It Pays to Set the Menu: Mutual Fund Investment Options in 401(k) plans

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. It Pays to Set the Menu: Mutual Fund Investment Options in 401(k) plans Veronika K. Pool, Clemens Sialm, and Irina Stefanescu 2014-96 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.

It Pays to Set the Menu: Mutual Fund Investment Options in 401(k) Plans∗ Veronika K. Pool Indiana University, Bloomington Clemens Sialm University of Texas at Austin and NBER Irina Stefanescu Board of Governors of the Federal Reserve System August 27, 2014 ∗Veronika K. Pool is at Indiana University, Bloomington. Email: vkpool@indiana.edu. Clemens Sialm is at the McCombs School of Business, University of Texas at Austin, Austin, TX 78712 and at the National Bureau of Economic Research. Email: clemens.sialm@mccombs.utexas.edu. Irina Stefanescu is at the Board of Governors of the Federal Reserve System. Email: irina.stefanescu@frb.gov. We thank Pierluigi Balduzzi, Keith Brown, Lauren Cohen, Van Harlow, Frank de Jong, Olivia Mitchell, Joshua Pollet, Jonathan Reuter, Paul Schultz, Laura Starks, Steve Utkus, Marno Verbeek, Scott Yonker, and seminar participants at DePaul University, Emory University, Indiana University, INSEAD, Yale University, Securities Exchange Commission,FederalReserveBoard,UniversityofGeorgia,UniversityofAlabama,CaliforniaStateUniversity Fullerton, at the American Economic Association Meeting in San Diego, the FIRS Conference in Croatia, the Humboldt University Conference on Recent Advances in Research on Mutual Funds, the IU-Notre Dame- Purdue Summer Symposium, the National Bureau of Economic Research Conference on Personal Retirement Challenges, the NETSPAR spring workshop, the Nova Finance Conference on Pensions and Retirement, the SecondMSUFCUConferenceonFinancialInstitutionsandInvestmentsatMichiganStateUniversity,andthe Society for Financial Studies Cavalcade in Miami for helpful comments, and NETSPAR, Indiana University, and the University of Texas at Austin for financial support. Clemens Sialm thanks the Stanford Institute for EconomicPolicyResearchforfinancialsupportduringhisSabbaticalleave. Theviewsexpressedinthispaper are those of the authors and do not reflect the views of the Board of Governors of the Federal Reserve System or its staff members.

It Pays to Set the Menu: Mutual Fund Investment Options in 401(k) Plans August 27, 2014 Abstract Thispaperinvestigateswhethermutualfundfamiliesactingasserviceprovidersin401(k) plans display favoritism toward their own funds. Using a hand-collected dataset on retirement investment options, we show that poorly-performing funds are less likely to be removed from and more likely to be added to a 401(k) menu if they are affiliated with the plan trustee. We find no evidence that plan participants undo this affiliation bias through their investment choices. Finally, the subsequent performance of poorlyperforming affiliated funds indicates that these trustee decisions are not information driven. JEL Classification: G23, J23 Keywords: 401(k), pension plans, trustee, favoritism, mutual funds 1

1 Introduction Employer-sponsored defined contribution (DC) accounts have gained significant importance around the world. In the United States, the value of 401(k) assets reached $4.2 trillion in 2013. The growth represents important business opportunities for mutual funds as they manage approximately half of the 401(k) investment pool.1 Moreover, mutual fund families often play an active role in creating the menu of investment options as – in addition to asset management services – they also provide administrative services in these employee benefit plans. Fund families involved in the plan’s design often face conflicting incentives. While they have an incentive to include their own proprietary funds on the menu even when more suitable options are available from other fund families,2 they are also pressured by plan sponsors to create menus that serve the interests of plan participants. Surprisingly, little is known about whether and how these conflicting incentives influence 401(k) menus. This is concerning given that DC accounts are the main source of retirement income for many of the beneficiaries. Inthispaper, weexaminetheconflictingincentivesofmutualfundcompaniesinthe401(k) industry. Building on Cohen and Schmidt (2009), we collect information on the identity of the trustee of employer-sponsored 401(k) plans. Focusing on menu changes, we hypothesize that these service providers are inclined to include their own funds on the investment menu and subsequently reluctant to remove them. Additionally, they may also be less sensitive to the performance of their own funds in menu altering decisions as they have an incentive to support poorly-performing proprietary funds or, more generally, those that are experiencing significant cash outflows (Coval and Stafford, 2007). 1Federal Reserve Statistical Releases and Investment Company Institute (ICI). 2See the U.S. Government Accountability Office (2011) report on “Improved Regulation Could Better Protect Participants from Conflicts of Interest”. 1

To investigate this favoritism hypothesis, we hand collect information on the menu of mutual fund options offered in a large sample of 401(k) plans for the period 1998 to 2009 from annual filings of Form 11-K with the U.S. Securities and Exchange Commission (SEC). Our sample includes plans that are trusteed by a mutual fund family as well as plans with nonmutual fund trustees. Most 401(k) plans in our sample adopt an open architecture whereby investment options include not only funds from the trustee’s family (“affiliated funds”) but those from other mutual fund families as well (“unaffiliated funds”). An interesting feature of our dataset is that a given fund often contemporaneously appears on several 401(k) menus that are administered by different fund families. This data feature provides us with an unique identification strategy and allows us to contrast how the very same fund is viewed across menus where the fund is affiliated with the trustee and menus where it is not. Our results reveal significant favoritism toward affiliated funds. Affiliated funds are more likely to be added and less likely to be removed from 401(k) plans. The biggest difference between how affiliated and unaffiliated funds are treated on the menu occurs for the worst performingfunds, whichhavebeenshowntoexhibitsignificantperformancepersistence(Carhart, 1997). For example, mutual funds ranked in the lowest decile based on past performance (among the universe of funds in the same style category over the prior 36 months), are approximately twice as likely to be deleted from those menus on which they are unaffiliated with the trustee than from those on which they are affiliated with the trustee. Protecting poorly-performing funds by keeping them on the menu helps mutual fund families to dampen the outflow of capital triggered by bad performance and, as a result, mitigates fund distress. Although the investment opportunity set of the plan is limited to the available menu choices, participants can freely allocate their contributions among these options. If participants are aware of provider biases or are simply sensitive to poor performance, they can at least partially undo favoritism in their own portfolios by, for instance, not allocating capital 2

to poorly performing affiliated funds. Therefore, to investigate whether menu favoritism has an impact on the overall allocation of plan assets, we examine the sensitivity of participant flows to the performance of affiliated and unaffiliated funds. Consistent with previous studies documenting that DC pension participants are naive and inactive (Benartzi and Thaler, 2001; Madrian and Shea, 2001; Agnew, Balduzzi, and Sunden, 2003), we show that participants are generally not sensitive to poor performance and do not undo the menu’s bias toward affiliated families. This in turn indicates that plan participants are affected by the biased behavior of mutual fund companies. Finally, whileourevidenceonfavoritismisconsistentwithadverseincentives, fundfamilies mayalsohavesuperiorinformationabouttheirownproprietaryfunds. Therefore, itispossible that they show a strong preference for these funds not because they are necessarily biased toward them, but rather, due to positive information they possess about these funds. To investigate this possibility, we examine future fund performance. For instance, if – despite lackluster past performance – the decision to keep poorly performing affiliated funds on the menu is information driven, then they should perform better in the future. We find that this is not the case: affiliated funds that rank poorly based on past performance but are not deleted from the menu do not perform well in the subsequent year. We estimate that, on average, they underperform by approximately 3.96% annually on a risk- and style-adjusted basis. Our results suggest that the menu bias we document in this paper has important implications for the employees’ income in retirement. Our study belongs to a nascent literature on the effect of business ties in DC plans. Davis and Kim (2007) and Cohen and Schmidt (2009) study conflicts of interest in the 401(k) industry and argue that to protect the valuable business relation that arises between the sponsoring company and mutual fund service providers, families cater to the sponsors while compromising their own fiduciary responsibilities. In particular, Cohen and Schmidt (2009) 3

find that trustee mutual fund families overinvest in the sponsoring company’s stock. They also show that when other mutual funds sell the stock, trustee funds tend to trade in the opposite direction thereby supporting the stock price of distressed firms. Davis and Kim (2007) document that mutual fund votes in shareholder meetings are influenced by 401(k) business ties. Our paper is also related to two additional areas of study. First, we contribute to the broader literature that focuses on the design and characteristics of DC plans.3 Second, our paper is related to the mutual fund literature on favoritism within fund families. Gaspar, Massa, and Matos (2006) show that mutual fund families strategically transfer performance acrossmemberfundstofavorthosefundsthataremorelikelytoincreaseoverallfamilyprofits. Reuter (2006) provides evidence that lead underwriters will use allocations of underpriced IPOs to reward those institutions with which they have strong business relationships.4 Our paper provides evidence that mutual fund families favor their own affiliated funds when they act as service providers of 401(k) pension plans. The rest of the paper is structured as follows. Section 2 provides information on the institutional, economic, and legal background of DC plans. Section 3 describes our data collection and provides summary statistics of our 401(k) plans as well as the mutual funds offered on the plans’ menu. Sections 4–6 discuss our results. Section 7 concludes. 3Benartzi and Thaler (2001), Madrian and Shea (2001), Choi et al. (2002, 2004), Del Guercio and Tkac (2002), Duflo and Saez (2002), Agnew, Balduzzi, and Sunden (2003), Huberman and Jiang (2006), Elton, Gruber,andBlake(2006,2007),Brown,Liang,andWeisbenner(2007),GoyalandWahal(2008),Carrolletal. (2009),Tangetal.(2010),BalduzziandReuter(2012),BrownandHarlow(2012),MitchellandUtkus(2012), Goldreich and Halaburda (2013), Christoffersen and Simutin (2014), and Sialm, Starks, and Zhang (2014) study the structure of pension plans and provide evidence that retirement savers are subject to behavioral biases and rarely adjust their portfolios. 4Several additional papers study favoritism within asset management companies. Kuhnen (2009) finds that fund directors and advisory firms that manage the funds hire each other preferentially based on the intensity of their past interactions. Bhattacharya, Lee, and Pool (2013) find that affiliated funds of mutual funds cross-subsidize funds in their complex that experience liquidity shortfalls. 4

2 Institutional Background 401(k) menus are jointly determined by the plan sponsor (i.e., employer) and the plan’s service providers. In this paper we use the term “service provider” to refer to those entities that provide administrative services to 401(k) plans. These services include trustee services (i.e., providing the safe holding of the plan’s assets in a trust), recordkeeping services (i.e., maintaining plan records, processing contributions and distributions, and issuing statements), participant education (i.e., online or face-to-face investment education), and compliance services (i.e., preparation of forms and legal services).5 These various services are often bundled and provided by a single entity. For example, over 90% of the mutual fund trustees in our sample are also recordkeepers of the same plan. In addition to these administrative services, mutual fund families often also serve as investment managers by offering their own funds as investment options on the menu. Service providers are selected by the plan sponsor and their compensation structure is negotiated along multiple dimensions. The first component of compensation is explicit and consists of administrative fees collected from the various investment options offered on the menu (i.e., asset-based fees), from sponsors (i.e., per plan fees), or from participants (i.e., per participant fees). In practice, per plan and per participant fees are less common. Instead, most administrative fees are asset-based and are typically built into the expense ratios paid by participants when investing in the funds offered by the plan. Whereas service providers can keep the management fees they generate from their own funds on the menu, they are often compensated by the unaffiliated funds through revenue sharing arrangements. Under these arrangements service providers will receive a fixed proportion of assets under management from the unaffiliated investment management companies (i.e., a portion of the expense ratio 5A description of the services provided is available at: http://www.ici.org/pdf/per19-04.pdf. 5

these companies collect from participants). For example, if the revenue sharing proportion is 20 basis points, then unaffiliated mutual funds will return 20 basis points of their expense ratio to the service providers.6 Such revenue sharing arrangements increase the incentives of service providers to include unaffiliated investment options in the plan. The second component is implicit compensation, which arises from the indirect benefits that fund families obtain from administering a 401(k) plan. These benefits include the ability to control the set of affiliated mutual fund options on the menu, as we document in this study. In addition, service providers obtain access to plan participants and can build a longterm relation with these employees. For example, such access allows them to motivate plan participantstoroll-overtheir401(k)planassetstoanaffiliatedIndividualRetirementAccount (IRA) after they retire or leave their jobs.7 A 2011 Deloitte survey of 401(k) fees finds that negotiations between sponsors and service providers include the number and type of investment options offered on the menu, the choice of offering proprietary vs. non-proprietary funds, or whether and what type of educational services may be offered to plan participants.8 Sponsors may benefit from structuring provider compensation in the form of asset-based fees in combination with implicit compensation arrangements, if their employees do not recognize the potential conflicts of interest in 401(k) plan design. Thus, sponsors may have the opportunity to reduce their own costs of administering a plan by allowing mutual fund providers to favor their own proprietary investment options on the menus. 6The U.S. Government Accountability Office (GAO) (2011) documents “revenue-sharing payments from hundreds of share classes of different investment funds that ranged from 5 to 125 basis points” (pages 16-17). 7The GAO (2013) report states that “the opportunity for service providers to sell participants their own retail investment products and services, such as IRAs, may create an incentive for service providers to steer participantstowardthepurchaseofsuchproductsandservicesevenwhentheymaynotservetheirparticipants’ best interests.” (page 22). 8See, www.ici.org/pdf/rpt 11 dc 401k fee study.pdf. 6

There are some safeguards that mitigate conflicts of interest in 401(k) plans. In particular, sponsors face constraints to offer 401(k) plans that satisfy legal and regulatory requirements. Employer-sponsored 401(k) plans are subject to regulatory and legal constraints imposed by the Employee Retirement Income Security Act (ERISA). ERISA has the requirement that plan fiduciaries act “solely in the interest of the participants and beneficiaries and (...) for the exclusive purpose of (...) providing benefits to participants and their beneficiaries.” ERISA fiduciary actions are those involving discretionary plan administration, asset or plan management, or investment advice. Over the last decade there have been numerous lawsuits filed against plan sponsors and service providers alleging excessive or hidden fees or improper monitoring of options.9 These legal and regulatory constraints and the sponsor’s involvement in the plan’s design significantly contribute to the prevalence of open architecture 401(k) plans.10 For example, mutual fund providers are motivated for legal reasons to outsource funds from unaffiliated mutualfundsfamiliesiftheirownfundofferingsarelimitedorspecialized, asERISAmandates plans to offer a diversified menu of options, or if their own fees are not competitive, as this reduces the risk of costly litigation. In the rest of the paper, we use an identification strategy that takes advantage of the existence of the open architecture plan design to investigate favoritism in 401(k) plans. 9ERISA rules are cited following Muir (2012) and are available at http://www.law.cornell.edu/uscode/text/29/chapter-18/subchapter-I/subtitle-B/part-4. The U.S. Department of Labor’s Employee Benefits Services Administration website includes additional information on fiduciary obligations in DC plans (http://www.dol.gov/ebsa/publications/fiduciaryresponsibility.html). A discussion of 401(k) lawsuits can be found in http://online.wsj.com/article/SB10001424052970204777904576651133452868572.html. 10See, for example Ruiz-Zaiko and Williams (2007) on the effect of growing litigation uncertainty in the industry. 7

3 Data and Summary Statistics This section describes the sample selection process and provides summary statistics for our sample of 401(k) menus. 3.1 Data collection We manually collect the investment options offered in 401(k) plans from Form 11-K filed with the U.S. Securities and Exchange Commission (SEC). A plan is required to file this form if it offers the stock of the sponsoring company as an investment option for participants. The filing provides an overall description of the plan, identifies the trustee of the plan, and lists the accumulated value of assets invested in the various investment options at the end of the fiscal year. We collect 26,624 links to 11-K filings but restrict this sample to companies covered by COMPUSTAT. From these documents we collect the tables that describe the “Schedule of Assets.” In most cases, the table reports the complete set of investment options offered by the plan, including the employers’ own stock, other common stocks, mutual funds, separate accounts, or commingled trusts. We supplement our Form 11-K information with plan level data from Form 5500 filed with the Department of Labor. The resulting dataset has more than 302,000 observations, containing information at the firm-year-plan-option level. To obtain information on the mutual funds included in DC plans, we match these data to the CRSP Survivorship Bias-Free U.S. Mutual Fund database. Since most plans do not identify the exact share class of the fund offered on the menu, we establish the link between our 401(k) sample and CRSP at the fund-level by combining information on the share classes into fund-level variables. Accordingly, fund age is calculated as the age of the oldest share class, fund size is the sum of the total net assets (TNA) of all share classes, and fund returns 8

and expense ratios are calculated as the TNA weighted average returns and expense ratios of the share classes, respectively. We also classify each mutual fund as “balanced,” “bond,” “domestic equity,” “international equity,” or “other.” We create separate dummy variables for money market, target date, and index funds. We manually group funds into target date and index fund categories based on fund name. Around 62% of the funds in the average plan in our sample are equity funds and 20% are bond funds. There is a steady increase in the numberof targetdatefunds over oursample period, especiallyafter thepassage ofthe Pension Protection Act (PPA) of 2006, also documented by Mitchell and Utkus (2012).11 3.2 Sample description Table 1 describes the composition of our final sample by year. Our data covers 2,494 distinct plans sponsored by 1,826 firms from 1998 to 2009.12 Overall, the final dataset has 13,367 plan-year observations. The number of plans is smaller during the early part of the sample as plan disclosures were generally less comprehensive. Similarly, our data for 2009 are potentially incomplete as they do not include late filers or filers with a late fiscal year end. Our sample is representative of the universe of plans offered by public companies filing Form 5500 with the DepartmentofLaborintermsofplansize,numberofparticipants,andindustrycomposition.13 In our sample, average plan size is approximately $324 million (with a median of $61 million). In 2009, our plans cover $376 billion in retirement assets and 9 million total participants. The typical account size is $42,107 and employees contribute $5,303 per year. The mean (median) percentage of assets invested in employer stock is 17% (10%). 11Following the PPA, the Department of Labor added a new fiduciary protection to ERISA for Qualified Default Investment Alternatives (QDIA), such as target-date funds, traditional balanced funds, and managed account advice services. 12Whenacompanysponsorsplanswithidenticalmenusweretainonlythelargestplaninordertopreserve the time series continuity required when defining deletions and additions. 13Our sample covers 30-35% of the 401(k) assets sponsored by publicly listed companies that report Form 5500. 9

The table also describes information on the structure of the plans. Around 76% of plans have trustees that are affiliated with mutual fund management companies. The sample has 112 distinct mutual fund trustees with, on average, 70 unique mutual fund trustees per year. The remaining plans are trusteed by commercial banks, consulting companies, individuals, or by the sponsoring company itself. We collectively refer to these other entities as “Non-Mutual Fund Trustees.” Non-mutual fund trustees are generally appointed by smaller plans. Tosummarizethegrowingpopularityofopenarchitecture,wereportthreemetrics. Trustee share represents the average proportion of total plan assets invested in mutual funds offered by the trustee family. The average trustee share amounts to around one-third in our sample.14 Additionally, we report the average number of management companies that offer at least one fund on the menu and the Herfindahl index of the menu calculated based on the dollar share of each of these management companies. These measures point to a decline in the share of the assets managed by trustee families and an increase in the number of families on the menu. Indeed, in 1998, 66.4% of mutual fund trusteed menus offered funds from more than one family, while the corresponding figure is 91.1% in 2009. The table also shows an increase in the number of funds offered in the average plan over time. Table 2 describes the characteristics of mutual funds that have been kept on, deleted from, or added to the menu by trustee affiliation. Standard errors of the difference between the mean characteristics of affiliated and unaffiliated funds are two-way clustered at the plan and fund levels.15 Our sample contains 134,789 fund-year observations involving funds that stay on the plan for at least two consecutive years, 18,474 fund deletions, and 29,688 fund additions. Thus, 14The average trustee share appears at first glance to be relatively low. However, this figure includes all plans in our sample, regardless of trustee type. Overall, we find that 47.1% of plans do not include affiliated funds. Trusteeshareamountsto62.4%ifweconditiononplansthatincludeatleastoneaffiliatedfundoption. 15Weinclude plan years inwhich atrustee changeoccurs inoursample andin theanalyses reported inthe paper. Our results are robust to excluding these plan years, as shown in Table A-6 in the Internet Appendix. 10

the unconditional probability for a fund deletion is around 12% per year. On average, each deleted affiliated (unaffiliated) fund accounts for 7.19% (7.60%) of plan assets. About 11.35% (14.57%) of all affiliated (unaffiliated) assets on the menu are deleted each year. By the end of the calendar year, affiliated and unaffiliated funds that are added to the menu during the year represent 14.35% and 20.74% of plan assets, respectively.16 Overall,fundsthataredeletedhavethelowestaverageperformanceacrossthethreegroups, as measured by their percentile performance among funds of the same style in the CRSP fund universe using the past three-year returns. Added funds are younger and come with better performance records than those that are kept or deleted. Thetablealsoshowsthataffiliatedfundstendtohavelowerexpenseratios, lowerturnover, and lower standard deviations of monthly returns. These differences occur as affiliated funds are more likely to be more basic investment options (such as standard domestic equity funds or passively-managed index funds), whereas unaffiliated funds are more likely to be specialized funds (such as international or sector funds). For example, approximately 13% of the affiliated funds in our sample are passively-managed index funds compared to 6% of unaffiliated funds. One reason why service providers outsource these more specialized funds is that they may not offer these investment options in their own product lineup. Nonetheless, the results in the table may point to a potential benefit of affiliated mutual fund options. These explicit benefitsmaycomeasaresultofincreasedimplicitcostshowever, asdescribedearlier. Wenext investigate the costs associated with including affiliated investment options on the menu. 16Simultaneous deletions and additions are the most common menu changes. In our sample, in 40.5% of the plan years the menu does not change, in 6.1% (17.1%) of the plan years we see fund exits (entries) but no entries (exits), and in the remaining 36.3% of the cases both entries and exits occur simultaneously. 11

4 Menu Changes Investmentallocationsin401(k)accountsaredrivenbytheplansponsor, theserviceproviders, and plan participants. In a first step, service providers along with the sponsor select the menu of investment options for the plan. In a second step, participants allocate their retirement savings and contributions across these options. To ensure that the plan continuously offers a suitable set of investment choices, 401(k) plans dynamically adjust 401(k) menus by deleting some investment options and adding others. In this section, we study these menu altering decisions to test whether mutual funds affiliated with the plan’s trustee are treated preferentially relative to funds from other mutual fund companies. 4.1 Univariate Relationship of Fund Deletions We first provide univariate analyses to investigate whether the propensity to delete a fund from the menu depends on whether the fund is affiliated with the trustee. Tomakethecomparisonbetweenthedeletionfrequenciesofaffiliatedandunaffiliatedfunds more meaningful, we also group funds into deciles based on past performance. In particular, we compute the percentile performance of each fund among funds of the same style in the CRSP fund universe.17 Funds are then grouped into decile portfolios based on their prior performance. Figure 1 reports the mean annual deletion frequencies by trustee affiliation for each performance decile using the prior 36 months to evaluate performance. We construct the figure by first computing the average deletion rates for each fund in each year in affiliated and unaffiliated 401(k) plans. We then average the deletion rates within the performance deciles by year. Finally, we average the decile deletion rates across time. Panel A shows the results 17We use the following style categories: “balanced,” “bond,” “domestic equity,” “international equity,” or “other.” 12

based on all funds in our sample. The numbers above the bars denote the differences between the affiliated and unaffiliated deletion frequencies.18 In Panel B, we focus only on those funds that contemporaneously appear on multiple 401(k) menus, at least once as an affiliated fund and at least once as an unaffiliated fund. By comparing the deletion probabilities of the same fund across plans managed by different trustees, our results are not contaminated by different fund characteristics or performance records. The figure shows that affiliated funds are less likely to be deleted from a 401(k) plan than unaffiliated funds regardless of past performance. For example, the average affiliated fund has a deletion rate of 13.7% across all performance deciles, whereas an unaffiliated fund has an average deletion rate of 19.1%. Furthermore, we find that the difference in deletion rates widens significantly if we focus on poorly performing funds. For example, funds in the lowest performance decile in Panel A have a probability of deletion of 25.5% for unaffiliated funds and a probability of deletion of only 13.7%foraffiliatedfunds. Indeedthedeletionrateofaffiliatedfundsinthelowestperformance decile is actually lower than the deletion rates of affiliated funds in deciles two through four. This is surprising provided that Carhart (1997) documents performance persistence among poorly performing funds. On the other hand, we find that in the top decile, affiliated funds are almost as likely to be deleted as unaffiliated funds. Overall, the difference in deletion rates between affiliated and unaffiliated funds is statistically significant for the nine lowest performance deciles. In addition, the difference between affiliated and unaffiliated deletion probabilities in the lowest decile is statistically significantly higher than the corresponding differences in each of the other nine deciles. Panel A of Table A-1 in the Internet Appendix tabulates the deletion frequencies for affiliated and unaffiliated 18Thenumberofobservationsintheindividualperformancedecilesrangesbetween407and867foraffiliated funds and 1,056 and 2,522 for unaffiliated funds. Significance levels for the differences in deletion rates are based on standard errors that are clustered at the fund level. 13

funds, as well as the difference between them. In addition to the 36 month performance evaluation horizon, it also reports results for performance ranks based on prior one and five years. Panel B shows similar results for the subsample of funds that are simultaneously offered as both affiliated and unaffiliated funds.19 In this analysis the funds in each decile are identical across the affiliated and unaffiliated groups. Thus, our results are not driven by differences in fund characteristics. Service providers have an incentive to protect their poorly performing affiliated funds, as many of these funds are experiencing outflows from other investors. For example, we find that investor money flows from the funds’ non-retirement clients equal -3.5% for affiliated funds in decile 1 and 22.8% for affiliated funds in decile 10.20 Thus, fund families reduce the volatility of fund flows by keeping those affiliated funds on the menu that otherwise experience large outflows. The deletion of an affiliated fund does not imply that the number of affiliated funds offered on the menu decreases. Although we do not observe where the assets of deleted funds are transferred, we find that service providers often introduce new affiliated funds when other affiliated funds are deleted. For example, if a plan deletes one or more affiliated funds, then there is a 95.7% probability that the plan will add at least one new affiliated fund during the same year. On the other hand, if a plan deletes one or more unaffiliated funds, then there is only a 43.2% probability that the plan will add at least one other fund from the deleted fund’s family. 19In both panels, standard errors are clustered at the fund level. For additional robustness, Panels C and D in Table A-1 report the corresponding deletion frequencies using the Fama-MacBeth methodology. 20WecomputeinvestormoneyflowsforDCandnon-DCinvestorsfollowingSialm,Starks,andZhang(2014) using information collected from surveys conducted by Pensions & Investments. Money flow (i.e., growth rate of new money (NMG)) by non-DC investors is computed as NMGNON−DC = [NonDCAssets − f,t f,t NonDCAssets ×(1+R )]/[NonDCAssets ×(1+R )], where R corresponds to the return of f,t−1 f,t f,t−1 f,t f,t fund f in year t. We winsorize these money flows at the 95% level. 14

These univariate results provide evidence that service providers favor their own funds when they adjust the investment menu. Favoritism is particularly pronounced for funds that experience poor prior performance. 4.2 Multivariate Relation of Fund Deletions To extend our univariate results in Section 4.1, we examine the performance sensitivity of affiliated and unaffiliated fund deletions using the following linear probability model: DEL = β +β AF +β LowPerf +β HighPerf p,f,t 0 1 p,f,t−1 2 p,f,t−1 3 p,f,t−1 + β AF ×LowPerf +β AF ×HighPerf 4 p,f,t−1 p,f,t−1 5 p,f,t−1 p,f,t−1 + Z(cid:48) γ +(cid:15) , (1) p,f,t−1 p,f,t where DEL is an indicator variable that takes the value of one if mutual fund f has been p,f,t deleted from plan p during year t and zero otherwise, AF is an indicator variable for p,f,t−1 whether the trustee of pension plan p is affiliated with the management company of mutual fund f at the end of year t−1, and Z is a vector of relevant lagged control variables. p,f,t−1 In our baseline model described in equation (1), we use two performance segments, evaluating deletion sensitivities to prior performance separately for below and above median funds. LowPerf and HighPerf are defined as LowPerf = min(Perf ,0.5) and p,f,t−1 p,f,t−1 HighPerf = max(Perf −0.5,0), where Perf is the performance percentile of p,f,t−1 p,f,t−1 p,f,t−1 mutual fund f over the previous three years. Performance percentiles are formed based on the performance of each fund among funds of the same style in the CRSP fund universe and range between zero and one. In some specifications we constrain the sensitivity of deletions to depend linearly on performance. For robustness, we also estimate our model using quintile-based performance segments following Sirri and Tufano (1998) in the Internet Appendix. To control for potential redundancies among menu options, which may lead to fund dele- 15

tions, we add an explanatory variable “MaximumCorr,” which captures the highest pairwise correlation between each option and all other mutual fund investment choices on the menu. The other control variables in Z include the natural logarithm of plan assets invested p,f,t−1 in the fund, the number of options offered on the menu, the expense ratio of the fund, the turnover of the fund, the natural logarithm of the fund’s size, fund age, the standard deviation of the fund’s return, and unreported indicator variables for specific fund types (e.g., domestic equity, international equity, balanced, bond, target date, index, and money market funds) and year fixed effects. Favoritism toward affiliated funds implies that, all else equal, affiliated funds are less likely to be delisted (i.e., β < 0) and that affiliated deletions are less sensitive 1 to prior performance (i.e., β > 0). 3 In the first two columns of Table 3 we report the coefficient estimates for the linear performance model and our baseline two-segment specification, respectively. We estimate equation (1) and the corresponding linear specification using a linear probability model, which allows for a straightforward interpretation of the piecewise linear terms and the corresponding interactions. The standard errors in the table are two-way clustered at the plan and fund levels.21 Consistent with Figure 1 we find that affiliated funds – and especially poorly performing affiliated funds – are significantly less likely to be deleted. For example, based on the results in column 1, affiliated funds are 10.5% less likely to be deleted than unaffiliated funds. Furthermore, a ten percentage point increase in the performance percentile for an unaffiliated fund decreases the probability of deletion by 1.67%, whereas the same performance increase for an affiliated fund decreases the probability of deletion by only 0.64%. Thus, the sensitivity of deletions to inferior fund performance is less than half of that of unaffiliated funds. The 21In Table A-2 of the Appendix we report the corresponding marginal effects using a probit specification. The interaction terms are calculated using the INTEFF command based on Ai and Norton (2003). Figures A-1–A-3 of the Appendix display the individual marginal effect estimates of the interaction terms for each observation of our sample along with the corresponding z-statistics. The findings in the table and the corresponding INTEFF graphs are qualitatively identical to those in Table 3. 16

two-segment specification summarized in the second column of Table 3 indicates that most of the performance sensitivity is driven by below-median funds.22 The additional control variables indicate that funds that are more correlated with other options on the menu are more likely to be deleted. Thus, the incumbent ensemble of the funds on the menu matters in deletion decisions. Additionally, funds with large plan investments are less likely to be deleted and plans with more investment options are less likely to delete a specific fund. Plan providers are also more likely to delete funds with high expense ratios, funds with high turnover, and smaller funds. Overall, our baseline results indicate that affiliated funds are significantly less likely to be deleted from 401(k) plans than unaffiliated funds and that this bias is particularly pronounced for poorly performing funds. As we discuss in Section 4.1 above, protecting poorly performing affiliated funds may be especially important as keeping these funds on the menu dampens the outflow of capital triggered by bad performance and, as a result, mitigates distress. To provide further evidence on the trustee’s incentive to support distressed funds, we now examine the role of non-DC money flows in more detail. In particular, we create an indicator variable that equals one if the fund experiences an outflow from its non-DC clients in the past year and zero otherwise, and an interaction term of this indicator variable with our ‘Affiliated Fund’ dummy. Non-DC flows are calculated as NonDCAssets −NonDCAssets ×(1+ f,t f,t−1 R ) based on footnote 20 above. f,t The last two columns of Table 3 report the results of adding these two additional explanatoryvariablestothemodelsincolumns1and2. Thecoefficientestimateson‘NegNonDCFlow’ are positive and highly significant suggesting that plans are more likely to delete those unaffil- 22InarobustnesstestreportedinTableA-3intheAppendix,wecomputetwoalternativerankingmethods, where the percentile performance of a fund is either measured relative to the other investment options in a specific 401(k) plan or relative to the other funds offered by the fund’s family. The results are consistent with thoseusingperformancerankingsbasedontheCRSPfunduniverse. Additionally,TableA-4intheAppendix estimates equation (1) for all three ranking methods using prior performance horizons of one and five years. 17

iated funds that are also shunned by outside investors. This implies that ‘NegNonDCFlow’ captures some aspects of the fund’s popularity among investors that are not captured by past performance or other fund characteristics. Interestingly, the interaction term is significantly negative indicating that affiliated funds receive support when they experience money outflows from their non-DC clients. Trustee support alleviates fund distress, which is costly for mutual funds.23 4.3 Subsample Analysis of Fund Deletions To analyze whether the incentives for fund deletions differ across different types of plans and across time, Table 4 shows the results of our linear probability model specified in equation (1) for various subsamples. In the first two columns, we compare the results for the three largest trustees and for all other trustees. The three largest trustees in our sample each manage over 10% of all 401(k) mutual fund assets.24 Large service providers have more in-house investment options and may have more bargaining power relative to small fund families. Our favoritism results hold for both subsamples albeit a little weaker for smaller trustees. To test whether our results are affected by economies of scale in plan management, we reestimate our model in columns 3 and 4 for below- and above-median sized plans, respectively. Sponsors with large 401(k) plans may have more negotiating power with service providers and may also monitor service providers more effectively. Our results are remarkably consistent across the two subsamples. The Pension Protection Act of 2006 (PPA) introduced comprehensive new legislation to 23Unfortunately, the Pensions & Investments survey is only available for a subsample of the mutual fund universe. Therefore, we do not include these variables in our baseline specification. 24The three largest trustees in terms of the dollar value of total plan assets in our sample are Vanguard, Fidelity, and State Street. 18

protect U.S. retirement plan participants. Although the reforms mainly concerned defined benefit plans, the PPA also affected DC plans by allowing companies to offer objective investment advice to participants and by requiring plans to provide specific benefit statements to participants.25 Furthermore, several class action lawsuits were filed in the mid 2000s against large employers for breaches of fiduciary obligations with respect to their 401(k) accounts.26 To investigate whether these lawsuits and regulatory reforms affect our results, we divide our sample into two subperiods (1998-2006 and 2007-2009). Columns 5 and 6 of Table 4 indicate that our key results do not differ between the two subperiods. We find that affiliated funds exhibit a lower propensity to be deleted from 401(k) menus and that deletions for affiliated funds are less sensitive to prior fund performance for both subperiods. Table A-6 in the Internet Appendix reports additional robustness analyses on fund deletions. For example, we show that the results are qualitatively unaffected if we include trustee fixed effects or fund fixed effects. Furthermore, the results are also robust if we focus only on plans with mutual fund trustees or if we delete target-date funds, index funds, or non-equity funds. 4.4 Fund Additions The previous sections provide evidence that trustees are substantially less likely to delete their own funds from the menus, and even more so when these funds are poorly performing. In this section we examine whether similar biases exist for fund additions as well. Toinvestigatehowafund’spropensitytobeaddedtoamenudependsonitsaffiliationwith the trustee, we determine the addition frequency of each fund in the CRSP fund universe as an affiliated and unaffiliated menu choice, respectively. Consistent with our deletion frequency 25The detailed regulations from the 2006 Pension Protection Act can be obtained from http://www.dol.gov/ebsa/pensionreform.html. 26See Ruiz-Zaiko and Williams (2007) for additional information on the lawsuits. 19

measures in Section 4.1, we define the affiliated addition frequency of a fund as the number of affiliated plans to which the fund is added as a new investment option during the year divided by the total number of affiliated menus to which it could be added (i.e., the number of affiliated plans in which the fund is not already offered as an option at the end of the previous year). Unaffiliated addition frequencies are defined analogously. The difference between the average addition frequencies of affiliated and unaffiliated funds is large. For example, in the overall sample, the average addition frequency is 1.33% for affiliated funds and just 0.02% for unaffiliated funds. We report the average addition frequencies by affiliation and performance in Figure A-4 and Table A-7 in the Appendix. While the difference between the groups is stark, it is difficult to assess the magnitude of favoritism for additions from these statistics alone. This is because addition frequencies implicitly assume that plan sponsors and trustees consider every fund in the CRSP universe when selecting new choices for their menus.27 Therefore, instead of emphasizing the level effect, in what follows, we rescale affiliated and unaffiliated addition probabilities by dividing each series by its own mean. Rescaling allows us to highlight the conditional effects instead. The scaled addition frequencies are depicted in Figure 2. As above, each year we sort funds into deciles according to their percentile performance among funds of the same style in the CRSP fund universe over the prior three years. Panel A summarizes the results using all existing mutual funds, whereas the average frequencies in Panel B are based on funds from only those families that act as trustees for at least one of our 401(k) plans during the year. Thus, Panel B excludes funds that could not be added as trustee funds during the year. This restriction allows us to examine the rescaled addition frequency of the same fund to an affiliated or unaffiliated menu, respectively. 27Thedifferenceinadditionfrequenciesissimilarlystarkwhenwelimitouranalysistoonlythoseinvestment styles in the CRSP universe that appear on 401(k) menus in our sample. 20

Figure 2 highlights several interesting results. First, for poorly performing funds, relative affiliated addition probabilities are higher than the corresponding unaffiliated probabilities. The difference is statistically significant in both panels. In the upper tail, the opposite is true: unaffiliated addition probabilities are significantly higher than affiliated probabilities. These results illustrate that the performance threshold unaffiliated funds have to meet to be included in the plan is significantly higher than that for affiliated funds. Moreover, while addition probabilitiesincreasewithperformanceforbothgroups, theyincreasedisproportionatelymore for unaffiliated funds than for affiliated funds, indicating that unaffiliated additions are more sensitive to performance. An improvement in performance from the lowest to the highest decile increases the rescaled addition probability for unaffiliated funds approximately eightfold from 0.29% to 2.31%. At the same time, an equivalent improvement in performance for affiliated funds results in only a three times larger value (from 0.55% to 1.67%). To provide additional evidence on favoritism in fund addition decisions, we also perform regression analyses using these rescaled variables. These results are reported in Table 5. The findings indicate that unaffiliated addition probabilities are more sensitive to performance than affiliated addition probabilities, even after controlling for various fund characteristics. Affiliated and unaffiliated addition probabilities also show very different sensitivities to fund distress and expense ratios. Distressed funds have a lower probability of being added to a menu as both affiliated and unaffiliated funds. More interestingly, the table shows that, consistent with our favoritism hypothesis, affiliated addition frequencies are significantly less sensitive to distress than are unaffiliated addition probabilities. Furthermore, addition rates are positively related to expense ratios for affiliated funds, whereas they are not significantly related to expense ratios for unaffiliated funds. Finally, we complement these findings with two additional results tabulated in the Internet Appendix. In Table A-8 we estimate a linear probability model for which the dependent 21

variable takes the value of one if the added fund is an affiliated fund, and zero otherwise. Since the sample used in this analysis includes only fund additions, it reflects the choice between selecting an affiliated fund over an unaffiliated fund. Consistent with menu favoritism, we find thataffiliatedfundadditionsareassociatedwithworsepastperformanceevenaftercontrolling for other fund characteristics. A similar result is conveyed by Figure A-5 which reports the distribution of fund additions by performance decile and fund affiliation. The figure reveals that the proportion of unaffiliated funds with strong past performance is larger compared to that of affiliated funds, while affiliated funds are more likely to come to the menu with a mediocre performance record. Overall, our results for both deletion and addition decisions provide evidence that trustees treat their own affiliated funds differently than unaffiliate funds. Affiliated funds are more likely to be added and are less likely to be deleted from a plan. Furthermore, fund additions and deletions are less sensitive to prior performance for affiliated than for unaffiliated funds. 5 Participant Flows While the investment opportunity set of the plan is determined by the menu choices selected by the employer and the service providers, participants can freely allocate their contributions within the opportunity set. If participants anticipate biases, they can offset favoritism in their own portfolios by, for instance, not allocating capital to poorly performing affiliated funds. In this section, we investigate whether menu favoritism has an impact on the overall allocation of plan assets by examining the sensitivity of participant flows to the performance of affiliated and unaffiliated funds. Our primary definition of the growth rate of new money of fund f held in 401(k) plan p 22

at time t is based on the following measure of fund flows: V −V (1+R ) p,f,t p,f,t−1 f,t NMG1 = . (2) p,f,t V (1+R ) p,f,t−1 f,t The numerator captures the dollar change in the value of participants’ investments (V ) p,f,t in fund f in plan p in year t after adjusting for the price appreciation of plan assets R f,t (i.e., fund return) during the year. The denominator is defined as the projected value of the lagged plan position in the fund without any new flow of money. If an investment option is deleted from a plan menu, then NMG1 equals exactly -100%. To remove outliers, we winsorize NMG1 at the 95% level.28 Since equation (2) is not defined for fund additions, we adopt two alternative measures for the growth rate of new money. Our second measure (NMG2) normalizes fund flows by the sum between beginning- and end-of-period assets: V −V (1+R ) p,f,t p,f,t−1 f,t NMG2 = . (3) p,f,t V +V (1+R ) p,f,t p,f,t−1 f,t Under this definition, new money growth takes a value in the interval [-1,1]. In particular, it equals -100% for deletions, as before, and +100% for a fund that is newly added to the employee benefit plan. More gradual inflows and outflows (i.e., participant flows) into the fund are represented by intermediate values. Finally, the denominator of our third measure (NMG3) is based on overall plan value at t−1 adjusted for fund returns. To remove outliers, we winsorize NMG3 at the 95% level: V −V (1+R ) p,f,t p,f,t−1 f,t NMG3 = . (4) p,f,t (cid:80) V (1+R ) f p,f,t−1 f,t These three definitions of new money growth allow us to decompose fund flows to menu options into components that are primarily driven by plan providers (i.e., flows due to fund 28Figure A-6 in the Appendix depicts histograms of the percentage flows. 23

additions and deletions) and components that are primarily driven by plan participants (i.e., all changes which are not driven by fund additions and deletions).29 To investigate the sensitivity of fund flows to prior performance, we estimate the following regression using the three alternative definitions of NMG: NMG = β +β ×AF +β ×LowPerf +β ×HighPerf p,f,t 0 1 p,f,t−1 2 p,f,t−1 3 p,f,t−1 + β ×AF ×LowPerf +β ×AF ×HighPerf 4 p,f,t−1 p,f,t−1 5 p,f,t−1 p,f,t−1 + Z(cid:48) γ +(cid:15) . (5) p,f,t−1 p,f,t Equation (5) is analogous to our two-segment baseline equation with two exceptions. First, ournewdependentvariableisNMG, acontinuousvariableunderallthreedefinitions. Second, if participants use the same allocation rule each year, growth occurs mechanically due to the additional money contributed to the accounts over time. To capture this mechanical feature of flows, we add contemporaneous plan growth as an additional control.30 The results are summarized in Table 6. The first three columns show the coefficient estimates for our full sample of NMG values. The full sample includes observations that capture fund deletion/addition as well as observations that reflect more gradual inflows and outflows by plan participants. The last three columns report coefficient estimates for participant flows only based on a subsample that excludes NMG observations that reflect fund additions and deletions. The main results in columns 1–3 using the full sample are consistent with the deletion 29Plan sponsors and service providers may not only affect flows through addition and deletion decisions. For example, the selection of default investment options, the freezing of existing investment options to new money, and the promotion of specific investment options during on-line or face-to-face educational activities are additional actions by plan sponsors or service providers that affect money flows. Unfortunately, we do not observethesedecisions. However,despiteournarrowdefinitionofmenuchangesinitiatedbyplansponsorsand service providers (which are only based on flows due to additions and deletions), we find that plan sponsors and service providers account for most of the variability of fund flows, as documented in Table 6. 30We calculate plan growth using information from Form 5500 on total contributions, total expenses, and total assets. 24

results from Table 3. Affiliated funds attract more new money than unaffiliated funds. We find that flows into various plan options increase with prior fund performance, consistent with Chevalier and Ellison (1997), Sirri and Tufano (1998), and Huang, Wei, and Yan (2007). The interaction effects indicate that overall flows are significantly less sensitive to poor performance for affiliated funds. For example, a ten percentage point increase in the past percentile performance of below-median funds increases flows over the next year by 5.54% for unaffiliated funds and by only 0.80% for affiliated funds. The additional control variables indicate that the growth rates are larger for funds that exhibit low correlations with existing menu options, for smaller investment options, for plans with higher growth rates, for funds with lower expense ratios and lower turnovers, and for larger funds. To investigate the importance of participant flows, we restrict our attention to the money flows of options that are not driven by deletions or additions in the last three columns of Table 6. We find that participant flows are generally higher for affiliated funds, although the coefficient estimates are smaller than the corresponding estimates in the first three columns of the table. Thus, the higher overall flows to affiliated funds in columns 1–3 are primarily driven by the decisions of plan sponsors and service providers. The coefficients on the two performance ranking segments indicate that participants chase prior fund performance. Comparing the coefficients in columns 4–6 to those in columns 1–3 reveals that most of the inflows into above-median performers are due to plan participants, whereas most of the outflows out of below-median performers are due to decisions by sponsors and service providers. The interaction effects between the affiliation dummy and the two performance segments indicate that plan participants do not offset the biased decisions of plan sponsors and trustees: if anything, they are also somewhat less sensitive to the performance of poorly performing affiliated funds. These results are consistent with previous studies that have documented that DC pension participants are naive and inactive (Benartzi and Thaler, 25

2001; Madrian and Shea, 2001; Agnew, Balduzzi, and Sunden, 2003). Ourresultsindicatethatdecisionsofplansponsorsandserviceprovidershaveasubstantial impact on flows to mutual funds. Affiliated mutual funds can benefit by obtaining higher money flows and by avoiding large outflows from their poorly performing funds. 6 Future Performance Our previous results indicate that 401(k) plans are less likely to delete affiliated funds from their menus and that deletions of affiliated funds are less sensitive to prior fund performance. We also document a similar behavior for fund additions. Finally, we show that participants do not direct flows away from the biased options offered by the trustee. Still, favoritismtowardaffiliatedfundsmaynothurtplanparticipantsiftheunderperforming affiliated funds exhibit superior subsequent performance. Indeed service providers may keep poor performers not because they are biased toward them, but rather, due to positive information they possess about the future returns of these funds. To investigate this hypothesis, we now examine the performance of affiliated and unaffiliated funds that are kept in, deleted from, or added to the plans using monthly fund returns. We restrict our sample to domestic equity funds in these analyses, since it is difficult to compare performance across different asset classes. At the end of each calendar year, we form equal-weighted portfolios of affiliated and unaffiliated funds separately based on whether the funds were kept, deleted, or added (“No Changes,” “Deletions,”and “Additions”) during the calendar year.31 This creates six portfolios. We then further subdivide these six groups based on past performance. In particular, “All Funds,” refers to all funds in the original six portfolios, while “Lowest Decile” and “Lowest Quintile” refer to subportfolios in each group that 31To avoid any lookahead biases, we do not include those plans in these analyses that have fiscal years ending before July of the calendar year. 26

contain only those funds that also rank in the lowest performance deciles or quintiles. We use percentile performance rankings during the prior three years as in our baseline specification. For example, “Affiliated Funds/Deletions/Lowest Decile” represents the portfolio of afffiliated funds in the worst performance decile that are deleted from a menu. We rebalance our portfolios at the end of each calendar year and calculate the portfolios’ return for each of the next 12 months keeping the portfolio composition fixed. Table 7 reports the abnormal returns of the various portfolios. Panels A, B, and C report the Carhart (1997) alphas, the Fama and French (1993) alphas, and the CAPM alphas, respectively. The future Carhart alpha for affiliated funds kept for at least two consecutive periods in the 401(k) plan is essentially 0 bps per month. Similarly, the corresponding alpha for unaffiliated funds is insignificantly different from zero at -6 bps per month. Consistent with the evidence on defined benefit plans provided by Goyal and Wahal (2008), we do not find that added funds on average perform significantly better than deleted funds. However, we find that affiliated funds that are kept in the 401(k) plans by their sponsors despite their poor performance exhibit significantly negative Carhart and Fama-French alphas. For example, affiliated funds ranked in the lowest performance decile over the prior three years exhibit a Carhart alpha of -0.33% per month. This represents a risk- and style-adjusted underperformance of 3.96% per year. The performance difference between affiliated and unaffiliated funds ranked in the lowest performance decile of 0.25% per month is also statistically significant at a 5% level. On the other hand, the results are less pronounced using CAPM alphas, which do not adjust for style effects, but the difference in performance between poorly performing affiliated and unaffiliated funds that are retained on the plans is similarly large. Our results in Table 7 confirm that the decision to retain poorly performing affiliated funds is not driven by private information about the future performance of these funds. Instead, consistent with Carhart (1997), poor performance persists, even after adjusting for 27

momentum factors. Overall, those plan participants who invest in these affiliated funds may cross-subsidizeotheremployeesbyshoulderingadisproportionateshareoftheplan’scostsand would have obtained a higher risk-adjusted performance had they switched their retirement savings from underperforming affiliated funds to other funds. 7 Conclusion While mutual fund families serving as service providers of 401(k) plans are expected to act in the best interest of participants, they also have a competing incentive to attract and retain retirement contributions in their own proprietary funds. Despite the increasing role of 401(k) plans as a retirement vehicle, little is known about how provider incentives influence the set of investment choices offered in the plans. This is surprising as small inefficiencies in the selection of investments options, especially early in the participants’ career, can have a significant impact on retirement savings outcomes. Ourpapertakesafirststeptoinvestigatethisquestion. Wedocumentsignificantfavoritism in 401(k) menu decisions. We show that affiliated funds are less likely to be removed from the menu relative to unaffiliated funds, independent of their performance record. Moreover, the difference in deletion propensities between affiliated and unaffiliated funds is largest among the worst performing funds. We find similar results for mutual fund additions. Interestingly, mutual fund affiliation does not affect how participants allocate their contributions, suggesting that participants do not offset these biases. We also show that the reluctance to remove poorly performing affiliated funds from the menu generates a significant subsequent negative abnormal return for participants investing in those funds. Since favoritism represents a form of implicit compensation for the plan’s service providers, these performance results imply that favoritism causes an unequal distribution of plan costs across the plan’s participants. 28

In sum, our paper provides a first look at the service providers in the 401(k) industry and their effect on plan design. Future research should explore and contrast additional costs and benefits of the various administrative arrangements of 401(k) plans. 29

References Agnew, J., P. Balduzzi, and A. Sunden (2003). Portfolio choice and trading in a large 401(k) plan. American Economic Review 93, 193–215. Ai, C. and E. C. Norton (2003). Interaction terms in logit and probit models. Economics Letters 80, 123–129. Balduzzi, P. and J. Reuter (2012). Heterogeneity in target-date funds and the pension protection act of 2006. Working paper. Benartzi, S. and R. H. Thaler (2001). Naive diversification strategies in defined contribution saving plans. American Economic Review 91(1), 79–98. Bhattacharya, U., J. H. Lee, and V. K. Pool (2013). Conflicting family values in mutual fund families. Journal of Finance 68, 173–200. Brown, J. R., N. Liang, and S. Weisbenner (2007). Individual account investment options and portfolio choice: Behavioral lessons from 401(k) plans. Journal of Public Economics 91, 1992–2013. Brown, K. C. and W. V. Harlow (2012). How good are the investment options provided by defined contribution plan sponsors? International Journal of Portfolio Analysis and Management 1, 3–31. Carhart, M. M. (1997). On the persistence of mutual fund performance. Journal of Finance 52(1), 57–82. Carroll, G.D., J.J.Choi, D.Laibson, B.C.Madrian, andA.Metrick(2009). Optimaldefaults and active decisions. Quarterly Journal of Economics 124, 1639–1674. Chevalier, J. A. and G. D. Ellison (1997). Risk taking in mutual funds as a response to incentives. Journal of Political Economy 105(6), 1167–1200. Choi, J. J., D. Laibson, B. C. Madrian, and A. Metrick (2002). Defined contribution pensions: Plan rules, participant decisions, and the path of least resistance. In J. M. Poterba (Ed.), Tax Policy and the Economy, pp. 67–113. Cambridge, MA: MIT Press. Choi, J. J., D. Laibson, B. C. Madrian, and A. Metrick (2004). For better or for worse. Default effects and 401(k) savings behavior. In D. A. Wise (Ed.), Perspectives on the Economics of Aging, pp. 81–121. Chicago, IL: University of Chicago Press. Christoffersen, S. K. and M. Simutin (2014). Plan sponsor oversight and benchmarking: Effects on fund risk-taking and activeness. Working Paper. Cohen, L. and B. Schmidt (2009). Attracting flows by attracting big clients. Journal of Finance 64(5), 2125–2151. Coval, J. and E. Stafford (2007). Asset fire sales (and purchases) in equity markets. Journal of Financial Economics 86, 479–512. Davis, G. F. and E. H. Kim (2007). Business ties and proxy voting by mutual funds. Journal of Financial Economics 85(2), 552–570. 30

Del Guercio, D. and P. Tkac (2002). The determinants of the flow of funds of managed portfolios: Mutual funds versus pension funds. Journal of Financial and Quantitative Analysis 37, 523–557. Duflo, E. and E. Saez (2002). Participation and investment decisions in a retirement plan: The influence of colleagues’ choices. Journal of Public Economics 85(1), 121–148. Elton, E. J., M. J. Gruber, and C. R. Blake (2006). The adequacy of investment choices offered by 401(k) plans. Journal of Public Economics 90, 1299–1314. Elton, E. J., M. J. Gruber, and C. R. Blake (2007). Participant reaction and the performance of funds offered by 401(k) plans. Journal of Financial Intermediation 16, 249–271. Fama, E. F. and K. R. French (1993). Common risk factors in the returns on stocks and bonds. Journal of Financial Economics 33(1), 3–56. Gaspar, J. M., M. Massa, and P. Matos (2006). Favoritism in mutual fund families? Evidence on strategic cross-fund subsidization. Journal of Finance 61, 73–104. Goldreich, D. and H. Halaburda (2013). When smaller plans are better: Variability in menusetting ability. Forthcoming: Management Science. Goyal, A. and S. Wahal (2008). The selection and termination of investment management firms by plan sponsors. Journal of Finance 63, 1805–1847. Huang, J., K. D. Wei, and H. Yan (2007). Participation costs and the sensitivity of fund flows to past performance. Journal of Finance 62, 1273–1311. Huberman, G. and W. Jiang (2006). Offering versus choice in 401(k) plans: Equity exposure and number of funds. Journal of Finance 61, 763–801. Kuhnen, C. M. (2009). Business networks, corporate governance, and contracting in the mutual fund industry. Journal of Finance 64, 2185–2220. Madrian, B.C.andD.F.Shea(2001). Thepowerofsuggestion: Inertiain401(k)participation and savings behavior. Quarterly Journal of Economics 116, 1149–1187. Mitchell, O. S. and S. Utkus (2012). Target-date funds in 401(k) retirement plans. Working Paper. Muir, D. M. (2012). Choice architecture and the locus of fiduciary obligation in defined contribution plans. Working Paper. Reuter, J. (2006). Are IPO allocations for sale? Evidence from mutual funds. Journal of Finance 61, 2289–2324. Ruiz-Zaiko,L.andB.Williams(2007). Plansponsorsbesiegedby401(k)feelawsuits. Pensions & Benefits Management Bridgebay Financial. Sialm, C., L. T. Starks, and H. Zhang (2014). Defined contribution pension plans: Sticky or discerning money? forthcoming Journal of Finance. Sirri, E. R. and P. Tufano (1998). Costly search and mutual fund flows. Journal of Finance 53(5), 1598–1622. Tang, N., O. S. Mitchell, G. R. Mottola, and S. P. Utkus (2010). The efficiency of sponsor and participant portfolio choices in 401(k) plans. Journal of Public Economics 94, 1073–1085. 31

Panel A: Overall Sample Panel B: Subsample of Funds on Both Affiliated and Unaffiliated Menus Figure 1: Fund Deletions by Affiliation. Thefiguredepictsmeanannualfunddeletionfrequencies by trustee affiliation and performance deciles. Panel A includes the full sample. Panel B includes the subsample of funds that appear contemporaneously on multiple 401(k) menus, at least once as an affiliated fund and at least once as an unffiliated fund. Every year, we calculate the ratio of the number of affiliated (unaffiliated) menus from which the fund is delisted during the year to the total number of affiliated (unaffiliated) menus associated with the fund. Performance deciles are created by grouping funds based on their percentile performance among funds of the same style in the CRSP fund universe over the prior three years. We then average across the funds’ deletion frequencies by performance and affiliation. The numbers above the bars are differences in the mean deletion rates between affiliated and unaffiliated funds. The corresponding standard errors are clustered at the fund level with significance levels denoted by *, **, ***, which correspond to 10%, 5%, and 1% levels, respectively. 32

Panel A: Overall Sample Panel B: Subsample of Funds on Both Affiliated and Unaffiliated Menus Figure 2: Rescaled Fund Additions by Affiliation. The figure depicts mean rescaled fund addition frequencies by trustee affiliation and performance deciles. Panel A includes the full sample. Panel B includes only those funds that are offered by fund families that serve as trustees for at least one plan in our sample. Every year, we calculate the ratio of the number of affiliated (unaffiliated) menus to which the fund is added during the year to the total number of affiliated (unaffiliated) menus that do not yet include the fund as an option. These affiliated and unaffiliated addition rates are then rescaled by dividing the raw addition rates by the corresponding mean raw addition rates. We then average across the funds’ rescaled addition frequencies by performance and affiliation. Performance deciles are created based on the fund’s percentile performance among funds with the same style in the CRSP fund universe over the prior three years. The numbers above the bars are differences in the mean rescaled addition rates between affiliated and unaffiliated funds. The corresponding standard errors are clustered at the fund level with significance levels denoted by *, **, ***, which correspond to 10%, 5%, and 1% levels, respectively. 33

.raeY yb scitsitatS evitpircseD elpmaS :1 elbaT derutpac srosnops nalp dna snalp fo rebmun eht troper 2 dna 1 snmuloC .raey yb scitsitats evitpircsed sedivorp elbat ehT elpmas ruo ni snalp fo egatnecrep eht troper ew 4 nmuloc nI .ezis nalp egareva swohs 3 nmuloC .ylevitcepser ,elpmas ruo ni sedulcni sihT .nalp eht fo erutcetihcra eht tuoba noitamrofni edivorp ew ,9-5 snmuloc nI .seetsurt dnuf lautum evah taht eht sa detaluclac erahs eetsurt eht ,snoitpo detailffia fo rebmun eht dna latot ni dereffo snoitpo dnuf lautum fo rebmun eht taht seinapmoc tnemeganam fo rebmun egareva eht ,sdnuf detailffia htiw detsevni stessa tnemeriter fo noitroporp llarevo erahs rallod eht no desab detaluclac unem eht fo xedni lhadnfireH eht dna ,unem eht no noitpo tnemtsevni eno tsael ta reffo .seinapmoc tnemeganam eseht fo hcae fo lhadnfireH rebmuN eetsurT rebmuN rebmuN htiw snalP egarevA rebmuN rebmuN raeY xednI .tmgM fo erahS detailffiA fo fo seetsurT FM eziS nalP fo fo seinapmoC )% ni( snoitpO snoitpO )% ni( )M$ ni( snalP srosnopS 76.0 69.2 10.43 83.2 10.7 13.06 62.682 317 816 8991 46.0 84.3 11.43 58.2 58.7 49.86 84.142 598 067 9991 95.0 00.4 86.53 35.3 92.9 12.37 34.592 400,1 928 0002 75.0 65.4 19.63 01.4 34.01 63.47 24.872 001,1 029 1002 45.0 10.5 62.73 06.4 05.11 95.67 72.052 032,1 210,1 2002 15.0 84.5 00.63 37.4 00.21 90.38 45.692 523,1 201,1 3002 84.0 98.5 58.33 81.5 91.31 33.38 83.723 413,1 601,1 4002 54.0 81.6 05.23 04.5 97.31 35.38 20.053 182,1 390,1 5002 44.0 92.6 65.13 18.5 75.41 21.87 35.104 522,1 430,1 6002 24.0 56.6 73.82 19.5 39.51 60.57 40.634 571,1 200,1 7002 24.0 80.7 99.82 94.6 02.71 04.57 74.223 621,1 079 8002 04.0 63.7 31.72 04.6 28.71 80.57 33.704 979 948 9002 15.0 14.5 30.33 87.4 55.21 95.57 34.423 411,1 149 egarevA 34

.scitsitatS yrammuS dnuF lautuM :2 elbaT ,elpmas ruo ni unem )k(104 a ot dedda dna ,morf deteled ,ni tpek era taht sdnuf eht ebircsed elbat eht fo C dna ,B ,A slenaP ni ,stessa nalp ot noitpo dnuf lautum egareva eht ni detsevni stessa latot fo oitar eht si eziS noitpO evitaleR .ylevitcepser ,deteled ,tpek era taht stessa )detailffianu( detailffia fo oitar eht si eziS noitpO latoT .)dedda ,deteled ,tpek( yrogetac hcae gniniamer ehT .unem eht no noitpo hcae ni )snoillim ni( stessa fo eulav rallod eht si eziS noitpO .raey hcae dedda ro eht ,tnemeganam rednu stessa latot yb derusaem sa )snoillib ni( ezis dnuf ,ega dnuf :selbairav level dnuf lautum era selbairav ecnamrofreP .selitnecrep ecnamrofrep naem ’sdnuf eht dna ,oitar esnepxe eht ,revonrut ,snruter dnuf ylhtnom fo ytilitalov .esrevinu dnuf PSRC eht ni elyts emas eht fo sdnuf no desab sraey eerht suoiverp eht revo detaluclac era )freP( selitnecrep detailffia rof detroper era segareva ehT .segatnecrep sa desserpxe era seulav lla ,ezis dnuf dna ega dnuf fo noitpecxe eht htiW rof slevel ecnacfiingiS .slevel dnuf dna nalp eht ta deretsulc yaw-owt era srorre dradnatS .yletarapes sdnuf detailffianu dna .ylevitcepser ,slevel %1 dna ,%5 ,%01 ot dnopserroc hcihw ,*** ,** ,* yb detoned era snaem ni ecnereffid eht fo stset segnahC oN :A lenaP .rY-3 roirP esnepxE revonruT nruteR dnuF dnuF noitpO latoT evitaleR rebmuN detailffiA ecnamrofreP oitaR )% ni( .veD .dtS eziS egA eziS eziS noitpO eziS noitpO .sbO fo dnuF )% ni( )% ni( )% ni( )B$ ni( )sraeY ni( )M$ ni( )% ni( )% ni( 42.06 49.0 45.67 89.3 05.51 16.91 29.8 24.58 65.8 055,28 0 91.85 75.0 21.25 83.3 30.21 92.71 74.31 46.88 06.7 932,25 1 ∗50.2− ∗∗∗73.0− ∗∗∗24.42− ∗∗∗06.0− 74.3− ∗23.2− ∗∗45.4 ∗∗∗12.3 ∗69.0− 987,431 ffiD snoiteleD :B lenaP .rY-3 roirP esnepxE revonruT nruteR dnuF dnuF noitpO latoT evitaleR rebmuN detailffiA ecnamrofreP oitaR )% ni( .veD .dtS eziS egA eziS eziS noitpO eziS noitpO .sbO fo dnuF )% ni( )% ni( )% ni( )B$ ni( )sraeY ni( )M$ ni( )% ni( )% ni( 92.15 60.1 43.39 80.4 03.8 91.81 66.6 75.41 06.7 981,41 0 73.15 08.0 86.08 84.3 10.7 45.71 95.9 53.11 91.7 582,4 1 80.0 ∗∗∗62.0− ∗∗66.21− ∗∗∗06.0− 92.1− 56.0− ∗∗29.2 ∗∗∗12.3− 14.0− 474,81 ffiD snoitiddA :C lenaP .rY-3 roirP esnepxE revonruT nruteR dnuF dnuF noitpO latoT evitaleR rebmuN detailffiA ecnamrofreP oitaR )% ni( .veD .dtS eziS egA eziS eziS noitpO eziS noitpO .sbO fo dnuF )% ni( )% ni( )% ni( )B$ ni( )sraeY ni( )M$ ni( )% ni( )% ni( 94.76 59.0 56.08 89.3 60.01 41.51 39.4 47.02 62.6 278,12 0 19.36 06.0 32.35 32.3 24.5 53.01 31.5 53.41 75.4 618,7 1 ∗∗∗85.3− ∗∗∗53.0− ∗∗∗24.72− ∗∗∗57.0− ∗46.4− ∗∗∗97.4− 02.0 ∗∗∗83.6− ∗∗∗96.1− 886,92 ffiD 35

Table 3: Linear Probability Model of Fund Deletions. The table reports the coefficient estimates for the following linear probability model: DEL = p,f,t β + β AF + β Perf + β AF × Perf + β NegNonDCFlow + β AF × 0 1 p,f,t−1 2 p,f,t−1 3 p,f,t−1 p,f,t−1 4 f,t−1 5 p,f,t−1 NegNonDCFlow + Z(cid:48) γ + (cid:15) and for the following and 2-segment piecewise linear model f,t−1 p,f,t−1 p,f,t DEL = β +β AF +β LowPerf +β HighPerf +β AF ×LowPerf + p,f,t 0 1 p,f,t−1 2 p,f,t−1 3 p,f,t−1 4 p,f,t−1 p,f,t−1 β AF ×HighPerf +β NegNonDCFlow +β AF ×NegNonDCFlow +Z(cid:48) γ+ 5 p,f,t−1 p,f,t−1 6 f,t−1 7 p,f,t−1 f,t−1 p,f,t−1 (cid:15) , where DEL is an indicator variable that takes the value of one if mutual fund f is deleted from p,f,t p,f,t plan p during year t and zero otherwise and AF is an indicator for whether the trustee of pension p,f,t−1 plan p is affiliated with the management company of fund f at the end of year t−1. Perf is the p,f,t−1 percentile performance rank of fund f over the previous three years based on funds in the same style in the CRSPfunduniverseandLowPerf andHighPerf aredefinedasLowPerf =min(Perf ,0.5)and p,f,t−1 p,f,t−1 HighPerf =max(Perf −0.5,0). ThevariableNegNonDCFlow isanindicatorvariablefor p,f,t−1 p,f,t−1 f,t−1 whether the non-DC flows of the fund in year t−1 are negative. The other lagged control variables Z include the maximum return correlation of the fund with existing menu options, the natural logarithm of option size, the number of options, the expense ratio, fund turnover, the natural logarithm of the fund’s size, fund age, thestandarddeviationofthefund’sreturn,andfundstyleandyearfixedeffects. Standarderrorsaretwo-way clustered at the plan and fund levels and are reported in parentheses. Significance levels are denoted by *, **, ***, which correspond to 10%, 5%, and 1% levels, respectively. Linear 2-Segment Linear 2-Segment Affiliated Fund −0.105∗∗∗ −0.150∗∗∗ −0.088∗∗∗ −0.118∗∗∗ (0.014) (0.019) (0.019) (0.027) Perf −0.167∗∗∗ −0.165∗∗∗ (0.014) (0.019) Perf*Affiliated Fund 0.103∗∗∗ 0.092∗∗∗ (0.018) (0.025) LowPerf −0.319∗∗∗ −0.343∗∗∗ (0.034) (0.044) HighPerf −0.051∗∗ −0.021 (0.023) (0.031) LowPerf*Affiliated Fund 0.249∗∗∗ 0.190∗∗∗ (0.045) (0.059) HighPerf*Affiliated Fund −0.006 0.013 (0.030) (0.040) Neg NonDC Flow 0.041∗∗∗ 0.041∗∗∗ (0.008) (0.008) Neg NonDC Flow*Affiliated Fund −0.028∗∗∗ −0.028∗∗∗ (0.010) (0.010) Maximum Corr 0.010∗∗∗ 0.010∗∗∗ 0.011∗∗∗ 0.011∗∗∗ (0.001) (0.001) (0.001) (0.001) Log(Option Size) −0.009∗∗∗ −0.009∗∗∗ −0.007∗∗∗ −0.007∗∗∗ (0.002) (0.002) (0.002) (0.002) No. of Options −0.002∗∗∗ −0.002∗∗∗ −0.002∗∗∗ −0.002∗∗∗ (0.000) (0.000) (0.000) (0.000) Expense Ratio 7.965∗∗∗ 7.626∗∗∗ 8.423∗∗∗ 7.600∗∗∗ (1.109) (1.106) (1.593) (1.595) Turnover 0.017∗∗∗ 0.017∗∗∗ 0.016∗∗∗ 0.016∗∗∗ (0.004) (0.004) (0.005) (0.005) Log(Fund Size) −0.019∗∗∗ −0.019∗∗∗ −0.017∗∗∗ −0.017∗∗∗ (0.002) (0.002) (0.003) (0.003) Fund Age 0.000 0.000 −0.000 0.000 (0.000) (0.000) (0.000) (0.000) Std. Dev. −0.088 36 −0.179 −0.274 −0.373∗ (0.200) (0.195) (0.207) (0.201) Observations 106,848 106,848 65,855 65,855 R-squared 0.077 0.079 0.080 0.082

.sisylanA elpmasbuS :snoiteleD dnuF rof ledoM ytilibaborP raeniL :4 elbaT snoiteled dnuf rof ledom ytilibaborp raenil esiweceip tnemges-2 enilesab ruo rof setamitse tneicffieoc eht stroper elbat ehT tsrfi eht ni seetsurt tsegral eerht eht rof ledom eht etamitse eW .atad ruo fo selpmasbus suoirav rof 3 elbaT ni debircsed snalprofstluserruoetamitse4dna3snmuloC .nmulocdnocesehtniraeyhcaeseetsurttsegraleerhtehtedulcxednanmuloc 6002-8991 sdoirepbus eht otni elpmas ruo edivid ew ,6 dna 5 snmuloc ni ,yllaniF .ezis tessa naidem evoba dna woleb htiw era elbat siht ni srorre dradnatS .stceffe dexfi raey dna elyts dnuf edulcni snoisserger ehT .ylevitcepser ,9002-7002 dna ,*** ,** ,* yb detoned era slevel ecnacfiingiS .sesehtnerap ni detroper era dna slevel dnuf dna nalp eht ta deretsulc yaw-owt .ylevitcepser ,slevel %1 dna ,%5 ,%01 ot dnopserroc hcihw retfA ot roirP egraL llamS edulcxE 3 poT 6002 7002 snalP snalP FM 3 poT FM seetsurT seetsurT ∗∗∗751.0− ∗∗∗521.0− ∗∗∗041.0− ∗∗∗471.0− ∗∗∗341.0− ∗∗∗381.0− dnuF detailffiA )520.0( )920.0( )320.0( )420.0( )420.0( )830.0( ∗∗∗523.0− ∗∗∗113.0− ∗∗∗133.0− ∗∗∗033.0− ∗∗∗803.0− ∗∗∗583.0− frePwoL )040.0( )150.0( )440.0( )730.0( )330.0( )960.0( 120.0 ∗∗∗611.0− 430.0− ∗∗∗270.0− ∗340.0− ∗∗311.0− frePhgiH )130.0( )030.0( )030.0( )420.0( )320.0( )640.0( ∗∗∗512.0 ∗∗∗642.0 ∗∗∗232.0 ∗∗∗972.0 ∗∗∗212.0 ∗∗∗013.0 dnuF detailffiA*frePwoL )850.0( )560.0( )450.0( )550.0( )060.0( )480.0( 120.0− 120.0 510.0− 600.0 730.0− 850.0 dnuF detailffiA*frePhgiH )040.0( )830.0( )730.0( )630.0( )040.0( )250.0( ∗∗∗110.0 ∗∗∗010.0 ∗∗∗010.0 ∗∗∗010.0 ∗∗∗010.0 ∗∗∗210.0 rroC mumixaM )100.0( )100.0( )100.0( )100.0( )100.0( )200.0( ∗∗∗010.0− ∗∗∗800.0− ∗∗∗120.0− 200.0− ∗∗∗010.0− ∗∗700.0− )eziS noitpO(goL )200.0( )200.0( )400.0( )200.0( )200.0( )300.0( ∗∗∗200.0− ∗∗∗200.0− ∗∗∗300.0− 100.0− ∗∗∗200.0− ∗∗∗400.0− snoitpO fo .oN )000.0( )100.0( )000.0( )100.0( )100.0( )100.0( ∗∗∗320.7 ∗∗∗614.9 ∗∗∗018.8 ∗∗∗428.6 ∗∗∗471.6 ∗∗∗921.51 oitaR esnepxE )322.1( )075.1( )494.1( )791.1( )771.1( )272.2( ∗∗∗410.0 ∗∗∗910.0 ∗∗∗520.0 ∗∗∗210.0 ∗∗∗410.0 ∗∗∗820.0 revonruT )500.0( )600.0( )500.0( )400.0( )400.0( )900.0( ∗∗∗610.0− ∗∗∗220.0− ∗∗∗310.0− ∗∗∗220.0− ∗∗∗910.0− ∗∗∗420.0− )eziS dnuF(goL )200.0( )300.0( )300.0( )200.0( )200.0( )400.0( 000.0 000.0 000.0 000.0 000.0 000.0 egA dnuF )000.0( )000.0( )000.0( )000.0( )000.0( )000.0( 740.0 ∗∗∗892.1− 301.0− 352.0− 410.0− ∗∗∗038.0− .veD .dtS )302.0( )093.0( )282.0( )502.0( )412.0( )072.0( 745,45 103,25 968,25 783,74 219,96 639,63 snoitavresbO 870.0 580.0 001.0 860.0 160.0 921.0 derauqs-R 37

.snoitiddA dnuF detailffianU dna detailffiA :5 elbaT :seicneuqerf noitidda detailffianu dna detailffia delacser rof ledom gniwollof eht fo setamitse tneicffieoc eht stroper elbat ehT detailffia rof yletarapes ledom eht etamitse ew erehw , (cid:15)+ γ (cid:48)Z+ wolFCDnoNgeN β+ freP β+ β = iETARDDA t,f 1−t,f 1−t,f 2 1−t,f 1 0 t,f detailffia fo rebmun eht sa denfied si ,t emit ta f dnuf fo etar noitidda detailffia ehT .)AU,A ∈ i ,.e.i( setar noitidda detailffianu dna ti hcihw ot sunem detailffia fo rebmun latot eht yb dedivid raey eht gnirud noitpo tnemtsevni wen a sa dedda si dnuf eht hcihw ot snalp .)raeysuoiverpehtfodneehttanoitponasadereffoydaerlatonsidnufehthcihwnisnalpdetailffiaforebmuneht ,.e.i(deddaebdluoc fo seicneuqerf noitidda delacser tneserper AUETARDDA dna AETARDDA .ylsuogolana denfied era seicneuqerf noitidda detailffianU t,f t,f .snaem gnidnopserroc rieht yb ,ylevitcepser ,setar noitidda detailffianu dna detailffia ruo gnidivid yb niatbo ew hcihw t emit ta f dnuf ta seetsurt sa evres taht seilimaf esoht ot gnoleb taht esrevinu PSRC eht ni sdnuf lla rof detaluclac era seicneuqerf noitidda ,yllaniF no desab sraey eerht suoiverp eht revo f dnuf fo knar ecnamrofrep elitnecrep eht si freP .elpmas ruo ni snalp eht fo eno ot tsael 1−t,f hcihw ni noitacfiiceps raenil esiweceip tnemges-2 a yb freP ecalper ew ,21–01 dna 6–4 snmuloc nI .PSRC ni elyts emas eht ni sdnuf 1−t,f ehT .)0,5.0− freP(xam= frePhgiHdna)5.0, freP(nim= frePwoLsadenfiedera frePhgiHdna frePwoL 1−t,f,p 1−t,f,p 1−t,f,p 1−t,f,p rehto ehT .evitagen era 1−t raey ni dnuf eht fo swofl CD-non eht rehtehw rof elbairav rotacidni na si wolFCDnoNgeN elbairav 1−t,f dradnats eht ,ega dnuf ,ezis s’dnuf eht fo mhtiragol larutan eht ,revonrut dnuf ,oitar esnepxe dnuf ,edulcni Z selbairav lortnoc deggal era dna level dnuf eht ta deretsulc era elbat siht ni srorre dradnatS .stceffe dexfi raey dna elyts dnuf dna ,nruter s’dnuf eht fo noitaived .ylevitcepser ,slevel %1 dna ,%5 ,%01 ot dnopserroc hcihw ,*** ,** ,* yb detoned era slevel ecnacfiingiS .sesehtnerap ni detroper tnemgeS-2 raeniL tnemgeS-2 raeniL ffiD AU A ffiD AU A ffiD AU A ffiD AU A ∗∗408.1−∗∗∗609.2 ∗∗201.1 ∗∗∗610.1−∗∗∗918.1 ∗∗∗408.0 freP )778.0( )968.0( )784.0( )752.0( )922.0( )371.0( 086.3− 710.5 833.1 031.0− 898.0 ∗∗867.0 frePwoL )916.3( )585.3( )440.1( )176.0( )126.0( )073.0( 241.0− 530.1 398.0 ∗∗918.1−∗∗∗556.2 ∗∗738.0 frePhgiH )659.3( )310.4( )289.0( )587.0( )767.0( )293.0( ∗∗∗555.3 ∗∗∗243.4− ∗∗∗787.0− ∗∗∗655.3 ∗∗∗443.4− ∗∗∗787.0− wolFCDnoNgeN )601.1( )331.1( )172.0( )901.1( )631.1( )172.0( ∗∗323.391 062.111− ∗∗360.28 ∗∗089.791 205.611− ∗∗874.18 ∗∗599.45 246.72− ∗∗353.72 ∗∗020.45 826.62− ∗∗293.72 oitaResnepxE )830.39( )672.59( )966.63( )976.89( )212.101( )447.63( )001.32( )643.32( )832.31( )462.32( )735.32( )102.31( ∗∗013.0− 580.0 ∗∗∗522.0− ∗∗023.0− 690.0 ∗∗∗422.0− 040.0− 050.0− ∗∗∗090.0− 740.0− 340.0− ∗∗∗090.0− revonruT )641.0( )541.0( )380.0( )741.0( )641.0( )380.0( )250.0( )540.0( )330.0( )350.0( )740.0( )330.0( ∗∗∗158.2−∗∗∗170.3 ∗∗022.0 ∗∗∗258.2−∗∗∗170.3 ∗∗022.0 ∗∗∗790.1−∗∗∗583.1 ∗∗∗882.0 ∗∗∗490.1−∗∗∗283.1 ∗∗∗882.0 )eziSdnuF(goL )846.0( )066.0( )090.0( )946.0( )166.0( )090.0( )271.0( )671.0( )530.0( )371.0( )771.0( )530.0( 140.0 120.0− 020.0 040.0 120.0− 020.0 700.0− 000.0 600.0− 600.0− 000.0− 600.0− egAdnuF )140.0( )040.0( )510.0( )140.0( )140.0( )510.0( )420.0( )420.0( )800.0( )420.0( )520.0( )800.0( ∗969.0− ∗∗011.1 241.0 ∗568.0− ∗∗499.0 921.0 211.0− ∗∗573.0 ∗∗∗362.0 561.0− ∗∗034.0 ∗∗∗562.0 .veD .dtS )415.0( )615.0( )971.0( )105.0( )305.0( )381.0( )261.0( )081.0( )290.0( )761.0( )481.0( )490.0( 500,4 500,4 500,4 500,4 500,4 500,4 417,91 417,91 417,91 417,91 417,91 417,91 snoitavresbO 601.0 541.0 660.0 601.0 441.0 660.0 150.0 121.0 640.0 150.0 121.0 640.0 derauqs-R 38

Table 6: Fund Flow Regressions. The table reports the coefficient estimates of the following linear regression: NMG =β +β AF + p,f,t 0 1 p,f,t−1 β LowPerf +β HighPerf +β AF ×LowPerf +β AF ×HighPerf + 2 p,f,t−1 3 p,f,t−1 4 p,f,t−1 p,f,t−1 5 p,f,t−1 p,f,t−1 Z(cid:48) γ+(cid:15) , wheretheexplanatoryvariablesoftheregressionareanalogoustothoseinTable3withthe p,f,t−1 p,f,t exception of Plan Growth, which is a new variable added in this table. Our first dependent variable (with correspondingresultsreportedincolumns1and4forallflowsandparticipantflows,respectively)isnewmoney growth defined as NMG1 = Vp,f,t−Vp,f,t−1(1+Rf,t), where V is the value of participants’ investments p,f,t Vp,f,t−1(1+Rf,t) p,f,t in fund f in plan p in year t and R is the fund’s return during the year. We use two additional definitions f,t for new money growth. NMG2 is new money growth defined as NMG2 = Vp,f,t−Vp,f,t−1(1+Rf,t), with p,f,t Vp,f,t+Vp,f,t−1(1+Rf,t) corresponding results reported in columns 2 and 5 for all flows and participant flows, respectively. Finally, NMG3 shares the numerator with the previous two definitions but replaces the denominator by lagged plan size adjusted for asset returns. Regression results using NMG3 as the dependent variable are reported in columns 3 and 6. Performance percentiles are calculated based on funds in the same style in the CRSP fund universe over the prior three years. The regressions include fund style and year fixed effects. Standard errors in this table are two-way clustered at the plan and fund levels and are reported in parentheses. Significance levels are denoted by *, **, ***, which correspond to 10%, 5%, and 1% levels, respectively. All Fund Flows Participant Flows Only NMG1 NMG2 NMG3 NMG1 NMG2 NMG3 Affiliated Fund 0.270∗∗∗ 0.200∗∗∗ 1.141∗∗∗ 0.076∗∗ 0.026∗∗ 0.057 (0.042) (0.024) (0.256) (0.035) (0.012) (0.133) LowPerf 0.554∗∗∗ 0.408∗∗∗ 3.769∗∗∗ 0.169∗∗∗ 0.077∗∗∗ 1.258∗∗∗ (0.074) (0.045) (0.517) (0.059) (0.021) (0.283) HighPerf 0.352∗∗∗ 0.146∗∗∗ 0.937∗∗∗ 0.343∗∗∗ 0.113∗∗∗ 0.855∗∗∗ (0.052) (0.025) (0.340) (0.046) (0.016) (0.254) LowPerf*Affiliated Fund −0.474∗∗∗ −0.286∗∗∗ −1.545∗∗ −0.143∗ −0.044 −0.601∗ (0.102) (0.057) (0.630) (0.085) (0.030) (0.354) HighPerf*Affiliated Fund 0.082 −0.033 −0.848∗ 0.026 −0.003 −0.131 (0.087) (0.039) (0.444) (0.078) (0.026) (0.297) Maximum Corr −0.012∗∗∗ −0.010∗∗∗ −0.051∗∗∗ 0.001 0.000 −0.020∗∗∗ (0.002) (0.001) (0.009) (0.001) (0.000) (0.005) Log(Option Size) −0.083∗∗∗ −0.068∗∗∗ −0.324∗∗∗ −0.113∗∗∗ −0.037∗∗∗ −0.082∗∗∗ (0.005) (0.001) (0.011) (0.004) (0.001) (0.011) Plan Growth 0.830∗∗∗ 0.161∗∗∗ 4.903∗∗∗ 0.836∗∗∗ 0.355∗∗∗ 7.569∗∗∗ (0.070) (0.037) (0.359) (0.066) (0.025) (0.384) No. of Options −0.001 −0.000 −0.014∗∗∗ −0.004∗∗∗ −0.002∗∗∗ −0.021∗∗∗ (0.001) (0.001) (0.004) (0.001) (0.000) (0.003) Expense Ratio −15.765∗∗∗ −8.794∗∗∗ −34.893∗∗∗ −6.392∗∗∗ −2.093∗∗∗ −4.392 (2.315) (1.146) (10.850) (1.770) (0.606) (6.228) Turnover −0.023∗∗∗ −0.013∗∗∗ −0.054∗∗ 0.001 0.001 −0.010 (0.008) (0.004) (0.021) (0.006) (0.002) (0.017) Log(Fund Size) 0.024∗∗∗ 0.024∗∗∗ 0.233∗∗∗ 0.001 0.002 0.088∗∗∗ (0.005) (0.003) (0.029) (0.005) (0.002) (0.017) Fund Age −0.001∗∗ −0.000 −0.002 −0.001∗∗∗ −0.000∗∗∗ −0.002 (0.000) (0.000) (0.003) (0.000) (0.000) (0.002) Std. Dev. −0.058 −0.197 3.546∗∗ 0.069 0.001 2.928∗∗ (0.457) (0.257) (1.540) (0.333) (0.118) (1.235) Observations 96,483 117,461 116,342 82,711 82,711 82,711 R-squared 0.159 0.515 0.138 0.250 0.221 0.108 39

Table 7: Abnormal Returns of Affiliated and Unaffiliated Funds. Panels A, B, and C of the table report the abnormal return α of fund portfolio f at time t using the f,t Fama-French-Carhart four-factor model (FFM), the Fama and French (1993) model, and the CAPM model, respectively, over our complete sample period using monthly fund return data. At the end of each calendar year,weformequal-weightedportfoliosofaffiliatedandunaffiliateddomesticequityfundsseparatelybasedon whether the funds were kept on, deleted from, or added to the 401(k) menu (“No Changes,” “Deletions,”and “Additions”) during the calendar year. This creates six portfolios. We then further subdivide these six groupsbasedonpastperformance. Inparticular, “AllFunds,” referstotheoverallsixportfoliosand“Lowest Quintile,” (“Lowest Decile”) refers to a sub-portfolio in each group that contains only those funds that also rank in the lowest performance quintile (decile) relative to funds in their style in CRSP during the prior three years. The performance measures are reported in % per month. Robust standard errors are reported in parentheses. Significance levels are denoted by *, **, ***, which correspond to 10%, 5%, and 1% levels, respectively. Panel A: Carhart Alphas No Changes Deletions Additions Affiliated Unaffiliated Affiliated Unaffiliated Affiliated Unaffiliated Funds Funds Funds Funds Funds Funds Lowest Decile −0.33∗∗ −0.08 −0.28∗ −0.15 −0.01 0.12 (0.14) (0.14) (0.17) (0.17) (0.28) (0.18) Lowest Quintile −0.20∗ −0.11 −0.19∗ −0.13 −0.11 −0.02 (0.11) (0.10) (0.11) (0.12) (0.14) (0.11) All Funds −0.00 −0.06 −0.07 −0.09 −0.00 −0.06 (0.04) (0.05) (0.05) (0.06) (0.05) (0.06) Panel B: Fama-French Alphas No Changes Deletions Additions Affiliated Unaffiliated Affiliated Unaffiliated Affiliated Unaffiliated Funds Funds Funds Funds Funds Funds Lowest Decile −0.33∗∗ −0.08 −0.28 −0.15 −0.02 0.13 (0.14) (0.16) (0.17) (0.19) (0.27) (0.19) Lowest Quintile −0.20∗ −0.10 −0.19∗ −0.13 −0.11 −0.02 (0.12) (0.11) (0.11) (0.14) (0.14) (0.12) All Funds −0.00 −0.06 −0.07 −0.09 −0.00 −0.06 (0.04) (0.05) (0.05) (0.06) (0.05) (0.06) Panel C: CAPM Alphas No Changes Deletions Additions Affiliated Unaffiliated Affiliated Unaffiliated Affiliated Unaffiliated Funds Funds Funds Funds Funds Funds Lowest Decile −0.09 0.22 −0.12 0.14 0.06 0.39 (0.17) (0.26) (0.20) (0.31) (0.31) (0.24) Lowest Quintile 0.03 0.12 −0.12 0.07 0.08 0.23 (0.18) (0.18) (0.14) (0.24) (0.18) (0.18) All Funds 0.02 −0.01 −0.03 −0.02 0.02 −0.01 (0.04) (0.05) (0.05) (0.06) (0.05) (0.08) 40

Internet Appendix to “It Pays to Set the Menu: Mutual Fund Investment Options in 401(k) Plans” Veronika K. Pool Clemens Sialm Irina Stefanescu 1

This internet appendix provides supplemental analyses to the main tables and figures in ‘It Pays to Set the Menu: Mutual Fund Investment Options in 401(k) Plans.’ 1 Data This section explains in more detail the data construction. We collect the investment options offered in 401(k) plans from Form 11-K filed with the U.S. Securities and Exchange Commission (SEC). All plans offering company stock as an investment option for plan participants are required to file this form with the SEC. The filing provides an overall description of the plan, identifies the trustee, all individual choices available to participants (the menu), and the accumulated value of assets invested in each of these vehicles at the end of the fiscal year. We manually collect these data as disclosure is not standardized across plans and firms. We start by webcrawling the SEC’s website from 1998 to 2009 to identify all companies that report Form 11-K. We collect 26,624 links to 11-K filings but restrict this sample to companies covered by COMPUSTAT.1 We eliminate filings that have been submitted to the SEC in duplicate and consolidate amendments with the corresponding original filings. From these documents we collect all tables that describe the “Schedule of Assets” of the plan. In most cases, the table reports the complete set of investment options offered by the plan, including the employers’ own stock, other common stocks, mutual funds, separate accounts, or commingled trusts, as well as the current value of investments in these options at the end of the fiscal year. Occasionally, the table describes only those investment options that capture more than 5% of the plan’s assets or alternatively, only mutual fund investments. To overcome the incomplete and non-standardized disclosure of these tables, we supplement our Form 11-K 1Our data collection initially included paper filings (not only pdfs of electronic documents). However, paper filings have been removed from public use on the SEC website while our data collection was still in progress. We only partially incorporate these plan year observations. 2

information with plan level data from Form 5500. The resulting dataset has more than 302,000 observations, containing information at the firm-year-plan-fund level. To obtain information on the characteristics of the mutual funds included in DC plans, we match all funds listed on the menus to the CRSP Survivorship Bias-Free U.S. Mutual Fund database. To aid our matching task, we proceed in several steps. We start by filtering our menu options for non-mutual fund assets. These include, for instance, common stocks, bonds, insurance products, or guaranteed investment contracts. In approximately 15% of the cases, the SEC Form 11-K contains information on the number of shares of each asset held by the plan in addition to the market value of the position. This allows us to calculate the net asset value (NAV) of the position on the report date. When the NAV information is available, we match the menu choice to the CRSP mutual fund files by NAV and date. For the rest of the sample, we hand match the 11-K funds to the mutual fund database by name. Since most plans do not identify the exact share class of the fund offered on the menu, we establish the link between our 401(k) sample and the CRSP Survivorship Bias-free Mutual Fund database at the fund-level, that is, we combine information on all available share classes of each fund in CRSP into fund-level variables. Accordingly, fund age is calculated as the age of the oldest share class, fund size is the sum of the total net assets of all share classes, and fund returns and expense ratios are calculated as the total net asset value weighted average returns and expense ratios of the share classes, respectively. We also classify each mutual fund in our sample as “balanced,” “bond,” “domestic equity,” “international equity,” or “other.” We create separate dummy variables for money market funds, target date funds, and index funds. We manually group funds into target date and index fund categories based on fund name. Finally, we perform two additional data steps to complete our sample. First, we assign unique plan IDs to create time-series at the plan level. Form 11-K does not always disclose the plan number. Companies occasionally sponsor multiple plans for different subsidiaries, salaried and hourly employees, or unionized and non-unionized workers. In order to track the same plan 3

over time, we collect the plan Employer Identification Number (EIN) and Plan Number (PN) by searching Form 5500 by plan name and assets. Once established, the link with Form 5500 allows us to collect additional information on total participants, active participants, employer and employee contributions, total assets, and whether the plan is collectively bargained or not. We manually collect the trustee name (and any trustee change occurring during the year) from the plan description available in Form 11-K. We supplement and cross check this information with the name of the trustee disclosed in Form 5500. 2 Menu Changes This section provides additional robustness tests for fund deletions and additions. 2.1 Fund Deletions by Performance Deciles Table A-1 summarizes mean annual deletion frequencies (as a %) by mutual fund affiliation. These deletion frequencies are analogous to those reported in Figure 1 in the paper, but also report the results based on performance percentile ranks that are determined by prior one and five year performance evaluation horizons. Panel A includes the full sample, Panel B includes only funds that appear contemporaneously as affiliated and unaffiliated funds. Standard errors in these panels are clustered at the fund level. Panes C and D report identical difference test but use the Fama-MacBeth methodology to calculate the deletion frequencies and corresponding standard errors. 4

2.2 Probit Model for Fund Deletions For robustness, we re-estimate the linear probability models in Table 3 in the paper using a probit specification. Table A-2 reports the estimated marginal effects for the linear and 2segment models of fund deletions. The interaction effects and the corresponding standard errors on the interaction variables between the affiliation dummy and the performance percentiles are estimated based on Ai and Norton (2003). The interaction effect is defined as the change in the predicted probability of a deletion for a change in both fund performance and fund affiliation. Figures A-1–A-3 depict the corresponding graphs. 2.3 Alternative Performance Rankings In the paper we rank mutual funds based on their prior performance relative to the universe of CRSP funds in the same style. We refer to this global ranking as “Overall Ranking” in this appendix. For robustness we also compute two alternative ranking methods, where the performance percentile of a fund is either measured relative to the other investment options in a specific 401(k) plan (“Plan Ranking”) or relative to the other funds offered by the fund’s family (“Family Ranking”). The overall ranking method captures the performance of a fund relative to the universe of available mutual funds in the U.S., which could be viewed as the most comprehensive metric. When a fund underperforms compared to the other investment choices included in the plan or the other options in the fund family, the plan may be pressured to remove the fund from the menu as underperformance in this setting is perhaps more transparent. Table A-3 summarizes the coefficient estimates when Perf is defined using the p,f,t−1 alternative ranking methodologies based on fund performance in the previous 36 months. The results are qualitatively and quantitatively similar to the base-case results reported in Table 5

3 in the paper. Thus, our findings are not affected by whether we benchmark mutual funds relative to the universe of mutual funds or relative to other funds included in the same 401(k) plan or other funds offered by the same fund family. 2.4 Robustness for Other Performance Ranking Horizons In Table A-4 we re-estimate our baseline 2-segment model for fund deletions using prior one and five year fund performance to create performance rankings. The table summarizes the results for all three alternative ranking methodologies introduced above (i.e., overall, plan, and family rankings). 2.5 Sensitivity to Extreme Performance To analyze in more depth the sensitivity of deletions to extreme performance, we estimate a specification using three piecewise linear segments instead of the two segments from equation (1) in the paper. The performance segments are 1) the lowest performance quintile, 2) the highest performancequintile,and3)thethreemiddleperformancequintiles,whicharepulledtogetherto represent a single performance segment. Following Sirri and Tufano (1998), the performance in the lowest quintile is given by LowPerfQ = min(Perf ,0.2), the performance in the p,f,t−1 p,f,t−1 three middle quintiles is given by MidPerfQ = min(Perf −LowPerfQ ,0.6), p,f,t−1 p,f,t−1 p,f,t−1 and the performance in the highest quintile is given by HighPerfQ = (Perf − p,f,t−1 p,f,t−1 LowPerfQ −MidPerfQ ). p,f,t−1 p,f,t−1 Table A-5 reports the estimates for the three piecewise linear segments using our alternative ranking methods, based on the three year performance ranking horizon. Consistent with the base-case specification from Table 3 in the paper, we find that deletions are less sensitive to poor and intermediate performance for affiliated funds. Interestingly, in our overall ranking 6

model, we find that the probability of deleting unaffiliated funds that rank in the highest performance quintile actually increases with the performance percentile. 2.6 Fund Deletions, Robustness Tests TableA-6showstheresultsofourlinearprobabilitymodelspecifiedinequation(1)usingvarious sample restrictions. In columns 1 and 2, we show that our results remain after controlling for trustee and fund fixed effects, respectively. In column 3 we use the Fama-MacBeth methodology to compute our coefficient estimates and corresponding standard errors. In column 4, we re-estimate our results using information only on those plans that are trusteed by a mutual fund family. In column 5, we only include mutual fund trustees and require that they offer at least 10 funds in their fund family. The rationale behind excluding trustees with only a few funds in their product lineup is that these trustees could be large financial conglomerates or banks with a small mutual fund arm. In column 6 we exclude all plan year observations when a trustee change occurs, as in these plan years fund exits and entries are likely driven by the plan sponsor. Finally, in columns 7-9 we restrict the sample of funds considered. In column 7 we exclude all target date funds since these funds are often used as default investment options. In column 8, we restrict our sample to equity funds, while in column 9 we only include actively managed funds. These results are very consistent with the results in our baseline specification. 2.7 Fund Additions by Performance Deciles To investigate how a fund’s propensity to be added to a menu depends on its affiliation, we determine the addition frequency of each fund in CRSP as an affiliated and unaffiliated menu choice,respectively,asdescribedinSection4.4inthepaper. WhileFigure2inthepaperdisplays 7

rescaled addition rates, we tabulate raw addition rates in Figure A-4 in this document. The corresponding average raw addition frequencies by affiliation and performance are summarized in Table A-7, which also extends the results for the one and five year performance evaluation horizons. Panel A includes the full sample, Panel B includes only those funds that belong to families that provide trustee services for at least one plan in our sample. Standard errors in these panels are clustered at the fund level. Panes C and D report identical difference test but use the Fama-MacBeth methodology to calculate the addition frequencies and corresponding standard errors. 2.8 Newly Added Funds by Affiliation This section provides some additional results on the determinants of fund additions. We investigate the characteristics of affiliated and unaffiliated funds based on our sample of newly added funds. Table 2 of the paper provides univariate evidence that newly listed affiliated funds exhibit lower past performance than unaffiliated funds in the same category. We confirm this finding in Figure A-5. The figure describes the distribution of affiliated and unaffiliated fund additions separately, by performance deciles. Fund performance is measured by the performance percentile of each fund in the universe of CRSP funds in the same style over the past three years. The results reveal that the proportion of unaffiliated funds with strong past performance is larger compared to that of affiliated funds, while affiliated funds are more likely to come to the menu with a mediocre performance record. To further explore the difference in past performance across newly added affiliated and unaffiliated funds, we estimate the following linear probability model for fund addition type: AFADD = β +β ×Perf +Z(cid:48) γ +(cid:15) , (1) p,f,t 0 1 p,f,t−1 p,f,t−1 p,f,t 8

where the dependent variable takes the value of one if fund f added to plan p at time t is an affiliated fund, and zero otherwise. Since the sample used in this analysis includes only fund additions, it reflects the choice between selecting an affiliated fund over an unaffiliated fund. Perf is the performance percentile of mutual fund f over the previous one, three, or five p,f,t−1 years based on overall rankings and it enters the analysis as a linear term. Our additional controls include various fund characteristics and plan level variables, such as the number of menu options and plan size. The results are reported in Table A-8 with standard errors two-way clustered at the plan and fund levels. Consistent with menu favoritism, affiliated fund additions are associated with worse past performance even after controlling for other fund characteristics. This is represented by our Perf coefficient estimates, which are significantly negative at the one percent p,f,t−1 level for each of our performance measures. 3 New Money Growth Figure A-6 provides histograms of the percentage flows into various plan options for affiliated and unaffiliated funds in the lowest performance quintile over the previous three years. 4 Future Performance In Section 6 of the paper, we compute the abnormal return α of fund portfolio f at time t f,t using the Fama-French-Carhart four-factor model (FFM) over our complete sample period using monthly fund return data from the CRSP Mutual Fund database: R −R = α +βM(R −R )+βSMB(R −R ) f,t TB,t f,t f,t M,t TB,t f,t S,t B,t 9

+βHML(R −R )+βUMD(R −R )+(cid:15) . (2) f,t H,t L,t f,t U,t D,t f,t Thereturnoffundportfoliof duringtimeperiodtisdenotedbyR . TheindexM corresponds f,t to the market portfolio and the index TB to the risk-free Treasury bill rate. Portfolios of small and large stocks are denoted by S and B, respectively; portfolios of stocks with high and low ratios between their book values and their market values are denoted by H and L, respectively; and portfolios of stocks with relatively high and low returns during the previous year are denoted by U and D, respectively. We obtain monthly factor returns and the risk-free rate from Kenneth French’s website. 4.1 Future Performance In Section 6 of the paper, we form equal-weighted portfolios of affiliated and unaffiliated domestic equity funds separately at the end of each calendar year, as described in the section. Table 7 in the paper reports the abnormal return (α) of these portfolios using the Fama-French- Carhart four-factor model (FFM), the Fama and French (1993) model, and the CAPM model, respectively, over our complete sample period using monthly fund return data. In Panels A, B, and C of Table A-9 we augment these results by reporting the difference in the abnormal returns of the affiliated and unaffiliated fund portfolios in each category. 10

References Ai, C. and E. C. Norton (2003). Interaction terms in logit and probit models. Economics Letters 80, 123–129. Fama, E. F. and K. R. French (1993). Common risk factors in the returns on stocks and bonds. Journal of Financial Economics 33(1), 3–56. Sirri,E.R.andP.Tufano(1998). Costlysearchandmutualfundflows. Journal of Finance 53(5), 1598–1622.

.seliceD ecnamrofreP yb snoiteleD dnuF :1-A elbaT slenaP .seliced ecnamrofrep dna noitailffia eetsurt yb )% a sa( seicneuqerf noiteled dnuf launna naem sezirammus elbat ehT )k(104 elpitlum no ylsuoenaropmetnoc raeppa taht sdnuf esoht ylno edulcni D dna B slenaP .elpmas lluf eht edulcni C dna A eht fo oitar eht etaluclac ew ,raey yrevE .dnuf detailffianu na sa ecno tsael ta dna dnuf detailffia na sa ecno tsael ta ,sunem detailffia fo rebmun latot eht ot raey eht gnirud detsiled si dnuf eht hcihw morf sunem )detailffianu( detailffia fo rebmun elitnecrep rieht no desab sdnuf gnipuorg yb detaerc era seliced ecnamrofreP .dnuf eht htiw detaicossa sunem )detailffianu( neht eW .sraey evfi dna ,eerht ,eno roirp eht revo esrevinu dnuf PSRC eht ni elyts emas eht fo sdnuf gnoma ecnamrofrep ni ecnereffid eht fo stset rof slevel ecnacfiingiS .noitailffia dna ecnamrofrep yb seicneuqerf noiteled ’sdnuf eht ssorca egareva srorre dradnats B dna A slenaP nI .ylevitcepser ,slevel %1 dna ,%5 ,%01 ot dnopserroc hcihw ,*** ,** ,* yb detoned era snaem dradnats tseW-yeweN dna stneicffieoc hteBcaM-amaF troper ew ,D dna C slenaP ni elihw ,level dnuf eht ta deretsulc era .3 fo htgnel gal a gnisu srorre sdnuF llA :A lenaP sraeY 5 sraeY 3 raeY 1 ecnamrofreP AU-A AU A AU-A AU A AU-A AU A eliceD ∗∗∗2.9− 8.32 6.41 ∗∗∗8.11− 5.52 7.31 ∗∗∗0.01− 2.22 1.21 1 ∗∗∗8.5− 8.32 0.81 ∗∗∗7.4− 5.32 8.81 ∗9.2− 1.12 3.81 2 ∗∗∗4.5− 5.12 1.61 ∗∗∗2.4− 2.22 0.81 ∗∗∗1.5− 6.02 5.51 3 ∗∗∗3.8− 7.12 4.31 ∗∗∗1.6− 6.02 5.41 ∗∗∗4.4− 5.02 1.61 4 ∗∗∗6.4− 4.81 8.31 ∗∗∗5.6− 6.81 0.21 ∗∗∗0.4− 5.71 5.31 5 ∗∗∗9.6− 2.91 3.21 ∗∗∗6.5− 7.71 1.21 ∗∗∗5.4− 2.71 7.21 6 ∗∗∗4.6− 2.81 8.11 ∗∗∗7.4− 4.61 7.11 ∗∗∗6.6− 1.71 4.01 7 ∗∗∗8.3− 1.51 3.11 ∗∗∗8.5− 4.61 6.01 ∗∗∗3.5− 2.71 0.21 8 ∗∗∗3.3− 6.41 4.11 ∗∗∗6.4− 9.41 3.01 ∗∗∗9.4− 0.61 1.11 9 0.2− 8.51 8.31 2.0− 3.51 1.51 ∗∗∗6.3− 0.61 4.21 01 suneM detailffianU dna detailffiA htoB no sdnuF fo elpmasbuS :B lenaP sraeY 5 sraeY 3 raeY 1 ecnamrofreP AU-A UA A AU-A UA A AU-A UA A eliceD ∗∗2.7− 5.32 4.61 ∗∗∗8.01− 5.42 6.31 ∗∗∗9.7− 1.91 1.11 1 ∗∗8.6− 7.42 9.71 8.1− 8.02 0.91 2.2− 8.91 6.71 2 ∗∗4.4− 1.91 7.41 ∗∗4.4− 8.02 4.61 4.2− 3.81 0.61 3 ∗1.3− 1.71 0.41 ∗∗3.3− 9.61 6.31 5.2− 1.81 5.51 4 ∗2.3− 8.51 6.21 6.0− 3.41 6.31 2.2− 7.61 6.41 5 ∗∗∗2.4− 7.61 5.21 ∗∗∗3.5− 5.71 2.21 ∗∗∗5.4− 2.51 7.01 6 ∗∗9.3− 5.51 5.11 ∗6.2− 0.31 3.01 ∗∗∗7.3− 9.21 2.9 7 ∗∗9.2− 0.31 1.01 ∗∗∗0.4− 5.31 5.9 0.2− 1.31 1.11 8 0.1− 8.01 9.9 ∗∗9.2− 3.21 4.9 ∗∗3.3− 8.31 5.01 9 4.0− 9.21 5.21 0.0 8.21 9.21 9.2− 8.21 9.9 01 12

stseT hteBcaM-amaF ,sdnuF llA :C lenaP sraeY 5 sraeY 3 raeY 1 ecnamrofreP AU-A AU A AU-A AU A AU-A AU A eliceD ∗∗∗6.9− 3.42 6.41 ∗∗2.11− 1.62 9.41 ∗∗∗1.9− 6.22 5.31 1 ∗∗∗1.6− 7.22 6.61 ∗∗6.4− 0.32 3.81 4.2− 5.02 1.81 2 ∗∗∗3.5− 9.02 6.51 ∗2.3− 5.12 3.81 ∗∗8.4− 4.02 6.51 3 ∗∗∗4.7− 6.12 1.41 ∗∗∗6.5− 2.02 6.41 ∗8.3− 3.02 4.61 4 ∗∗8.3− 0.81 2.41 ∗∗∗2.5− 9.71 6.21 ∗∗∗9.3− 2.71 4.31 5 ∗∗∗7.6− 9.81 1.21 ∗∗∗6.5− 4.71 8.11 ∗∗∗7.4− 0.71 3.21 6 ∗∗∗0.6− 6.71 6.11 ∗∗∗5.4− 9.51 4.11 ∗∗∗9.5− 1.61 2.01 7 ∗∗∗2.4− 8.41 6.01 ∗∗∗9.5− 9.51 0.01 ∗∗∗9.4− 9.61 0.21 8 ∗∗4.3− 3.41 9.01 ∗∗∗8.4− 8.41 0.01 ∗∗∗9.4− 8.51 9.01 9 2.2− 7.51 5.31 1.1− 4.51 4.41 ∗∗1.4− 0.61 9.11 01 stseT hteBcaM-amaF ,suneM detailffianU dna detailffiA htoB no sdnuF fo elpmasbuS :D lenaP sraeY 5 sraeY 3 raeY 1 ecnamrofreP AU-A AU A AU-A AU A AU-A AU A eliceD ∗∗∗2.21− 4.72 2.51 ∗∗∗5.21− 1.62 6.31 ∗∗∗6.7− 6.02 9.21 1 9.6− 6.22 6.51 1.2− 0.02 9.71 0.1− 9.71 9.61 2 ∗∗∗6.3− 6.81 9.41 ∗7.4− 0.12 4.61 9.0− 0.71 1.61 3 1.2− 5.61 4.41 1.3− 9.61 8.31 6.2− 7.71 1.51 4 3.2− 3.51 0.31 1.0 7.31 8.31 5.2− 0.71 4.41 5 ∗6.4− 0.71 4.21 ∗∗2.5− 4.71 2.21 ∗∗8.4− 0.51 2.01 6 5.3− 9.41 3.11 ∗∗7.2− 1.31 3.01 ∗∗4.3− 6.21 1.9 7 ∗∗∗0.4− 3.31 3.9 ∗∗2.4− 2.31 0.9 9.1− 0.31 1.11 8 2.1− 7.01 5.9 ∗3.3− 3.21 0.9 6.3− 1.41 4.01 9 5.1− 7.31 2.21 5.1− 7.31 1.21 ∗∗8.3− 5.31 7.9 01 13

Table A-2: Probit Model for Fund Deletions. Columns 1 and 2 report estimated marginal effects for the linear performance probit model for fund deletions: Pr(DEL = 1|X) = Φ(β + β AF + β Perf + β AF × Perf + p,f,t 0 1 p,f,t−1 2 p,f,t−1 3 p,f,t−1 p,f,t−1 β NegNonDCFlow +β AF ×NegNonDCFlow +Z(cid:48) γ+(cid:15) ),whilecolumns3and4tabu- 4 f,t−1 5 p,f,t−1 f,t−1 p,f,t−1 p,f,t latecorrespondingmarginaleffectsforthe2-segmentprobitmodel: Pr(DEL =1|X)=Φ(β +β AF + p,f,t 0 1 p,f,t−1 β LowPerf +β HighPerf +β AF ×LowPerf +β AF ×HighPerf + 2 p,f,t−1 3 p,f,t−1 4 p,f,t−1 p,f,t−1 5 p,f,t−1 p,f,t−1 β NegNonDCFlow +β AF ×NegNonDCFlow +Z(cid:48) γ), where DEL is an indicator 6 f,t−1 7 p,f,t−1 f,t−1 p,f,t−1 p,f,t that takes the value of one if fund f is deleted from plan p in year t and zero otherwise and AF is an p,f,t−1 indicator for whether the trustee of plan p is affiliated with the family of fund f at the end of year t−1. Perf isthepercentileperformancerankoff overthepriorthreeyearsbasedonfundsinthesamestylein p,f,t−1 the CRSP fund universe and LowPerf and HighPerf are defined as LowPerf =min(Perf ,0.5) p,f,t−1 p,f,t−1 and HighPerf = max(Perf −0.5,0). NegNonDCFlow is an indicator for whether the p,f,t−1 p,f,t−1 f,t−1 non-DCflowsofthefundinyeart−1arenegative. TheotherlaggedcontrolvariablesZ includethemaximum returncorrelationofthefundwithothermenuoptions, thelogarithmofoptionsize, thenumberofoptions, the expense ratio, fund turnover, the logarithm of the fund’s size, fund age, the standard deviation of the fund’s return, and fund style and year fixed effects. The marginal effects for the interaction terms are computed using the INTEFF command based on Ai and Norton (2003). Standard errors are clustered at the plan level and are in parentheses. Significance levels are denoted by *, **, ***, corresponding to 10%, 5%, and 1% levels, respectively. Linear 2-Segment AffiliatedFund −0.069∗∗∗ −0.059∗∗∗ −0.085∗∗∗ −0.061∗∗∗ (0.007) (0.009) (0.009) (0.011) Perf −0.128∗∗∗ −0.119∗∗∗ (0.007) (0.008) Perf*AffiliatedFund 0.042∗∗∗ 0.028∗∗ (0.012) (0.013) LowPerf −0.215∗∗∗ −0.109∗∗∗ (0.013) (0.014) HighPerf −0.051∗∗∗ −0.029∗∗ (0.012) (0.014) LowPerf*AffiliatedFund 0.102∗∗∗ 0.064∗∗ (0.024) (0.028) HighPerf*AffiliatedFund −0.010 −0.037 (0.023) (0.026) NegNonDCFlow 0.032∗∗∗ 0.044∗∗∗ (0.004) (0.004) NegNonDCFlow*AffiliatedFund −0.009 −0.012∗ (0.006) (0.006) MaximumCorr 0.009∗∗∗ 0.010∗∗∗ 0.010∗∗∗ 0.011∗∗∗ (0.001) (0.001) (0.001) (0.001) Log(OptionSize) −0.008∗∗∗ −0.007∗∗∗ −0.008∗∗∗ −0.007∗∗∗ (0.001) (0.001) (0.001) (0.001) No. ofOptions −0.002∗∗∗ −0.002∗∗∗ −0.002∗∗∗ −0.002∗∗∗ (0.000) (0.000) (0.000) (0.000) Exp. Ratio 6.972∗∗∗ 7.143∗∗∗ 6.735∗∗∗ 7.384∗∗∗ (0.624) (0.657) (0.628) (0.656) Turnover 0.012∗∗∗ 0.011∗∗∗ 0.012∗∗∗ 0.010∗∗∗ (0.002) (0.002) (0.002) (0.002) Log(FundSize) −0.018∗∗∗ −0.016∗∗∗ −0.018∗∗∗ −0.018∗∗∗ (0.001) (0.001) (0.001) (0.001) FundAge 0.000 −0.000 0.000 0.000 (0.000) (0.000) (0.000) (0.000) Std. Dev. −0.074 −0.232∗∗ −0.145 −0.554∗∗∗ (0.117) (0.109) (0.115) (0.106) Observations 106,848 65,855 106,848 65,855 R-squared 0.0939 0.0996 0.0948 0.0950 14

Table A-3: Linear Probability Model for Fund Deletions: Alternative Rankings. The table reports the coefficient estimates for the linear and piecewise linear 2-segment fund deletion models estimated in Table 3 in the paper for two alternative performance rankings. Under “plan ranking” we calculate the performance percentile rank of each fund on the menu relative to the other investment options in the 401(k) plan. Under “family ranking”, performance percentile ranks are calculated relative to the other funds in the fund’s family. In both cases, we use fund performance in the prior 36 months to compute performance ranks and include fund style and year fixed effects. Standard errors are two-way clustered at the plan and fund levels and are reported in parentheses. Significance levels are denoted by *, **, ***, which correspond to 10%, 5%, and 1% levels, respectively. Plan Ranking Family Ranking Linear 2-Segment Linear 2-Segment Affiliated Fund −0.085∗∗∗ −0.102∗∗∗ −0.072∗∗∗ −0.093∗∗∗ (0.013) (0.015) (0.013) (0.018) Perf −0.142∗∗∗ −0.089∗∗∗ (0.017) (0.015) LowPerf −0.262∗∗∗ −0.160∗∗∗ (0.023) (0.029) HighPerf −0.015 −0.029 (0.022) (0.025) Perf*Affiliated Fund 0.086∗∗∗ 0.055∗∗∗ (0.020) (0.019) LowPerf*Affiliated Fund 0.144∗∗∗ 0.122∗∗∗ (0.033) (0.043) HighPerf*Affiliated Fund 0.032 −0.003 (0.032) (0.037) Maximum Corr 0.010∗∗∗ 0.011∗∗∗ 0.011∗∗∗ 0.011∗∗∗ (0.001) (0.001) (0.001) (0.001) Log(Option Size) −0.009∗∗∗ −0.009∗∗∗ −0.009∗∗∗ −0.009∗∗∗ (0.002) (0.002) (0.002) (0.002) No. of Options −0.002∗∗∗ −0.002∗∗∗ −0.002∗∗∗ −0.002∗∗∗ (0.000) (0.000) (0.000) (0.000) Exp. Ratio 8.570∗∗∗ 8.558∗∗∗ 8.574∗∗∗ 8.469∗∗∗ (1.098) (1.100) (1.110) (1.111) Turnover 0.016∗∗∗ 0.016∗∗∗ 0.017∗∗∗ 0.017∗∗∗ (0.004) (0.004) (0.004) (0.004) Log(Fund Size) −0.019∗∗∗ −0.019∗∗∗ −0.021∗∗∗ −0.021∗∗∗ (0.002) (0.002) (0.002) (0.002) Fund Age 0.000 0.000 0.000 0.000 (0.000) (0.000) (0.000) (0.000) Std. Dev. −0.213 −0.300 −0.287 −0.319 (0.203) (0.202) (0.194) (0.197) Observations 107,355 107,355 107,175 107,175 R-squared 0.075 0.077 0.070 0.071 15

Table A-4: Linear Probability Model for Fund Deletions: Different Horizons. The table reports the OLS coefficient estimates of our baseline piecewise linear 2-segment fund deletion model describedinequation(1)inthepaperfortwoalternativeperformanceevaluationhorizons. Incolumns1,3, and 5 we calculate the performance percentile rank of each fund over the previous one year based on either overall rankings (column 1), plan rankings (column 3), or fund family rankings (column 5). In columns 2,4, and 6 we report corresponding results for the three ranking methods using the fund’s performance in the previous five years to calculate percentile ranks. The regressions include fund style and year fixed effects. Standard errors are two-way clustered at the plan and fund levels and are reported in parentheses. Significance levels are denoted by *, **, ***, which correspond to 10%, 5%, and 1% levels, respectively. Overall Ranking Plan Ranking Family Ranking 1 Year 5 Years 1 Year 5 Years 1 Year 5 Years Affiliated Fund −0.108∗∗∗ −0.107∗∗∗ −0.088∗∗∗ −0.103∗∗∗ −0.075∗∗∗ −0.090∗∗∗ (0.016) (0.023) (0.014) (0.016) (0.014) (0.018) LowPerf −0.183∗∗∗ −0.197∗∗∗ −0.183∗∗∗ −0.160∗∗∗ −0.100∗∗∗ −0.134∗∗∗ (0.030) (0.037) (0.019) (0.021) (0.025) (0.033) HighPerf −0.017 −0.142∗∗∗ 0.025 −0.165∗∗∗ −0.013 −0.066∗∗ (0.024) (0.024) (0.021) (0.019) (0.025) (0.026) LowPerf*Affiliated Fund 0.171∗∗∗ 0.099∗ 0.119∗∗∗ 0.089∗∗∗ 0.071∗∗ 0.109∗∗ (0.037) (0.055) (0.033) (0.026) (0.034) (0.046) HighPerf*Affiliated Fund −0.036 0.099∗∗∗ −0.002 0.107∗∗∗ 0.033 0.006 (0.032) (0.034) (0.031) (0.026) (0.034) (0.039) Maximum Corr 0.011∗∗∗ 0.010∗∗∗ 0.011∗∗∗ 0.010∗∗∗ 0.011∗∗∗ 0.011∗∗∗ (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) Log(Option Size) −0.009∗∗∗ −0.009∗∗∗ −0.009∗∗∗ −0.009∗∗∗ −0.009∗∗∗ −0.009∗∗∗ (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) No. of Options −0.002∗∗∗ −0.002∗∗∗ −0.002∗∗∗ −0.002∗∗∗ −0.002∗∗∗ −0.002∗∗∗ (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) Exp. Ratio 8.132∗∗∗ 8.073∗∗∗ 8.806∗∗∗ 8.232∗∗∗ 8.593∗∗∗ 8.668∗∗∗ (1.117) (1.143) (1.114) (1.139) (1.120) (1.111) Turnover 0.017∗∗∗ 0.018∗∗∗ 0.016∗∗∗ 0.018∗∗∗ 0.017∗∗∗ 0.017∗∗∗ (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) Log(Fund Size) −0.021∗∗∗ −0.018∗∗∗ −0.021∗∗∗ −0.017∗∗∗ −0.021∗∗∗ −0.020∗∗∗ (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) Fund Age 0.000 −0.000 0.000 −0.000 0.000 0.000 (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) Std. Dev. −0.478∗∗ −0.078 −0.568∗∗∗ −0.042 −0.451∗∗ −0.177 (0.198) (0.194) (0.201) (0.196) (0.206) (0.195) Observations 106,848 106,848 107,355 107,355 107,175 107,175 R-squared 0.072 0.075 0.071 0.075 0.069 0.071 16

Table A-5: Linear Probability Model for Fund Deletions: Alternative Functional Forms. The table reports the coefficient estimates of the model for fund deletions described in equation (1) but replaces our baseline 2-segment model with a 3-segment piecewise linear specification. In the 3-segment specification the performance segments are 1) the lowest performance quintile, 2) the highest performance quintile, and 3) the three middle performance quintiles, which are pulled together to represent a single performance segment. Following Sirri and Tufano (1998), the performance in the lowest quintile is given by LowPerfQ = min(Perf ,0.2), the performance in the three middle quintiles is given by p,f,t−1 p,f,t−1 MidPerfQ = min(Perf −LowPerfQ ,0.6), and the performance in the highest quintile p,f,t−1 p,f,t−1 p,f,t−1 is given by HighPerfQ =(Perf −LowPerfQ −MidPerfQ ), where Perf is p,f,t−1 p,f,t−1 p,f,t−1 p,f,t−1 p,f,t−1 the performance percentile of mutual fund f over the previous three years based on either overall rankings (column 1), 401(k) plan rankings (column 2), or fund family rankings (column 3). The regressions include fund style and year fixed effects. Standard errors in this table are two-way clustered at the plan and fund levels and are reported in parentheses. Significance levels are denoted by *, **, ***, which correspond to 10%, 5%, and 1% levels, respectively. Performance Ranking Overall Plan Family Affiliated Fund −0.172∗∗∗ −0.124∗∗∗ −0.116∗∗∗ (0.035) (0.025) (0.030) LowPerfQ −0.521∗∗∗ −0.699∗∗∗ −0.449∗∗∗ (0.122) (0.089) (0.109) MidPerfQ −0.187∗∗∗ −0.105∗∗∗ −0.065∗∗∗ (0.018) (0.015) (0.020) HighPerfQ 0.207∗∗∗ 0.107 −0.044 (0.074) (0.088) (0.074) LowPerfQ*Affiliated Fund 0.468∗∗ 0.318∗∗ 0.320∗ (0.188) (0.129) (0.164) MidPerfQ*Affiliated Fund 0.110∗∗∗ 0.068∗∗∗ 0.031 (0.025) (0.020) (0.026) HighPerfQ*Affiliated Fund −0.167 0.049 0.098 (0.102) (0.121) (0.121) Maximum Corr 0.010∗∗∗ 0.011∗∗∗ 0.011∗∗∗ (0.001) (0.001) (0.001) Log(Option Size) −0.009∗∗∗ −0.009∗∗∗ −0.009∗∗∗ (0.002) (0.002) (0.002) No. of Options −0.002∗∗∗ −0.002∗∗∗ −0.002∗∗∗ (0.000) (0.000) (0.000) Exp. Ratio 7.607∗∗∗ 8.593∗∗∗ 8.417∗∗∗ (1.094) (1.098) (1.109) Turnover 0.018∗∗∗ 0.015∗∗∗ 0.017∗∗∗ (0.004) (0.004) (0.004) Log(Fund Size) −0.019∗∗∗ −0.019∗∗∗ −0.021∗∗∗ (0.002) (0.002) (0.002) Fund Age 0.000 0.000 0.000 (0.000) (0.000) (0.000) Std. Dev. −0.186 −0.332 −0.322 (0.191) (0.203) (0.198) Observations 106,848 107,355 107,175 R-squared 0.079 0.078 0.071 17

.stseT ssentsuboR :snoiteleD dnuF rof ledoM ytilibaborP raeniL :6-A elbaT eht ni )1( noitauqe ni debircsed ledom noiteled dnuf tnemges-2 enilesab ruo rof setamitse tneicffieoc eht stroper elbat ehT eetsurt rof gnillortnoc retfa niamer stluser ruo taht wohs ew ,2 dna 1 snmuloc nI .atad ruo fo selpmasbus suoirav rof repap setamitse tneicffieoc ruo etupmoc ot ygolodohtem hteBcaM-amaF eht esu ew 3 nmuloc nI .ylevitcepser ,stceffe dexfi dnuf dna dnuf lautum ro seetsurt dnuf lautum edulcni ylno 5 dna 4 snmuloC .3 fo htgnel gal a htiw srorre dradnats tseW-yeweN dna nmuloC .segnahc eetsurt gnidulcxe yb stluser ruo setamitseer 6 nmuloC .seilimaf rieht ni sdnuf net tsael ta evah taht seetsurt deganam ylevitca rof stluser stroper 9 nmuloc dna ,sdnuf ytiuqe-non lla sedulcxe 8 nmuloc ,sdnuf etad tegrat sedulcxe 7 eht ta deretsulc yaw-owt era elbat siht ni srorre dradnatS .stceffe dexfi raey dna elyts dnuf edulcni snoisserger ehT .sdnuf ,%01 ot dnopserroc hcihw ,*** ,** ,* yb detoned era slevel ecnacfiingiS .sesehtnerap ni detroper era dna slevel dnuf dna nalp .ylevitcepser ,slevel %1 dna ,%5 ylnO ylnO edulcxE edulcxE seetsurTFM FMylnO amaF edulcnI edulcnI evitcA ytiuqE etaDtegraT seetsurT tsaeLtahtiW seetsurT hteBcaM EFdnuF EFeetsurT sdnuF sdnuF sdnuF segnahC sdnuF01 ∗∗∗931.0− ∗∗∗621.0− ∗∗∗341.0− ∗∗∗931.0− ∗∗∗971.0− ∗∗∗271.0− ∗∗∗011.0− ∗∗∗941.0− ∗∗∗661.0− dnuFdetailffiA )220.0( )420.0( )120.0( )910.0( )220.0( )120.0( )910.0( )410.0( )810.0( ∗∗∗273.0− ∗∗∗793.0− ∗∗∗073.0− ∗∗∗313.0− ∗∗∗713.0− ∗∗∗523.0− ∗∗∗272.0− ∗∗∗772.0− ∗∗∗513.0− fePwoL )430.0( )630.0( )430.0( )530.0( )340.0( )930.0( )660.0( )120.0( )330.0( ∗940.0− ∗∗∗560.0− ∗340.0− 830.0− ∗∗160.0− ∗940.0− ∗170.0− ∗∗∗830.0− ∗∗∗850.0− frePhgiH )520.0( )420.0( )420.0( )420.0( )130.0( )720.0( )430.0( )410.0( )220.0( ∗∗∗902.0 ∗∗∗661.0 ∗∗∗602.0 ∗∗∗842.0 ∗∗∗162.0 ∗∗∗862.0 ∗∗821.0 ∗∗∗902.0 ∗∗∗332.0 dnuFdetailffiA*frePwoL )050.0( )550.0( )840.0( )440.0( )250.0( )940.0( )350.0( )920.0( )340.0( 800.0 140.0 030.0 120.0− 600.0− 110.0− 410.0 ∗∗∗150.0 800.0 dnuFdetailffiA*knaRhgiH )730.0( )730.0( )430.0( )030.0( )830.0( )430.0( )210.0( )910.0( )920.0( ∗∗∗110.0 ∗∗∗410.0 ∗∗∗110.0 ∗∗∗900.0 ∗∗∗110.0 ∗∗∗110.0 ∗∗∗410.0 ∗∗∗810.0 ∗∗∗110.0 rroCmumixaM )100.0( )100.0( )100.0( )100.0( )100.0( )100.0( )300.0( )100.0( )100.0( ∗∗∗110.0− ∗∗∗110.0− ∗∗∗010.0− ∗∗∗700.0− ∗∗∗010.0− ∗∗∗010.0− ∗∗∗900.0− ∗∗∗800.0− ∗∗∗010.0− )eziSnoitpO(goL )200.0( )200.0( )200.0( )200.0( )200.0( )200.0( )100.0( )100.0( )100.0( ∗∗∗200.0− ∗∗∗200.0− ∗∗∗200.0− ∗∗∗200.0− ∗∗∗200.0− ∗∗∗200.0− ∗∗∗300.0− ∗∗∗300.0− ∗∗∗300.0− snoitpOfo .oN )000.0( )000.0( )000.0( )000.0( )000.0( )000.0( )000.0( )000.0( )000.0( ∗∗∗224.7 ∗∗∗924.6 ∗∗∗609.7 ∗∗∗065.6 ∗∗∗231.8 ∗∗∗749.7 ∗∗∗575.7 ∗∗753.4− ∗∗∗868.8 oitaR .pxE )432.1( )391.1( )322.1( )570.1( )992.1( )952.1( )928.0( )781.2( )040.1( ∗∗∗410.0 ∗∗∗420.0 ∗∗∗410.0 ∗∗∗710.0 ∗∗∗020.0 ∗∗∗710.0 ∗∗210.0 ∗∗600.0 ∗∗∗510.0 revonruT )400.0( )700.0( )400.0( )400.0( )500.0( )400.0( )400.0( )300.0( )400.0( ∗∗∗020.0− ∗∗∗120.0− ∗∗∗910.0− ∗∗∗810.0− ∗∗∗020.0− ∗∗∗810.0− ∗∗∗910.0− ∗∗∗940.0− ∗∗∗120.0− )eziSdnuF(goL )300.0( )200.0( )200.0( )200.0( )200.0( )200.0( )100.0( )400.0( )200.0( 000.0 000.0− 000.0 000.0 000.0 000.0 000.0 100.0− 000.0 egAdnuF )000.0( )000.0( )000.0( )000.0( )000.0( )000.0( )000.0( )300.0( )000.0( 111.0 722.0 320.0 340.0− 632.0− 171.0− 930.0− ∗∗∗997.0− ∗∗272.0− .veD .dtS )591.0( )612.0( )881.0( )902.0( )722.0( )312.0( )462.0( )441.0( )531.0( 722,18 550,17 532,29 861,69 978,27 167,68 777,601 848,601 091,101 snoitavresbO 370.0 190.0 570.0 470.0 390.0 780.0 880.0 381.0 011.0 derauqs-R 18

)delacsnu( seliceD ecnamrofreP yb snoitiddA dnuF :7-A elbaT C dna A slenaP .seliced ecnamrofrep dna noitailffia yb )% a sa( seicneuqerf noitidda dnuf launna naem sezirammus elbat ehT seetsurt sa evres taht seilimaf dnuf yb dereffo era hcihw sdnuf esoht ylno edulcni D dna B slenaP .elpmas lluf eht edulcni ot sunem )detailffianu( detailffia fo rebmun eht fo oitar eht etaluclac ew ,raey yrevE .elpmas ruo ni nalp eno tsael ta rof eht edulcni tey ton od taht sunem )detailffianu( detailffia fo rebmun latot eht ot raey eht gnirud dedda si dnuf eht hcihw elitnecrep rieht no desab sdnuf gnipuorg yb detaerc era seliced ecnamrofreP .raey eht fo gninnigeb eht ta noitpo na sa dnuf eht ssorca egareva neht eW .sraey eerht roirp eht revo esrevinu dnuf PSRC eht ni elyts emas eht fo sdnuf gnoma ecnamrofrep no desab era snaem ni ecnereffid eht fo stset rof slevel ecnacfiingiS .noitailffia dna ecnamrofrep yb seicneuqerf noitidda ’sdnuf ot dnopserroc hcihw ,*** ,** ,* yb detoned era dna B dna A slenaP ni level dnuf eht ta deretsulc era taht srorre dradnats dradnats tseW-yeweN dna stneicffieoc hteBcaM-amaF troper ew ,D dna C slenaP nI .ylevitcepser ,slevel %1 dna ,%5 ,%01 .3 fo htgnel gal a gnisu srorre sdnuF llA :A lenaP sraeY 5 sraeY 3 raeY 1 ecnamrofreP AU-A AU A AU-A AU A AU-A AU A eliceD ∗∗∗666.0 500.0 176.0 ∗∗∗247.0 500.0 747.0 ∗∗∗051.1 900.0 061.1 1 ∗∗∗705.0 600.0 315.0 ∗∗∗495.0 700.0 206.0 ∗∗∗067.0 010.0 077.0 2 ∗∗∗450.1 010.0 460.1 ∗∗∗078.0 900.0 978.0 ∗∗∗561.1 310.0 871.1 3 ∗∗∗328.0 210.0 538.0 ∗∗∗131.1 310.0 441.1 ∗∗∗852.1 610.0 472.1 4 ∗∗∗529.0 610.0 149.0 ∗∗∗941.1 710.0 561.1 ∗∗∗381.1 020.0 302.1 5 ∗∗∗422.1 810.0 242.1 ∗∗∗954.1 220.0 184.1 ∗∗∗403.1 320.0 723.1 6 ∗∗∗526.1 320.0 846.1 ∗∗∗525.1 520.0 055.1 ∗∗∗272.1 220.0 492.1 7 ∗∗∗986.1 130.0 127.1 ∗∗∗573.1 820.0 204.1 ∗∗∗824.1 720.0 554.1 8 ∗∗∗923.2 540.0 473.2 ∗∗∗799.1 240.0 930.2 ∗∗∗827.1 430.0 267.1 9 ∗∗∗252.2 540.0 792.2 ∗∗∗112.2 440.0 552.2 ∗∗∗508.1 830.0 348.1 01 suneM detailffianU dna detailffiA htoB no sdnuF fo elpmasbuS :B lenaP sraeY 5 sraeY 3 raeY 1 ecnamrofreP AU-A AU A AU-A AU A AU-A AU A eliceD ∗∗∗166.0 010.0 176.0 ∗∗∗737.0 010.0 747.0 ∗∗∗141.1 810.0 061.1 1 ∗∗∗505.0 800.0 315.0 ∗∗∗095.0 110.0 206.0 ∗∗∗657.0 410.0 077.0 2 ∗∗∗740.1 710.0 460.1 ∗∗∗468.0 410.0 978.0 ∗∗∗751.1 120.0 871.1 3 ∗∗∗318.0 220.0 538.0 ∗∗∗221.1 220.0 441.1 ∗∗∗842.1 620.0 472.1 4 ∗∗∗019.0 130.0 149.0 ∗∗∗331.1 330.0 561.1 ∗∗∗071.1 330.0 302.1 5 ∗∗∗802.1 430.0 242.1 ∗∗∗444.1 730.0 184.1 ∗∗∗092.1 730.0 723.1 6 ∗∗∗116.1 730.0 846.1 ∗∗∗705.1 340.0 055.1 ∗∗∗852.1 730.0 492.1 7 ∗∗∗776.1 340.0 127.1 ∗∗∗163.1 240.0 204.1 ∗∗∗014.1 540.0 554.1 8 ∗∗∗603.2 860.0 473.2 ∗∗∗089.1 850.0 930.2 ∗∗∗017.1 250.0 267.1 9 ∗∗∗232.2 560.0 792.2 ∗∗∗391.2 260.0 552.2 ∗∗∗597.1 840.0 348.1 01 19

stseT hteBcaM-amaF ,sdnuF llA :C lenaP sraeY 5 sraeY 3 raeY 1 ecnamrofreP AU-A AU A AU-A AU A AU-A AU A eliceD ∗∗846.0 600.0 356.0 ∗∗717.0 600.0 327.0 ∗∗899.0 010.0 700.1 1 ∗∗∗405.0 600.0 015.0 ∗∗∗735.0 800.0 545.0 ∗∗∗017.0 110.0 127.0 2 ∗∗∗420.1 110.0 530.1 ∗∗∗348.0 010.0 358.0 ∗∗∗961.1 510.0 481.1 3 ∗∗∗997.0 310.0 218.0 ∗∗∗901.1 510.0 421.1 ∗∗∗482.1 910.0 303.1 4 ∗∗∗698.0 810.0 419.0 ∗∗∗511.1 910.0 431.1 ∗∗∗551.1 220.0 771.1 5 ∗∗∗691.1 120.0 712.1 ∗∗∗434.1 520.0 954.1 ∗∗∗282.1 520.0 703.1 6 ∗∗∗475.1 520.0 006.1 ∗∗∗294.1 720.0 025.1 ∗∗∗442.1 520.0 962.1 7 ∗∗∗966.1 530.0 407.1 ∗∗∗653.1 130.0 783.1 ∗∗∗934.1 030.0 964.1 8 ∗∗∗303.2 150.0 453.2 ∗∗∗459.1 740.0 100.2 ∗∗∗117.1 930.0 947.1 9 ∗∗∗902.2 050.0 952.2 ∗∗∗341.2 940.0 391.2 ∗∗∗008.1 340.0 348.1 01 stseT hteBcaM-amaF ,suneM detailffianU dna detailffiA htoB no sdnuF fo elpmasbuS :D lenaP sraeY 5 sraeY 3 raeY 1 ecnamrofreP AU-A AU A AU-A AU A AU-A AU A eliceD ∗∗446.0 900.0 356.0 ∗∗317.0 900.0 327.0 ∗∗299.0 610.0 700.1 1 ∗∗∗305.0 800.0 015.0 ∗∗∗435.0 110.0 545.0 ∗∗∗807.0 410.0 127.0 2 ∗∗∗810.1 710.0 530.1 ∗∗∗938.0 510.0 358.0 ∗∗∗261.1 220.0 481.1 3 ∗∗∗097.0 220.0 218.0 ∗∗∗201.1 220.0 421.1 ∗∗∗672.1 720.0 303.1 4 ∗∗∗288.0 230.0 419.0 ∗∗∗101.1 330.0 431.1 ∗∗∗541.1 230.0 771.1 5 ∗∗∗481.1 330.0 712.1 ∗∗∗124.1 730.0 954.1 ∗∗∗072.1 730.0 703.1 6 ∗∗∗265.1 730.0 006.1 ∗∗∗774.1 240.0 025.1 ∗∗∗332.1 630.0 962.1 7 ∗∗∗166.1 340.0 407.1 ∗∗∗543.1 240.0 783.1 ∗∗∗524.1 440.0 964.1 8 ∗∗∗682.2 860.0 453.2 ∗∗∗349.1 950.0 100.2 ∗∗∗796.1 350.0 947.1 9 ∗∗∗691.2 360.0 952.2 ∗∗∗331.2 950.0 391.2 ∗∗∗597.1 840.0 348.1 01 20

Table A-8: Linear Probability Model for Affiliated Fund Additions. The table reports the coefficient estimates of the following model for affiliated fund additions: AFADD = p,f,t β +β ×Perf +Z(cid:48) γ+(cid:15) , where AFADD is an indicator variable equal to one if mutual 0 1 p,f,t−1 p,f,t−1 p,f,t p,f,t−1 fund f added to the plan p during year t is affiliated with the management company acting as the plan’s trustee and zero otherwise. Perf is the performance percentile of mutual fund f over the previous one, p,f,t−1 three, or five years and is included as a percentage. The overall performance rank of each fund depends on the performance of the fund relative to other funds in CRSP in the same style. The other lagged control variables Z include the number of options, the expense ratio, fund turnover, the natural logarithm of the fund’s size, fund age, the standard deviation of the fund’s return (all measured during the previous year), and unreported indicator variables for specific fund styles, and year and trustee fixed effects. Standard errors are two-way clustered at plan and fund levels and are reported in parentheses. Significance levels are denoted by *, **, ***, which correspond to 10%, 5%, and 1% levels, respectively. 1 Year 3 Years 5 Years Perf (1 YR) −0.140∗∗∗ (0.026) Perf (3 YR) −0.201∗∗∗ (0.036) Perf (5 YR) −0.228∗∗∗ (0.041) No. of Options −0.001 −0.001∗ −0.001∗ (0.000) (0.000) (0.000) Log(Plan Assets) −0.017∗∗∗ −0.017∗∗∗ −0.016∗∗∗ (0.003) (0.003) (0.003) Exp. Ratio −0.134∗∗∗ −0.131∗∗∗ −0.127∗∗∗ (0.023) (0.022) (0.022) Turnover −0.001 −0.002 −0.002 (0.010) (0.010) (0.010) Log(Fund Size) −0.007 −0.005 −0.001 (0.007) (0.007) (0.007) Fund Age −0.002∗∗∗ −0.002∗∗∗ −0.002∗∗∗ (0.001) (0.001) (0.001) Std. Dev. 0.014 0.041 0.044 (0.033) (0.032) (0.032) Observations 20,925 20,925 20,925 R-squared 0.723 0.725 0.726 21

Table A-9: Differences in the Abnormal Returns of Affiliated and Unaffiliated Funds. Panels A, B, and C of the table report the difference in abnormal returns (α) across the affiliated and unaffiliated portfolios using the Fama-French-Carhart four-factor model (FFM), the Fama and French (1993) model, and the CAPM model, respectively, over our complete sample period using monthly fund return data. At the end of each calendar year, we form equal-weighted portfolios of trustee and non-trustee domestic equity funds separately based on whether the funds were kept on, deleted from, or added to the 401(k) menu (“No Changes,” “Deletions,”and “Additions”) during the calendar year. This creates six portfolios. We then further subdivide these six groups based on past performance. In particular, “All Funds,” refers to the overall six portfolios and “Lowest Quintile,” (“Lowest Decile”) refers to a sub-portfolio in each group that contains only those funds that also rank in the lowest performance quintile (decile) relative to funds in their style in CRSP during the prior three years. The performance measures are reported in % per month. Robust standard errors are reported in parentheses. Significance levels are denoted by *, **, ***, which correspond to 10%, 5%, and 1% levels, respectively. Panel A: Carhart Alpha Differences No Changes Deletions Additions Lowest Decile −0.25∗∗ −0.13 −0.10 (0.12) (0.16) (0.22) Lowest Quintile −0.10 −0.06 −0.09 (0.07) (0.12) (0.13) All Funds 0.06∗∗ 0.02 0.06∗ (0.02) (0.04) (0.03) Panel B: Fama-French Alpha Differences No Changes Deletions Additions Lowest Decile −0.26∗∗ −0.13 −0.15 (0.13) (0.17) (0.23) Lowest Quintile −0.10 −0.06 −0.09 (0.07) (0.13) (0.14) All Funds 0.06∗∗ 0.02 0.06∗ (0.02) (0.03) (0.03) Panel C: CAPM Alpha Differences No Changes Deletions Additions Lowest Decile −0.32∗ −0.25 −0.37 (0.16) (0.21) (0.32) Lowest Quintile −0.09 −0.19 −0.15 (0.07) (0.18) (0.21) All Funds 0.03 −0.01 0.02 (0.03) (0.04) (0.06) 22

Panel A: Interaction Effects for Linear Performance (Table A-2, column 1) Panel B: Interaction Effects for Linear Performance (Table A-2, column 2) Panel C: Interaction Effects for Negative Non-DC flows (Table A-2, column 2) Figure A-1: Marginal effects between the indicator variable for affiliated funds and the performance ranks: Linear performance model. The graphs display the marginal effects and corresponding z-statistics by observation on the interaction variables between the affiliation dummy and the below- and above-median performance ranks in Table A-2, estimated using Ai and Norton (2003). 23

Panel D: Interaction Effects for Below-Median Performance (Table A-2, column 3) Panel E: Interaction Effects for Above-Median Performance (Table A-2, column 3) Figure A-2: Marginal effects between the indicator variable for affiliated funds and the performance ranks: 2-Segment model (Specification 1). The graphs display the marginal effects and corresponding z-statistics by observation on the interaction variables between the affiliation dummy and the below- and above-median performance ranks in Table A-2, estimated using Ai and Norton (2003). 24

Panel F: Interaction Effects for Below-Median Performance (Table A-2, column 4) Panel G: Interaction Effects for Above-Median Performance (Table A-2, column 4) Panel H: Interaction Effects for Negative Non-DC flows (Table A-2, column 4) Figure A-3: Marginal effects between the indicator variable for affiliated funds and the performance ranks: 2-Segment model (Specification 2). The graphs display the marginal effects and corresponding z-statistics by observation on the interaction variables between the affiliation dummy and the below- and above-median performance ranks in Table A-2, estimated using Ai and Norton (2003). 25

Panel A: Overall Sample Panel B: Subsample of Funds on Both Affiliated and Unaffiliated Menus Figure A-4: Fund additions by affiliation. The figure depicts mean annual fund addition frequencies by affiliation and performance deciles. Panel A includes the full sample. Panel B includes only those funds that are offered by fund families that serve as trustees for at least one plan in our sample. Every year, we calculate the ratio of the number of affiliated (unaffiliated) menus to which the fund is added during the year to the total number of affiliated (unaffiliated) menus that do not yet include the fund as an option. Performance deciles are created by grouping funds based on their percentile performance among funds of the same style in the CRSP fund universe over the prior three years. We then average across the funds’ addition frequencies by performance and affiliation. 26

Figure A-5: The distribution of mutual fund additions by performance decile and fund affiliation. The figure shows the distribution of the funds that are added to a 401(k) menu at some point during our sample period by performance decile and affiliation. The dark line shows the fraction of affiliated funds in the various performance deciles, while the grey line provides the corresponding values for unaffiliated funds. Performance deciles are created from percentile performance ranks. These are calculated using overall rankings, in which fund performance is ranked relative to all other mutual funds in CRSP with the same style over the prior 36 months. 27

Panel A: Affiliated Funds Panel B: Unaffiliated Funds (Overall Ranking) (Overall Ranking) Panel C: Affiliated Funds Panel D: Unaffiliated Funds (Plan Ranking) (Plan Ranking) Figure A-6: New money growth of lower performance quintiles for affiliate and unaffiliated Funds. The figure displays the distribution of fund flows to poorly performing mutual funds on the menu by affiliation. Fund flows, or the growth rate of new money NMG of fund f held in 401(k) plan p at time t is defined by NMG = [V − p,f,t p,f,t p,f,t V (1+R )]/[V (1+R )]. The numerator captures the dollar change in the value p,f,t−1 f,t p,f,t−1 f,t of participants’ investments (V ) in fund f in plan p in year t after adjusting for the price p,f,t appreciation R during the year. The denominator is defined as the projected value of the f,t lagged plan position in the fund without any new flow of money. If an investment option is deleted from a plan menu, then NMG equals exactly -100%. In Panels A and B, the distributions describe fund flows to those affiliated and unaffiliated funds, respectively, that fall into the worst performance decile of the universe of mutual funds in the same style. Panels C and D depict the distributions of the corresponding flows using performance rankings based on only those mutual funds that are offered on the same 401(k) menu. 28