Small Price Responses to Large Demand Shocks

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Small Price Responses to Large Demand Shocks Etienne Gagnon and David Lopez-Salido 2014-18 NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Small Price Responses to Large Demand Shocks (cid:3) Etienne Gagnon David L(cid:243)pez-Salido Federal Reserve Board Federal Reserve Board February 11, 2014 Abstract We study the pricing response of U.S. supermarkets to large demand shocks triggered by labor con(cid:135)icts, mass population relocation, and shopping sprees around major snowstorms and hurricanes. Our focus on demand shocks is novel in the empirical literature that uses large datasets of individual data to bridge micro price behavior and aggregate price dynamics. We (cid:133)nd that large swings in demand have, at best, modest e⁄ects on the level of retail prices, consistent with (cid:135)at short- to medium-term supply curves. This (cid:133)nding holds even when shocks are highly persistent and even though stores adjust prices frequently. We also uncover evidence oftit-for-tatbehaviorbywhichretailerswithradicallydi⁄erentdemandshocksnonethelessseek to match their local competitors(cid:146)pricing movements and recourse to sales and promotions. JEL classi(cid:133)cation: E30, L11. Keywords: Demandshocks,in(cid:135)ation,sales,laborcon(cid:135)icts,masspopulationdisplacement,severe weather events. (cid:3)The views in this paper are solely the responsibility of the authors and should not be interpreted as re(cid:135)ecting the views of the Board of Governors of the Federal Reserve System or any other person associated with the Federal Reserve System. We are grateful to Information Resources Inc. (IRI) for providing the scanner data. All estimates andanalysesinthispaperbasedonIRIdataarebytheauthorsandnotbyIRI.WethankJasonSockinforproviding superb research assistance, and Laura Ardila, Andrew B. Gi¢ n, and Teyanna Munyan for further assistance with the data. We also thank our colleague Charlie Gilbert for assisting us with the weekly seasonal (cid:133)ltering procedure. Comments and suggestions can be directed to etienne.gagnon@frb.gov and david.lopez-salido@frb.gov.

1 Introduction The e⁄ect of movements in aggregate demand on prices and quantities remains a lingering question in macroeconomics. A key reason is that prices and quantities are simultaneously determined; analysesbasedoncorrelationsbetweenthesetwovariablesmaythussu⁄erfrombiasesincoe¢ cients and not be used to support causal claims. Ultimately, of course, it is the causal relationship that most interests scholars.1 Inthispaper,wedocumentanumberofinstancesinwhichlargedemandshockscanbeidenti(cid:133)ed and show that, in those instances, the level of prices and retailing strategies more broadly react proportionally little. In particular, we study the pricing response of U.S. supermarkets to large demand shocks(cid:151)that is, shocks that either profoundly reshape the customer base or that alter shoppers(cid:146)willingness to consume(cid:151)brought about by labor con(cid:135)icts, mass population relocation, and shopping sprees caused by severe weather events. As we shall argue, we see the occurrence of our shocks as exogenous to retailers(cid:146)pricing strategies and supply factors, allowing us to use them as instruments to assess the causal e⁄ect of broad-based variation in demand on prices.2 We perform our analysis using a large dataset of weekly scanner data made available to researchersbyInformationResourcesInc.(IRI).Thisdatasetcontainspriceandquantityinformation from 2001 through 2011 on 29 personal care, housekeeping, and food products, and is well suited for our endeavor for several reasons. First, it covers 50 U.S. metropolitan markets, making possible thestudyofshocksthathadalargeimpactondemandataregionallevelbutalimitedin(cid:135)uenceon national aggregates. Second, its exceptionally large size(cid:151)about 4 million individual observations per week after our trimming(cid:151)permits the computation of reasonably accurate statistics at the product category(cid:150)market level; this feat would not be achievable using the considerably smaller micro datasets collected for the computation of the o¢ cial CPI and national account statistics. Third, it contains the universe of items sold by stores within each product category, thus allowing us to track the evolution of spending on these categories. In conjunction with the price data, the spending data permit us to derive real (same-store) quantity measures using a methodology similar to that employed for the national accounts. Fourth, the weekly data frequency allows us to zoom in on shocks that have a short life-span. Fifth, the availability of data from multiple retail chains and stores permits us to contrast the pricing behavior of local competitors facing di⁄ering shocks to demand. Ourmostcrucialtaskistoidentifyeventsthatcanbroadlymovedemandbutthatareotherwise exogenous to supply factors. We (cid:133)rst consider two labor con(cid:135)icts that both began in October 1The literature on the identi(cid:133)cation of aggregate demand shocks and their e⁄ects is too vast for us to o⁄er a comprehensive review in these pages. See Shea (1993) for a discussion of identi(cid:133)cation issues along with evidence of manufacturing demand shocks derived from shifts in input-output linkages. For monetary policy shocks, see the complementary approaches in Christiano and Eichenbaum (1999), Romer and Romer (2004), and Bernanke, Mihov, and Eliaz (2005). For (cid:133)scal policy shocks, see Hall(cid:146)s (2009) survey. For evidence on the aggregate demand e⁄ects of uncertainty shocks, see Leduc and Liu (2013). 2One could argue that both labor con(cid:135)icts and natural disasters constitute variation in the treatment(cid:151)that is, the variable (changes in demand) whose causal e⁄ect we would like to understand(cid:151)as if they were generated by a random experiment (see, for instance, Angrist and Imbens (1995)). 2

2003. The con(cid:135)ict in St. Louis, MO, was relatively short (less than a month) while the con(cid:135)ict in Southern California dragged on nearly (cid:133)ve months to become the longest supermarket strike and lockout in U.S. history. A⁄ected stores remained open during both con(cid:135)icts. Because many shoppers would rather take their business elsewhere than cross picket lines, strikes and lockouts can profoundly reshaped store frequentation within a market. Many stores in our sample saw their revenues collapse 50 percent or more for the duration of these con(cid:135)icts while others experienced correspondingly large increases. Stores whose employees were not on strike or locked-out faced no supply constraints, making the rise in their demand akin to an exogenous demand shock. At the end of the con(cid:135)icts, sta⁄numbers normalized but store frequentation did not always revert to pre-con(cid:135)ict levels, providing additional opportunities to study the e⁄ects of variation in demand. We next consider the mass population displacement brought about by Hurricane Katrina, the most expensive natural disaster in U.S. history. The hurricane displaced about 1 million persons, many of whom took years to resettle. Most displaced households moved in with relatives or friends, boostingpopulationdensityandstorefrequentationinneighborhoodsthatwerelessa⁄ectedbythe storm. Stores in our sample from the New Orleans, Louisiana, market and the Mississippi market experienced a persistent rise in sales volumes of about 20 percent, on average, in the wake of the hurricane. Finally, we consider shopping sprees around major snowstorms and hurricanes not associated with mass population displacement. Like Hurricane Katrina, their occurrence is unquestionably exogenous to retailing activities. But contrary to Hurricane Katrina, the typical e⁄ects of these storms on demand are short-lived and operate primarily through a rise in the demand of existing shoppers rather than through a reshu› ing of the consumer base. Storms that result in the closing of schools and workplaces force households to consume a larger fraction of their meals at home, thus boosting demand for food items. Similarly, the demand for personal care and housekeeping products may rise as households engage in more home production or take advantage of their trip to the supermarket to purchase items other than food. Our key (cid:133)nding is that large swings in demand appear to have, at best, a modest e⁄ect on the level of prices, consistent with a (cid:135)at short- to medium-term supply curve in the retail industry. This (cid:133)nding holds even in the case of our most persistent shocks for which stores adjusted the price of most items multiple times (or at least a couple of times in the case of regular prices). Put di⁄erently, the lack of a signi(cid:133)cant price response to large swings in demand seems inconsistent with the marginal cost of retailers being sensitive to the level of demand because of (cid:133)xed factors of production, or to retailers making large alterations to their markups in response to changes in demand. Ourevidenceisfurtherconsistentwithtit-for-tatbehaviorinwhichretailerswithradically di⁄erent demand shocks nonetheless seek to match their local competitors(cid:146)pricing movements and recourse to sales and promotions. Our two labor con(cid:135)icts feature the most dramatic shocks to demand in our study. For the Southern California strike and lockout, we split establishments into two broad groups: those that saw a drop in revenues in excess of 10 percent and those that saw an increase in excess of 10 3

percent (henceforth generically referred to as (cid:147)on strike(cid:148)and (cid:147)not on strike,(cid:148)respectively). The group on strike saw quantities sold drop by half relative to the pre-strike period while the group not on strike saw quantities increase by a third. Despite this remarkable di⁄erence, the level of prices was similarly behaved between the two groups throughout the con(cid:135)ict, rising about two percent in the (cid:133)rst couple months and erasing half of that increase in the remaining months. These price movements are not especially large relative to the historical volatility of the sample. Importantly, they are modest relative to quantity developments. Perhaps remarkably given limited sta¢ ng numbers, stores on strike put almost the same number of items on sale during the con(cid:135)ict, supporting the view that engaging in price discrimination is a central activity of retailers. At the end of the Southern California con(cid:135)ict, stores that had been on strike recouped only 4 out of every 5 dollars in lost business. To regain market share, these stores increased the number of sales and promotions, a move that was largely mimicked by their local competitors. This strategy graduallybroughtcustomersbackeventhoughrelativelyaggressivediscountsdidnottranslateinto lower transaction prices on average. This (cid:133)nding points to marketing having an in(cid:135)uence on sales volumes that goes beyond what can be inferred from movements in price indexes alone. The shorter St. Louis con(cid:135)ict had similarly large e⁄ects on quantities at stores on strike and stores not on strike than the longer Southern California con(cid:135)ict. However, the end of the strike brought an immediate return to pre-con(cid:135)ict activity levels even though the strike had arguably lasted long enough to induce consumers deterred by picket lines to shop at una⁄ected establishments. Perhaps as a result, stores that had been on strike did not launch a sales and promotion campaign in the wake of the con(cid:135)ict. Overall, these (cid:133)ndings suggest that there is some stickiness in consumer preferences for particular points of purchase, that those preferences are not eroded by shopping elsewhere for a few weeks, but that longer displacements such as during the Southern California strike can lead to permanent changes. We break down the e⁄ects of Hurricane Katrina on retail activities in two phases. In the week of the storm, store revenues in our New Orleans sample jumped by nearly a quarter from their average of the previous couple months. The jump was especially sizable for a number of personal care products such as toothbrushes, a sad reminder of the personal tragedies endured by many displaced families. The second phase extends from the weeks after the storm, when store revenues were about 20 percent above average, to well over a year later as displaced households slowly returned home or settled elsewhere. We (cid:133)nd little if any evidence that retailers responded to the persistent increase in demand by raising prices. Although food prices rose a little more in New Orleans than the corresponding national average in the months after the disaster, the price of personal care and housekeeping products rose by somewhat less. Moreover, we cannot exclude the possibility that supply disruptions created some upward pressure on food prices in New Orleans, further reducing the contribution of higher demand to the observed food price increase. Ourlastsetofdemandshocks(cid:151)shoppingspreestriggeredbymajorsnowstormsandhurricanes(cid:151) has no apparent e⁄ect on prices and other elements of our retailers(cid:146)pricing strategies. This (cid:133)nding is perhaps unsurprising because weather forecasts su⁄er from considerable uncertainty even for 4

horizons as short as a few days. Moreover, the e⁄ect of these shocks on demand rapidly dissipates. The (cid:135)eeting character of these shocks may simply leave retailers insu¢ cient time to adjust prices accordingly. Nonetheless, they invite us to qualify earlier (cid:133)ndings that retail prices tend to fall around periods of peak demand. Warner and Barsky (1995), MacDonald (2000), Chevalier, Kashyap, and Rossi (2003) report that prices of household appliances and some food items tend to decline around major holidays, which are also short-lived peaks in demand.3 This evidence has been used by a number of authors as suggestive of counter-cyclical markups at the macroeconomic level. Our conjecture is that the high predictability of demand peaks due to holidays or the passing of the seasons allows retailers to adjust their marketing strategies accordingly, making these events somewhat uninformative about how retailers might respond to other high-demand episodes. To our knowledge, our focus on broad-based demand shocks is novel in the empirical literature that uses large datasets of individual data to bridge micro price behavior and aggregate price dynamics. Perhaps one exception is Coibion, Gorodnichenko, and Hong (2012), who look at the link between local labor market conditions and pricing. Otherwise, previous work has considered a number of supply-related factors such as changes in sales taxes (see, for example, Dhyne et al. (2005) and the references therein, Karadi and Rei⁄(2012), and Gagnon, L(cid:243)pez-Salido, and Vincent (2013)), imported good price shocks (for example, Gopinath and Itskhoki (2010)), and commodity price shocks (for example, Nakamura (2008) and Hong and Li (2013)). This literature has typically found moderate pass-through rates that contrast with the muted price responses to our shocks. Our use of natural disasters to identify exogenous changes in demand is also somewhat unusual as the literature has typically been interested in their disruptive e⁄ects on supply. In a related paper, Cavallo, Cavallo, and Rigobon (2013) use online price data to show that a 2010 earthquake in Chile and the 2011 Sendai earthquake and tsunami in Japan both led to widespread product unavailability but not to higher prices. The paper is organized as follows. Section 2 details the construction of our dataset and presents keyfeaturesofretailactivitiesinnormaltimes, whichwewilluseasareferencetoassesstheimpact ofourdemandshocks. Section3analyzescustomerbasedisplacementduetolaborcon(cid:135)icts. Section 4 investigates shifts in demand due to major weather events, starting with the year-long population displacementcausedbyHurricaneKatrinaandproceedingwithtransitorybooststodemandrelated to major snowstorms and hurricanes. Section 5 concludes. 2 Retail pricing in normal times Our price and quantity analysis is performed using weekly scanner data made available to researchers for a nominal fee by IRI. The content of the dataset is detailed in Kruger and Pagni (2008) and Bronnenberg, Kruger, and Mela (2008), so we shall give only a brief exposition. The data come from a large sample of over 1,500 U.S. supermarket stores belonging to a variety of retail 3This(cid:133)nding also appearsto apply attheproduct-category levelin ourIRIsample. Formostproductcategories, we (cid:133)nd a negative correlation between movements in our price and quantity indexes after removing variations at frequencies higher than 53 weeks. 5

chains and operating in 50 U.S. markets. An observation corresponds to the information about an item (that is, a barcode sold in a particular store) in a given week. The number of observations is exceptionally large at about 300 million per year. Available item information includes the number ofunitssoldandtotalrevenueduringtheweek. Asiscustomarywithscannerdata,wederiveaunit price by dividing total revenue by the number of units sold. Although the data are limited to 29 food, housekeeping, and personal care products, they have the appealing feature of encompassing all item transactions within those product categories.4 To make the micro data suitable for our purposes, we drop items with suspiciously large price adjustments and apply various (cid:133)lters to extract regular, sales, and reference price series. We then construct price and quantity indexes for each IRI market(cid:150)product category combination using a methodology that closely matches that employed by the Bureau of Labor Statistics (BLS) and the Bureau of Economic Analysis (BEA) for the U.S. CPI and the U.S. National Income and Product Accounts, respectively. We relegate these technical details to an appendix. 2.1 Central features of retail pricing in normal times We begin by describing the general behavior of retail prices in the dataset to provide us with a benchmark against which to compare pricing behavior in the presence of large demand shocks. The key message is threefold: retail prices are adjusted frequently, price adjustments are primarily associated with sales and promotions, and the bulk of barcode-level price variation is common across stores belonging to the same retail chain. Fact #1: Retail prices are adjusted frequently Table 1 presents some key statistics on posted price adjustments broken down by product category.5 On average, we are able to compute a posted price change for nearly 4 million items each week, which are associated with weekly revenues of about $110 million. Of these items, 30:4 percent experience a (nonzero) posted price adjustment.6 The weekly frequency of price changes is nontrivial for all product categories, ranging from 13:7 percent for sugar substitutes to 43:9 percent for carbonated beverages. When we use only weekly price observations corresponding to the 15th day of each month, we derive a mean monthly frequency of price changes of 42:3 percent. Again, all product categories display frequent price adjustments, with sugar substitutes having the lowest monthly frequency, at 27:5 percent, and frozen dinner prices having the highest monthly frequency, at 55:2 percent. The mean monthly frequency of price changes in our sample is noticeably higher than corresponding estimates for the U.S. CPI, which range from about 26 percent in Bils and Klenow (2004) and Nakamura and Steinsson (2008), to about 36 percent in Klenow and Kryvtsov (2008). This di⁄erence is not due to the IRI sample being tilted toward product categories that 4TheIRIdatasetalsohasinformationoncigarettesandphotographicsupplies;weexcludetheseproductcategories because of regulatory restrictions on pricing and their gradual obsolescence, respectively. 5Our product-category statistics di⁄er slightly from those reported by Coibion, Gorodnichenko, and Hee Hong (2012)forIRIdataduetosmalldi⁄erencesin(cid:133)ltering. Moreover,oursamplerunsthrough2011whereastheirsstops in 2007. 6Our sample-wide statistics aggregate product-category statistics by weekly sales. 6

have relatively high frequencies of price changes; if we restrict the CPI sample to personal care and processed food categories, we (cid:133)nd a mean frequency similar to that of the full IRI basket. Instead, the IRI sample(cid:146)s relatively high frequency of price changes likely re(cid:135)ects a selection of stores that are actively changing prices. Fact #2: Retail price adjustments are primarily driven by sales A majority of price changes in our sample are associated with sales. To establish this fact, we identifysalesusingaversionofthe(cid:133)lterproposedbyNakamuraandSteinsson(2008)thateliminates temporary downward price movements. (See the appendix for the details of our implementation.) Table 2 shows that a relatively modest share (11:2 percent) of regular prices are adjusted in a typical week, consistent with two out of every three price adjustments in our sample being driven by sales and promotions. The pre-eminence of sales in price adjustments is found in all product categories. Onamonthlybasis, 23:7percentofregularpricesareadjusted, sothatsalesaccountfor about half of monthly price adjustments. The fact that sales represent a noticeably smaller fraction of monthly price adjustments than of weekly price adjustments indicates that many sales-related price movements are missed when data are collected monthly. We also experimented with identifying low-frequency price movements using the concept popularized by Eichenbaum, Jaimovich, and Rebelo (2011) of a (cid:147)reference price,(cid:148)which we de(cid:133)ne as an item(cid:146)s modal price over a centered 13-week window. Deviations from reference prices are even more common than deviations from regular prices. At a weekly frequency, references prices are adjusted only 4:5 percent of the time, implying that roughly six out of seven price adjustments in our sample are transitory deviations from the reference prices. At a monthly frequency, reference prices are adjusted 15:6 percent of the time, consistent with over two thirds of monthly price adjustments being transitory deviations from reference prices. Fact #3: Most prices are set at the retail chain level Price behavior in the IRI sample is consistent with pricing decisions originating principally at the retail chain level. To establish this fact, we de(cid:133)ne a weekly (cid:147)chain price(cid:148)as the modal price at which a barcode is sold across stores belonging to the same retail chain. Because we cannot link retail chains across IRI markets, our chain price measure is necessarily market-speci(cid:133)c even though retail chains may be active in several IRI markets. To ensure that we have a meaningfully large sample of stores to compute a chain price, we require that a barcode(cid:146)s weekly posted price be observed at a minimum of (cid:133)ve stores belonging to the same retail chain, in addition to imposing that items are frequently traded (see the appendix for the details of our data cleaning). If there is more than one modal price, then we use the largest one as the chain price. These methodological choices are immaterial for our (cid:133)nding that most pricing decisions do not originate at the store level. Table 3 presents our key statistics on retail chain pricing. Overall, 71:2 percent of posted prices equaltheircorrespondingchainprice, aclearindicationthatthelevelofpostedpriceiscoordinated to some degree across stores belonging to the same retail chain. This conclusion applies to all product categories in our sample, with the lowest proportion being a still-substantial 65:8 percent in the case of carbonated beverages. It also applies to regular prices, for which we (cid:133)nd similar 7

proportions of prices matching the modal regular price of their barcode across stores belonging to the same retail chain. We also (cid:133)nd evidence of an appreciable degree of coordination in the timing of price adjustments. As noted earlier, about 30 percent of posted prices are adjusted in a typical week. However, the probability of an item experiencing a price change jumps to over 80 percent when we condition on a change in the item(cid:146)s chain price. (Here we exclude the item(cid:146)s own price from the computation of the chain price to avoid creating a spurious increase in the price change probability.) In sharp contrast, if there is no adjustment in the chain price, then the item has only a 12:3 percent probability of being adjusted. As was the case with the level of prices, the (cid:133)nding of strong coordination in price changes across stores belonging to the same retail chain applies to all product categories and is found for both posted and regular prices. Summing up, the strong synchronization across stores in the level of prices and the timing of their adjustments indicates that stores play a relatively minor role in the determination of prices in our sample. This (cid:133)nding is consistent with Nakamura (2008), who explores similar issues using a multi-market sample of about 100 barcodes in 2004 from AC Nielsen. She reports that 65 percent of the variation in the level of prices is common across stores belonging to the same retail chain, while only 17 percent is speci(cid:133)c to the store and product. Additionally, because the time series variation in item prices is greater than the variation in barcode-level averages, she argues that the large shocks driving retail prices generally do not arise at the manufacturer level. Other aspects of the data Tables1and2showthatdownwardnominalpriceadjustmentsarenearlyascommonasupward price adjustments, a (cid:133)nding that holds true whether one focuses on either posted prices or regular prices. In addition to being frequent, posted price adjustments are also large. On average, the mean weekly posted price increase is 17:6 percent whereas the mean posted price decrease is a little larger at 18:7 percent. Regular price increases and decreases are somewhat smaller, at 12:0 percent and 14:5 percent, respectively. 3 Labor con(cid:135)icts as exogenous demand shocks Labor con(cid:135)icts can have substantial e⁄ects on store frequentation because many shoppers would rather take their business elsewhere than cross employee picket lines. Thus, a⁄ected stores can lose most of their demand whereas una⁄ected stores may experience a jump in sales volume. This section looks at such rearrangements of the customer base during two major labor con(cid:135)icts in the supermarket industry. The key (cid:133)nding is that pricing strategies at stores whose demand rose largely mimicked those at stores whose demand declined. This (cid:133)nding is consistent with the e⁄ects of radically di⁄erent demand shocks on pricing being trumped by a desire to keep up with pricing movements by local competitors. Because labor con(cid:135)icts are man-made and a⁄ect sta¢ ng levels, it is appropriate to take a moment to discuss their exogeneity with respect to pricing strategies and their likely e⁄ects on stores(cid:146)ability to supply goods. For stores whose employees are not on strike or locked-out, there 8

were no disruption to sta¢ ng numbers.7 The observed rise in their frequentation was thus akin to an exogenous demand shock, supporting our use of the strike as an instrument. The normalization of sta¢ ng levels at the end of our con(cid:135)icts provides other opportunities to study the e⁄ects of variation in demand in the absence of constraints on supply because store frequentation did not always revert back to pre-crisis levels. In the case of stores whose employees were on strike or locked-out, the ability to meet any given level of demand was certainly hindered by low sta¢ ng levels over the duration of the con- (cid:135)icts. However, as we will see shortly, the con(cid:135)ict was also accompanied by a large and arguably more important adverse shock to demand as many shoppers opted not to cross picket lines and took their business elsewhere. Strikes and lockouts are often due to disagreements over employee compensation, which enters a grocers(cid:146)marginal cost and should thus in(cid:135)uence prices. Due to the censoring of store and retail chain information in the IRI sample, we cannot assess the importance of compensation for our establishments.8 However, we know from looking at (cid:133)nancial statements of publicly-traded U.S. supermarket chains that compensation accounts for a relatively small share of overall supermarket costs. For instance, the largest two publicly-traded U.S. retail chains with activities concentrated in supermarket products, Kroger and Safeway, reported goods acquisition costs equivalent to 77.5 percent of their combined revenues in 2012. By contrast, operating and administrative expenses, which include employee compensation as well as spending on store management, utilities, local advertising, etc., accounted for only 18.3 percent of revenues. If these (cid:133)gures are re(cid:135)ective of the cost structure of establishments in our sample, then the di⁄erence in bargainingpositionbetweenemployersandemployeesareunlikelytohaveexceededasmallfraction ofoverallcosts. Inaddition,thesettlementoflaborcon(cid:135)ictsisrarelyone-sided,withcost-increasing measures often being bargained in exchange of cost-saving measures. Furthermore, labor disputes may involve non-price factors such as the management of employee schedules. For all these reasons, we see labor disputes as ultimately creating only limited uncertainty regarding desired prices. That said, even small changes to employee compensation can have a large impact on store pro(cid:133)tability because supermarkets typically have small pro(cid:133)t margins (operating pro(cid:133)ts averaged 2.6 percent of Kroger and Safeway(cid:146)s combined revenues in 2012). 3.1 2003-2004 Southern California supermarket strike and lockout On October 11, 2003, about 70,000 unionized supermarket workers in Southern California either went on strike or were put on lockout due to a disagreement with several retail chains over bene(cid:133)ts andcompensation. Thedisputeendednearly(cid:133)vemonthslater,onFebruary29,2004,whenworkers votedinfavorofanegotiatedcontract,resolvingwhathadbecomethelongestsupermarketstrikein U.S.history. Storeswhoseemployeeswereonstrikeoronlockoutremainedinoperationthroughout the con(cid:135)ict, in part as retail chains reallocated managerial resources to the aisles and pursued 7In addition, we have not come across evidence that the labor supply of employees at una⁄ected stores was disrupted in other ways. 8In addition, the terms of our contract with IRI prevent from reporting any information that could lead to the identi(cid:133)cation of particular retail chains or establishments. 9

alternative arrangements to ensure continued product availability. Our dataset contains establishments that experienced marked drops in revenues during the strike as well as establishments that experienced marked increases. As noted earlier, the identity of stores and retail chains is censored in the sample and, in compliance with our terms of usage, we make no attempt to uncover them. To further preserve the anonymity of establishments and retail chains involved, we pool all data from the Los Angeles and San Diego markets and look at two broad groups of stores: those whose revenues dropped more than 10 percent (generically labeled (cid:147)on strike(cid:148)) and those whose revenues rose more than 10 percent (generically labeled (cid:147)not on strike(cid:148)) during the strike relative to the corresponding period of the previous year. Because our identi(cid:133)cation strategy uses only observed movements in revenues but no retail chain information, the two groups may not perfectly overlap with the sets of stores that were actually a⁄ected and una⁄ected by the strike. Some stores also saw revenues change by less than 10 percent; we ignore them to focus on stores subject to large demand shocks. We (cid:133)nally compute separate statistics for each group to contrast the situation of stores with large positive and large negative shocks to demand. 3.1.1 Overall impact on quantities and prices Salient features of the strike(cid:146)s impact on retailing activities are presented in (cid:133)gure 1. As the top panel shows, stores whose employees were on strike experienced, on average, a staggering 50 percent drop in sales volumes over the con(cid:135)ict(cid:146)s duration. Some stores even saw sustained declines in quantities close to 80 percent. The drop was somewhat more pronounced in the (cid:133)rst few weeks, suggesting that some shoppers who had initially declined to cross picket lines or to limit their purchases quickly returned. This modest rebound aside, the drop in revenues at establishments on strikewaslargeandsustainedforthedurationofthelaborcon(cid:135)ict. Newsorganizationsreportingon the con(cid:135)ict indicated that shelves were generally fully stocked but that few shoppers were strolling the aisles. For this reason, we believe that supply constraints made a negligible contribution to the large drop in quantities observed at stores on strike. In sharp contrast, establishments in our group of stores that bene(cid:133)ted from the strike witnessed large increases in revenues(cid:151)over 30 percent, on average, with some stores even seeing their revenues more than double. Contrary to stores on strike, those that bene(cid:133)tted continued to have their regular employees present to serve consumers, replenish the shelves, and move products from warehouses to stores. In the absence of disruptions totheirabilitytosupplyproducts,thesuddenriseindemandforthesestoresisthusunambiguously interpretable as a demand shock. The top panel of (cid:133)gure 1 further suggests that a majority of consumers displaced by the strike returned to their previous shopping location at the end of the con(cid:135)ict: Sales volumes immediately rebounded at stores that had been on strike and fell sharply at stores that had bene(cid:133)ted from the strike. Thenormalizationwasincomplete,however. Intheyearthatfollowedtheendofthecon(cid:135)ict, sales volumes at stores that had been on strike were still 10 percent below sales volumes before the strike, whereas stores that had bene(cid:133)ted from the strike retained some of the customers displaced 10

by the strike. This fact could be consistent with theories that emphasize consumer loyalty to stores and chains such as switching costs (for example, Kleshchelski and Vincent, 2009).9 The immediate returnofsalesvolumestowardtheirpre-con(cid:135)ictlevelssuggeststhatconsumerpreferencesforpoints of purchase may persist even after consumers have switched stores for nearly (cid:133)ve months. It is also possible that other factors were at play. For instance, stores with signi(cid:133)cantly higher sales volumes during the strike may have been able to o⁄er fresher produce as a result. Such bene(cid:133)ts may, in turn, have helped them retain consumers at the end of the con(cid:135)ict. As the middle panel shows, price movements during the con(cid:135)ict, at only a couple of percentage points, were more than an order of magnitude smaller than swings in sales volumes, suggestive of a (cid:135)at supply curve. For example, if we attribute the observed rise in prices at stores not on strike entirely to a positive demand shock, then the estimates in table 4 imply a supply elasticity equal to log(1:021)=log(1:336) = 0:07. Thismodestestimatefallsslightlyifwecontrolforin(cid:135)ationoverthe strike period by measuring the demand shock(cid:146)s e⁄ect as the rise in excess of that for the price index of the full IRI sample.10 These supply elasticity estimates for the retail sector are markedly lower than the 0:18 (cid:133)gure reported by Shea (1993) for the U.S. manufacturing sector. One explanation for the di⁄erence could be that Shea(cid:146)s (1993) estimate applies to a one-year horizon whereas the Southern California strike ended after (cid:133)ve months, leaving less time for prices to adjust to higher demand. However, our evidence for the longer post-strike period, which features persistently large di⁄erences in demand and no supply disruptions at both groups of stores, also points to little if any price response.11 It is also apparent that price movements were similar between stores on strike and stores not on strike before, during, and after the con(cid:135)ict. After rising a couple percent in the spring and summer of 2003, prices declined a little in the months prior to the strike, then rose over 3 percent in the (cid:133)rst half of the con(cid:135)ict before retracing half of that rise in the second half. Overall, divergences in price movements between the two groups of stores were short-lived and not exceptionally large in comparison to other relative price movements over our sample period. After the strike, constraints on stores(cid:146)ability to adjust prices and to supply goods vanished. Fair pricing motives that could have made retailers reluctant to boost prices amid exceptionally high demand would have greatly eased now that customers could shop freely at all stores. Despite some persistent di⁄erences in the quantity indexes, our price indexes show little economically meaningful divergence between the two groups of stores. To provide a more direct comparison of the level of prices between stores that saw their demand 9Theevidencecannotbeinterpretedassupportiveofmodelswithproduct-levelloyalty,suchasdeephabits(Ravn, Schmitt-GrohØ, and Uribe 2006), unless one reinterprets the stock of habits as pertaining to speci(cid:133)c establishments or retail chains instead of speci(cid:133)c goods and services. 10We could further measure the supply elasticity by comparing relative price movements to relative quantity movements between stores on strike and stores not on strike, with the caveat that stores on strike may have suffered from supply disruptions in addition to a relative demand shock. The resulting elasticity is essentially zero at log(1:336=0:512)=log(1:021=1:013)=0:01. 11When we measure the medium-run elasticity of supply using relative movements in prices and quantities from the period before to the period after the strike, we obtain a mildly negative slope estimate, log(1:009=1:012)=log(1:036=0:904)= 0:02. (cid:0) 11

soar and stores that saw their demand collapse, we next look at the cost of purchasing identical baskets of goods. We consider three such baskets. The (cid:133)rst basket (the (cid:147)(cid:133)xed(cid:148)basket) consists of all barcodes continuously available at both groups of stores over the two-year period displayed in (cid:133)gure 1, that is, over a period starting 26 weeks before and ending 78 weeks after the beginning of the con(cid:135)ict. The number of units purchased for each barcode is set to the average weekly number of units sold across all stores over the two-year period. This (cid:133)xed-quantity basket is reasonably representative of overall purchases in Southern California, accounting for nearly 70 percent of total revenues over the period. At the product category level, the coverage of the basket ranges from 27 percent of total revenues for razors to 95 percent for peanut butter. The second basket (the (cid:147)on-strike(cid:148)basket) corresponds to the number of units purchased at stores on strike, again using only barcodes that are continuously available at both groups of stores. Similarly, the third basket (the (cid:147)not-on-strike(cid:148)basket) consists of the number of units sold at stores not on strike. Contrary to the (cid:133)xed basket, the composition of the on-strike and not-on-strike baskets varies from week to week in line with shoppers(cid:146)actual consumption. The lower panel of (cid:133)gure 1 reports the cost of purchasing each of the three baskets at the mean transaction price observed at stores on strike relative to that at stores not on strike. (See the appendix for the exact formulas.) All three ratios tell the same story: The cost of purchasing any of our baskets at stores on strike relative to stores not on strike hovered near its pre-strike level for the duration of the con(cid:135)ict. A slight increase in the relative price of the baskets at stores that were on strike is apparent several months after the end of the con(cid:135)ict. We are reluctant to attribute this rise in the ratios to a price response to persistently lower demand given the historical variability of the sample. We note that shoppers at stores on strike could have purchased identical baskets of goods for an equal or even lower price at stores that were not on strike at any point over our two-year period, a (cid:133)nding that suggest some insensitivity to the level of prices on the part of consumers. This conclusion comes with a number of caveats. Because the identity of stores and retail chains is censored, we cannot control for di⁄erences in factors such as income and sales taxes that may a⁄ect the level of prices across areas. Also, our baskets comprise solely barcodes that are simultaneously available in both groups of stores; by ignoring roughly 30 percent of store revenues in our product categories, we may be overlooking the e⁄ect of private labels on the e⁄ective costs of a typical basket. Finally, retailers sell items in product categories that are not covered by the IRI sample. 3.1.2 Price adjustments, price discounts, and the labor con(cid:135)ict Figure 2 presents some key statistics on the broader pricing strategies of retailers before, during, and after the strike. The mean of each statistic over these periods is reported in table 4. On average, stores on strike saw their weekly frequency of posted price changes drop 4:3 percentage points during the con(cid:135)ict, with a reduction in regular price adjustments accounting for most of the drop. The fraction of items on sale (the middle-left panel of (cid:133)gure 2) declined 1:2 percentage point to 22:0 percent while the mean discount conditional on a sale (the middle-right panel) was 12

unchanged. In short, stores on strike broadly managed to maintain the importance of sales during the con(cid:135)ict. Stores not on strike experienced a somewhat smaller drop in the frequency of posted price changes, 3:0 percentage points, that re(cid:135)ected small declines in both sales-related and regular price adjustments. On the one hand, the continued use of sales by stores whose demand tumbled suggests that engaging inprice discrimination is animportant endeavorof retailers independently oftheir level of demand. Perhaps it also re(cid:135)ects a desire from stores on strike to project a business-as-usual image, or an implicit promise to o⁄er bargain shoppers some opportunities to buy a portion of their basket at a discounted price every week. Indeed, it is conceivable that price-sensitive shoppers who continued to patronize stores on strike would have found it unfair to see their loyalty rewarded with higher prices. On the other hand, the reduction in regular price adjustments seems consistent with limited human resources hindering stores(cid:146)ability to reprice. It is also possible that, given markedly lower sales volumes, stores on strike would have tolerated larger deviations of regular prices from their optimum because (cid:133)xed repricing costs would have been spread over a smaller number of units sold.12 We also note that the level of individual prices continued to be coordinated within stores belonging to the same retail chain (the lower-right panel) despite variation across stores in the magnitude of the drop in demand and ability to use managerial sta⁄ to (cid:133)ll positions previously held by striking employees. In fact, the share of prices equal to the chain price edged up a couple of percentage points during the strike at both stores on strike and stores not on strike. As noted above, stores on strike recouped only 4 out of every 5 dollars in lost business once the con(cid:135)ict ended. To win back customers, they increased the frequency and depth of sales in the ensuing year. The middle panels of (cid:133)gure 2 show that the share of discounted barcodes and the mean discount both rose a couple of percentage points before slowly edging back.13 To further explore the strategic use of sales during and after the con(cid:135)ict, we break down discounts into 10-percentage-point bins, starting with discounts that are below or equal to 10 percent, then below or equal to 20 percent but greater than 10 percent, and so on. Figure 3 displays the contribution of discounts in each bin to total revenues (the top row of panels) and to total savings (the lower row of panels).14 The upper-left panel shows that the contribution of items to total revenues generally was declining in the size of the discounts extended. Before the strike, discounts up to 10 percent accounted for almost 8 cents out of every dollar in revenues at 12See Golosov and Lucas (2007) for an illustration. The widening of the price inaction region at stores whose demand fell should, in addition to lowering the frequency of price adjustments, boost the absolute size of their price changes. We do (cid:133)nd an large increase (4.4 percentage points) in the average size of regular price adjustments. However, a notable increase (1.5 percentage points) is also present at stores that saw their demand jump. 13Wede(cid:133)nethemeandiscountastheaveragesavingontheentirebasketofgoodspurchased. Formally,wecompute (preg p )q = pregq , where preg is the regular price of item j in week tidenti(cid:133)ed by our sales (cid:133)lter. j j;t (cid:0) j;t j;t j j;t j;t j;t (cid:16) 14To construct (cid:133)g(cid:17)ure(cid:16)3, we (cid:133)rst p(cid:17)ool allweekly observations in each of our three time periods (before, during, and P P after the strike). For the top panels, we then weigh weekly observations by their share of total revenues (that is, ! =p q = p q , where is the set of weeks in the period of interest). The area under the curve i;t i;t i;t t 2T j j;t j;t T integrates to th(cid:16)e share of total re(cid:17)venues accounted by items on sale. Similarly, for the lower panels, we weigh weekly P P observationsbytheircontributiontototalsavingsovertheperiod(thatis,! =(preg p )q = pregq ). i;t i;t (cid:0) i;t i;t t 2T j j;t j;t The area under the curve then integrates to the mean discount o⁄ered by retailers. (cid:16) (cid:17) P P 13

stores that were subsequently a⁄ected by the con(cid:135)ict, while items o⁄ered at discounts in excess of 50 percent accounted for only about 2 cents out of every dollar in revenues. The lower-left panel shows that discounts in the 40 percent to 50 percent range, which were both substantial and frequent, allowed shoppers to lock in the largest savings. By contrast, very small and very large discounts contributed little to actual consumer savings because they were either too small or too infrequent to have a large impact. The distribution of contributions to total revenues and to total savings of discounts had similar shapes in the pre-con(cid:135)ict period at stores una⁄ected by the strike than at stores a⁄ected by the strike, although discounts of medium sizes played a somewhat lesser role in their overall marketing strategy. Duringthestrike,thenumberofitemsonsalesdeclinedonlyalittleata⁄ectedstoresduringthe strike. The upper-left panel of (cid:133)gure 2 shows that the importance of small and medium discounts for total revenues fell most while that of discounts in excess of 50 percent actually rose some. This shift toward deep discounts may have been part of a marketing strategy to lure shoppers to stores despite the strike, as heavy discounts make for good advertising. That said, the contribution to total savings of heavily discounted items was o⁄set by the lesser importance of small and medium discounts, leaving the mean saving on the entire basket about unchanged. For stores that were not on strike, we also witness a decline in the importance of small and medium discounts during the con(cid:135)ict period but do not (cid:133)nd a corresponding rise in the importance of very large discounts. Once the strike was over, the contribution of small and medium discounts to total revenues and to total savings rose above its pre-strike level at both groups of stores. The importance of discounts in excess of 50 percent also rose at stores that had not been on strike, catching up with the similar increase documented at stores on strike during the con(cid:135)ict. In sum, the evidence is consistent with stores that had been on strike o⁄ering more frequent discounts(cid:151)small, medium, and large(cid:151)than usual to win back customers, and stores that had avoided the strike responding with more frequent sales of all sizes to retain them. Set against the background of di⁄ering swings in demand over the strike and the post-strike periods, these similarities in the recourse to sales suggest that retailers attach much value to matching changes in their local competitors(cid:146)pricing strategies. 3.2 A shorter labor con(cid:135)ict: the 2003 St. Louis grocery strike and lockout The2003St.Louisgrocerystrikeandlockoutpaintsabroadlysimilarportraitasthelongercon(cid:135)ict in Southern California: Pricing strategies of stores on strike and stores not on strike broadly resembled each other despite highly diverging demand. It started on October 7 when employees at several supermarket chains went on strike or were locked-out after voting down a tentative labor agreement. The dispute ended 24 days later when negotiating teams reached an agreement that proved acceptable to all parties. The main e⁄ects of the con(cid:135)ict on prices and quantities are shown in (cid:133)gure 4. Due to the smaller size of the St. Louis sample (22 stores compared to 107 stores in Southern California), we compare establishments whose revenues dropped more than 10 percent (generically referred to as (cid:147)on strike(cid:148)) to all other establishments in the sample (generically referred to as (cid:147)not on strike(cid:148)). 14

Picketing was again e⁄ective at deterring shopping activities, with stores on strike seeing sales volumes plunge about 45 percent, on average, while stores not on strike experienced a modest increase. Reading through the weekly volatility, prices at stores not on strike rose a touch more around the time of the con(cid:135)ict than prices at stores on strike. However, this slight divergence was part of a broader trend over the two-year period displayed and thus di¢ cultly attributable to the strike. Moreover, price movements on a like-for-like basis were much more similar between the two groups of stores over the period displayed and around the strike in particular; our relative price measures controlling for the composition of the basket, which are shown at the bottom of (cid:133)gure 4 and employ the same methodology as those computed earlier for Southern California, were very stable. Ofnote,salesvolumesatstoresonstrikefullyrecoveredassoonasthelaborcon(cid:135)ictwasover,in contrast with the customer base erosion apparent at stores on strike in Southern California. Given limitedhouseholdinventories, a24-daycon(cid:135)ictisarguablytoolongforcustomersunwillingtocross picket lines at their usual supermarket not to go shopping elsewhere. The recovery in sales volumes thus suggests some stickiness in consumer preferences for particular establishments that are not capturedinstandardmodelsofstoreswitchingcostssuchasKleshchelskiandVincent(2009),where store changes are permanent. In any case, the quick recovery in sales volumes probably explains why we did not observe a rise in the frequency and depth of discounts (shown in the middle panels of (cid:133)gure 5 and documented in table 5) in the wake of the St. Louis con(cid:135)ict. 4 Major weather events as exogenous demand shocks OurnextsetofdemandshockswerecreatedbyMotherNature;theiroccurrencewasunambiguously exogenous to retail activities in general and to supermarkets(cid:146)pricing strategies in particular. We (cid:133)rst look at Hurricane Katrina that, in the span of a few tragic weeks in the summer of 2005, led to massive population displacement. Stores located in areas that received an in(cid:135)ow of refugees experienced a sharp rise in store frequentation that persisted for well over a year. We then look at shopping sprees triggered by snowstorms and hurricanes. Contrary to strikes and Hurricane Katrina, these storms do not feature a recon(cid:133)guration of retailers(cid:146)customer base but rather a temporary increase in the demand of all customers. 4.1 Hurricane Katrina In August 2005, Hurricane Katrina created an estimated $108 billion in property damages (in 2005 dollars), making it the most expensive natural disaster in U.S. history. It was also directly responsible for the tragic loss of about 1,200 lives and the displacement of roughly 1 million persons.15 The city of New Orleans, Louisiana, sustained the most extensive damage due to the failure of its levee system, which led to the (cid:135)ooding of approximately 80 percent of the city. The (cid:135)ooding of residential areas made it impossible for many displaced households to return home for several 15These estimates are taken from Blake and Gibney (2011). 15

months or years, and some households even chose to permanently relocate elsewhere. According to research conducted at the U.S. Census Bureau and reported in Geaghan (2011), as of late 2009, 31,500 households in the New Orleans metropolitan area (7 percent of the area(cid:146)s total) did not consider themselves permanently resettled.16 ThehurricanehaddisruptiveconsequencesonretailactivitiesinNewOrleansandothera⁄ected states. Of the 23 New Orleans stores that participated in the IRI sample on the eve of the tragedy, (cid:133)ve exited the sample as soon as the storm hit while one store ceased to report data for a period of eight months. Similarly, (cid:133)ve out of the nine stores in the Mississippi sample did not report data for a week or two around the hurricane. Although the IRI dataset contains no information that would permit us to ascertain that these sample exits and missing reports were caused by the hurricane, we interpret their coincidental timing as strongly suggestive that they were. As(cid:133)gure6shows,thehurricanealsoa⁄ectedretailactivitiesatstoresthatremainedinbusiness. In the weeks and months that followed the disaster, stores that continued to report data to the IRI experienced sales volumes that were, on average, about 20 percent higher than before the disaster. We found an equally large rise in the smaller Mississippi sample (not shown); a smaller e⁄ect is also apparent for the Houston, Texas, sample (also not shown). We interpret this persistent rise in sales volumes at continuing stores as evidence that mass relocation boosted the demand for supermarket products in some areas that were relatively una⁄ected by the storm. Supporting our interpretation is the fact that, according to Geaghan (2011), 74 percent of displaced New Orleans householders reported living with an acquaintance.17 In addition, the persistent increase in sales volumes of food productsaswellashousekeepingandpersonalcareproducts,whichareshownseparatelyin(cid:133)gure6, were of a similar proportion, consistent with mass relocation boosting the demand of supermarkets(cid:146) entireproducto⁄ering. Indeed, weobserveincreasesinall29productcategoriesinthesample. The upper-leftpanelof(cid:133)gure6alsofeaturesashort-livedbutoutsized60-percentsurgeinsalesvolumes of housekeeping supplies and personal care products the precise week that Hurricane Katrina hit. Revenues from product categories such as toothbrushes and razors witnessed transitory increases in excess of 100 percent, again consistent with population displacement being a key driver of retail activities over the period. Our empirical evidence provides little if any support to the view that retailers took advantage of higher demand brought about by the hurricane to raise prices, either initially when spending on personal care products skyrocketed or over the medium run when store frequentation was boosted by mass relocation. The upper-right panel of (cid:133)gure 6 shows that the price of food products and of personal care and housekeeping products both rose in the weeks that followed the storm before 16Three weeks after Hurricane Katrina, authorities in Texas ordered 1.8 million persons to evacuate coastal areas along the Gulf of Mexico ahead of Hurricane Rita(cid:146)s landfall. Transitory jumps in sales volumes are apparent for the main markets that hosted refugees. We exclude this storm from our analysis because the e⁄ects of mass population relocationcannotbedisentangledfromthoseofshoppingspreesaheadofthestormbyhouseholdsshelteringinplace. We will explore this latter phenomenon in section 4.2. 17The fact that most refugees sought shelter with family members and friends suggests a similarity in consumer characteristics between displaced and hosting households. See Lach(cid:146)s (2007) analysis of a mass migration from the former Soviet Union to Israel for evidence that large in(cid:135)ux of consumers with price elasticities and search costs di⁄ering from the native population can a⁄ect retail prices. 16

erasing some of these gains. When we average over the period covered by the federal emergency declaration, we (cid:133)nd that prices were 1:4 percent higher overall than they were over the 26-week period before the hurricane. As table 6 reports, this increase is modestly larger than the average rise across IRI markets not directly a⁄ected by Hurricane Katrina (that is, excluding New Orleans, Mississippi, and Houston). Muted price movements translate into small estimates of the short-term elasticity of supply. If we measure the price impact as the rise in the New Orleans price index in excess of the average rise for IRI markets not directly a⁄ected by Hurricane Katrina, then our estimate of the (short-run) elasticity of supply is 0:03. When we compute the elasticity of supply using price and quantity movements over the longer period after the federal emergency relative to the pre-storm period, we get a supply elasticity estimate of 0:13. However, we are reluctant to interpret this medium-run estimate as suggestive that retailers took advantage of higher demand to boost prices for two main reasons. First, this estimatemaybebiasedupwardbyhurricane-relateddisruptionstotheregion(cid:146)sfoodsupply. Second, price pressure was not broad based, as the price of housekeeping and personal care products rose in line with the national average. Theremainingpanelsof(cid:133)gure6providefurtherresultsregardingtheimpactofthehurricaneon retailers(cid:146)broader pricing strategy. The frequency of price changes edged down during the federal emergency period, and reverted to its pre-hurricane average for about a year before sliding in the fall of 2006. The hurricane had a somewhat more apparent e⁄ect on the mean discount, which temporarilyslidfrombetween5and6percentoftheregularpriceintheweeksbeforethehurricane to a low near 2 percent, before rebounding at the end of the federal emergency declaration. Much of this decline re(cid:135)ects a lower proportion of items on sales, especially in late September and early October, rather than a shift in customer spending toward items with lesser or no discounts, as hinted by the relative stability of both the mean discount and the share of items on sale in the (cid:133)rst couple weeks after the storm. 4.2 Major snowstorms and hurricanes Majorweathereventssuchaslargesnowstormsandhurricanescana⁄ectthedemandforsupermarket products through several channels. Storms that result in the closing of schools and workplaces force households to consume a greater proportion of their meals at home, thus boosting demand for food items. Similarly, the demand for personal care and housekeeping products may rise as households engage in greater home production or take advantage of their trip to the supermarket to purchase items other than food. Storms may also displace consumption across time periods by making it di¢ cult or impossible to shop on certain days. In particular, some households may seek to build their domestic inventories in anticipation of a storm to ensure continued supplies, while others may need to replenish their inventories once the storm is over. 17

4.2.1 Identi(cid:133)cation We have identi(cid:133)ed 59 combinations of an IRI market and a major snow episode whose disruptive consequences were favorable to the triggering of a shopping spree. Many of these combinations feature a peak in average store revenues of 10 percent or more relative to the previous few weeks. While there were hundreds of smaller snowstorms in our sample, we expressly choose to leave them aside because storms whose disruptive e⁄ects last only a couple of days are less likely to leave a clear imprint in supermarket data collected weekly. In identifying disruptive snowstorm episodes, we account for the fact that some localities have a greater ability to cope with snowfall than others. Snow accumulations that have crippling e⁄ects in Southern states, where snowstorms are scarce and snow plowing equipment is in short supply, may have only limited disruptive e⁄ects in Northern states, where local authorities are accustomed to clearing snow o⁄ the streets rapidly. To do so, we match our IRI scanner data with the U.S. NationalOceanicandAtmosphericAdministration(cid:146)sRegionalSnowfallIndex(RSI)andtheFederal Emergency Management Agency(cid:146)s (FEMA) list of federal disaster declarations. The RSI controls fordi⁄erencesinhistoricalsnowprecipitationandtheauthorities(cid:146)abilitytocopewithsnowthrough the setting of precipitation thresholds that are speci(cid:133)c to nine U.S. regions (the West Coast of the UnitedStatesisnotcoveredduetoinsu¢ cientlyfrequentsnowstorms). Stormswhosesocialimpact is roughly in the top half of historical storms in their region are given a rank between 1 and 5. Manyofthesnowepisodesinoursamplewereranked(cid:147)3(cid:150)major,(cid:148)(cid:147)4(cid:150)crippling,(cid:148)or(cid:147)5(cid:150)exceptional,(cid:148) placing them in the top 5 percent of storms in terms of regional-level disruptiveness. Many storms werealsograntedan(cid:147)emergency(cid:148)ora(cid:147)disasterrelief(cid:148)statusbyFEMAbecausetheywereof(cid:147)such severity and magnitude that e⁄ective response [was] beyond the capabilities of the State and the local governments and that Federal assistance [was] necessary.(cid:148)18 We validate our list of snow episodes against daily snowfall measurements reported by local weather stations to avoid situations in which a disruptive storm at the regional level results in little snow accumulation or passes as rain in a particular market. We also use local daily snowfall measurements to include a number of storms that likely had large localized e⁄ects but whose regional impact, as measured by the RSI, was small. In a few cases, our snow episodes cover two snowstorms rather than one because the separate meteorological systems are indistinguishable in weekly data. Finally, we incorporate a few major snowstorms from the U.S. West Coast for which RSI scores are not available. The list of snowstorms, along with their cumulative snowfall, RIS classi(cid:133)cation, and FEMA declaration, is provided in an online appendix. We follow a similar strategy for identifying hurricanes that are likely to induce shopping sprees. We look at all (cid:147)emergency(cid:148)and (cid:147)major disaster(cid:148)declarations by FEMA that are attributed to a hurricane. We then validate our list against daily rainfall and maximum wind speed measurements from local weather stations.19 In total, we have identi(cid:133)ed 21 combinations of an IRI market and a 18Disasterdeclarationsputintomotionshort-tolong-termfederalrelief,someofwhichmaybedirectedtoindividuals. Emergencydeclarationsaremorelimitedinscopeandseektomeetspeci(cid:133)cemergencyneedsortohelpprevent major disasters from occurring. Both types of declaration require presidential approval. 19We exclude Hurricane Katrina and Hurricane Rita due to their exceptional mass population displacement, a 18

hurricane, which we also list in our online appendix along with their characteristics. 4.2.2 Illustration: 2009(cid:150)2010 winter in Washington, D.C. Figure7illustratessomee⁄ectsofmajorstormsonretailactivitiesusingtwosnowepisodesthathit Washington, D.C., during the 2009(cid:150)2010 winter. The (cid:133)rst episode began on Friday, December 18, 2009, and left 41.7 centimeters (16.4 inches) of snow at D.C.(cid:146)s Reagan National Airport. Federal o¢ ces were closed the following Monday and operated on an unscheduled leave basis for two more days due to impracticable roads in parts of the metropolitan area. The second snow episode was more disruptive, consisting of two back-to-back blizzards that together blanketed the U.S. capital with 72.6 centimeters (28.6 inches) of snow. Federal o¢ ces closed early as the (cid:133)rst blizzard moved in on Friday, February 5, 2010, remained closed through February 12, and then operated on an unscheduled leave basis through February 16. As is typical of the episodes in our sample, the National Weather Service and local media began reporting on the approaching snowstorms several days ahead of their occurrence. The upper-left panel of (cid:133)gure 7 shows that sales volumes peaked 20 to 45 percent above their trendintheDecember2009andFebruary2010snowepisodes. Thetimingofthesurgeinquantities di⁄ers between the two episodes, although the use of weekly retail data limits our ability to identify their timing precisely (note that IRI weeks run from Monday to Sunday). As is apparent for both episodes, the quantity of food products and of personal care and housekeeping products spiked around the storms, supporting our treatment of major snowstorms as shocks to the overall demand of supermarkets rather than as shocks speci(cid:133)c to some product categories. In the (cid:133)rst episode, quantities of both groups of products rose 10 percent in the week of the storm relative to their recent trend, and rose even more in the ensuing week. In the second episode, sales volumes surged 27 percent for personal care and housekeeping products in the week encompassing the beginning of the second episode and 45 percent for food products. The demand for food products remained above its recent trend during the ensuing week. The large impact of the second storm may be due to the anticipation of greater snow totals (and, if our experience is representative, it could also re(cid:135)ect some learning from the (cid:133)rst episode that shopping in advance of the storm avoids some headaches). The two snowstorm episodes left no apparent imprint on price indexes, as shown in the upperrightpanel. Similarly,thefrequencyofpricechanges,themeandiscount,theshareofitemsonsales, and the share of store revenues derived from items on sales were all within the range experienced over the 2009(cid:150)2010 winter. Our econometric analysis below, on the broader sample of snowstorms and hurricanes, con(cid:133)rms these impressions. phenomenon that we analyzed separately in section 4.1. We also drop most observations related to Hurricane Ike andHurricaneLilibecausetheycameontheheelofotherdisruptivehurricanesthatrequiredfederalassistance. Our regression results are robust to keeping these latter hurricanes in the sample. 19

4.2.3 Econometric analysis There is some uncertainty regarding the precise timing of when weekly retail activities should feel the e⁄ects of storms most strongly. Storms that hit early in the week or whose disruptive e⁄ects are anticipated may a⁄ect retail activities prior to the storm. Similarly, storms that hit late in the week or that require domestic inventory rebuilding by households afterwards could have some e⁄ect in the week after the storm. As a (cid:133)rst step into our investigation, we ignore this timing issue and compare various retailing statistics in the week corresponding to the observed peak in quantities to their respective average in the weeks prior to the storm. More precisely, for a storm beginning in week t, we identify the peak in quantity over the weeks t 1, t, and t+1 and then compare the (cid:0) statistic of interest for that peak week by its average over the weeks t 4 to t 2 (the (cid:147)pre-storm (cid:0) (cid:0) period(cid:148)).20 Our sample of snow episodes supports the patterns apparent in (cid:133)gure 7, namely that major snowstorms boost consumer spending on supermarket items while having little if any in(cid:135)uence on pricing. The mean peak in quantities around snowstorms is 12:9 percent higher than the average overthepre-stormperiod, whereasthepeakinpricesisonly0.1percenthigher. Asimplestatistical test that the population mean of the ratio across snow episodes equals 1 is rejected for quantities but not for prices. The peak quantity and price responses, as well as their statistical signi(cid:133)cance, are nearly identical for hurricanes. For both types of storms, we also (cid:133)nd similar statistics on broader features of our stores(cid:146)pricing strategies between the week of the peak in quantities and the pre-storm period. This remarkable (cid:133)nding suggests that retailers broadly retain their usual pricing strategies. We next follow a regression-based approach similar to that used by Chevalier, Kashyap, and Rossi (2003) in their study of demand peaks around holidays. We regress our statistics of interest (illustrated here with the log of our market-speci(cid:133)c price index) on a quadratic time trend and a pre during set of dummies marking the week immediately before (W ), the week during (W ), and the c;t c;t after week after (W ) a storm, c;t log PF = (cid:11) +(cid:11) t+(cid:11) t2+(cid:12) W pre +(cid:12) W during +(cid:12) W after +" : c;t c;0 c;1 c;2 1 c;t 0 c;t 1 c;t c;t (cid:0) (cid:0) (cid:1) The estimated coe¢ cients on the week dummies can be interpreted as the average movement in the left-hand side statistic (relative to its trend) across the various storm episodes in the sample. To ensure that our quadratic trend (cid:133)ts the level of the individual series properly, we retain only data from the (cid:133)rst week in October through the last week in March in the case of snowstorms and from the (cid:133)rst week of July to the last week of December in the case of hurricanes (there are no qualifying storms outside of these broadly-de(cid:133)ned snowstorm and hurricane seasons). We next (cid:133)t a quadratic time trend that is speci(cid:133)c to each IRI market and 6-month season combination, represented by the subscript c in the above equation. We then run separate regressions for snowstorms and hurricanes. 20We normalize the quantity and price indexes in each market by dividing their value during the quantity peak week by their average during the reference period. 20

Table 7 reports the results for our key statistics of interest. For major snowstorms, we (cid:133)nd that the boost to quantities occurs almost entirely in the week of the storm. We also observe a statistically-signi(cid:133)cant small decline in quantities of about 1:9 percent in the week after the storm. This decline is suggestive that greater shopping activities during snowstorms result in the bringing forwardofsomehouseholdexpenditures, leadingtoasmallpullbackintheweekimmediatelyafter. The corresponding estimates for prices point to a small increase during the week of the storm that is not statistically signi(cid:133)cant. In the case of hurricanes, we (cid:133)nd a statistically signi(cid:133)cant boost to spendinginboththeweekimmediatelybeforeandtheweekofthestorm. This(cid:133)ndingsuggeststhat hurricanesprimarilypullforwardconsumptionexpenditures,perhapsduetoagreaterpredictability oftheseeventsandagreaterriskthathouseholdscouldexperiencereducedsuppliesforaprotracted period. A pull back is also observed in the week immediately after a hurricane, although it is not statistically signi(cid:133)cant at standard con(cid:133)dence levels. 4.2.4 Discussion Our analysis suggests that supermarkets do not take advantage of transitory peaks in demand brought about by major snowstorms and hurricanes to boost prices. More broadly, they appear to implement pricing strategies during and around storms that are highly similar to those in other periods. In this sense, the response to peaks in demand brought about by storms di⁄ers from the (cid:133)nding that prices tend to fall during periods of peak demand around holidays (see Warner and Barsky (1995), MacDonald (2000), and Chevalier, Kashyap, and Rossi (2003)). This di⁄erence could re(cid:135)ect a number of factors. Notably, whereas the timing of holidays is perfectly predictable, the occurrence of a major snowstorm or hurricane can be anticipated at most a week or so in advance and only with great uncertainty. This limited predictability may not leave enough time for manufacturers to adjust production or for retailers to alter their pricing strategy; circulars may already have been printed and, in any case, there can be lags of several months in the planning of sales and promotions (see Anderson et al. (2013) for a discussion). Moreover, one cannot exclude thatfairpricingmotivesofthekinddescribedbyRotemberg(2005,2011)couldbeatplay. Retailers maywanttoavoidbeingperceivedasunjustlypro(cid:133)tingo⁄theircustomers(cid:146)unusuallyhighmarginal utility for fear of losing their business in the future. That said, our sample excludes products whose consumption is essential during storms, such as de-icing salt, batteries, and snow shovels. Our sample instead contains products consumed year-round and for which several brands are typically available. A retailer that would want to take advantage of temporarily high demand would thus have to raise e⁄ective prices over a broad product o⁄ering. This feat could be achieved by reducing the number or depth of sales rather than by raising regular prices but our sample does not contain support for this channel. 21

5 Conclusion We have shown that the level of supermarket prices responds little to large swings in demand brought about by labor con(cid:135)icts, mass population relocation, and shopping sprees around storms and hurricanes. This evidence is consistent with (cid:135)at short- to medium-term supply curves in the retail sector. In particular, it seems inconsistent with the marginal cost of retailers being sensitive to the level of demand because of (cid:133)xed factors of production, a hypothesis that is often made in macro models. And when prices did (cid:135)uctuate some, we have found that variations in the frequency and depth of sales were often important channels of price adjustment, thus cautioning against focusing solely on regular prices to understand the transmission of shocks. A number of authors have reported that, contrary to the textbook treatment, prices tend to fall in periods of peak demand, including for the kind of goods present in our sample. On the surface, this evidence seems supportive of models featuring countercyclical markups. However, the absence of a price response to major snowstorms and hurricanes suggests that the perfect predictability of holidays and the passing of seasons make their associated peaks in demand of a very di⁄erent nature than those triggered by shocks that are di¢ cult to forecast. In addition, if retailers respond little to demand shocks, they seem concerned with keeping up with the pricing strategies of their local competitors. This fact is most clearly seen from price movements during and after the labor con(cid:135)icts in our sample, when retailers with radically di⁄erent demand shocks nonetheless tracked their local competitors(cid:146)pricing movements and recourse to sales and promotions. These observations invite a reconsideration of the place occupied by the retail sector in macro models. Many modelers con(cid:135)ate the notion of producers and retailers, and then calibrate their model to match pricing facts of retailers only. Our analysis suggests that the retail sectors(cid:146)shortto medium-term supply curve is quite (cid:135)at relative to that of other sectors such as manufacturing, a (cid:133)nding in line with the relatively low-margin, high-volume nature of the industry. Although we do not observe (marginal) costs in our dataset, it seems sensible to conjecture that our low supply elasticity estimates could re(cid:135)ect relatively steady marginal costs and markups, a possibility more directly suggested by Eichenbaum, Jaimovich, and Rebelo(cid:146)s (2011) evidence that item-level retail markupsvarylittlearoundtheirmean. Ifso, theninvestigationsofmarkupbehavioranddeviations from constant returns to scale at lower levels of the production chain seem much needed. 22

References [1] Anderson, Eric, Emi Nakamura, Duncan Simester, and J(cid:243)n Steinsson (2013). (cid:147)Informational Rigidities and the Stickiness of Temporary Sales,(cid:148)NBER Working Papers 19350, National Bureau of Economic Research, Inc. [2] Angrist, Joshua D. and Guido W. Imbens (1995). (cid:147)Two-Stage Least Squares Estimation of the AverageCausalE⁄ectsinModelswithVariableTreatmentIntensity,(cid:148)Journal of the American Statistical Association, vol. 90(430), pages 431(cid:150)442. [3] Barro, Robert (1972). (cid:147)A Theory of Monopolistic Price Adjustment,(cid:148)Review of Economic Studies, vol. 39(1), pages 17(cid:150)26. [4] Bernanke, Ben, Jean Boivin, and Piotr S. Eliasz (2005). (cid:147)Measuring the E⁄ects of Monetary Policy: A Factor-augmented Vector Autoregressive (FAVAR) Approach,(cid:148)Quarterly Journal of Economics, vol. 120(1), pages 387(cid:150)422. [5] Bils, Mark and Peter J. Klenow (2004). (cid:147)Some Evidence on the Importance of Sticky Prices,(cid:148) Journal of Political Economy, vol. 112(5), pages 947(cid:150)985. [6] Blake, Eric S. and Ethan J. Gibney (2011). (cid:147)The Deadliest, Costliest, and Most Intense United States Tropical Cyclones from 1851 to 2010 (and Other Frequently Requested Hurricane Facts),(cid:148)NOAA Technical Memorandum NWS NHC-6, U.S. National Oceanic and Atmospheric Administration. [7] Bradley, Ralph (2012). (cid:147)Price Index Estimation Using Price Imputation for Unsold Items,(cid:148) chapter in: Scanner Data and Price Indexes, NBER Studies in Income and Wealth, vol. 64, pages 349(cid:150)382, University of Chicago Press. [8] Bronnenberg, Bart J., Michael W. Kruger, and Carl F. Mela (2008). (cid:147)Database Paper: The IRI Marketing Data Set,(cid:148)Marketing Science, vol. 27(4), pages 745(cid:150)748. [9] Cavallo, Alberto, Eduardo Cavallo, and Roberto Rigobon (2013). (cid:147)Prices and Supply Disruptions during Natural Disasters,(cid:148)NBER Working Papers 19474, National Bureau of Economic Research, Inc. [10] Chevalier,JudithA.andAnilK.Kashyap(2013).(cid:147)BestPrices: PriceDiscriminationandConsumer Substitution,(cid:148)NBER Working Papers 16680, National Bureau of Economic Research, Inc. [11] Chevalier, Judith A., Anil K. Kashyap, and Peter E. Rossi (2003). (cid:147)Why Don(cid:146)t Prices Rise During Periods of Peak Demand? Evidence from Scanner Data,(cid:148)American Economic Review, vol. 93(1), pages 15(cid:150)37. [12] Christiano, Lawrence J., Martin Eichenbaum, and Charles L. Evans (1999). (cid:147)Monetary policy shocks: What have we learned and to what end?,(cid:148)Handbook of Macroeconomics, in: John B. Taylor and Michael Woodford (ed.), vol. 1, pages 65(cid:150)148, Elsevier. [13] Cleveland, William P. and Stuart Scott (2007), (cid:147)Seasonal Adjustment of Weekly Time Series with Application to Unemployment Insurance Claims and Steel Production,(cid:148)Journal of O¢ cial Statistics, vol. 23, pages 209(cid:150)221. 23

[14] Coibion, Olivier, Yuriy Gorodnichenko, and Gee Hee Hong (2012). (cid:147)The Cyclicality of Sales, RegularandE⁄ectivePrices: BusinessCycleandPolicyImplications,(cid:148)NBERWorkingPapers 18273, National Bureau of Economic Research, Inc. [15] Dhyne,Emmanuel,LuisJ.`lvarez,HervØLeBihan,GiovanniVeronese,DanielDias,Johannes Ho⁄mann, Nicole Jonker, Patrick L(cid:252)nnemann, Fabio Rumler, and Jouko Vilmunen (2005). (cid:147)Price Setting in the Euro Area: Some Stylized Facts from Individual Consumer Price Data,(cid:148) Working Paper 524, European Central Bank. [16] Eichenbaum, Martin, Nir Jaimovich, and Sergio Rebelo (2011). (cid:147)Reference Prices, Costs, and Nominal Rigidities,(cid:148)American Economic Review, vol. 101(1), pages 234(cid:150)62. [17] Federal Emergency Management Agency, (cid:147)Declared Disasters by Year or State,(cid:148)last accessed December 2, 2011, http://www.fema.gov/news/disaster_totals_annual.fema. [18] Gagnon, Etienne, David L(cid:243)pez-Salido, and Nicolas Vincent (2013). (cid:147)Individual Price Adjustment along the Extensive Margin,(cid:148)chapter in: NBER Macroeconomics Annual 2012, vol. 27, pages 235(cid:150)281. [19] Geaghan, Kimberly A. (2011). (cid:147)Forced to Move: An Analysis of Hurricane Katrina Movers 2009 American Housing Survey: New Orleans,(cid:148)Social, Economic, and Housing Statistics Division Working Paper Number 2011-17, U.S. Census Bureau. [20] Golosov, Mikhail and Robert E. Lucas Jr. (2007). (cid:147)Menu Costs and Phillips Curves,(cid:148)Journal of Political Economy, vol. 115, pages 171(cid:150)199. [21] Gopinath,GitaandOlegItskhoki(2010).(cid:147)FrequencyofPriceAdjustmentandPass-Through,(cid:148) Quarterly Journal of Economics, vol. 125 (2), pages 675(cid:150)727. [22] Hall, Robert E. (2009). (cid:147)By How Much Does GDP Rise If the Government Buys More Output?,(cid:148)Brookings Papers on Economic Activity, vol. 40(2), pages 183(cid:150)249. [23] Hong, Gee Hee and Nicholas Li (2013). (cid:147)Market Structure and Cost Pass-Through in Retail,(cid:148) Working Papers tecipa-470, University of Toronto, Department of Economics. [24] Karadi, Peter and Adam Rei⁄ (2012).(cid:147)Large shocks in menu cost models,(cid:148)Working Paper Series 1453, European Central Bank. [25] Klenow, Peter J. and Oleksiy Kryvtsov (2008). (cid:147)State-Dependent or Time-Dependent Pricing: DoesItMatterforRecentU.S.In(cid:135)ation?,(cid:148)Quarterly Journal of Economics, vol.123(3), pages 863(cid:150)904. [26] Kleshchelski, Isaac and Nicolas Vincent (2009). (cid:147)Market Share and Price Rigidity,(cid:148)Journal of Monetary Economics, vol. 56(3), pages 344(cid:150)352. [27] Kruger, Michael W. and Daniel Pagni (2008). IRI Academic Data Set Description, version 1.31, Chicago: Information Resources Incorporated. [28] Lach, Saul (2007). (cid:147)Immigration and Prices,(cid:148)Journal of Political Economy, vol. 115(4), pages 548(cid:150)587. [29] Liu, Zheng and Sylvain Leduc (2013).(cid:147)Uncertainty Shocks Are Aggregate Demand Shocks,(cid:148) 2013 Meeting Papers 270, Society for Economic Dynamics. 24

[30] Lee, Lung-Fei and Mark M. Pitt (1986). (cid:147)Microeconometric Demand Systems with Binding Nonnegativity Constraints: The Dual Approach,(cid:148)Econometrica, Econometric Society, vol. 54(5), pages 1237(cid:150)1242. [31] MacDonald, James M. (2000). (cid:147)Demand, Information, and Competition: Why Do Food Prices Fall at Seasonal Demand Peaks?,(cid:148)Journal of Industrial Economics, vol. 48(1), pages 27(cid:150)45. [32] Nakamura,Emi(2008).(cid:147)Pass-ThroughinRetailandWholesale,(cid:148)American Economic Review: Papers & Proceedings, vol. 98(2), pages 430(cid:150)437. [33] Nakamura, Emi and J(cid:243)n Steinsson (2008). (cid:147)Five Facts about Prices: A Reevaluation of Menu Cost Models,(cid:148)Quarterly Journal of Economics, vol. 123(4), pages 1415(cid:150)1464. [34] National Climate Data Center, (cid:147)NCDC Climate Data Online: Daily Surface Data,(cid:148)last accessed December 5, 2011, http://cdo.ncdc.noaa.gov/pls/plclimprod/somdmain.somdwrapper. [35] Ravn, Morten, Stephanie Schmitt-GrohØ, and Mart(cid:237)n Uribe (2006). (cid:147)Deep Habits,(cid:148)Review of Economic Studies, Oxford University Press, vol. 73(1), pages 195(cid:150)218. [36] Romer,ChristinaD.andDavidH.Romer(2004).(cid:147)ANewMeasureofMonetaryShocks: Derivation and Implications,(cid:148)American Economic Review, vol. 94(4), pages 1055(cid:150)1084. [37] Rotemberg, Julio J. (2005). (cid:147)Customer anger at price increases, changes in the frequency of price adjustment and monetary policy,(cid:148)Journal of Monetary Economics, vol. 52(4), pages 829(cid:150)852. [38] Rotemberg, Julio J. (2011). (cid:147)Fair Pricing,(cid:148)Journal of the European Economic Association, vol. 9(5), pages 952(cid:150)981. [39] Shea, John (1993). (cid:147)Do Supply Curves Slope Up?,(cid:148)Quarterly Journal of Economics, vol. 108(1), pages 1(cid:150)32. [40] Warner, Elizabeth J. and Robert B. Barsky (1995). (cid:147)The Timing and Magnitude of Retail Store Markdowns: Evidence from Weekends and Holidays,(cid:148)Quarterly Journal of Economics, vol. 110(2), pages 321(cid:150)352. 25

Appendix A: Cleaning and organizing the IRI data A.1 Data trimming and (cid:133)ltering We perform a number of data cleaning steps to create a sample suitable for our purposes. We only use observations pertaining to grocery stores to ensure maximum comparability across stores and retail chains. The IRI sample also includes observations from drugstores but the number of such stores is more limited than for grocery stores. We exclude two product categories from the sample, cigarettes and photo supplies, because their prices and advertising are heavily regulated and they gradually became obsolete over our sample period, respectively. As is common with scanner data, we obtain a weekly transaction price by dividing total revenue by the number of units sold. In a tiny proportion of cases, the division yields a price with fractional cents (e.g., $4.8573). For New York City, fractional prices represent less than 0.2 percent of posted prices. The origin of these fractional prices is unclear; they may be related to reporting errors, price adjustments during the week, membership card usage, etc. Upon inspection, we opted to keep them in the sample as they were very close to prices in neighboring periods. Excluding them would be inconsequential for our results. In addition to analyzing posted prices, we consider measures of regular prices and reference prices to capture lower-frequency movements in prices. Our regular price (cid:133)lter is based on that proposed by Nakamura and Steinsson (2008) in section A of their technical appendix. It removes temporary price drops that are followed by an increase to a price at or above the previous price or to a new regular price. Our implementation uses the parameters J = 6, K = 3, and L = 3.21 Our reference price (cid:133)lter is a variant of that of Eichenbaum, Jaimovich, and Rebelo (2011). The referencepriceofaniteminweektcorrespondstothemodalpriceobservedovera13-weekcentered window. To compute monthly posted, regular, and reference price statistics, we retain only weekly observations that encompass the 15th day of each month to match the BLS(cid:146)practice of sampling in the same narrow time window every month. Finally, all weekly and monthly statistics employ only observations from frequently traded items that are free of anomalous price movements, as we now detail in the context of deriving price and quantity indexes. A.2 Constructing indexes of total revenues, prices, and quantities We (cid:133)rst compute indexes of total store revenues at the IRI market(cid:150)product category level (the (cid:147)stratum(cid:148)level). This task is made possible by the IRI dataset(cid:146)s comprehensive coverage of items available for sale at participating stores within each of the 29 product categories. This feature makes it possible to track the evolution of a store(cid:146)s total revenues by product category over time and to monitor the entry and exit of items in its o⁄ering. Occasionally, all observations from a 21NakamuraandSteinsson(cid:146)s(2008)technicalappendixdoesnotdiscusstheissueofmissingprices,whicharequite commoninourweeklyscannerdata. Wecarriedforwardthelastobservedpriceto(cid:133)llinmissingpricesbeforerunning the (cid:133)lter. Regular prices imputed that way are censored after running the (cid:133)lter. 26

storemaybecodedasmissing. Insuchproblematicweeks, weimputemissingstore(cid:146)stotalrevenues usingaveragerevenuegrowthforthesubsetofstoresinthesameIRImarketwhoserevenuesforthe productcategoryareobserved. Formally,letsindexstoresinmarketmandproductcategorycthat are present in the sample at week t and for which total sales in week t 1 are known (because they (cid:0) m;c were either directly observed or imputed from the preceding period). Let also I be an indicator s;t that store s(cid:146)data are not systematically missing in the period. Whenever all observations are missing at some stores, we (cid:133)rst compute the revenue growth for stores that are reporting correctly in period t, m;c m;c m;c g = revenues = revenues ; t 0 s;t 1 0 s;t (cid:0) 1 1 sIm;c=1 sIm;c=1 jXs;t jXs;t @ A @ A m;c m;c and set total revenues at problematic stores equal to g revenues . The estimated growth t (cid:1) s;t 1 (cid:0) rate of revenues in the stratum is then m;c revenues revenues m;c s;t t = s : revenues m;c Prevenues m;c t 1 s;t 1 (cid:0) s (cid:0) P Thisestimateistransformedintoanindexoftheleveloftotalrevenuesbycomputingitscumulative product over time. The index is scaled so that its geometric mean in 2005 matches average weekly nominal sales observed in the sample that year. We follow a methodology similar to that employed by the BLS for the U.S. CPI to compute disaggregated price indexes for the IRI sample. We (cid:133)rst obtain an estimate of in(cid:135)ation at the IRI market(cid:150)product category level using a geometric mean formula, p wi;t m;c i;t (cid:25) = ; t p i 2Y I t m;c(cid:18) i;t (cid:0) 1 (cid:19) m;c where (cid:25) is (gross) in(cid:135)ation in period t for market m and product category c; w is the relative t i;t m;c weight of item i, and I is the set of admissible items. In practice, the BLS adjusts item weights t to ensure that the CPI sample is representative of U.S. households(cid:146)shopping habits as established throughitsPoint-of-PurchaseSurvey. FortheIRIsample, weuseuniformweightsacrossadmissible itemsinpartduetothegreatersimilarityinoursampleofstores, whichallpertaintoaretailchain. Like the BLS, we restrict the set of items admissible to compute in(cid:135)ation to those that are actively traded, arestrictionwhich, inthecaseoftheIRIsample, hastheaddedbene(cid:133)toflimitinginference issues created by the censoring of price observations that have no transactions. In particular, we restrict admissible price observations to those pertaining to blocks of 52 consecutive weeks with at most 15 percent of missing observations (that is, at most 8 missing weekly prices). Like the BLS, we impute missing item prices by incrementing their previous weekly price by the average log price increase of items pertaining to the same stratum whose prices are not missing in the current 27

period.22 To limit measurement error, we exclude all items featuring a movement in their weekly log price in excess of 1.9 (such a movement corresponds to a drop in excess of 85 percent in the case of negative adjustments). We also discard every item(cid:146)s (cid:133)rst and last 12 price observations, thus shaving o⁄ roughly one quarter worth of price changes at each end of item price histories. This lattertrimmingseekstopreventselectione⁄ectsassociatedwiththeIRI(cid:146)spracticeofdroppingfrom its sample observations with missing prices at either the beginning or the end of price trajectories. Unlessindicatedotherwise, allstatisticsreportedinthepaperarecomputedusingthesubsampleof frequentlytradeditemsthatarefreeofsuspiciouslylargepricemovementsandwhosepricehistories are trimmed at both ends. We transform our in(cid:135)ation estimates at the stratum into measures of the price level by calculating these estimates(cid:146)cumulative product over time. We then scale the price indexes so that their geometric mean equals unity in 2005. We (cid:133)nally obtain a real same-store quantity index at the stratum level by de(cid:135)ating our total revenue index by our geomeans price indexes.23 This approach of computing a quantity index by e⁄ectively pooling price and spending data from separate surveys mimics that employed by the BEA for disaggregated real personal consumption expenditures in the U.S. National Income and Product Accounts. Also echoing the BEA methodology, we aggregate stratum statistics across product categories using a chain-weighted approach. We compute a Fisher quantity index, QF = QPQL, as the t t t geometric average of a Paasche quantity index, p p q QP = j j;t j;t ; t p q Pj j;t j;t 1 (cid:0) P and a Laspeyres quantity index, p q QL = j j;t (cid:0) 1 j;t : t p q Pj j;t 1 j;t 1 (cid:0) (cid:0) The index j applies to all stratum statistics bPeing aggregated. Similarly, we calculate a Fisher price index, PF = PPPL, as the geometric average of a Paasche price index, t t t p p q PP = j j;t j;t ; t p q Pj j;t 1 j;t (cid:0) P and a Laspeyres price index, p q PL = j j;t j;t (cid:0) 1 :24 t p q Pj j;t 1 j;t 1 (cid:0) (cid:0) 22We also experimented with carrying forward anPitem(cid:146)s last observed price whenever its current price is missing. The resulting price indexes were nearly identical. 23Theinterpretationofourindexesiscomplicatedbyachangeinthetreatmentoffrequentshoppercarddiscounts. Starting January 2002, IRI is reporting dollars spending net of loyalty program discounts for the subset of stores that makes use of these programs and report the value of these rebates to IRI. Before that date, participating stores onlyreportedtotalsalesbeforerebatesareapplied. Weignorefornow possiblebreaksinourindexescreated bythis methodologicalchangebutouranalysisofnaturaldisasterswill(cid:135)agthehandfulcasesforwhichthischangecouldbe problematic. 24Contrary to geomeans price indexes, Fisher price indexes have the attractive property of taking into account 28

A.3 Comparing the level of prices One attractive feature of the IRI sample is that we observe item barcodes, which allows us to directly compare the level of prices across stores for speci(cid:133)c items or for baskets of items. Let I~ t be a set of items sold simultaneously in stores s and s in period t. Let q~ be the number of units i;t 0 of item i in the comparison basket.25 The price of purchasing the basket in store s relative to the price of purchasing it in store s is given by 0 p q~ R t k;k 0 = i 2 I~ t p i;s;t q~ i;t ; Pi I~ t i;s 0 ;t i;t 2 P where p denotes the price of item i in store s at time t. This formula can be used to compare i;s;t the level of price across groups of stores by rede(cid:133)ning p and p as the average unit price of i;s;t i;s;t 0 item i in groups of stores s and s , respectively. 0 A.4 Adjustments for breaks and seasonal patterns Several IRI product categories display a jump in the number of barcodes per store at the beginning of 2007. The origin of this jump is not explicated in IRI documentation but it likely re(cid:135)ects a broadeningofproductcategoryde(cid:133)nitionsratherthananactualincreaseinthenumberofbarcodes o⁄ered by stores. To prevent spurious breaks in our disaggregated total sales and quantity indexes, we impute total sales growth in the problematic week at the stratum level using average total sales growth of items that are present in both the last week of 2006 and the (cid:133)rst week of 2007.26 In addition, a number of stores exiting the IRI sample early report exceptionally large sales(cid:151) sometimes over ten times normal volumes(cid:151)during their last couple of weeks in the sample. The reason for such skyrocketing reported sales is unclear to us but their occurrence seems largely orthogonal to the object of our study. For this reason, we systematically drop the last two periods in the sample of stores that exit early to remove any associated breaks in the level of our series. Inadditiontocontrollingforbreaks,weoftenwanttocontrolforseasonalvariationandholidays when comparing retailing activities during, say, natural disasters, to those during periods in which they are absent. To this end, we (cid:133)lter our stratum-level indexes with a procedure developed by Cleveland and Scott (2007) for weekly time series. The (cid:133)lter allows us to account for the precise variation in spending shares across periods. We contemplated the computation of Fisher formulas directly on micro price and quantity data rather than on our disaggregated indexes. Unfortunately, the presence of weeks without transactions and, more generally, missing observations makes this approach largely impractical on a large scale. See LeeandPitt(1986)foradiscussionofzerotransactiondatainanindustrialorganizationcontextandBradley(2003) for an exploration of alternative imputation approaches. 25The computation of our geomeans price indexes assumed that the comparison basket was (cid:133)xed over time. Here we allow the composition to vary to be able to compare, say, how much the basket purchased by the consumers of store s would have cost had they instead shopped in store s0. 26The imputation proceeds by regressing total sales growth for all items on a constant and total sales growth for continuingitemsinthestratumoveracentered52-weekwindowaroundproblematicperiods. Weexcludeobservations in the window for which the di⁄erence in total sale growth between the full sample and the subsample of continuing observationsexceed5percent. Wethencomputethebestlinearpredictorforproblematicperiods. The(cid:133)rstandlast 12 weekly observations ofeach item history are excluded from the calculation ofsales growth for continuing items to avoid selection e⁄ects related to the trimming of items with no transactions. 29

timing of holidays and to exclude particular periods from the computation of the seasonal factors. Our analysis of labor con(cid:135)icts excludes the strike periods while that of Hurricane Katrina drops the week immediately before the storm and the entire federal disaster declaration period. For shopping sprees around major snowstorms and hurricanes, we exclude the weeks immediately before, during, and after the storms. National aggregates use all weekly observations. Price and quantity indexes use multiplicative seasonal factors whereas all other indicators, which are in percent, use additive factors. 30

Table 1: Posted price adjustments statistics Weekly Statistics Monthly Statistics (in percent) (in percent) Mean Weekly Mean Weekly Share Mean Mean Share Mean Mean Product Category Frequency Frequency Observations Sales decreases increase decrease decreases increase decrease Beer 196,259 10,346,913 21.6 47.4 8.2 8.6 36.9 46.7 9.5 9.9 Blades 46,388 817,634 17.7 47.7 18.5 20.9 31.6 45.6 17.8 21.3 Carbonated Beverages 360,592 15,307,421 43.9 49.2 20.0 20.5 51.3 48.1 19.5 20.3 Coffee 155,122 3,609,792 25.1 48.3 16.9 18.2 42.2 46.5 17.3 18.9 Cold Cereal 274,289 8,381,493 28.8 48.0 23.0 25.1 43.8 47.6 24.6 27.4 Condiments 64,706 980,701 17.8 47.8 14.7 15.9 36.6 49.9 23.5 25.2 Deodorant 138,584 862,847 21.4 50.0 24.0 25.2 41.4 50.7 12.1 12.9 Diapers 58,779 1,642,245 27.7 50.2 12.3 12.8 44.3 48.8 18.8 20.4 Facial Tissue 35,411 1,133,320 30.4 49.4 17.3 18.2 52.8 49.4 23.6 24.7 Frozen Dinner 387,087 6,781,188 36.6 49.2 23.2 24.0 55.2 49.2 21.4 22.4 Frozen Pizza 131,892 3,676,668 40.6 49.0 21.3 22.1 34.8 48.3 17.4 19.5 Hot Dogs 52,276 2,105,301 36.8 48.1 27.1 28.8 47.1 46.8 25.5 28.2 Household Cleaners 82,487 969,090 18.8 48.8 15.8 17.4 43.9 49.1 19.9 21.3 Laudry Detergent 111,344 3,666,854 29.7 48.8 19.7 21.1 42.8 45.7 19.7 22.2 Margarine/Butter 68,568 1,644,178 26.4 47.5 18.3 20.0 34.3 42.7 17.5 21.3 Mayonnaise 47,705 1,410,940 22.1 46.1 18.4 21.0 41.2 45.6 11.3 12.9 Milk 113,946 13,607,991 26.1 47.2 10.2 11.4 31.0 46.2 16.5 18.7 Paper Towels 40,895 2,664,537 27.3 48.5 15.4 16.6 40.8 47.7 17.1 18.7 Peanut Butter 49,848 1,180,936 23.6 47.6 13.8 14.9 39.1 46.9 16.0 17.6 Razors 7,444 112,878 23.1 49.1 16.8 18.3 38.5 48.5 16.7 19.2 Salty Snack 364,220 9,506,203 30.2 48.8 19.8 20.7 42.1 48.3 20.6 21.5 Shampoo 132,957 976,285 23.8 50.0 21.6 22.7 40.4 50.0 21.3 22.9 Soup 316,460 4,448,335 23.7 47.9 23.2 25.1 37.9 47.1 23.7 26.2 Spaghetti Sauce 131,143 2,068,199 29.0 49.0 20.6 21.6 46.3 48.6 21.8 23.1 Sugar Substitutes 26,218 438,704 13.7 48.0 10.1 11.5 27.5 46.4 12.0 14.1 Toilet Tissue 53,117 4,067,906 32.6 48.2 15.1 16.2 46.1 47.4 16.9 18.3 Tooth Brushes 69,569 519,586 22.4 49.8 24.7 26.1 36.3 50.1 26.1 28.1 Tooth Paste 111,224 1,114,115 23.8 49.5 21.8 23.0 38.6 49.4 22.7 24.2 Yogurt 246,642 4,940,942 34.2 49.5 19.4 19.8 46.7 49.0 19.7 20.7 Total 3,875,174 108,983,204 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. Mean 133,627 3,758,042 26.9 48.6 18.3 19.6 41.1 47.8 19.0 20.8 Weighted Mean n.a. n.a. 30.4 48.4 17.6 18.7 42.3 47.6 18.7 20.3 Notes: (cid:147)Mean weekly observations(cid:148)refers to the number of observations for which the posted and regular prices in both the current and preceding weeks are non-missing. (cid:147)Mean(cid:148)statistics across product categories use uniform weights whereas (cid:147)weighted mean(cid:148)statistics use mean weekly sales as weights. Monthly statistics use observations for weeks that encompass the 15th day of each month. 31

Table 2: Regular price adjustments statistics Weekly Statistics Monthly Statistics (in percent) (in percent) Mean Weekly Mean Weekly Share Mean Mean Share Mean Mean Product Category Frequency Frequency Observations Sales decreases increase decrease decreases increase decrease Beer 196,259 10,346,913 9.3 46.4 6.6 7.3 21.2 44.6 8.1 8.7 Blades 46,388 817,634 5.9 47.7 10.5 17.5 16.4 42.5 11.6 18.0 Carbonated Beverages 360,592 15,307,421 18.6 49.2 16.1 17.2 32.0 47.4 16.6 17.9 Coffee 155,122 3,609,792 9.9 47.1 11.3 14.2 24.8 44.1 13.0 15.3 Cold Cereal 274,289 8,381,493 9.3 46.6 13.0 17.6 21.1 44.9 14.8 20.0 Condiments 64,706 980,701 7.2 46.5 10.9 13.5 18.3 43.9 13.3 16.5 Deodorant 138,584 862,847 5.6 57.5 15.0 22.3 15.6 52.0 16.6 22.3 Diapers 58,779 1,642,245 8.5 55.8 8.4 11.1 21.0 53.5 9.2 11.2 Facial Tissue 35,411 1,133,320 10.8 49.9 11.9 14.6 23.4 47.9 14.3 17.2 Frozen Dinner 387,087 6,781,188 10.7 50.1 15.1 18.2 24.7 49.2 16.7 20.0 Frozen Pizza 131,892 3,676,668 13.0 49.3 14.8 17.3 27.6 48.5 16.4 18.9 Hot Dogs 52,276 2,105,301 10.2 44.9 16.3 20.5 15.7 46.9 12.9 17.7 Household Cleaners 82,487 969,090 6.0 48.9 10.9 15.5 21.6 42.2 16.8 21.9 Laudry Detergent 111,344 3,666,854 9.5 50.5 12.6 16.1 22.2 49.4 13.8 16.9 Margarine/Butter 68,568 1,644,178 9.1 43.2 10.8 13.9 21.6 40.6 12.9 16.1 Mayonnaise 47,705 1,410,940 8.8 41.7 10.8 14.9 20.5 37.8 11.9 16.4 Milk 113,946 13,607,991 12.1 44.9 6.7 8.5 26.1 42.9 7.8 9.8 Paper Towels 40,895 2,664,537 10.1 48.7 10.1 12.9 22.3 46.4 12.3 15.2 Peanut Butter 49,848 1,180,936 9.6 45.2 9.5 11.3 21.4 44.2 11.2 13.3 Razors 7,444 112,878 7.5 56.4 11.5 17.5 20.3 50.6 13.0 18.2 Salty Snack 364,220 9,506,203 10.3 49.5 14.7 16.9 19.8 47.3 16.2 18.7 Shampoo 132,957 976,285 6.6 58.0 14.3 20.3 18.1 52.9 15.5 20.4 Soup 316,460 4,448,335 8.2 46.7 13.9 17.8 20.3 44.8 15.6 19.6 Spaghetti Sauce 131,143 2,068,199 9.6 48.3 13.1 15.8 23.2 46.9 14.9 17.6 Sugar Substitutes 26,218 438,704 4.7 46.2 7.4 11.4 13.2 43.1 9.4 13.6 Toilet Tissue 53,117 4,067,906 13.5 47.8 10.1 12.1 26.3 45.9 12.1 14.3 Tooth Brushes 69,569 519,586 7.8 54.8 18.7 23.2 17.2 52.3 20.8 25.7 Tooth Paste 111,224 1,114,115 6.4 52.4 14.0 19.2 16.7 49.6 15.9 20.5 Yogurt 246,642 4,940,942 9.3 49.6 12.1 14.3 21.0 48.0 13.1 15.4 Total 3,875,174 108,983,204 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. Mean 133,627 3,758,042 9.2 49.1 12.1 15.6 21.2 46.6 13.7 17.1 Weighted Mean n.a. n.a. 11.2 48.0 12.0 14.5 23.7 46.1 13.3 16.0 Notes: (cid:147)Mean weekly observations(cid:148)refers to the number of observations for which the posted and regular prices in both the current and preceding weeks are non-missing. (cid:147)Mean(cid:148)statistics across product categories use uniform weights whereas (cid:147)weighted mean(cid:148)statistics use mean weekly sales as weights. Monthly statistics use observations for weeks that encompass the 15th day of each month. 32

Table 3: Basic chain prices statisics Posted Prices Regular Prices Adjustement probability (in percent) Adjustement probability (in percent) Equal to Equal to Product Category chain price Conditional Conditional chain price Conditional Conditional (in percent) Unconditional on chain on no chain (in percent) Unconditional on chain on no chain price change price change price change price change Beer 66.6 24.9 73.0 10.3 67.3 12.0 49.2 6.2 Blades 77.0 18.6 70.6 9.2 76.1 9.8 39.7 4.1 Carbonated Beverages 65.8 42.8 90.0 17.4 68.6 17.6 67.9 11.4 Coffee 71.5 25.1 79.4 11.9 72.3 11.9 54.1 6.2 Cold Cereal 68.1 27.2 83.7 11.8 72.0 9.7 57.9 5.4 Condiments 78.6 24.1 73.8 10.2 74.5 11.2 30.9 4.1 Deodorant 76.1 27.8 78.4 13.7 75.5 11.0 43.1 6.2 Diapers 70.0 27.4 84.4 13.2 73.1 10.4 60.5 6.3 Facial Tissue 72.8 34.6 86.5 13.6 74.8 10.7 55.1 6.4 Frozen Dinner 71.3 37.5 87.7 17.6 74.4 11.9 58.4 8.3 Frozen Pizza 74.4 19.2 72.8 9.3 75.5 8.8 44.4 3.7 Hot Dogs 72.6 35.5 90.5 13.1 79.5 9.3 69.5 5.8 Household Cleaners 72.0 29.1 81.9 13.0 73.5 10.5 49.0 6.3 Laudry Detergent 71.3 24.6 84.0 9.3 74.2 9.3 63.2 4.2 Margarine/Butter 70.2 21.0 79.9 8.7 72.5 9.6 58.2 4.6 Mayonnaise 67.6 25.1 81.6 11.6 70.4 13.3 64.8 6.1 Milk 72.3 17.9 74.7 8.9 74.0 9.1 53.6 4.1 Paper Towels 71.2 23.7 83.2 11.9 74.6 9.5 61.4 5.6 Peanut Butter 69.1 22.4 78.9 10.7 71.1 10.7 57.1 5.3 Razors 77.1 26.5 75.4 12.7 73.5 13.0 34.0 5.9 Salty Snack 76.4 29.5 89.6 12.4 81.4 10.0 67.5 6.3 Shampoo 79.9 25.4 77.0 10.8 77.1 10.3 35.5 4.5 Soup 72.2 22.9 79.8 9.7 74.4 9.6 54.7 4.4 Spaghetti Sauce 72.2 27.9 83.1 11.8 73.9 10.2 55.3 5.6 Sugar Substitutes 73.8 14.0 67.0 7.0 75.2 7.4 44.9 2.7 Toilet Tissue 65.8 29.2 85.2 13.7 69.5 12.9 65.3 7.3 Tooth Brushes 70.2 28.2 78.0 16.1 68.7 16.0 48.7 8.6 Tooth Paste 77.8 24.6 78.6 11.1 77.0 9.2 40.4 4.5 Yogurt 72.6 32.3 87.5 12.0 76.9 8.8 61.5 5.1 Mean 72.4 26.5 80.6 11.8 73.8 10.8 53.3 5.7 Weighted Mean 71.2 28.4 82.3 12.3 73.2 11.3 58.0 6.4 Observations n.a. 1,193,827,794 838,593,893 355,233,901 n.a. 1,193,827,794 1,058,562,164 135,265,593 Notes: We de(cid:133)ne a weekly (cid:147)chain price(cid:148)as the modal posted price of a barcode across stores belonging to the same retail chain and IRI market. We require a minimum of (cid:133)ve store prices to compute a chain price. We exclude an item(cid:146)s own price from the computation of the chain price before computing its adjustment probability conditional on a chain price adjustment or conditional on no chain price adjustment. 33

Table 4: Mean Statistics before, during, and after the 2003(cid:150)2004 Southern California supermarket strike On Strike Not on Strike Statistic Before During After Before During After Quantity index 1.000 0.513 0.904 1.000 1.336 1.036 (0.004) (0.020) (0.004) (0.005) (0.018) (0.006) Price index 1.000 1.013 1.012 1.000 1.021 1.009 (0.002) (0.003) (0.001) (0.002) (0.002) (0.001) Frequency of posted 45.9 41.6 48.1 34.1 31.1 32.5 price changes (0.4) (0.6) (0.3) (0.3) (0.4) (0.3) Frequency of regular 20.0 16.1 19.4 15.2 13.4 12.7 price changes (0.3) (0.4) (0.3) (0.2) (0.3) (0.2) Share of items 23.2 22.0 25.0 18.1 16.8 18.9 on sale (0.3) (0.5) (0.2) (0.2) (0.3) (0.2) Mean discount 9.4 9.4 10.2 8.3 6.8 8.5 (0.1) (0.2) (0.2) (0.1) (0.2) (0.1) Share of revenues 30.4 28.2 31.4 27.3 24.2 28.1 from items on sale (0.4) (0.7) (0.3) (0.4) (0.4) (0.3) Share of prices 53.2 55.8 51.9 59.9 62.4 63.5 equal to chain price (0.4) (0.5) (0.3) (0.3) (0.3) (0.2) Addendum Price index, IRI sample 1.000 1.004 1.019 1.000 1.004 1.019 ex. striking markets Notes: The table reports the mean of the time-series statistics shown in (cid:133)gures 1 and 2, along with theirstandarddeviationinparentheses, overthreedistinctperiods. Thelabel(cid:147)before(cid:148)corresponds to the 26-week period immediately before the strike, the label (cid:147)during(cid:148)to the 20-week period of the strike, and the label (cid:147)after(cid:148)to the 59-week period immediately after the strike. The price and quantityindexesarescaledsuchthattheirgeometricmeanequals1inthe(cid:147)before(cid:148)period; allother statistics are in percent. The price index in addendum applies to the full IRI sample excluding the Los Angeles, San Diego, and St. Louis markets. 34

Table 5: Mean Statistics before, during, and after the 2003 St. Louis supermarket strike On Strike Not on Strike Statistic Before During After Before During After Quantity index 1.001 0.566 0.967 1.002 1.053 0.939 (0.007) (0.060) (0.004) (0.011) (0.022) (0.008) Price index 1.000 1.001 1.005 1.000 1.011 1.032 (0.001) (0.002) (0.001) (0.001) (0.001) (0.001) Frequency of posted 24.7 22.5 24.9 22.6 21.1 22.6 price changes (0.3) (3.1) (0.2) (0.3) (0.8) (0.2) Frequency of regular 6.3 5.6 6.5 5.6 3.9 5.1 price changes (0.3) (1.4) (0.2) (0.3) (0.3) (0.1) Share of items 19.7 17.9 19.5 17.8 18.4 18.2 on sale (0.2) (0.5) (0.1) (0.3) (0.1) (0.2) Mean discount 9.7 9.1 10.3 9.6 10.8 10.8 (0.2) (0.5) (0.1) (0.1) (0.4) (0.1) Share of revenues 28.5 23.5 28.8 26.5 28.3 28.1 from items on sale (0.4) (0.9) (0.2) (0.4) (0.5) (0.3) Share of prices 94.3 92.7 94.0 87.5 87.3 87.7 equal to chain price (0.2) (1.1) (0.2) (0.4) (0.4) (0.3) Addendum Price index, IRI sample 1.000 1.001 1.017 1.000 1.001 1.017 ex. striking markets Notes: The table reports the mean of the time-series statistics shown in (cid:133)gures 4 and 5, along with theirstandarddeviationinparentheses, overthreedistinctperiods. Thelabel(cid:147)before(cid:148)corresponds to the 26-week period immediately before the strike, the label (cid:147)during(cid:148)to the 4-week period of the strike, and the label (cid:147)after(cid:148)to the 75-week period immediately after the strike. The price and quantityindexesarescaledsuchthattheirgeometricmeanequals1inthe(cid:147)before(cid:148)period; allother statistics are in percent. The price index in addendum applies to the full IRI sample excluding the Los Angeles, San Diego, and St. Louis markets. 35

Table6: Meanstatisticsbefore,during,andafter2005HurricaneKatrinainNewOrleans,Louisiana Housekeeping All Food and Personal Care Statistic Before During After Before During After Before During After Quantity index 1.001 1.220 1.135 1.001 1.224 1.136 1.001 1.188 1.131 (0.008) (0.019) (0.009) (0.008) (0.021) (0.010) (0.010) (0.049) (0.007) Price index 1.000 1.014 1.030 1.000 1.015 1.031 1.000 1.004 1.025 (0.001) (0.002) (0.002) (0.001) (0.002) (0.002) (0.002) (0.003) (0.003) Frequency of posted 30.6 28.9 30.0 30.9 29.4 30.3 28.7 25.6 28.4 price changes (0.3) (0.5) (0.3) (0.4) (0.6) (0.3) (0.6) (0.8) (0.5) Mean discount 5.2 3.9 5.0 4.9 4.0 4.8 7.2 3.2 6.6 (0.1) (0.3) (0.1) (0.1) (0.4) (0.1) (0.4) (0.5) (0.2) Share of items 14.1 13.4 14.1 14.1 13.6 13.9 13.7 11.3 15.5 on sale (0.3) (0.6) (0.1) (0.3) (0.6) (0.2) (0.4) (0.8) (0.3) Share of revenues 18.1 15.7 17.4 17.6 15.8 16.6 21.9 15.2 22.9 from items on sale (0.4) (1.0) (0.2) (0.4) (1.1) (0.2) (0.8) (1.3) (0.5) Addendum Price index, IRI 1.000 1.007 1.017 1.000 1.007 1.016 1.000 1.007 1.026 sample ex. Katrina Notes: The table reports the mean of each statistic shown in (cid:133)gure 6, along with their standard deviation in parentheses, over three distinct periods. The label (cid:147)before(cid:148)corresponds to the 26week period immediately before Hurricane Katrina struck New Orleans, the label (cid:147)during(cid:148)to the 10-week federal emergency period declared by FEMA, and the label (cid:147)after(cid:148)to the 69-week period immediately after the lifting of the federal emergency. The price and quantity indexes are scaled such that their geometric mean equals 1 in the (cid:147)before(cid:148)period; all other statistics are in percent. The price index in addendum applies to the full IRI sample excluding the New Orleans and Mississippi markets. 36

Table 7: E⁄ect of major snowstorms and hurricanes on retailing activities Frequency of Share of Share of Frequency of Mean Share of items Quantity index Price index regular price revenues fromprices equal to price changes discount on sale (in logs*100) (in logs*100) changes items on sale chain price (percent) (percent) (percent) (percent) (percent) (percent) Major snowstorms Quantity peak analysis Pre storm average 0.00 0.00 31.96 10.77 8.91 16.97 26.86 70.02 Peak week value 12.88 0.12 31.53 10.30 8.46 16.85 25.93 70.15 t test of zero difference 14.14 1.51 1.36 2.70 3.46 0.68 3.11 0.37 Regression results Estimated coefficients 0.74 0.09 0.64 0.20 0.17 0.29 0.47 0.29 Week before storm (1.47 ) (1.85 ) ( 2.81 ) ( 1.52 ) ( 1.51 ) ( 1.86 ) ( 1.92 ) (1.09 ) 11.70 0.00 0.33 0.49 0.32 0.03 0.73 0.22 Week during storm (23.36 ) ( 0.02 ) ( 1.47 ) ( 3.69 ) ( 2.88 ) ( 0.17 ) ( 3.04 ) (0.84 ) 1.85 0.07 0.61 0.44 0.26 0.36 0.91 0.38 Week after storm ( 3.69 ) (1.28 ) ( 2.69 ) ( 3.27 ) ( 2.34 ) ( 2.29 ) ( 3.78 ) (1.42 ) p value, joint significance 0.00 0.20 0.00 0.00 0.00 0.05 0.00 0.33 Number of snowstorms 59 59 59 59 59 59 59 50 Observations 1,456 1,456 1,456 1,456 1,456 1,456 1,456 1,213 Hurricanes Quantity peak analysis Pre storm average 0.00 0.00 30.19 11.05 7.22 15.73 24.08 72.91 Peak week value 12.84 0.06 30.64 11.60 7.22 16.10 24.54 71.84 t test of zero difference 9.14 0.45 0.97 1.22 0.01 1.35 1.18 2.33 Regression results Estimated coefficients 6.44 0.18 0.20 0.41 0.09 0.10 0.35 0.12 Week before hurricane (7.36 ) ( 2.31 ) ( 0.55 ) (1.68 ) (0.49 ) ( 0.44 ) ( 0.91 ) ( 0.29 ) 6.03 0.13 0.71 0.82 0.34 0.34 0.64 1.02 Week during hurricane (6.90 ) (1.67 ) (1.96 ) (3.33 ) ( 2.59 ) ( 1.50 ) ( 1.66 ) ( 2.60 ) 0.97 0.09 0.99 1.05 0.39 0.39 1.38 0.17 Week after hurricane ( 1.11 ) (1.17 ) (2.72 ) (4.29 ) ( 2.24 ) ( 1.71 ) ( 3.61 ) ( 0.43 ) p value, joint significance 0.00 0.01 0.01 0.00 0.01 0.20 0.00 0.08 Number of hurricanes 21 21 21 21 21 21 21 14 Observations 476 476 476 476 476 476 476 300 Notes: For a storm beginning in week t, we de(cid:133)ne the (cid:147)peak week(cid:148)as the week corresponding to the observed peak in quantities over the period t 1, t, and t + 1. Our (cid:147)response at peak (cid:0) analysis(cid:148)reports the mean across storms of each statistic during the peak week and during a (cid:147)prestorm period(cid:148)covering weeks t 4 to t 2. The mean of the logged quantity and price indexes (cid:0) (cid:0) are normalized to zero during the pre-storm period. The t-statistics test the hypothesis that the di⁄erence between the value during the peak week and the average during the pre-storm period is zero across storms in the sample. The (cid:147)regression results(cid:148)are reported for a regression of each statistic on a quadratic time trend speci(cid:133)c to each 6-month snowstorm or hurricane season and market and a set of dummies for the weeks around the storms. The p-values are for F-tests of the hypothesis that the coe¢ cients on the week before, week during, and week after dummies are jointly di⁄erent from zero. Not all markets have a su¢ ciently large number of stores per retail chain to allow for the computation of chain prices. 37

Figure 1: Impact of 2003 Southern California supermarket strike on quantities and prices Quantity index 1.4 1.2 1 0.8 0.6 on strike 0.4 not on strike 7/03 10/03 1/04 4/04 7/04 10/04 1/05 4/05 Month/Year Price index 1.06 on strike not on strike 1.04 1.02 1 0.98 7/03 10/03 1/04 4/04 7/04 10/04 1/05 4/05 Month/Year Relative price controlling for basket composition (cost in stores on strike relative to cost in stores not on strike) 1.15 1.1 1.05 1 fixed basket 0.95 on strike basket not on strike basket 0.9 7/03 10/03 1/04 4/04 7/04 10/04 1/05 4/05 Month/Year Notes: The shaded areas indicate the strike period and the tick marks correspond to the (cid:133)rst day of each month. All series combine data from the San Diego and Los Angeles markets and, with the exception of the price ratios, are seasonally adjusted. The series labeled "on strike" and "not on strike" correspond to stores that experienced a drop in revenues of 10 percent or more and an increase in revenues of 10 percent or more, respectively. The price and quantity indexes are scaled so that their geometric mean equals 1 in the 26-week period immediately before the strike. See the appendix for methodological details. 38

Figure 2: Some key pricing features of the 2003 Southern California supermarket strike Frequency of price changes Frequency of regular price changes 50 22 20 45 18 tn e c 40 on strike tn e c 16 re not on strike re P 35 P 14 12 30 10 25 7/03 10/03 1/04 4/04 7/04 10/04 1/05 4/05 7/03 10/03 1/04 4/04 7/04 10/04 1/05 4/05 Month/Year Month/Year Share of items on sale Mean discount 28 14 26 12 24 tn 22 tn 10 e e c c re 20 re P P 8 18 16 6 14 7/03 10/03 1/04 4/04 7/04 10/04 1/05 4/05 7/03 10/03 1/04 4/04 7/04 10/04 1/05 4/05 Month/Year Month/Year Share of revenues from items on sale Share of prices equal to chain price 40 70 65 35 tn tn 60 e e c 30 c re re P P 55 25 50 20 45 7/03 10/03 1/04 4/04 7/04 10/04 1/05 4/05 7/03 10/03 1/04 4/04 7/04 10/04 1/05 4/05 Month/Year Month/Year Notes: The shaded areas indicate the strike period and the tick marks correspond to the (cid:133)rst day of each month. All series combine data from the San Diego and Los Angeles markets and seasonally adjusted. The series labeled "on strike" and "not on strike" correspond to stores that experienced a drop in revenues of 10 percent or more and an increase in revenues of 10 percent or more, respectively. The price and quantity indexes are scaled so that their geometric mean equals 1 in the 26-week period immediately before the strike. See the appendix for methodological details. 39

Figure 3: Importance of discounts before, during, and after the 2003 Southern California supermarket strike Contribution to total revenues Contribution to total revenues (stores on strike) (stores not on strike) 8 8 before 7 during 7 after s 6 s 6 tn tn io 5 io 5 p p e e g 4 g 4 a a tn tn e 3 e 3 c c re re P 2 P 2 1 1 0 0 0 20 40 60 80 0 20 40 60 80 Percent discount Percent discount Contribution to total savings Contribution to total savings (stores on strike) (stores not on strike) 3.5 3.5 3 3 s tn 2.5 s tn 2.5 io io p e 2 p e 2 g g a a tn 1.5 tn 1.5 e e c c re 1 re 1 P P 0.5 0.5 0 0 0 20 40 60 80 0 20 40 60 80 Percent discount Percent discount Notes: The panels show the distribution of discounts measured in percent of the regular price identi(cid:133)edbyoursales(cid:133)lter. Thediscountsareaggregatedin10-percentage-pointbins,startingwith discounts in the interval (0,10] percent, then (10,20] percent, and so on. Observations in the top panelsareweightedbyeachitem(cid:146)scontributiontototalrevenues(thatis, ! = p q = p q ), i;t i;t i;t j j;t j;t so that the area under the curve integrates to the share of total revenues derived from items on P sale. Observations in the lower panels are weighted by each item(cid:146)s total dollar savings over the reg reg period (that is, ! = (p p )q = p q ), so that the area under the curve integrates to i;t i;t (cid:0) i;t i;t j i;t j;t the average discount o⁄ered by retailers. All statistics combine data from the San Diego and Los P Angeles markets and do not control for seasonality. The series labeled "on strike" and "not on strike" correspond to stores that experienced a drop in revenues of 10 percent or more and an increase in revenues of 10 percent or more, respectively. The label (cid:147)before(cid:148)corresponds to the 26week period immediately before the strike, the label (cid:147)during(cid:148)to the 21-week period of the strike, and the label (cid:147)after(cid:148)to the 58-week period immediately after the strike. 40

Figure 4: Impact of 2003 St. Louis supermarket strike on pricing decisions Quantity index 1.2 1.1 1 0.9 0.8 0.7 0.6 on strike 0.5 not on strike 7/03 10/03 1/04 4/04 7/04 10/04 1/05 4/05 Month/Year Price index 1.06 on strike not on strike 1.04 1.02 1 0.98 7/03 10/03 1/04 4/04 7/04 10/04 1/05 4/05 Month/Year Relative price controlling for basket composition (cost in stores on strike relative to cost in stores not on strike) fixed basket 1.15 on strike basket not on strike basket 1.1 1.05 1 0.95 0.9 7/03 10/03 1/04 4/04 7/04 10/04 1/05 4/05 Month/Year Notes: All series are seasonally adjusted. The shaded areas indicate the strike period. The price and quantity indexes are scaled so that their geometric mean equals 1 in the 26-week immediately period before the strike. The series labeled "on strike" and "not on strike" correspond to stores that experienced a drop in revenues of 10 percent or more and an increase in revenues of 10 percent or more, respectively. See the appendix for methodological details. 41

Figure 5: Impact of 2003 St. Louis supermarket strike on pricing decisions Frequency of price changes Frequency of regular price changes 30 10 9 8 25 tn e tn e 7 c c re re 6 P P 20 5 4 on strike not on strike 3 15 7/03 10/03 1/04 4/04 7/04 10/04 1/05 4/05 7/03 10/03 1/04 4/04 7/04 10/04 1/05 4/05 Month/Year Month/Year Share of items on sale Mean discount 14 22 13 12 20 tn tn 11 e e c c re 18 re 10 P P 9 16 8 14 7 7/03 10/03 1/04 4/04 7/04 10/04 1/05 4/05 7/03 10/03 1/04 4/04 7/04 10/04 1/05 4/05 Month/Year Month/Year Share of revenues from items on sale Share of prices equal to chain price 100 34 32 95 30 tn tn e e c 28 c 90 re re P P 26 85 24 22 80 7/03 10/03 1/04 4/04 7/04 10/04 1/05 4/05 7/03 10/03 1/04 4/04 7/04 10/04 1/05 4/05 Month/Year Month/Year Notes: Allseriesareseasonallyadjusted. Theshadedareasindicatethestrikeperiod. Thepriceand quantityindexesarescaledsothattheirgeometricmeanequals1inthe26-weekperiodimmediately before the strike. See the appendix for methodological details. 42

Figure 6: Impact of 2005 Hurricane Katrina on pricing decisions in New Orleans, Louisiana Quantity index Price index 1.6 housekeeping and 1.06 personal care 1.5 food 1.4 1.04 1.3 1.02 1.2 1.1 1 1 0.98 0.9 4/05 7/05 10/05 1/06 4/06 7/06 10/06 1/07 4/05 7/05 10/05 1/06 4/06 7/06 10/06 1/07 Month/Year Month/Year Frequency of price changes Mean discount 12 35 10 tn 30 tn 8 e e c c re re 6 P 25 P 4 20 2 0 4/05 7/05 10/05 1/06 4/06 7/06 10/06 1/07 4/05 7/05 10/05 1/06 4/06 7/06 10/06 1/07 Month/Year Month/Year Share of items on sale Share of revenues from items on sale 22 30 20 18 25 tn 16 tn e e c c 20 re 14 re P P 12 15 10 10 8 4/05 7/05 10/05 1/06 4/06 7/06 10/06 1/07 4/05 7/05 10/05 1/06 4/06 7/06 10/06 1/07 Month/Year Month/Year Notes: All series are seasonally adjusted. The price and quantity indexes are scaled so that their geometric mean equals 1 in the 26-week period immediately prior to the strike. The shaded areas indicate the period (August 29, 2005, to November 1, 2005) covered by FEMA(cid:146)s major disaster declaration. See the appendix for methodological details. 43

Figure 7: Illustration of snowstorms e⁄ects: winter 2009(cid:150)2010 in Washington, D.C. Quantity index Price index housekeeping and 1.4 1.02 personal care food 1.015 1.3 1.01 1.2 1.005 1 1.1 0.995 1 0.99 0.985 11/09 12/09 1/10 2/10 3/10 4/10 11/09 12/09 1/10 2/10 3/10 4/10 Month/Year Month/Year Frequency of price changes Mean discount 40 12 10 35 8 tn tn e e c 30 c 6 re re P P 4 25 2 20 0 11/09 12/09 1/10 2/10 3/10 4/10 11/09 12/09 1/10 2/10 3/10 4/10 Month/Year Month/Year Share of items on sale Share of revenues from items on sale 20 35 18 30 tn 16 tn e e c c 25 re re P 14 P 20 12 10 15 11/09 12/09 1/10 2/10 3/10 4/10 11/09 12/09 1/10 2/10 3/10 4/10 Month/Year Month/Year Notes: The leftmost shaded areas mark the December 18(cid:150)19, 2009, nor(cid:146)easter while the rightmost shaded areas mark a pair of blizzards that hit the U.S. capital on February 5(cid:150)10, 2010. The price and quantity indexes are scaled so that their geometric mean equals 1 in November 2009. All series are seasonally adjusted. See the appendix for methodological details. 44