230B: Public Economics Cash Welfare Programs and Program Take-up

Transcription

1 230B: Public Economics Cash Welfare Programs and Program Take-up Hilary Hoynes Berkeley 1

2 Thanks to others The slides I am using this term are the result of a collaboration across lots of other faculty teaching PE including: Emmanuel and Gabriel Raj Chetty Day Manoli John Friedman Nathan Hendren Owen Zidar Amy Finkelstein 3

3 OUTLINE 1) Canonical cash welfare model; predictions 2) US system overview 3) Empirical literature: welfare and welfare reform 4) Quantile treatment effects and heterogeneous impacts of welfare reform [Bitler, Gelbach and Hoynes, AER 2006] 5) SSI, Deshpande ) The importance of take-up 2

4 OUTLINE 1) Canonical cash welfare model; predictions 2) US system overview 3) Empirical literature: welfare and welfare reform 4) Quantile treatment effects and heterogeneous impacts of welfare reform [Bitler, Gelbach and Hoynes, AER 2006] 5) SSI, Deshpande ) The importance of take-up 3

5 Government Transfer Programs Public Assistance versus Social Insurance Cash versus In-kind Universal versus targeted In-work versus out-of-work Conditional Cash Transfers

6 Social insurance vs public assistance Key Distinction: Social insurance programs seek to fill missing insurance markets; universal programs tied to prior work experience Public assistance programs seek to redistribute resources from rich to (some) poor; depends on current resources but not prior work experience Of course some blurring of the lines (e.g., Medicaid) but in general literatures are quite separate We are examining public assistance programs (social insurance mostly covered in 230A)

7 Framing: Optimal Policy Problem Seek to balance equity and efficiency concerns: [equity] governments value redistribution (from rich to poor); increasing incomes of the poor [efficiency] redistribution is costly in terms of efficiency loss due to disincentives of taxes (to raise revenue) and transfers (embedded in the program) [efficiency] categorical programs generate another form of (potential) efficiency loss through distorting choices of the outcomes that categorical programs are defined on (e.g. single mothers and welfare) Protection vs distortion

8 Canonical out of work benefit program cash welfare Goal is to provide some income floor, a basic level of assistance below which income will not fall. Must be phased out otherwise the benefit must be given to everyone (and then it is not means tested) The benefit formula: B = G t(e D) Program parameters are G (guaranteed income) and t (tax rate). E is earnings (w*h) and D is allowable deductions (practically speaking I will ignore D)

9 B Most countries have t very high and G set to be desired income floor

10 Cash welfare: effects on labor supply Extensive margin Reservation wage? goes up due to G Market wage? goes down due to t; w(1-t) Probability of work decreases Intensive margin Income effect less work (due to G, t) Substitution effect less work (due to t) Hours of work decreases Effects on income? Mixed, positive for those with low h. Distortion: leaky bucket given transfer B leads to an increase in income of less than B.

11 Lowering the tax rate: $ of consumption per year 25,000 18,000 A B 2 Note: It is very common for t to be very high in these programs. In US in AFDC during much of the history t=100% G = 9,000 B 1 D C 560 1,280 2,000 Hours of leisure per year

12 changes in cash welfare G t expand or contract? expand expand Costs? Caseload? Effect on extensive margin? Reservation wage no change Market Wage no change Overall effect on probability of work Effect on Intensive margin? income effect substitution effect no change Overall effect on hours of work Effect on income protection at no or very low earnings levels Effect on income supplementation at low to moderate earnings levels Mixed no change at no work

13 General lessons for cash welfare programs Basic structure discourages work Basic structure provides protection / insurance when not in work. A high phase-out can generate large moral hazard, work disincentives Efforts to encourage work (reducing t) can have unintended effects on new entrants and increasing the cost and size of the program

14 Key research questions How does welfare affect labor supply? What drives take-up? Stigma? How does welfare affect other outcomes (family formation, fertility) Does welfare provide protection against shocks (labor market, other)

15 OUTLINE 1) Canonical cash welfare model; predictions 2) US system overview 3) Empirical literature: welfare and welfare reform 4) Quantile treatment effects and heterogeneous impacts of welfare reform [Bitler, Gelbach and Hoynes, AER 2006] 5) SSI, Deshpande ) The importance of take-up 5

16 Evolution of Government Transfers in the U.S. 1930s Social Security AFDC Unemployment Insurance Great Society 1960s-1970s Food Stamps Medicare Medicaid Disability Civil Rights Act 1990s Welfare Reform Rise of the EITC 2010 Obamacare

17 Evolution of Government Transfers in the U.S. 1930s Social Security AFDC Unemployment Insurance Great Society 1960s-1970s Food Stamps Medicare Medicaid Disability Civil Rights Act 1990s Welfare Reform Rise of the EITC 2010 Obamacare

18 CBPP Low Income programs not driving Nation s Long-Term Fiscal Problem 2/21/17

19

20

21 Review articles Recent book with overview papers on social programs (The Economics of Mean-Tested Transfer Programs in the U.S.) Welfare (Ziliak) Nutritional Support (Hoynes and Schanzenbach) Supplemental Security Income (SSI) (Duggan and Kearney) Medicare (Buchmueller, Ham, Shore-Sheppard) Employment and Training (Smith) Housing (Ludwig) EITC (Rothstein) Early Education (Heckman)

22 OUTLINE 1) Canonical cash welfare model; predictions 2) US system overview 3) Empirical literature: welfare and welfare reform 4) Quantile treatment effects and heterogeneous impacts of welfare reform [Bitler, Gelbach and Hoynes, AER 2006] 5) SSI, Deshpande ) The importance of take-up 7

23 AFDC and TANF in a snapshot AFDC Aid to Families with Dependent Children 1935 SSA Eligibility depends on income & asset tests Categorical eligibility: single mothers Entitlement, cost sharing between states and fed States comply with federal guidelines but can set G Tax rate 100% (much of its history) Funded with matching formulas TANF Temporary Assistance for needy families 1996 Federal welfare reform (PRWORA) Fixed block grant States can spend grant on things other than B States can set G, t States must set time limits and work requirements

24 Low benefits: maximum benefits averaged about 39% of the poverty guideline Huge variation in G across states Tax rate: pre % % %

25 What do we know: Effects of AFDC (pre reform) WORK: Very high benefit reduction rates. Consistent evidence that AFDC reduced employment and earnings Evidence: reduces labor supply by 10-50% (off a low base). $1 to get 63 cents in increased income. DEMOGRAPHICS: Significant concerns that AFDC was a big part of the large and growing rate of female headed households and rising rate of nonmarital births Evidence: does not explain much PROTECTION: Even though it is funded at a low rate, prior to welfare reform AFDC provided important protection against deep poverty (a low floor) OTHER BENEFITS: Not much work on this Empirical strategies: Best work uses quasi-experimental design, difference-in-difference comparing the outcomes using variation across states and over time in the generosity of their state policies (the G). Also some structural analysis, jointly modeling hours and participation.

26 Welfare reform in 1990s Reforms in the 1990s addressed long-standing criticisms that AFDC discourages work and marriage, and causes long term dependence. Two periods of reform (1) State waivers ( ) States request HHS to waive specific eligibility and benefit requirements. Between 1992 and 1996, 28 states were granted major waivers. required by law to be evaluated with RCT Rich variation in timing and nature of waivers (2) Federal reform, PRWORA 1996

27 Elements of federal welfare reform Funding converted to block grant (uncapped entitlement severed); set nominally Time limits: Limit to using federal funds for five year time lifetime time limit Work requirements Financial sanctions for failing to adhere to work requirements Flexibility: states could set G and t Flexibility: state can spend block grant on things other than B (cash benefits) Flexibility: states can change categorical eligibility

28 AFDC and TANF in a snapshot AFDC Aid to Families with Dependent Children 1935 SSA Eligibility depends on income & asset tests Categorical eligibility: being a single mother Entitlement, cost sharing between states and fed States comply with federal guidelines but can set G Tax rate 100% (much of its history) Funded with matching formulas TANF Temporary Assistance for needy families 1996 Federal welfare reform (PRWORA) Fixed block grant States can spend grant on things other than B States can set G, t States must set time limits and work requirements

29 IMPACT OF WELFARE REFORM: GENERAL PREDICTIONS 1. reduction in welfare caseloads 2. extensive margin labor supply increase; intensive margin? 3. poverty? Could increase or decrease 4. family structure is unclear

30 IMPACT OF WELFARE REFORM: Time Limits Mechanical effect is to eliminate welfare when recipient reaches the time limit leading to an increase in labor supply. Anticipatory response is to bank welfare and exit prior to time limits.

31 Empirical models of welfare reform Empirical approaches to evaluate welfare reform Experiments: RCTs of waivers in states prior to federal reform (CT, FL, MN, etc) Quasi-experimental DD analysis of waivers: rich variation across states and over time in implementation and types of reforms Quasi-experimental DD analyses of TANF: little temporal variation (all states adopt within 18 months) but significant variation in type of reform implemented In the end we have 50 states, 50 separate programs. Hard to tell what elements of reform led to what changes And, need to pay attention to disentangle effects: welfare reform, the increase in the EITC and the strong economy of the late 1990s

32 The decline of welfare Source: U.S. Department of Health and Human Services Indicators of Welfare Dependence. 2013

33 Per Capita Real Expenditures on Social Safety Net Source: Bitler and Hoynes (2011) updated to 2014.

34 Source: Hoynes and Rothstein 2016, Current Population Survey ASEC Note: Employed is measured as whether an individual worked at all in the past 12 months. 38

35 Effects of welfare reform (tons of studies) Led to increase in labor supply, reduction in caseloads (post 1993 EITC and economy take on larger share of the changes) Welfare reform along doesn t show much evidence of increases in income (e.g. increases in earnings not enough to offset reduction in benefits)

36 Studies focused on time limits EX: Grogger and Michaelopoulus (JPE 2003), Grogger (RESTAT 2003) Point out that standard DD models not well suited to tease out impact of a particular element of reform. Reform appears as bundles. Model time limits in forward looking model Idea is to use age of women s youngest child to create untreated group (if youngest child is 14 then doesn t matter if you face a 4 year TL) Prediction for impact of TLs: Families with younger children have the greatest incentive to reduce welfare participation to preserve insurance value in the future. Use this to difference out the main effects of reform and isolate the impact of the time limits; need to assume other policies have homogeneous impacts by child age which is probably wrong Find large impacts of time limits

37 OUTLINE 1) Canonical cash welfare model; predictions 2) US system overview 3) Empirical literature: welfare and welfare reform 4) Quantile treatment effects and heterogeneous impacts of welfare reform [Bitler, Gelbach and Hoynes, AER 2006] 5) SSI, Deshpande ) The importance of take-up 10

38 Bitler, Gelbach and Hoynes What Mean Impacts Miss AER 2006 Develop quantile treatment effect estimator (QTE) and apply it to welfare reform Starting point: many states reduced tax rates as part of their welfare reform packages; no evidence in the DD literature of the prediction for reductions in earnings for some and increases in earnings for others. Methodologies used not well suited to reveal heterogeneous impacts Goal: what can be estimated in experimental context without any further assumptions, maintaining nonparametric appeal of the experimental estimators In our application, the theory predicts negative impacts on labor supply for some and positive impacts on labor supply for others; good setting for using a distributional estimator.

39 Setting: RCT using Connecticut s waiver program, Jobs First Why Jobs First? Most TANF-like of all programs (e.g. TL) evaluated using experimental design The heterogeneous labor supply predictions we predict are not new but JF has most dramatic change to work incentives that I know of Advantages of experimental analysis Identification of impacts of welfare reform using nonexperimental methods is less clear; especially for TANF (Blank 2002) Here design is clean and transparent Rich set of covariates including welfare history and earnings history 42

40 Comparison of Jobs First and AFDC Policies Time Limit CT Jobs First 21 months [6 mo. extension] AFDC NONE Earnings Disregard B=G-t(E-D) 100% of earnings disregarded up to poverty line Tax rate=0% Mo 1-4: $120+1/3 Mo 4-12: $120 Mo >12: $90 Tax rate=100% Other changes in Jobs First: strengthening work requirements and financial sanctions 43

41 Expected Effects of Jobs First on Labor Supply Income Poverty line Figure 1: Stylized Budget Constraint for AFDC and Jobs First Jobs First AFDC G Predictions: Employment increases Incentive to increase hours at the bottom of the distribution (H=H1) Incentive to decrease hours higher up the distribution (H>H1) Bottom line: Predicted effects of JF are heterogeneous Mean effects may mask positive and negative effects H2 H1 0 Hours 44

42 A QUICK PRIMER ON EXPERIMENTAL METHODS AND THE EVALUATION PROBLEM Evaluation research seeks to estimate the impact of a policy or treatment on an outcome. Treatment is dichotomous: received training, faced different welfare program, etc. The task of evaluation research, therefore, is to devise methods to reliably estimate their effects on outcomes, so that informed decisions about program expansion and termination can be made. 12

43 Bitler, Gelbach and Hoynes What Mean Impacts Miss AER 2006 Develop quantile treatment effect estimator (QTE) and apply it to welfare reform Starting point: many states reduced tax rates as part of their welfare reform packages; no evidence in the DD literature of the prediction for reductions in earnings for some and increases in earnings for others. Methodologies used not well suited to reveal heterogeneous impacts Goal: what can be estimated in experimental context without any further assumptions, maintaining nonparametric appeal of the experimental estimators In our application, the theory predicts negative impacts on labor supply for some and positive impacts on labor supply for others; good setting for using a distributional estimator.

44 Setting: RCT using Connecticut s waiver program, Jobs First Why Jobs First? Most TANF-like of all programs (e.g. TL) evaluated using experimental design The heterogeneous labor supply predictions we predict are not new but JF has most dramatic change to work incentives that I know of Advantages of experimental analysis Identification of impacts of welfare reform using nonexperimental methods is less clear; especially for TANF (Blank 2002) Here design is clean and transparent Rich set of covariates including welfare history and earnings history 42

45 Comparison of Jobs First and AFDC Policies Time Limit CT Jobs First 21 months [6 mo. extension] AFDC NONE Earnings Disregard B=G-t(E-D) 100% of earnings disregarded up to poverty line Tax rate=0% Mo 1-4: $120+1/3 Mo 4-12: $120 Mo >12: $90 Tax rate=100% Other changes in Jobs First: strengthening work requirements and financial sanctions 43

46 Expected Effects of Jobs First on Labor Supply Income Poverty line Figure 1: Stylized Budget Constraint for AFDC and Jobs First Jobs First AFDC G Predictions: Employment increases Incentive to increase hours at the bottom of the distribution (H=H1) Incentive to decrease hours higher up the distribution (H>H1) Bottom line: Predicted effects of JF are heterogeneous Mean effects may mask positive and negative effects H2 H1 0 Hours 44

47 Definitions: Potential Outcome Framework Angrist, Imbens, and Rubin (ECMA 1994) Evaluation research seeks to estimate the impact of a policy or treatment on an outcome. Treatment is dichotomous: received training, faced different welfare program, etc. The task of evaluation research, therefore, is to devise methods to reliably estimate their effects on outcomes, so that informed decisions about program expansion and termination can be made. 14

48 Definitions: Potential Outcome Framework Angrist, Imbens, and Rubin (ECMA 1994) Potential Outcomes: Important framework for evaluation of programs. Notation: D i = treatment indicator Y i1 = outcome if treated Y i0 = outcome if not treated i = Y i1 Y i0 = effect of program Fundamental Problem: never observe both Y i1 and Y i0 Possible treatment effects of interest: Treatment on the Treated (TOT) = E [ D = 1] Average Treatment Effect (ATE) = E [ ] 15

49 Bitler, Gelbach and Hoynes AER 2006 RCT provides a solution to the evaluation problem; given random assignment, the average effect of treatment on the treated can be estimated by comparing means in the treatment and control groups BGH AER 2006 generalize the Angrist et al. framework; estimate the effects of the treatment on the distn; goal is to get estimates of the distributional effects without further assumptions beyond random assignment Define q (QTE) = Y q (1) Y q (0) as the quantile treatment estimate Where Y q ( ) is the qth quantile in the outcome distribution E.g., Y 50 (1) is the median of the distribution of Y (1) QTE estimates the impact of the treatment on the distribution of outcomes. Interpretation: change in expected value of the outcome at the qth quantile when we take a randomly chosen, previously untreated person and give them the treatment. The impact of the treatment on the distribution can be estimated without further assumptions beyond random assignment 16

50 Note that estimating each of the q does not tell us about the distribution of i Requires further assumptions such as rank preservation Not obviously met in many cases

51 Jobs First Experiment and Data: Existing welfare recipients and new entrants randomly assigned to Jobs First or AFDC program Random assignment 1/96-2/97; 4 year follow-up Evaluation in New Haven & Manchester offices MDRC performed evaluation; we use their public use data: 4,803 single parent cases Administrative data (2 years before, 4 years after): Earnings: from UI (quarterly) Transfers: AFDC, Food Stamps (monthly) Demographic data from pre-ra interview: education, age, race, marital status, number and ages of children 45

52 Is assignment to Jobs First random? [SKIP] Levels Demographics Jobs First AFDC Difference White Black Hispanic HS Dropout HS Grad More than HS >2 children * >=2 children Youngest Child <= Never Married Mom< Mom Mom> Recipient (stock) Sample ** Joint test for equality p=0.16 Can not reject that assignment was random. Source: Bitler, Gelbach and Hoynes AER

53 Randomization in JF (cont) [SKIP] Levels Jobs First AFDC Difference Average pre-treatment quarterly earnings *** Average pre-treatment quarterly welfare ** Fraction of pre-treatment quarters with earnings ** Fraction of pre-treatment quarters with welfare ** Source: Bitler, Gelbach and Hoynes AER

54 Estimation details The Jobs First group has lower earnings and higher welfare use prerandom assignment. We control for this using inverse propensity score weighting (Firpo, 2007). We predict program assignment with demographic controls, and a full set of quarterly values of pre-ra earnings, AFDC payments, and Food Stamps payments, and dummies for their receipt. We weight by a normalized version of Di (1 Di) ˆ i pˆ ˆ i (1 pi) All standard errors are block bootstrapped to account for estimation of propensity scores and within-person variation. Confidence intervals use the percentile method (95% confidence intervals shown) Bottom line: Adjusting for observables affects mean impacts, but not the quantile treatment effects 48

55 Earnings QTE for the full sample (quarters 1-7) 1400 Findings: ATE QTE Range of QTE is large QTE large relative to ATE Pattern consistent with theory; especially decline at the top Percentile Solid line is QTE, dashed line is mean difference and dotted line is 95% CI. Source: Bitler, Gelbach and Hoynes AER

56 Is there an impact of time limits? Source: Bitler, Gelbach and Hoynes AER 2006

57 QTE for earnings before and after time limits Consistent with labor supply predictions Source: Bitler, Gelbach and Hoynes AER 2006

58 QTE for transfers before and after time limits Before TL: gains in transfers for low utilizers After TL: loss for high utilizers Source: Bitler, Gelbach and Hoynes AER 2006

59 QTE for income before and after time limits Before TL: gains at the top of the distn After TL: losses at the bottom Source: Bitler, Gelbach and Hoynes AER 2006

60 Distributional Estimates Often a very important feature of empirical analysis Depends critically on excellent data, otherwise a mess Other examples: Kline and Tartari (2016) attempt to relax the rank preservation assumption using bounds from theory Athey and Imbens (2006) generalizes dif-in-dif for distributions (Changes-in-Changes) Poor man s distributional estimate: series of dummy variables E.g., Replace actual earnings with dummy variable for $ > X Estimate for all X 18

61 Welfare and TANF: Summary Consensus that PRWORA reforms reduced caseloads (alongside Economic conditions and eitc) Changes increased work and decreased income (though perhaps not consumption) Effects focused at the very bottom of the distribution Little evidence on marriage or teen births despite program focus Literature has tapered off in recent years and huge literature Less evidence on protection, much more on distortion (Gruber JPUBE 2000, Bitler and Hoynes 2015, 2016) 19

62 Longer term changes due to welfare reform [SKIP] Things to look over on your own

63 WHAT HAPPENED? No longer a cash assistance program (No longer targeted on most disadvantaged) 26% of funds used For cash assistance Source: Bitler and Hoynes, The Hamilton Project.

64 Many states with less than 10% of spending for cash assistance Source: Bitler and Hoynes, The Hamilton Project.

65 Source: Bitler and Hoynes, The Hamilton Project.

66 FINDING: Loss of automatic stabilizer role Source: Bitler and Hoynes, JOLE 2016.

67 FINDING: Less protection for families in poverty Welfare reform Source: Bitler and Hoynes, The Hamilton Project.

68 OUTLINE 1) Canonical cash welfare model; predictions 2) US system overview 3) Empirical literature: welfare and welfare reform 4) Quantile treatment effects and heterogeneous impacts of welfare reform [Bitler, Gelbach and Hoynes, AER 2006] 5) SSI, Deshpande ) The importance of take-up 21

69 Social Security Disability Income (SSDI or DI) $130.6B (2012) (plus $70+B for Medicare) Social insurance, eligible workers through Social Security 10.1 M individuals/month (2012) Cash income Also qualify for Medicare Individual benefits Federal program Supplemental Security Income (SSI) $54B (2013) (plus $133B for Medicaid) Means tested program Low income elderly, disabled, including children 8.4 M individuals/month (2013) Cash income Also quality for Medicaid Individual (elig household) Federal program, with state add-ons

70 Social Security Disability Income (SSDI or DI) Financed by payroll tax (1.8% of earnings up to SS cap) Benefits depend on work history (need 10 years) and earnings, higher replacement rate for lower earnings (indexed as social security is; convert to retirement benefits at 65) Ave benefit $1100 (2012) Supplemental Security Income (SSI) Financed by general funds Max benefits = $710 (2013) (indexed)

71 SSI benefits (like canonical welfare, G & t) Eligibility depends on income test, asset test and categorical eligibility (being aged or disabled) B = G (UN $20) 0.50 (E - $65) Unearned income taxed at 100% after $20, earned income taxed at 50% after $65; G is $721 (higher than TANF) Annual break-even point for earned income about $18,000 Asset limit is $2,000 for an individual and $3,000 for a couple and excludes the value of a home and one vehicle. Medically determined physical or mental disability that limits the ability to engage in substantial gainful activity (SGA) and further demonstrate that this disability will last at least 12 months or result in death SSI children: income and assets of the parents are used for eligibility (only the portion of earnings above the SSI adult threshold are deemed to the child). Once a child turns 18, own income and assets are used

72 SSI Children Changes over time, rise of child caseload. Result of 1990 Zebley decision that expanded the medical eligibility criteria Zebley: Statutorily discriminated against children since child applicants did not have the option of demonstrating a disability using a vocational assessment, as could adults Overlap with TANF (transitions from AFDC/TANF to SSI with welfare reform). State incentives to move onto SSI since SSI is federal 1996 welfare reform in turn contracted the eligibility definition disability determination: a medically determined physical or mental impairment which results in marked and severe functional limitations, which can be expected to lead to death or which has been or can be expected to last for a continuous period of not less than 12 months

73 Source: Duggan, Kearney & Rennane 2016.

74 Deshpande, AER nearly all of the growth coming from mental conditions other than intellectual disability, including ADHD, speech delay, and autism spectrum disorders

75 Research Questions Protection versus distortion Once on disability, very large work disincentives (if income>sga then lose benefit) Individual decision about application: benefit cost calculation; as wage & employment declines opportunity cost declines greater incentive to apply Child disability does this affect where they are placed in school (mental/intellectual issues)

76 Deshpande AER 2016 Focus on child SSI participation Examines distortions of two sorts: Effects of SSI on the child s labor supply (when they reach 18) Effects of SSI on the parent s labor supply (when the child reaches 18) Takes advantage of a policy change due to 1996 Welfare Reform act (response to rising rolls post-zebley) Child turning 18 before 8/22/96: easier to convert to adult SSI on reaching 18 Child turning 18 after 8/22/96: harder to convert to adult SSI on reaching 18 This creates a sharp change in policy, triggered by whether you turn 18 before or after a single date RD To do this work, she had to work at SSA to get administrative data to analyze these behaviors

77 Regression discontinuity Sharp change in policy, usually not over time, but instead across some characteristic (e.g. birthdate, age) Running variable the characteristic under which there is the sharp change; ideally an immutable characteristic So AGE not income, for example Empirical Model: single difference design (dummy for > discontinuity) but in addition a flexible function of the running variable, allowing for differences on both sides of the discontinuity. General statement of model (assume age a is running variable) y ia = h a + D a π + X ai β + ε ai y=outcome variable h(a): smooth function of age, parameters differ on either side of discontinuity (Ex: polynomial) D a = dummy for age > discontinuity, π is treatment effect

78 Assumptions and checks on data Need to confirm that there is no confounding factors also changing at the discontinuity. Usual check: See that other Xs are smooth through the discontinuity Need to confirm that there is no manipulation of running variable (e.g. income, might adjust to be eligible). Usual check: See if density is smooth through discontinuity These models are very data intensive; typically need administrative data (large samples) and quality measurement of running variable Often you zoom in to a range close to the discontinuity. Tradeoff: closer in is better comparison in T vs C, but less sample size (power)

79 Deshpande s RD model Running variable Birth date in weeks Pre/Post dummy: Turn 18 after August 22, 1996 Empirical Model (reduced form) DOB = polynomial of order n, different polynomial for post=0 and post=1

80 Is there a treatment? Check to see that indeed medical reviews increase after PRWORA date. More unfavorable reviews, ever.

81 Standard RD graph Means for each bin Plus estimated polynomial Control group to capture any seasonality in births

82 Each point is a separate regression, RD coef on SSI enrollment

83 IV RD IV = first stage is SSI participation, instrument is when you turn 18 (sample includes those with an 18 th birthday within 37 weeks of 8/22/96 cutoff) RD = the instrument is in an RD setting Second stage effect of SSI on an outcome (earnings, income, etc)

84 IV-RD Effects on own earnings and income Each point is a separate regression of the IV-RD model with y measured at a different time period

85 IV-RD Effects on household earnings and income

86 Results SSI youth who are removed earn $4,000 per year, an increase of $2,600 relative to the earnings of those who remain on the program But loss of $7,700 in SSI large drop in income due to relatively small increase in earnings Also, income volatility increases Little evidence for the hypothesis that SSI holds recipients back from self-sufficiency (and long follow up period) Parental earnings respond to loss of SSI, but again does not make up the loss in income Still more to learn, does not answer the question of how the length of time in SSI (pre-age 18) affects the outcomes Welfare analysis shows that for reasonable values of risk aversion SSI provides a substantial income stabilization benefit to recipients.

87 OUTLINE 1) Canonical cash welfare model; predictions 2) US system overview 3) Empirical literature: welfare and welfare reform 4) Quantile treatment effects and heterogeneous impacts of welfare reform [Bitler, Gelbach and Hoynes, AER 2006] 5) SSI, Deshpande ) The importance of take-up 23

88 Incomplete take-up is common across programs Well documented that take-up of social transfer programs is incomplete Examples of take-up rates (Currie (2006) review and other more recent data) EITC 75-85%; SNAP 85% for US (66% for CA); 8-14% for SCHIP Three main explanations offered for limited take-up: Informational barriers to take-up (eligibility, benefits, application process) Transaction costs associated with enrollment Stigma associated with participation (could be a form of transaction cost) Optimizing models (take-up if expected benefits > expected cost) and non-optimizing models Moffitt (AER 1983) first to present model of joint labor supply/program participation; structural estimation Research today is focused on understanding role of three explanations 24

89 Some key questions Positive / descriptive What are key barriers to take-up? Relative roles of information, transaction costs and stigma Who is the marginal person deterred by barriers? Normative implications: Is low take-up bad? Welfare implications of increasing take-up levels Welfare implications of self-targeting (=own sign up) 25

90 Is increased take-up a goal? Policy makers and advocates talk about goal of increasing take-up Private welfare gain from increased take-up depends critically on whether individuals are making optimal decisions (e.g. trading off costs of acquiring information or applying against expected benefits). If yes, no first order welfare gain from increasing take-up by reducing barriers But if individuals are (sub-optimally) unaware / inattentive / failing to apply, could have first order welfare gain Social welfare: Incomplete take-up may actually be a desired (constrained) optimum With imperfect information about indivdiual s type, take-up barriers may improve self-targeting efficiency of redistributive program (or they may not) If barriers to claiming a credit discourage applications from those of low economic need, then such barriers may be efficient. if barriers reduce claiming by those with high need, then barriers may reduce welfare. 26

91 Critical for assessing the welfare implications of low take-up is a deeper understanding of why exactly those who are eligible for benefits fail to claim.

92 Self-targeting: Nichols and Zeckhauser AER 1982 Their set up: optimal tax / transfer problem; assume that truly needy (ability) is not observable Key insight: If demand for specific goods is correlated with unobserved characteristic, can use this to target desired group Example: in kind vs. cash: general view is that cash dominates but NZ argue that in-kind may improve self-targeting Weighing productive efficiency (usual efficiency concern) and target efficiency (who gets it) 27

93 Self-targeting: Ordeals NZ (1982) implies may be optimal to have ordeals in transfer programs: Tedious administrative procedures; stigma May enhance target efficiency if benefits from transfers vary across potential recipients Example: apply make people fill out lots of forms / wait in long lines to Pure deadweight loss / ordeal Nevertheless, may be a good screen for those whose marginal utility of receipt is low Self-targeting can work if you can offer a transfer that appeals only to the intended recipients. 28

94 An alternative take on ordeals Bertrand, Mullainthan and Shafir (AEA P&P 2004): hassle costs (e.g. 36 page food stamp application with confusing questions) deter the low ability people you want to transfer to Key question: Is marginal person deterred given current program rules someone who looks like we wouldn t want to redistribute to them (NZ) or someone we would like to (BMS) Use observables to see how they change as screen is changed (e.g. education, etc) Mullainathan and Shafir (2013) Scarcity Ordeals screen out those with limited bandwidth / consume cognitive resources Poverty as a bandwith tax: poor face many concerns and have to tunnel attention on a few 29

95 Alatas et al JPE 2016 RCT in Indonesia; 400 villages; test automatic screening (status quo) against self application with varying application costs (=distance to application center) They find that self-application improves targeting But, marginal increases in application costs (via distance) do not further improve targeting. Why? Long tail of people with low probability of passing screen; so large mass of people w very small probability of receipt get weeded out by small application cost 30

96 Deshpande and Li NBER wp 2017 Natural experiment: leverage timing of closing of 125 out of 1230 Social Security field offices between 2000 and 2014 apply for DI in field office (or over phone or online) field offices process DI applications Study how closings affect level (and characteristics) of application and enrollment (what are characteristics of marginal lost applicants) To deal with potential endog of closing offices, compare within set of offices that close (using differences in timing of closing). Requires that timing of closings, rather than the closings themselves, be as good as random Outcomes: counts of applicants, characteristics of applicants 31

97 Source: Deshpande and Yi NBER wp 2017 Closings lead to reductions in applications (good pre-trend)

98 Results on marginal applicants who are lost Compelling evidence of role of transaction costs in deterring applications and enrollment (e.g. apps decline) Applications and enrollment decline more for lower SES (lower education, lower pre-application earnings), those that file in person or phone (vs online), those without address, and those without representation No increase in acceptance rates Closings create modest increase in travel time to nearest open office (10 minutes by drive; 36 minutes by public transit) and congestion costs (increased time to be served in neighboring offices); not easily justified by time cost Bottom line: Consistent with Mullaintathan and Shafir that increased costs screen out low SES (or more needy) 33

99 Information Barriers: Bhargava and Manoli AER 2015 Moving beyond roles of hassles (application costs) on take-up and targeting; what about role of limited information? Costs involved in learning about eligibility and application rules Motivated by wanting to understand about psychological frictions associated with non-claiming:confusion, complexity, inattention Randomized experiment on incomplete take-up of EITC among childless [sidepoint: takeup in EITC is not low 75-80%; but it is low for childless recipients (56%) and others that are eligible for low levels of benefits) Setting is low transaction cost (claiming requires that recipient sign and return a one-age worksheet provided in stamped envelope)] 34

100 Setting and design: Bhargava and Manoli AER 2015 IRS sends out reminders to FILERS who look eligible for EITC but didn t fill out EITC form mailing includes letter and information, a worksheet with screeners, fill out and sign and get benefit their universe starts with those who have not responded to first mailer (but have filed which is a big hurdle) Sample includes 35,000 tax filers in CA Treatments: 1. Complexity interventions: of letter, of worksheet 2. Informational interventions - info about program, eligibility, costs ( less than X minutes ) and benefits 3. Stigma interventions - ( hard work and 4 of 5 claim their refund ) 35

101 Source: Bhargava and Manoli AER 2015

102 Source: Bhargava and Manoli AER 2015

103 Source: Bhargava and Manoli AER 2015

104 Bhargava and Manoli AER 2015 Results Take-up is sensitive to frequency, salience and simplicity with which information is provided Second mailing - just months after first - increases take-up by 22 percentage points! Nature of mailing has effects Simplification (e.g. visually more appealing notice or shorter worksheet) raises enrollment; poorest individuals most deterred by complexity (Figure 6) Stigma treatments have little effect. (cant tell if stigma not important or intervention didn t change it) they conclude that: confusion, program complexity, and lack of program awareness play a significant role in the failure to take-up, while stigma, and high perceived economic costs of claiming, do not 37

105 Bhargava and Manoli AER 2015 Interpretation Interpret results as evidence of low awareness of eligibility and benefits Survey in which participants reviewed experimental interventions and then beliefs assessed suggest interventions shaped behavior by influencing beliefs (about eligibility and benefit size) and increasing attention paid to forms Difficult to rationalize with a traditional / rational model of take-up in which eligible individuals balance accurately perceived expectations of benefits and costs Large impact of second notice Large impact of reducing complexity or changing salience Survey evidence suggested interventions increase awareness and reduce confusion Conclude there are psychological frictions and more work is needed to model and understand them 38

106 Hassles vs Information Bettinger et al. AER 2012; RCT in H&R Block aimed at increasing college enrollment Experimental design on getting FAFSA assistance: Some individuals offered personalized aid estimates and immediate assistance filing forms Others just offered personalized aid estimates Controls (status quo) Result: information by itself has no effect, immediate completion and sending out leads to increased apps / enrollment / persistence 39

107 Finkelstein and Notowidigdo (new wp on SNAP enrollment RCT) Food stamp take-up particularly low among elderly (40% compared to 80% overall) SNAP is an area with much interest in increasing take-up: (a) important program, (b) administered at state level (so in their control) but paid for federally (so state doesn t incur cost of benefits) Non-profit partner Benefits Data Trust (wants to increase take-up); identifies those not on SNAP but likely eligible (on Medicaid) Experimental design (30,000 elderly) 10,000 randomized into High Touch (Info + Assistance) [mailer plus access to call bank that facilitates enrollment and elig determination] 10,000 randomized into Low Touch (Info only) [mailer] 10,000 not contacted (control) bring in normative analysis 40

108 Finkelstein and Notowidigdo RESULTS information and transaction costs are barriers to take-up Treatment increases enrollment, more with assistance (+5pp info only; +12pp info + assistance) Interventions decrease targeting: marginal entrants are more advantaged (less sick, higher income, more likely white, english speaking) Bottom line: Consistent with NZ that info and transaction costs improve targeting Elderly population: could part of this be cognitive? 41

109 Summary of take-up Much interest in understanding role of information, hassles, and stigma; making progress that info and hassles are important Inconsistent evidence on how screening affects who take up Alatas et al suggest small hassle has good screening property but little/no targeting return to marginal increase in hassle Deshpande and Li suggest marginal increases in hassles may screen out right (low severity) and wrong (low SES) types Bhargava and Manoli suggest complexity screens out wrong types Finkelstein and Notowidigdo suggest that hassles screen out more advantaged 42

110 Attractive features of this area Areas for future work Rich, interesting and inconclusive theory Relative paucity of empirical evidence Positive and normative questions Fertile ground for research Impact of reducing barriers on take-up, screening, and welfare (should we have auto enrollment?) Recertifications RCTs feasibile in this space Letters are cheap; partners interested in improving or demonstrating their efficacy Yet implementing and expositing compelling quasi-experimental design in this space very valuable 43