NEIGHBORHOOD EFFECTS IN SAVINGS POLICY: EVIDENCE FROM THE SAVER S CREDIT

Transcription

1 NEIGHBORHOOD EFFECTS IN SAVINGS POLICY: EVIDENCE FROM THE SAVER S CREDIT Raj Chetty, Harvard University and NBER John N. Friedman, Harvard University and NBER April 2013 Abstract A growing literature suggests that peers may be an important determinant of savings behavior. We do not know, however, whether these peer effects lead to aggregate savings effects at a macroeconomic level, nor do we understand how peers affect responses to government policies on savings. To answer these questions, we analyze responses to the Saver s Credit using selected records from U.S. tax returns ( ). Using non-parametric identification techniques, we see that responses to the Saver s Credit vary tremendously both across regions of the country and within states. High-response areas tend to be those in which underlying IRA take-up is high, suggesting an information agglomeration mechanism. We then study the behavior of households who move from one city to another to isolate the causal effect of neighborhood characteristics as separate from persistent sorting of individuals across cities. The results point to an important role of information and peer effects in the take-up of savings incentives. We thank Sarah Griffis, Shelby Lin, and Heather Sarsons for their excellent research assistance. This research was supported by the U.S. Social Security Administration through grant #5RRC to the National Bureau of Economic Research as part of the SSA Retirement Research Consortium. The findings and conclusions expressed are solely those of the authors and do not represent the views of SSA, any agency of the Federal Government, or the NBER.

2 I Introduction Many proposals for social security reform in the United States include an element of individual choice, often over not only the allocation of savings across accounts but also the amount saved. Savings policy in the United States also attempts to affect individual decisions in private savings accounts, for instance by allowing for tax-deferred accounts and subsidizing contributions to them. Understanding the determinants of individual savings decisions is therefore a critical step in evaluating the welfare consequences of both current savings policy and potential reforms. Traditional economic models assume that two factors determine an individual s response to savings policies: individual preferences and the policy parameters. But a growing economic literature suggests that such responses may depend on context in important ways. Madrian and Shea (2001) and Choi et al. (2002), for instance, illustrate that default rules in 401(k) plans have strong and lasting effects on asset allocation. Duflo and Saez (2002, 2003), use a randomized intervention to demonstrate peer effects in savings decisions within a large university. Duflo et al. (2006, 2007) additionally use field experiments at H&R Block to suggest that information about a savings program, in addition to the incentives included in such a program, determine take-up and impact. Peer effects can be especially powerful because they can generate a social multiplier that amplifies the effects of policies within a group (Glaeser et al. 2002). In these cases, it is important to distinguish between peer effects in individual types, which affect behavior only at a micro-level, and peer effects in actions that combine to affect aggregate behavior. This paper studies these issues in the context of the Saver s Credit, a federal tax credit designed to encourage low-income families to contribute to tax-deferred accounts such as 401(k)s or IRAs. Because the Saver s Credit is a federal program, there is no cross-sectional variation in the incentives to which individuals are exposed. In addition, the incentives to contribute have not changed since 2002 when the credit was designed. Instead, we follow Chetty et al. (2013) and look for spatial differences in knowledge about the program. We do so using selected data from U.S. population tax records spanning , which include over 57 million unique individuals who are eligible for the Saver s Credit and 542 million annual observations on their income and savings behavior. Our empirical work proceeds in two steps. We first develop a proxy for local information about the exact incentives generated by the Saver s Credit. With the right survey tool, one might be able to directly measure the beliefs of residents about the incentives generated by the Saver s Credit, or whether locals have any knowledge of the program at all. Lacking such a survey, we instead proxy 1

3 for the information locally available using the extent to which individuals adjust their planned IRA contributions to take maximal advantage of the discontinuities in the schedule for the Saver s Credit. We show that these responses differ by an order of magnitude across 3-digit ZIP codes in the U.S., with substantial variation at both the regional and local level. For instance, response to the Saver s Credit is more than twice as large in Fort Myers, FL than in Tampa, FL. We also show that response is highest in areas with higher latent contribution rates to the tax-deferred accounts in the relevant income range, suggesting a role for information agglomeration. We then study the causal effect of location on savings decisions by analyzing the behavior of households who move from one city to another. We find that people who move from a low-response to a high-response neighborhood significantly increase the likelihood of responding to the Saver s Credit. The key identification assumption on which this conclusion relies is that households do not selectively move into high-response locations when they already intend to increase their own response to the Saver s Credit. We show non-parametrically that this is unlikely to bias our results. Together, these results suggest that information plays a crucial role in taxpayers responses to the Saver s Credit. There are two sides to this fact. More positively, individuals seem quite sensitive to the local information environment in their response to this tax incentive. Since the Saver s Credit intends to offset other distortions to encourage low-income households to save, the government may be able to significantly increase the effectiveness of this credit at lower cost by reducing the subsidy while focusing on information and knowledge. More negatively, the data suggest that there are many households that locate at points on the budget frontier which are likely to be dominated. Chetty et al. (2013) show that information is important even for programs such as the EITC which are long-standing, pay out a large credit as a fraction of recipient incomes, and affect many individuals. Therefore it is not surprising to find a larger effect of information on a program that is relatively new, pays out credits worth a much smaller fraction of recipients incomes, and affects only a small share of the population (low-income families making contributions to tax-deferred accounts). The remainder of the paper is organized as follows. Section II describes the institutions that we study and the U.S. tax data within which we conduct the analysis. Section III develops our proxy for local information. Section IV tests for the causal impact of place by examining a subsample of individuals who move between neighborhoods. We conclude in Section V by discussing policy implications and direction for future research. 2

4 II Institutions and Data II.A Saver s Credit Program Structure The Saver s Credit is a non-refundable tax credit on the first $2,000 contributed to qualified retirement plans or eligible deferred compensation plans (401(k), 403(b), SIMPLE IRA, SEP, etc.), or individual retirement arrangements (IRAs). The Saver s Credit was first implemented in 2002 as a temporary incentive for retirement savings for low- and middle-income taxpayers and was permanently extended in The percent rate of the Saver s Credit ranges from 10% to 50% for eligible filers based on adjusted gross income (AGI), and is zero above a threshold that varies with filing status. Table 1 displays the full eligibility and credit rate details for the Saver s Credit from 2002 to 2006 (as in Duflo et al. 2007). Column 3 displays Saver s Credit rates for joint filers the credit rate is 50% for adjusted gross incomes below $30,000, 20% for incomes between $30,000 and $32,500, and 10% in the lowest eligible range up to $50,000. Similarly, Columns 4 and 5 display Saver s Credit cutoffs for head of household and single filers, respectively. These thresholds were held constant from their inception in 2002 through 2006; the thresholds have then increased in subsequent years. In years following 2006, the first threshold for married filing jointly households increased to $31K, $32K, $33K, $33.5K, $34K, and finally $34.5K in tax year As in , in all subsequent years thresholds for those filing as heads of households and single were 75% and 50% of the married filing jointly threshold, respectively. The Saver s Credit, particularly the 50% credit rate below the first cutoff, offers strong incentives to make retirement contributions. A married household filing jointly with AGI below $30,000 in 2002 that contributes $2,000 to their retirement savings account would receive a $1,000 tax credit. As a result, the out-of-pocket cost of that contribution is only $1,000; the Saver s Credit in this range is effectively a 100% matching incentive on the first $1,000 contributed to retirement savings. In the two higher brackets, as the credit rate falls to 20% and 10%, the implicit match rate falls to 25%, and 11%, respectively. The credits and the equivalent match rates are shown in Columns 1 and 2 of Table 1. The Saver s Credit is a non-refundable credit, so tax-payers may only benefit to the extent that they have a positive tax liability to offset. Many low income households already qualify for substantial tax credits such as the EITC or the Additional Child Tax Credit, which are refundable, as well as others (such as the Child Tax Credit) which are not. If these other credits already offset 3

5 all positive tax liability, then households become effectively ineligible for the Saver s Credit. 1 II.B Sample and Variable Definitions We use data from the universe of U.S. administrative tax returns from For more details of the construction of our sample dataset from the raw population tax data, see Appendix A of Chetty et al. (2013). Variable Definitions. Income is defined at the household level because the Saver s Credit is based on household income, but we conduct our analysis using individual-level panel dataset to account for potential changes in marital status. We define adjusted gross income (AGI), the measure used to calculate the Saver s Credit, as the sum of wage earnings on third-party W-2 forms filed by employers and any net self-employment earnings reported on 1040 tax forms. ZIP code in year t is defined as the ZIP code from which the individual filed a tax return in year t or the ZIP code to which a W-2 was mailed if he did not file a tax return. Our neighborhood analysis is based on the first three digits of ZIP code (ZIP-3). Retirement savings contributions are recorded on 1040 tax forms (line 32) by filers reporting eligible deductible contributions to IRAs, SEP or SIMPLE retirement accounts. For expository simplicity, throughout our analysis we will refer to an indicator for non-zero retirement contributions as an indicator of IRA contributions. Core Sample. Our core analysis sample includes all annual observations between 1996 and 2009 for individuals who in at least one year between 2002 and 2009 satisfy two conditions: they must (1) file a tax return as a primary or secondary file (in the case of married filers) and (2) have earnings below $55,000 (in 2010 dollars) with positive tax liability. These restrictions limit the full sample to individuals potentially eligible for the Saver s Credit at least once between 2002 and Adjusted gross income in years with no reported earnings activity is reported as 0. This core sample includes over 2.28 billion observations and over 163 million unique individuals. Our empirical analysis utilizes two different subsamples of this sample. Cross-Sectional Analysis Sample. Our first research design compares IRA contribution rates across the earnings distribution in repeated cross-sections. We limit the core sample described above to annual observations in which the individual files a tax return and has income in the Saver s Credit eligible range for the relevant filing status. In our analysis we normalize our income measure for head of household and single filers to align with the Saver s Credit rate structure of married 1 See IRS Publication 590 for more details on the Saver s Credit structure and eligibility. 2 We have data from W-2 forms available starting in

6 joint filers. Movers Sample. Our second research design tracks individuals over time as they move across neighborhoods. We limit the core sample to annual observations in which an individual files a tax return and has income in Saver s Credit eligible range for the relevant filing status. We then further restrict the sample to individuals who move 3-digit ZIP codes (ZIP-3s) in some year between 2002 and If individuals move more than once during this time span, we include only the first move. Descriptive Statistics. Table 2 presents summary statistics for our cross-sectional analysis sample. Mean total earnings (AGI) are $20,081. Across the sample, 2.20 percent of individuals make positive IRA contributions reported on Form 1040s and almost 70 percent of tax returns are filed by a professional tax preparer. The bottom panel of Table 2 presents summary statistics for neighborhood (ZIP-3) characteristics in 2002, the first year of the Saver s Credit program, weighted by the number of observations per ZIP-3 in our sample. Saver s Credit bunching, formally defined in Section 3, is the ratio of IRA participation for incomes directly below the 100% match rate cutoff relative to a base level of average neighborhood IRA participation in a lower income range. Neighborhood education levels are for adults age 25 and older from the 2000 Census. III III.A Neighborhood Variation in Response to the Saver s Credit Aggregate Response to the Saver s Credit To investigate responses to the Saver s Credit, we begin by studying the influence of the nonlinear subsidy schedule on contribution rates (as in Duflo et al. 2007). For instance, as described in Section 2.1, couples filing jointly in 2002 with adjusted gross income (AGI) of $30,000 or less qualified for a credit equal to 50% of any contributions to tax-deferred accounts up to $2,000. But if individuals have AGI of $30,001, the credit falls to just 20% of contributions. But because IRA contributions are deductible from AGI, this non-linear credit generates a notch in the tax schedule for households making contributions to tax-deferred plans. For instance, a household with AGI of $30,001 planning to make a $1,000 to an IRA would lower their tax liability by $300 by contributing $1 extra to the IRA. Theory predicts that such a notch will generate bunching in distribution of AGI for those contributing to tax-deferred plans. Households may face uncertain income streams or may be unaware of the exact threshold at which the credit rate jumps up. In many other cases, such noise generates diffuse heaping around the tax incentive rather than sharp bunching (as in Chetty et al and Chetty et al. 2013). In this case, however, individuals can (and often do) make IRA or 5

7 other tax-deductible contributions after the completion of a given tax year. By law, individuals have until April 15 of the following tax year (when tax returns are due) to make contributions out of a given tax year s income. In fact, many households may choose the size of their contribution at the time of tax preparation. Many professional tax preparation firms offer filers the option to open an IRA account and contribute as they are filing their return; for instance, H&R Block offers customers an express IRA, which has been studied by Duflo et al. (2007). As a result, we expect to find large spikes in the fraction of taxpayers contributing to tax-deferred accounts, and especially IRAs, just below the match rate thresholds, with matching holes of lower contribution rates just above the thresholds. We demonstrate this bunching in Figure 1 by plotting the fraction of taxpayers in each $500 AGI bin making a contribution to an IRA. In order to combine data from across years, we normalize AGI relative to the 100% match threshold. In order to combine taxpayers from all different filing statuses on the same graph, we multiply incomes from individuals and heads of households by 2 and 4/3, respectively, in order to align the thresholds. The responses to the Saver s Credit are evident in Figure 1, which shows large and visually evident spikes in the contribution rates just below the 100% and 25% match thresholds. For taxpayers just below the 100% match threshold, the contribution rate is nearly 3.5% on average, which is nearly double the contribution rate of 2% for those with AGI just above the threshold. There is a smaller but still visually evident spike just below the 25% threshold. Because the 100% match threshold is substantially larger in nearly all subsamples, we focus our analysis on this notch; our results still remain, with less precision and magnitude, when repeated using the 25% threshold notch. The sharpness of the spikes in Figure 1 make it unlikely that they are driven by anything other than the tax incentives generated by the Saver s Credit. Nevertheless, in order to confirm this supposition, we plot the fraction contributing across AGI bins in as a placebo series. Because the Saver s Credit was not implemented until tax year 2002, there should be no spikes of any kind in this series. As predicted, the series is entirely smooth. In the analysis that follows, we use the size of the bunch just below the 100% threshold as a statistic measuring the intensity of response to the Saver s Credit. Specifically, we calculate the ratio of the fraction contributing in the spike to the average fraction contributing between $2,000 and $4,000 below the spike. Denoting r i as the fraction of taxpayers contributing to an IRA in 6

8 $500 AGI bin i, formally we measure the size of the bunch as b = 1 5 r i= 4000 r i In Figure 1, b = 0.87 for taxpayers affected by the Saver s Credit and just b = 0.12 for the placebo series. 1. III.B Spatial Heterogeneity in Response to the Saver s Credit The large aggregate response to the Saver s Credit in Figure 1 masks substantial heterogeneity. We analyze responses to the Saver s Credit by neighborhood, calculating the size of the bunch b in each 3-digit ZIP code (ZIP-3). To give a sense of the variation, Figure 2 replicates Figure 1 for taxpayers from ZIP-3s from the highest and lowest deciles of bunching. There is a dramatic difference in the bunching. In ZIP-3s in the top decile, the size of the bunch is b = 1.87, in contrast with just b = 0.30 in the bottom decile. Note that contribution rates are higher in all AGI bins in top decile areas, not just those near the bunch; this suggests that the magnitude of response to the Saver s Credit depends, in large part, on the number of taxpayers who are already contributing to tax-deferred accounts. We return to study this pattern more thoroughly in the next section. The three panels in Figure 3 display the geographic distribution of bunching in response to the Saver s Credit by ZIP-3 across the United States in 2002, 2005, and Darker areas represent more response; the definitions of the color-bins are held constant across all three maps. There are clear regional patterns of response. For instance, taxpayers appear to be most responsive to the Saver s Credit in the Upper Midwest in states such as Wisconsin, Minnesota, and the Dakotas. Responses are smallest in states in the Southern Atlantic seaboard, including Delaware, Maryland, and South Carolina. There is also a great deal of highly local variation between otherwise similar cities. For instance, Tampa is just 125 miles north of Fort Myers on the Gulf Coast of Florida, and yet in 2002 response to the Saver s Credit is more than twice as large in Fort Myers at b = 1.1, compared to b = 0.52 in Tampa. There are also clear patterns across years. Response in 2002 is most highly concentrated in the Upper Midwest, but the higher levels of response then spreads somewhat to neighboring regions such as the Great Lakes and Central Plains states in In 2008, the magnitude of responses shrinks almost everywhere, as taxpayers become somewhat less sensitive in the recession. This reduction may be somewhat driven by the increased liquidity constraints in the recession which discouraged many individuals from saving in illiquid account such as those eligible for the Saver s Credit. 7

9 III.C Cross-Sectional Correlations and Information As we saw in Figure 2, neighborhoods which exhibited the largest response to the Saver s Credit also had the highest counterfactual contribution rates to tax-deferred accounts. Figure 4 explores this more thoroughly. In Figure 4a, we plot the contribution rate at the top of the spike (y-axis) against the contribution rate in the [ 4000, 2000] control range, which we label the base contribution rate. In order to construct this binned-scatterplot (as well as the ones that follow), we group all ZIP-3s into 20 equal-sized bins, weighted by population and ordered on the x-axis variable. We then plot the mean of the x-axis variable (base contribution rate) against the mean of the y-axis variable (contribution rate in the spike). The reported coefficient and standard error come from a weighted OLS regression on the microdata. Neighborhoods with higher base contribution rates demonstrate higher contribution rates at the peak of the bunch below the cutoff. A linear relationship would lead to a constant b across the distribution of base contribution rates; but the relationship is instead convex, and so b increases substantially as the base contribution rate in a neighborhood increases. Figure 4b shows this relationship directly, plotting our measure of bunching b against the base contribution rate. Neighborhoods with a base contribution rate of 5%, near the top of the national distribution, exhibit nearly twice the response to the Saver s Credit, as measured by b, as compared with neighborhoods with base contribution rate of 1%. Table 3 presents regressions that formalize the relationships in Figure 4. The first column regresses b on the neighborhood base contribution rate, corresponding exactly to Figure 4b. For each one percentage point increase in the base contribution rate, the response to the Saver s Credit b increases by This relationship is highly significant, with a t-statistic over 6. Adding demographic controls in Column 2 reduces the magnitude of this coefficient by about 40%, but the relationship is still highly statistically significant. This relationship mirrors that found in Chetty et al. (2013). That paper finds that geographic heterogeneity in response to the EITC is driven in large part by information, which is in turn driven by information agglomeration from exposure to others who share similar tax incentives and from whom one may learn about the program. The correlation between the response to the Saver s Credit and the base contribution rate suggests a similar informational interpretation of these findings. In order to test further the possibility that information drives response to the Saver s Credit, we examine the relationship between intensity of response and other variables that may proxy for 8

10 local information. One natural variable is the fraction of taxpayers who file using professional tax preparers. Tax preparers may provide a substantial amount of information to taxpayers, especially about tax credit such as the Saver s Credit that is not as well known as other programs for lowincome individuals (such as the EITC). Figure 5a plots this relationship. A 10 percentage point increase in the fraction using tax preparers implies an increase of 0.06 in Saver s Credit bunching, with a t-statistic of 3. A second natural variable is education. Well-educated individuals may be more able to understand complicated aspects of the tax system, including the effective match rate implied by the Saver s Credit. Figure 5b plots the relationship between bunching and the fraction of individuals within a ZIP-3 with at least a high school degree. A 10 percentage point increase in the high school graduation rate increases Saver s Credit bunching by 0.09, with a t-statistic of more than 4. Table 3 explores these correlations further. Columns 3 and 4 correspond exactly to the plots in Figure 5a and 5b. Column 5 includes all three core explanatory variables together in a single regression. Controlling for the other effects actually increases the magnitudes of the effects of the local professional tax preparation rate and the relationship with high school graduation rate. Column 6 includes the same demographic controls as in Column 2; the magnitude of the coefficients fall somewhat, but all remain statistically significant. The results in Figures 4 and 5 and Table 3 suggest that information may drive the local heterogeneity in the response to the Saver s Credit. We now turn to an examination of the behavior of movers to rule out fixed sorting of individuals across neighborhoods as an alternative explanation of these findings. IV Movers and the Saver s Credit We now evaluate whether the differences in response to the Saver s Credit across neighborhoods are driven by differences in people or differences in places. The ideal experiment would randomly assign eligible households to different locations across the country. Because of the random assignment of place, there could not be any correlation between unobservable household characteristics and Saver s Credit response; any differences in behavior across places would therefore be the result of the causal impact of place. We operationalize this test in our observational data by examining the behavior of households that move from one ZIP-3 to another. In order for this test to be valid, we require that individuals do not select into new cities when moving based on their knowledge of or propensity to respond 9

11 to the Saver s Credit. We believe this identification assumption is reasonable, as such matters are not likely to be high on the list of characteristics driving location choice. Nevertheless, we present non-parametric evidence below to substantiate this assumption. We implement this test within our sample of Saver s-credit-eligible movers, as defined in Section 2.2, which include all individuals in our core sample who move at least once across ZIP-3s at some point between 2002 and This sample includes 23.6 million unique individuals and 288 million observations spanning 1999 to We define the degree of bunching for prior residents as the amount of response b in each ZIP-3 the year before the move. We then divide the ZIP-3-by-year cells into deciles based on the Saver s Credit response of prior residents by splitting the household-level observations in the movers sample into ten equal-sized groups by ZIP-3. We begin by studying households who begin in 5th-decile ZIP-3s and move either up to the highest-bunching decile of ZIP-3s or down to the lowest-bunching decile ZIP-3s. Figure 6 plots the fraction of households in each AGI bin around the threshold in each of these two groups, both before and after the move. Figure 6a, which plots the contribution rates in the year before the move (t = 1), shows broadly similar patterns, both at and away from the threshold, for the two different groups. Formally, we measure b = 0.65 for those who move to high-bunching ZIP-3s and b = 0.59 for those moving to low-bunching ZIP-3s. This figure suggests that the identification assumption described above is satisfied; conditional on pre-move ZIP-3 decile, individuals do not appear to select into destination ZIP-3 based on propensity to respond to the Saver s Credit. Figure 6b then plots contribution rates after the move (t = 1). Contribution rates remain similar, especially a distance above the thresholds, where there are only small incentives to contribute about which one might become aware. The spike in contribution rates is much higher for those who move to high-knowledge neighborhoods (b = 0.60) than for those who move to low-knowledge neighborhoods (b = 0.41). In addition, we see that contribution rates are also higher to the left of the threshold. These individuals have AGI strictly below the threshold, and so we do not expect bunching; but these individuals are still exposed to a neighborhood with more knowledge of the very high match rates for which they are now eligible. Figure 7 presents the same data in an event-study format. For households moving from the 5thdecile to either a top- or bottom-decile neighborhood, Figure 7 plots bunching b in each event year, relative to that measured for households who move from a 5th-decile neighborhood to another 5thdecile neighborhood. Before the move, average bunching levels are very similar across groups. But bunching then increases sharply for households moving to top-decile neighborhoods, and decreases 10

12 sharply for those moving to bottom-decile neighborhoods. We quantify the magnitude of these effects in Figure 8, where we use not just movers to extreme top- or bottom-decile neighborhoods but instead make use of all movers in our sample. In order to control for starting neighborhood, we reweight movers to each destination decile to look similar to movers to the 5th decile following the procedure developed by DiNardo, Fortin and Lemieux (1996). Figure 8a presents a binned scatterplot of the relationship between bunching among movers (y-axis) and neighborhood bunching among prior residents (x-axis). We estimate that bunching among new arrivals increases by 59.7% of the bunching among prior residents. This effect is highly statistically significant. As a placebo test, we also run the same regression but instead examining the behavior of movers in the year before the move. As suggested by the Figures 6a and 7, we find no statistically significant relationship between bunching at t = 1 and the characteristics of the neighborhood to which individuals will move. More importantly, the coefficient in Figure 8a is statistically different from the placebo coefficient in Figure 8b. We conclude that much of the differences in response to the Saver s Credit across neighborhoods reflects the causal effect of place, rather than the systematic sorting of individuals. In their paper on the effect of information on intensive-margin responses to the EITC, Chetty et al. (2013) examine asymmetry among movers as an additional test for the effects of information. Specifically, they show that individuals who move to higher-information neighborhoods respond more, since they learn more about the details of the tax code. In contrast, movers to lowerinformation neighborhoods do not respond less, since they do not forget the information they already learned from their prior neighbors. Unfortunately, we do not possess the power in our mover s sample to conduct a similar test in this setting. Crucially, the fraction of eligible taxpayers contributing to IRAs to claim the Saver s Credit rarely exceeds 5%; in contrast, roughly 75% of eligible tax payers claim the EITC. V Conclusion This analysis shows that information mediates the impact of the Saver s Credit on the savings decisions of low-income families. Despite featuring tax parameters that do not vary across the country, some ZIP-3s exhibit an order-of-magnitude larger response to the Saver s Credit than others. These high-response places tend to be locations in which households are more likely to contribute to tax-deferred accounts, which suggests that the Saver s Credit may exacerbate existing regional and local differences in savings rates. In areas in which individuals are already contributing, 11

13 the Saver s Credit increases saving further; but in areas where few people contribute, the policy has little effect. There are a number of directions for future research. Recent research suggests that incentives for savings in Denmark generate very little additional savings for two reasons. First, many savers do not respond at all to the incentives, perhaps because they are unaware of the incentives as in this analysis. Second, others respond to savings incentives by shifting savings between accounts rather than by increasing total savings or cutting consumption (Chetty et al. 2012). It would be interesting to investigate this issue explicitly in the context of the Saver s Credit. Low-income households who are the focus of the Saver s Credit may have more trouble understanding subtle or complex tax incentives than more highly educated households; professional tax preparers may mitigate this to some extent, but only if they focus on savings incentives. In contrast, policies that mandate savings or default taxpayers into savings have a larger impact on total savings. For instance, the government could offer low-income households the option of receiving some portion of their EITC refund as a contribution to a tax-deferred savings account. These findings also have relevance for any future reforms of social security that place greater emphasis on individual accounts, and it would be interesting to investigate individual knowledge of the social security system and the effects of this knowledge (or lack thereof) on behavior. For instance, Liebman and Luttmer (2011) show that informing individuals of the link between benefits and taxes paid can increase labor supply among those nearing retirement. Further work on the role of information in the choice to claim social security before the normal retirement age may yield additional insights. 12

14 References Chetty, Raj, John N. Friedman, and Emmanuel Saez Using Differences in Knowledge Across Neighborhoods to Uncover the Impacts of the EITC on Earnings. American Economic Review, (forthcoming). Chetty, Raj, John N. Friedman, Soren Leth-Petersen, Torben Nielsen, and Tore Olsen Active vs. Passive Decisions and Crowdout in Retirement Savings Accounts: Evidence from Denmark. NBER Working Paper No Chetty, Raj, John N. Friedman, Tore Olsen, and Luigi Pistaferri Adjustment Costs, Firm Responses, and Micro vs. Macro Labor Supply Elasticities: Evidence from Danish Tax Records. The Quarterly Journal of Economics, 126 (2): Choi, James J., David Laibson, Brigitte C. Madrian, and Andrew Metrick Defined Contribution Pensions: Plan Rules, Participant Choices, and the Path of Least Resistance. in Tax Policy and the Economy, Volume 16 NBER Chapters, National Bureau of Economic Research, pp DiNardo, John, Nicole M Fortin, and Thomas Lemieux Labor Market Institutions and the Distribution of Wages, : A Semiparametric Approach. Econometrica, 64 (5): Duflo, Esther and Emmanuel Saez Participation and investment decisions in a retirement plan: the influence of colleagues choices. Journal of Public Economics, 85 (1): Duflo, Esther and Emmanuel Saez The Role Of Information And Social Interactions In Retirement Plan Decisions: Evidence From A Randomized Experiment. The Quarterly Journal of Economics, 118 (3): Duflo, Esther, William Gale, Jeffrey Liebman, Peter Orszag, and Emmanuel Saez Saving Incentives for Low- and Middle-Income Families: Evidence from a Field Experiment with H&R Block. The Quarterly Journal of Economics, 121 (4): Duflo, Esther, William Gale, Jeffrey Liebman, Peter Orszag, and Emmanuel Saez Savings Incentives for Low- and Moderate-Income Families in the United States: Why is the Saver s Credit Not More Effective? Journal of the European Economic Association, 5 (2-3): Glaeser, Edward L., David Laibson, and Bruce Sacerdote An Economic Approach to Social Capital. Economic Journal, 112 (483): Liebman, Jeffrey B. and Luttmer, Exro F.P Would People Behave Differently if They Better Understood Social Security? Evidence from a Field Experiment. NBER Working Paper No Madrian, Brigitte C. and Dennis F. Shea The Power of Suggestion: Inertia in 401(k) Participation and Savings Behavior. The Quarterly Journal of Economics, 116 (4):

15 TABLE 1 Saver s Credit Program Structure, Saver s Credit Rate, t Equivalent Match Rate, t/(1 t) Adjusted Gross Income Married Filing Head of Single or Other Jointly Household (1) (2) (3) (4) (5) 50% 100% $0-$30,000 $0-$22,500 $0-$15,000 20% 25% $30,001-$32,500 $22,501-$24,375 $15,001-$16,250 10% 11% $32,501-$50,000 $24,375-$37,500 $16,251-$25,000 0% 0% >$50,000 >$37,500 >$25,000 14

16 TABLE 2 Summary Statistics for Cross-Sectional Analysis Sample, Variable Mean SD (1) (2) Income Measures Total Earnings (AGI) $20,091 $10,784 Wage Earnings $18,308 $12,537 Self-Employment Income $1,770 $6,074 Has Non-Zero Self-Employment Income 19.6% Number of W-2 s Tax Characteristics Has IRA Contributions 2.20% Tax Professional Usage 69.9% Demographics Age Number of Children Married 30.3% Female (for single filers) 73.0% Neighborhood (ZIP-3) Characteristics Saver s Credit Bunching Base IRA Participation 2.41% 0.95% High School Education 79.9% 6.98% College Education 23.3% 8.28% Number of Observations 219,742,011 15

17 TABLE 3 Cross-Sectional Correlates of Saver s Credit Bunching Dep. Var.: % Base IRA Participation % Professional Tax Prep % High School Degree Neighborhood Saver s Credit Bunching (1) (2) (3) (4) (5) (6) (0.015) (0.019) (0.002) (0.002) (0.018) (0.002) (0.003) Demographic Controls X X (0.019) (0.002) (0.003) Number of ZIP-3s

18 FIGURE 1 IRA Participation by Income Relative to Saver s Credit Cutoff Percent Contributing to IRA % Match 25% Match 11% Match Percent Contributing to IRA Income Relative to Saver s Credit Cutoff

19 FIGURE 2 IRA Participation by Income For Top and Bottom Bunching Decile ZIP-3s Percent Contributing to IRA Income Top Bunching Decile Bottom Bunching Decile 18

20 FIGURE 3 ZIP-3 Saver s Credit Bunching by Year a) 2002 b) 2005 c)

21 FIGURE 4 Saver s Credit Bunching by Post-Move Neighborhood IRA Participation a) Percent Contributing to IRA for Incomes Directly Below Cutoff Percent Contributing to IRA Directly Below Cutoff Post-Move Neighborhood Base IRA Participation Level (%) b) Saver s Credit Bunching Saver's Credit Bunching Post-Move Neighborhood Base IRA Participation Level (%) 20

22 FIGURE 5 Correlates of Saver s Credit Bunching a) Fraction with Professional Tax Preparation Saver's Credit Bunching Post-Move Neighborhood Professional Tax Preparation (%) b) Fraction with High School Degree Saver's Credit Bunching Post-Move Neighborhood High School Graduates (%) 21

23 FIGURE 6 IRA Participation for Movers to Top and Bottom Decile ZIP-3s 3 a) Pre-Move: IRA Participation by Income Relative to Saver s Credit Cutoff Percent Contributing to IRA Income Relative to Saver s Credit Cutoff Movers to Low Bunching Decile Movers to High Bunching Decile 3 b) Post-Move: IRA Participation by Income Relative to Saver s Credit Cutoff Percent Contributing to IRA Income Relative to Saver s Credit Cutoff Movers to Low Bunching Decile Movers to High Bunching Decile 22

24 FIGURE 7 Event Study: Saver s Credit Bunching for Movers to Top and Bottom Bunching Deciles Saver's Credit Bunching Year Lowest Bunching ZIP-3s High Bunching ZIP-3s 23

25 FIGURE 8 Saver s Credit Bunching by Post-Move Neighborhood 0.3 a) Post-Move 0.2 Saver's Credit Bunching β = (0.150) P-value = Post-Move Neighborhood Saver's Credit Bunching 0.3 b) Pre-Move Saver's Credit Bunching β = (0.150) Post-Move Neighborhood Saver's Credit Bunching 24