oN oN oN stce↵e dexfi dlohesuoH latot dna htlaew diuqiL .erusaem ecneirepxe lanosrep eht si ”)lanosreP( .pxE“ dna ,erusaem ecneirepxe cimonoceorcam eht si ”)orcaM( .pxE“ .setoN edulcni slortnoc emocnI .serusaem ecneirepxe 41 t dna ,serusaem ecneirepxe 21 t ,01 t eht esu yletarapes eW .tfard niam eht ni sa denfied era htlaew DISP neewteb sraey pag roF .serusaem ecneirepxe eht esu ew raey eht dna 2 t neewteb emocni latot ylimaf egareva eht dna emocni latot ylimaf 1 t eht dna ,level noitacude ,sutats latiram ,ecar ,redneg ’sdaeh dlohesuoh eht ,ezis ylimaf edulcni slortnoc cihpargomeD .emocni raey-roirp esu ew ,sevaw yevrus *** ,** ,* .trohoc yb deretsulc era )sesehtnerap ni( srorre dradnats tsuboR .selbairav htlaew dna emocni lla fo mhtiragol eht ekat eW .sutats tnemyolpme .ylevitcepser ,ecnacfiingis %1 dna ,%5 ,%01 etoned 82

A.2 Robustness using CEX-Nielsen Synthetic Panel Data In order to keep the advantages of panel analysis but also exploit the comprehensivenessoftheCEX,wematchthetwodatasetsandcreateasyntheticpanel. Specifically, we match a CEX household i with a Nielsen household j on a set of common covariates (characteristics) x = (x ,x ,...,x ) and x = (x ,x ,...,x ), which i i,1 i,2 i,p j j,1 j,2 j,p include age, income, marital status, household size, education, race, region of residency, employment status, as well as their consumption of non-durable items, using the nearest-neighbor matching estimator from Rosenbaum and Rubin (1983) and Abadie and Imbens (2011). The distance between x and x is the vector norm i j x x = ((x x )S 1(x x ))1/2, where S is a symmetric, positive-definite i j S i j 0 i j k k matrix. We find the set of nearest-neighbor indices for observation i in the CEX as ⌦ = (j t = 1 t , x x < x x ,t = 1 t ,l = j) in Nielsen. In words, i j i i j S i l S l i | k k k k 6 the nearest-neighbor propensity-score matching chooses for each observation in the CEX an observation in Nielsen that has the closest estimated propensity score. Table A.14 provides summary statistics for the matched sample. In the matched dataset, the distributions on total and durable consumption are comparable to those of the underlying CEX data, which is indicative of successful matching. For an average household, its share of durable consumption makes up 10% of total spending, while non-durable consumption amounts to 69% of total spending. Table A.14: Summary Statistics (Nielsen-CEX Matched Data) Variable Mean SD p10 p50 p90 N Total consumption expenditure 4,508 4,919 1,838 3,371 7,111 866,819 Durable consumption 1,078 4,466 0 117 1,460 866,819 Non-durable consumption 2,612 1,178 1,423 2,400 4,025 866,819 Non-durable consumption (Nielsen) 2,139 1,602 618 1,757 4,083 3,171,833 Experience (Macro) 5.9 0.2 5.8 5.9 6.2 866,819 Notes. The sample period runs from 2004 to 2012. Observations are quarterly and not weighted. Table A.15 shows results from re-estimating specification (11) using the matched CEX-Nielsen sample. In columns (1) and (4), we use total expenditures as the outcome variable, in columns (2) and (5) durable consumption spending, and in columns (3) and (6) non-durables. As before we show the results both without household fixed e↵ects (columns 1 to 3) and with fixed e↵ects (columns 4 to 6). 83

)elpmaS dehctaM XEC-nesleiN( noitpmusnoC :51.A elbaT selbarud-noN selbaruD latoT selbarud-noN selbaruD latoT 330.0- ***697.0- ***662.0- **022.0- ***797.0- ***853.0- )orcaM( ecneirepxE )820.0( )541.0( )150.0( )910.0( )221.0( )830.0( seY seY seY seY seY seY lortnoc emocnI seY seY seY seY seY seY slortnoc dlohesuoH seY seY seY seY seY seY stce↵e dexfi egA seY seY seY seY seY seY stce↵e dexfi emiT seY seY seY seY seY seY stce↵e dexfi aera-tekraM seY seY seY oN oN oN stce↵e dexfi dlohesuoH 918,668 918,668 918,668 918,668 918,668 918,668 snoitavresbO 960.0 800.0 020.0 752.0 350.0 381.0 derauqs-R tnedneped eht sa noitpmusnoc elbarud-non dna ,noitpmusnoc elbarud ,erutidnepxe noitpmusnoc latot )gol( htiw snoissergeR .setoN lanoitan fo ecneirepxe emitefil s’dlohesuoh( tnemyolpmenu fo erusaem ecneirepxe cimonoceorcam eht si )orcaM( ecneirepxE .selbairav stce↵e dexfi emiT .ecar dna ,noitacude ,ezis dlohesuoh ,sutats tnemyolpmenu edulcni scitsiretcarahc dlohesuoH .)setar tnemyolpmenu morf snur doirep elpmas ehT .nesleiN morf sthgiew gnilpmas dlohesuoh yb dethgiew era snoissergeR .stce↵e dexfi retrauq-raey edulcni .ylevitcepser,ecnacfiingis%1dna,%5,%01etoned***,**,* .trohocybderetsulcera)sesehtnerapni(srorredradnatstsuboR .2102ot4002 84

For all outcome variables – durable, non-durable, and total consumption – we continue to estimate highly significant negative experience e↵ects. Households that have experienced worse unemployment conditions during their lifetime spend significantly less in total, and also specifically on durable and on non-durable items. One exception is non-durables in the case where we identify only within household; here the coecient becomes small and insignificant. Otherwise, the coecients are stable across specifications, and the economic magnitudes are large: a one standard deviation increase in lifetime unemployment experience is associated with a $38 decline in monthly non-durable consumption and $108 decline in monthly total consumption (using the estimates of columns 3 and 1 respectively). The new estimate for durable consumption is large and highly significant across specifications. A one standard deviation increase in lifetime unemployment experience is associated with a $57 decline in monthly durable consumption. Appendix B Model We implement the empirical model of Low, Meghir, and Pistaferri (2010) with a few minor adjustments to our setting. All key equations are retained and, when possible, all parameters are set to the same values. As in Low et al., some parameters are set separately for high- and low-education groups, including the probability of job destruction and job o↵ers. B.1 Parameters governing the income process and utility maximization The utility function and lifetime expected utility are defined in equations (4) and (5) in Section IV as U(c,P) = (c ⇥ e⌘P)1 and U(c ,P )+E L s tU(c ,P ) , 1 i,t i,t t s=t+1 i,s i,s respectively. Inthesimulations,wefollowLowetal. andtakheriskaversionparameteir P = 1.5 from Attanasio and Weber (1995), use the estimates for ⌘ from their Table 2, and set the discount factor = 1/R in the value function. For the gross quarterly income w h, we also follow Low et al. in setting the i,t number of hours worked per quarter to h = 500. In the wage process lnw = i,t d + x + u + a , we recover the parameters ↵, , and governing the t 0i,t i,t i,j,t0 1 2 85

deterministic component, d +x = ↵+ age+ age2, from the parameters t 0i,t 1 · 2 · in the Fortran code published alongside Low et al. In the permanent component u = u +⇣ , ⇣ is i.i.d. normal with mean 0 and variance 2, and we use the i,t i,t 1 i,t i,t ⇣ value of given in Table 1 of Low et al.. The consumer-firm job match component, ⇣ a , is drawn from a normal distribution with mean 0 and variance 2, and we use i,j,t0 a the value of given in Table 1 of Low et al.. a We obtain the values for the probabilities of job destruction , of a job o↵er when employed (1 )e, and of a job o↵er when unemployed n from Table 2 in Low et al. (2010). Note that, while the probability of job destruction is constant across time for a given household, the probability of receiving a job o↵er varies depending on whether or not an agent is employed. B.2 Budget constraint The intertemporal budget constraint for a working individual i in period t is given by A = R[A c ]+(w h(1 ⌧ ) F )P i,t+1 i,t i,t i,t w i,t i,t + B IUI(1 IDI)+D IDI (1 P )+T IT i,t i,t i,t i,t i,t i,t i,t i,t where A is beginning-of-period-t assets, R is the interest factor, ⌧ a tax, F the i,t w fixed cost of working, P an indicator for whether an individual is working, B are unemployment benefits, D disability benefits, T food stamp benefits, c is consumption, and the I variables are indicators of receiving the associated social insurance. As in Low et al. (2010), we assume that individuals cannot borrow and thus A 0 t. Also as in Low et al. (2010), we set r = .15 and define R = 1+r. We i,t 8 use the estimates for F from their Table 2. In Low et al. (2010), ⌧ is a variable of w interest and solved for, albeit as fixed percentage (not progressive or regressive). As we do not focus on the value of social insurance programs, including the tax revenues to be raised to fund them and their relation with consumption, we normalize ⌧ = 0. w During retirement individuals receive social security equal to the value of disability, so the budget constraints simplifies to A = R[A +D c ]. i,t+1 i,t i,t i,t 86

B.3 Social Insurance programs As in Low et al. (2010), we implement three social insurance programs, unemployment insurance, food stamps, and disability insurance. Unemployment Insurance. Unemployment Insurance is paid only during the quarter following job destruction. Unemployment benefits are given by bw h if bw h < B , i,t 1 i,t 1 max B = i,t 8 B if bw h B . < max i,t 1 max where b is the replacement rat:io, and B is the cap on unemployment benefits. We max set b = .75 as in Low et al. (2010) and B to the value used in the associated code. max Food Stamps (Means-Tested Social Insurance). Defining gross income as ygross = w hP + B IUI(1 IDI)+D IDI (1 P ), i,t i,t i,t i,t i,t i,t i,t i,t i,t and net income as y = (1 ⌧ )ygross d, w the amount of food stamps allocated to agent i in period t is T .3 y ify y i,t i,t T = ⇥  i,t 8 0 otherwise, < : where T is a maximum payment and y is a poverty line. One important implication of this definition is that there is no disincentive to hold assets. Adjusting to quarterly values, we set T to the maximum food stamp allotment for a couple in the US in 1993, y to the maximum food stamp allotment for the US in 1993, and d to the standard deduction for a couple in the US in 1993. Disability. As in Low et al. (2010), individuals above 50 can apply for disability when they are unemployed, and are accepted with a fixed probability of .5. If an application is successful, disability becomes an absorbing state for the remainder of the person’s working life. If a person is not accepted, they can only reapply in a future bout of unemployment, after having worked again for at least one year. As a 87

disincentive to applying, the individual must be unemployed in both the period they apply and the period after. We also impose that individuals must have a suciently low u and not be working or have a job o↵er at the time of application. The formula for disability benefits is .9 w if w a i i 1 ⇥  8 .9 a +.32 (w a ) if a < w a D = > > ⇥ 1 ⇥ i 1 1 i  2 i,t > > > > < .9 ⇥ a 1 +.32 ⇥ (a 2 a 1 )+.15 ⇥ (w i a 2 ) if a 2 < w i  a 3 .9 a +.32 (a a )+.15 (a a ) if w > a > > ⇥ 1 ⇥ 2 1 ⇥ 3 2 i 3 > > > > wherea , a , an:da arefixedthresholdsfromlegislation, and w isthemeanearnings 1 1 3 i priortoapplication. SimilartoLowetal. (2010), weassumew canbeapproximated i using the agent’s value of u at the time of application. i,t B.4 Implementation Appendix-Table B.1 details all parameters referenced above and their sources. As discussed, most values are obtained directly from Low et al. (2010), and some are retrieved from examining the associated Fortran 90 code published with the paper. In cases where we were unable to ascertain values in either source, as is the case for several welfare values, we use actual values from 1993, the year in which the SIPP survey used in Low et al. for hourly wage data begins. This is also the closest year in the SIPP survey to the PSID data, and the values are consistent with the model values. When we combine the high- and low-education data, we use 70% low- and 30% high-education observations, roughly corresponding to recent US census estimates of those without and with a bachelor’s degree.36 Like Low et al. (2010), we solve the model numerically. In the last period, all agents consume the entirety of their assets. We then iteratively solve backwards for consumption and other relevant decisions that maximize the agents’ value functions. Further details of the model solution can be found in Low et al. (2010). 36 The percent of the US population with at least a bachelor’s degree has increased over the last three decades. It was closer to 25% in 2007 and 20% in 1995. We opted for the more recent estimates to err, if anything, on the side of a greater inclusion of high-education individuals. 88

Table B.1: Model Parameters Used in Simulations Parameter Low Education High Education Source (from Low, Meghir, and Pistaferri (2010)) 1.5 1.5 Text 0.226 0.229 Table 1 a 0.095 0.106 Table 1 ⇣ P(⇣) .25 .25 Text .049 .028 Table 2 e .67 .72 Table 2 n .76 .82 Table 2 b .75 .75 Text r (yearly) .015 .015 Text 1/(1+r) 1/(1+r) Text F 1088 1213 Table 2 ⌘ -.55 -.62 Table 2 h 500 500 Text b .75 .75 Text UI Cap 3178 3178 Code P(Disability .5 .5 Text Acceptance) a 1203 1203 Code 1 a 7260 7260 Code 2 a 16638 16638 Code 3 ↵ 1.0583 .642 Code .0486 .0829 Code 1 -0.0004816 -.0007768 Code 2 Parameter Low Education High Education Source d 6200/4 Standard couple deduction in 1993a y (6970+2460)/4 Actual poverty line in 1993 for coupleb T 203 3 Actual max food stamp al- ⇥ lotment for US 1993c a See https://web.archive.org/web/20190228193856/https://www.irs.gov/pub/irs-pr ior/f1040a--1993.pdf. b See https://web.archive.org/web/20190228194017/https://aspe.hhs.gov/prior-hhs -poverty-guidelines-and-federal-register-references. c See https://web.archive.org/web/20190228193653/https://fns-prod.azureedge.net /sites/default/files/Trends1999-2005.pdf. Accessed via https://web.archive.org/web/ 20190228195514/https://www.fns.usda.gov/snap/trends-food-stamp-program-participat ion-rates-1999-2005. 89

Figure B.1: Average Life-Cycle Consumption Notes. Average consumption for rational learners and experience-based learners (with = 1) in the low-education group, based on 10,000 lifetime simulations for each type. 90

Figure B.2: Average Life-Cycle Consumption for Agents with Good Realizations Early in Life Notes. Average consumption for rational learners and experience-based learners (with = 1) in the low-education group, based on 10,000 lifetime simulations for each type and then restricted to those simulations where agents have, or in the rational case would have, a believed delta of 0.025 or less at period 30. 91

Figure B.3: Average Life-Cycle Consumption Patterns for Agents with Bad Realizations Early in Life Notes. Average consumption for rational learners and experience-based learners (with = 1) in the low-education group, based on 10,000 lifetime simulations for each type and then restricted to those simulations where agents have, or in the rational case would have, a believed delta of 0.1 or greater at period 30. 92

Figure B.1 depicts the resulting average consumption trends of rational and experience-based learners during their working years, which are the years used in the regressions. The graph hints at a pattern that, early in life, experience-based learners underestimate theprobability ofjob destruction, spend more, and must then save more towards the end of their working life. Figure B.2 provides an amplified illustration of the di↵erences. In this figure, we only consider the subset of experience-based learners in the simulation who, at period 30, have a believed delta of 0.025 or less and, in the rational case, the subset of agents who would have a believed delta of 0.025 or less at period 30 if they were experience-basedlearners. Sincethetrueprobabilityofjobdestructionis0.049, these agents were “lucky” early in life. For these consumers, the trend of over-consumption among experienced based learners in the early periods is much more pronounced. Figure B.3 illustrates the opposite scenario. Here, we only consider the subset of experience-based learners in the simulation who, at period 30, have a believed delta of 0.1 or greater, as well as the corresponding rational agents. In light of the true probability of job destruction of 0.049, these agents have had bad luck early in life. This “unlucky” group of experience-based learners has a markedly di↵erent savings pattern. They consistently consume less than their rational counterparts for almost their entire lives. Moreover, the illustration hints at an additional prediction, wealth build-up due to excess frugality . 93