Give Credit Where? The Incidence of Child Care Tax Credits Luke P. Rodgers∗ University of Texas at Austin November 16, 2016 Job Market Paper

Abstract

Child care tax credits are intended to relieve the financial burden of child care expenses for working families, yet the benefit incidence may fall on child care providers if they increase prices in response to credit generosity. Using policyinduced variation in the Child and Dependent Care Credit and multiple datasets in both difference-in-differences and instrumental variable frameworks, I find evidence of substantial pass-through: between $0.73 - $0.90 of every dollar is passed through to providers in the form of higher prices and wages. Robustness checks confirm the pattern that the bulk of credits are crowded out by increased prices. Furthermore, the relative inelasticity of child care suppliers implies that increased non-refundable credit generosity may have the unintended effect of making child care less affordable for low-income families, though the magnitude of this effect is tempered by heterogeneous pass-through rates. JEL Classification: J13, H22 Keywords: Child Care, Tax Credits, Incidence

∗ I am grateful to Jason Abrevaya, Marika Cabral, Naomi Feldman, Mike Geruso, John Hatfield, Richard Lowery, Dayanand Manoli, Jonathan Meer, Richard Murphy, Gerald Oettinger, Steve Trejo, and the seminar participants at UT Austin for their comments and helpful feedback. All errors are my own. The latest version of this paper will always be posted on www.lukeprodgers.com. Please contact me at [email protected] with any comments or questions.

1

“And, right now, in 31 states, high-quality child care costs more than a year of tuition at a state university. Think about that...we’re going to offer a tax cut of up to $3,000 per child per year. I don’t want anybody being ‘daycare poor.’ ” President Barack Obama

1

Introduction

Child care is a central component of the modern economy, largely due to increased labor market attachment of mothers and to the work requirements of numerous poverty-relief programs. Child care subsidies are a common policy tool intended to incentivize labor force participation of parents and provide tax relief to working families. Indeed, prior research has shown that parents are sensitive to the price of child care (Baker, Gruber, and Milligan, 2008; Blau and Tekin, 2007; Gelbach, 2002) and that child care costs can claim up to 18 percent of a family’s income (Herbst, 2015). Absent from this area of research is an estimate of the benefit incidence of child care tax credits.1 This paper asks the question: how much of child care tax credits are captured by child care providers in the form of higher prices and wages? Child care has been at the center of policy discussions recently, with President Obama and both major political party presidential candidates in the 2016 election each offering their prescription for what should be done to alleviate the burden of child care expenses.2 These discussions often highlight how rising child care prices discourage labor market participation, effectively working against programs like the Earned Income Tax Credit, and how high quality child care may be an effective tool to improve child outcomes and facilitate intergenerational mobility. Despite this increased attention, little is known about how effective child care tax credits are at relieving the burden of child care costs. This paper attempts to fill this crucial gap in policy evaluation. Understanding how much of the over $5 billion in annual child care tax credit expenditures remains with the parents and how much of it is passed-through to child care providers is of central importance to policymakers and parents alike. 1 Other papers have focused on the distributional incidence of these credits, or which types of families benefit most from these programs. See Altshuler and Schwartz (1996), Gentry and Hagy (1996), and Eiler and Hrung (2003). 2 President Obama has proposed large increases in the existing Child and Dependent Care Credit, as highlighted in the above quote. Hillary Clinton has various child care policy proposals, the central goal being the limitation of child care expenses to no more than 10% of a household’s budget. Donald Trump has proposed a tax deduction of child care expenses capped at the average annual cost of child care.

2

Estimating benefit incidence presents identification challenges that I address using multiple research designs and datasets. The expansion of the Child and Dependent Care Credit (CDCC) in 2003 will serve as the primary policy change in these natural experiments. First, I use the Bureau of Labor Statistics Occupational Employment Statistics (OES) in a difference-in-differences (DD) framework. Since child care wages make up nearly 80 percent of the cost of child care they can serve as a proxy for child care prices (Blau, 2001). I compare the wages of child care workers to the wages of workers in similar occupations before and after the largest child care credit expansion in recent history. If wages in the child care industry increase relative to the control group after the credit expansion, this could be evidence that providers raise prices in response to increased credit generosity. In the second research design I use instrumental variables (IV) and the Survey of Income and Program Participation (SIPP) to estimate the change in child care prices due to plausibly exogenous variation in credit generosity over time and across states. The federal expansion of the CDCC triggered different increases in state child care tax credits, most of which are various percentages of the federal credit, resulting in different total credit increases between states. An IV strategy is necessary because child care tax credits are endogenous to the price of child care. A simple regression of prices on child care tax credits would produce upward biased results, as child care expenses are the primary input to the credit formula. A family that purchased expensive child care would therefore receive a larger credit than an identical family that purchased the same amount of child care at a lower price. The IV approach is intended to reveal by how much, if at all, child care tax credits cause child care prices to increase. Both analyses reveal evidence of substantial pass-through. My results show that between $0.73 - $0.90 of every dollar in child care tax credits are passed through to the child care provider in the form of higher prices and wages. These estimates are consistent across research designs and datasets despite requiring different identification assumptions. Robustness checks alter the exact magnitude and significance of the pass-through rate but generally confirm that well over half of the credits are captured by child care providers. Heterogeneity analysis suggests that pass-through may increase along the wage distribution. This pattern is consistent with earlier research which found that wealthier families, who are the primary beneficiaries of child care tax credits, have experienced larger price increases over the last few decades than their low-income counterparts (Herbst, 2015). 3

When considered in the context of prior work and general incidence theory, I show that even small differences in relative price elasticities can have a large impact on the efficacy of this program at different parts of the income distribution. Policymakers may find these results useful in crafting the next era of child care subsidies and I provide back-of-the-envelope estimates of how increasing the non-refundable CDCC, a proposal floated by many politicians, may have the unintended consequence of making child care less affordable for low-income families. If child care providers are able to increase prices in response to credit generosity, low-income families who are ineligible for the full credit amount may experience a net increase in child care expenses when credits are expanded. From a welfare standpoint, there may be a better, more targeted use of the annual $5 billion in federal child care credits that are largely captured by child care providers and primarily claimed by middle- and high-income families, the children of whom have been found most likely to be negatively impacted by child care (Baker et al., 2008; Havnes and Mogstad, 2015; Kottelenberg and Lehrer, 2016). This paper is organized as follows. Section 2 provides background information on child care tax credits in the United States and benefit incidence theory. Section 3 describes the data. Section 4 describes the empirical strategies and Section 5 presents the results. Section 6 synthesizes the results with incidence theory while Section 7 provides evidence of heterogeneity within the child care market. Section 8 includes robustness checks. Section 9 offers distributional implications of increased credit generosity given the estimated pass-through rates. Section 10 concludes.

2

Background

2.1

Child Care Credits in the United States

The primary child care tax credit in the United States is the Child and Dependent Care Credit (CDCC).3 First offered in 1976, this credit was most recently expanded in 2003. To be eligible for this credit, the head of household (for single parent homes) or both parents (for joint filers) must be employed. Additionally, child care payment must be made to someone outside of the household; e.g. a teenage dependent who watches the younger sibling does not qualify. The household does 3

A household can use this credit to offset the cost of caring for a non-child dependent, such as a parent who needs a personal aide. The primary use is for children, however.

4

not need to itemize to claim this credit. When filing their taxes, a household may claim up to a capped amount of child care expenses for the previous tax year. The cap was $2,400 per child before 2003 and $3,000 after the expansion. A household can claim expenses for up to two children, for a total limit of $4,800 or $6,000. Next, a reimbursement rate which is a function of income, shown in Table 1, is multiplied by the capped expenses, thereby generating an “unadjusted” credit. Lastly, the non-refundable nature of the credit is accounted for: the actual or adjusted credit the household receives is the minimum of the unadjusted credit or their income tax liability.4 The nonrefundability of this credit means that many low-income families receive smaller adjusted credits than wealthier families despite a larger reimbursement rate; the former simply does not owe enough federal income tax while the latter typically owes enough that the non-refundability never comes into play. This interaction between statutory progressivity in the form of more generous reimbursement rates for low-income households and non-refundability has been the topic of papers focusing on the distributional incidence of these tax credits.5 These policy characteristics will be critical for evaluating the distributional implications of increased credit generosity later in the paper. Important additional variation comes from state child care credits. Twenty-two states currently have supplemental child care credits and all but three of them are a function of the federal credit. More importantly, these state credits are different percentages of the federal credit. For example, eligible Ohio residents receive a state credit equal to the full federal credit amount while an identical resident of Kentucky would receive a state credit equal to twenty percent of the federal credit. There are additional differences beyond percentages in state child care credits, including refundability and eligibility based on income thresholds. The federal expansion in 2003 thus differentially expanded all of the state credit amounts, the combination of which generates substantial exogenous variation in total credit generosity (state plus federal). As shown in Figure 1, a household in a state with no additional state child care credit would have received a maximum total credit increase of $330 while the identical household in a generous state, such as New York, would have received a maximum total credit increase of $693. Figure 2 offers another visualization of how the exogenous increase in federal CDCC generosity triggered 4

Non-refundable credits are claimed in a specific order with the CDCC second only to the Foreign Tax Credit. To be clear: I investigate how much of these credits get passed through to child care suppliers while other papers consider what type of family claims the credits. See Altshuler and Schwartz (1996), Gentry and Hagy (1996), and Eiler and Hrung (2003). 5

5

different total credit expansions based on pre-existing state policies. Figure 3 shows the geographic variation in increased total credit generosity, with no obvious concentration in any particular part of the United States. They key point is that federal expansion of the CDCC caused different increases in total child care credits across the country. It is important to note that child care tax credits are just one of many policies at both the state and federal levels intended to facilitate labor force participation of parents, reduce the cost of child care, and provide early education. Notable examples are Head Start and the Child Care and Development Fund (CCDF) program, the latter of which provides child care vouchers to very lowincome families. Since this may affect the outcomes of interest in the analysis below, it is necessary to control for funding changes and subsidy receipt whenever possible. Some businesses also offer tax-deductible spending accounts that can be used for child care, the use of which directly reduces the value of the CDCC. Such accounts are becoming more common but are typically only offered by large businesses whose employees’ marginal tax rates are high enough to make the deduction more valuable than the CDCC. Child care comes in many forms in the United States. Aside from direct parental care, children are often cared for by relatives such as grandparents. Formal care is the focus of this paper and is defined as any arrangement where parents pay a non-family member for child care. Child care centers are for-profit or non-profit businesses that specialize in caring for young children. Family day care homes offer care in a residential environment, usually by a mother who cares for her child alongside the children of her customers. Pre-schools combine care and educational instruction for children between the ages of two and five. Nannies or domestic workers are another option for parents who prefer to pay for one-on-one care in the child’s home environment. The CDCC can be claimed for any formal child care expenses.

2.2

Benefit Incidence

As mentioned in Section 1, this is the first paper to quantify the incidence of child care tax credits. Previous work in other areas, including college tax credits (Turner, 2011; 2012), dependent health care mandates (Goda et al., 2016), and wage subsidies like the Earned Income Tax Credit (EITC) (Rothstein, 2010; Leigh, 2010), has repeatedly demonstrated the potential for offsetting behaviors

6

that reduce, or even reverse, a policy’s impact.6 Basic tax incidence theory provides us with a simple framework for understanding how the child care market would be affected by an increase in child care tax credits. Referring to Figure 4, a tax subsidy (or credit), c, shifts the demand curve out, resulting in a new equilibrium price, P1 . The less elastic side of the market bears the bulk of the tax burden or, in this case, claims the majority of the benefit. The producer’s burden, or pass-through rate, can be written as: ∆P D = ∆T S − D

(1)

where the change in price resulting from a tax subsidy is a function of supply and demand elasticities. If this relationship is equal to unity then the full subsidy is passed-through to the supplier, if it is zero the subsidy stays entirely with the demand side, and any fraction between (0, 1) represents the portion captured by the supplier. The pass-through rate can be calculated if one has estimated structural elasticities for both sides of the market.7 My work directly estimates the left-hand side of this equation by exploiting exogenous variation in the denominator. The empirical pass-through rates calculated in this paper will reveal only relative elasticities of the child care market, though I offer a discussion of how my estimates compare to previous work in Section 7.

3

Data

3.1

Wage response using the OES

My first research design uses the Bureau of Labor Statistics Occupational Employment Statistics (OES). Wages make up nearly eighty percent of the cost of providing child care (Herbst, 2015) and have been previously used as a proxy for prices (Blau, 2001). I use this dataset because it permits heterogeneity analysis (see Section 7) and is designed for cross-occupation wage comparisons over time. The OES data are generated from surveys of establishments that are aggregated to the occupation-state-year level. My primary outcome variable is median hourly wage. I perform the same analysis using the American Community Survey and estimate very similar results despite 6

Other notable incidence topics include the mortgage interest deduction (Hanson, 2012), vehicle credits and promotions (Sallee, 2011; Busse et al., 2006), and payroll tax reductions (Gruber, 1997). 7 Blau and Currie (2004) provide a detailed survey of various elasticities related to child care.

7

additional self-response noise, shown in Appendix A.1. Thinking in terms of the difference-in-differences design, I need to compare child care workers to occupations with similar employee characteristics before and after the CDCC expansion. The OES uses nested occupation classifications with child care workers belonging to “Personal Care and Service Occupations.” While this group is somewhat diverse, there are a handful of occupations that are plausible control occupations. My final control group is made up of teaching assistants in elementary or secondary school, kindergarten teachers, residential advisors, cashiers, and waiters, though I vary this group in the robustness checks to ensure that no control occupation is driving my results.8 Pre-school expenses qualify for the CDCC so pre-school teachers are included in the treated group along with child care workers. Summary statistics for the OES data are located in Table 2. Child care workers earn lower hourly wages than the control group, largely due to the inclusion of kindergarten teachers. The rest of the OES sample is based on statewide data so there is nothing else different between the two groups.9 Unemployment rates are from the Bureau of Labor Statistics, population, state income, and mother’s labor force participation rates are from the Census, state minimum wages were compiled by the Tax Policy Center, and Child Care and Development Fund (CCDF) grant funding is from the Office of Child Care at the Department of Health and Human Services.10 The median hourly wage trends are shown in Figures 5. The control group’s mean wages are higher than the treated group and proceed in a similar pattern in the years before 2003, bolstering the DD assumption of parallel trends. Formally testing if these groups have different slopes in the pre-period suggests this is a reasonable assumption.11 The OES chart displays a noticeable increase in median hourly wage for the child care workers relative to the control group starting in 2003, the year of the CDCC expansion. Thus, the occupations I’ve selected as a control group appear to be similar intuitively and in their pre-period wage behavior. 8

This control group aligns very closely to the similar occupations listed on the O*NET website. Table A1 confirms the validity of the control group comparison along other dimensions. 10 I focus on mandatory and discretionary CCDF federal funds since matching funding (state and federal) may be endogenous. Similarly, TANF funding repurposed for child care is likely endogenous. 11 The OES pre-trends tests are much better when I drop 1999 from the sample (0.82 versus 0.05), but my analysis uses the full range of years as this generates the most conservative results, i.e. smaller magnitudes and similar significance. 9

8

3.2

Price response using the SIPP

In the second research design I use the Survey of Income and Program Participation (SIPP) and the NBER TAXSIM calculator. There are a handful of datasets that I investigated as potential candidates for this portion of the analysis, however the SIPP was by far the best suited for this analysis.12 The detailed topical modules on child care provide extensive information that helps generate the best price calculation I could find, while the financial information provides accurate inputs for the TAXSIM calculator. I pool the 1996, 2001, 2004, and 2008 panels of the SIPP.13 My sample is restricted to households with at least one child under the age of six in paid formal child care: daycare center, family daycare home, nanny, or nursery school. I drop married couples where only one parent works since they are ineligible for the credit, though my results are unaffected if I include them in the analysis.14 From the topical modules, I extract information on how many hours every child was in some sort of child care as well as the total cost, thereby allowing me to calculate the hourly price of child care. The main modules include demographic and income information that I use as inputs to the TAXSIM calculator in order to estimate tax liability and tax credits. It is important to discuss how the CDCC is calculated, both generally and with respect to the SIPP data. The CDCC is a function of child care expenses, income, tax liability, number of children in child care, and state of residence: Ci (Ei , Ii , τi , Ki , Si ). I use the TAXSIM calculator with the appropriate SIPP variables to calculate adjusted gross income (AGI), federal and state credits, and tax liability.15 Because the child care credit is non-refundable, it is important to use the “adjusted” credit amount which is merely the estimated tax liability before credits or the predicted credit, whichever is smaller.16 12

The Consumer Expenditure Survey (CES) has been used in child care papers in the past (Miller and Mumford, 2015) but the smaller sample size and additional assumptions necessary to calculate the price of child care were problematic for this analysis. Despite these limitations, early results using the CES generated similar results to those using the SIPP, namely that there is substantial pass-through of child care tax credits. The IRS Statistics of Income Public Use Microdata Sample is ideal for tax calculations but the censored information on child care expenditures once again tipped the scales in favor of the SIPP. Lastly, Appendix A.3 provides details on a custom dataset that may still prove useful. 13 I only use observations after 1997 due to major welfare reform that likely affected child care throughout 1996 and 1997. 14 I also drop households where the parent is living with the grandparents, which is most often the case with teenage parents. This is due to complications identifying the filing status of these multiple generation households as well as the fact that there is a very different environment of informal child care. Households that received outside assistance to pay for child care are not included as their price responses are likely inaccurate. 15 I follow the document by Dominic Coey on the NBER website that suggests ways to link SIPP variables to TAXSIM inputs. 16 Consider three taxpayers with the same predicted credit. The first taxpayer has zero tax liability, so her adjusted

9

State credits provide additional variation and are almost always a function of the federal credit. The percentages vary across states, so households receiving the same federal CDCC will receive different state credits. My independent variable is total adjusted credits, which is simply the sum of the federal and state credits. This should be the most relevant quantity since it is the total impact we are concerned with, not the level of credits separately. TAXSIM calculates the size of the state child care credit, but it does not provide any information on pre-credit state income tax liability. This is not a problem for those states with refundable child care tax credits, but when a household has zero tax liability and a non-refundable state credit it is impossible to tell if the lack of state tax liability is the result of or regardless of the state child care credit. For non-refundable state credits, therefore, the adjusted credit is defined as the predicted amount if there is positive state income tax liability and zero otherwise.17 Figure 6 demonstrates the variation in total (state plus federal) child care credits from my main SIPP sample. The key takeaway is that total child care credits vary substantially between states, largely triggered by the federal expansion in 2003. Table 3 includes summary statistics for the restricted SIPP subsample. The three credit variables, which will be elaborated on in Section 4.2, show that a family receives on average between $575 and $645 per year in total child care credits after adjusting for tax liability. There is a lot of variation in prices and income, with the mean hourly price of income at $3.60 and the mean annual household income around $75,000. Children are in formal child care on average nearly 32 hours a week, which is just slightly above the full-time threshold for most facilities. About twenty percent of the sample are single parents and almost all of the sample has at least a high school degree. Overall, my restricted sample looks very similar to the larger sample of families with children under the age of 6. child care credit is zero. The second tax payer has tax liability greater than the predicted child care credit, so her adjusted credit is the full predicted amount. A third taxpayer with positive tax liability less than the predicted credit will have an adjusted credit the size of the remaining tax liability. 17 This is the conservative approach since it biases my results to zero.

10

4

Empirical Strategy

4.1

Wage Response using the OES

The first research design approaches the incidence question from the supply side by quantifying the impact of child care credits on wages. Using a difference-in-differences (DD) design and the control group discussed in Section 3.1, my primary specification is: 0 Wjst = α + φ1(j ∈ CC industry)1(t ≥ 2003) + Xjst β + jst

(2)

where j is the occupation, s is the state, and t is the year. X includes state level controls for the fraction of population under the age of five, log of median income, unemployment rate, log of CCDF funding, labor force participation of mothers with children under the age of 5, and minimum wage. The coefficient of interest, φ, is on the interaction between an indicator for the post-expansion period and an indicator for a job being in the child care industry. This coefficient thus represents the change in wages for the treated group (i.e. the child care industry) relative to the control group after child care credits were increased. A triple-difference (DDD) specification that takes advantage of the state credit variation is:

Wjst = α + φ1(j ∈ CC industry)1(t ≥ 2003) + δ1(j ∈ CC industry)1(t ≥ 2003)1(s ∈ generous) 0 + ρ1(t ≥ 2003)1(s ∈ generous) + ν1(j ∈ CC industry)1(s ∈ generous) + Xjst β + jst (3)

The main coefficient of interest remains φ, which captures the average wage change for the child care industry after the credit expansion relative to the control group. The DDD coefficient, δ, captures the differential wage response of the treated group to the policy change in states with larger, or more “generous,” increases in total (state + federal) child care credits. There are various ways to define this discrete category, as well as continuous credit specifications, but given that the results are qualitatively very similar across specifications I present the simplest approach. A state is “generous” if the average total credit increase due to the expansion is above the median.18 18 Using a baseline of families that filed for the CDCC in the IRS Statistics of Income Public Use Microdata Sample, I adjusted inputs for inflation and ran them through the NBER TAXSIM calculator in both 2002 and 2003. A constant spending distribution isolates the policy-induced changes in total credits and accounts for non-refundability.

11

Compared to the DD, the DDD sacrifices power in order to tease out whether or not states with larger credit increases experienced different wage responses. The identifying assumption is that wages in the child care industry would have evolved in the same way as the control group were it not for the increased child care tax credits. The parallel trends assumption appears to be reasonable, as discussed in Section 3.1. I estimate these regressions by ordinary least squares (OLS) and cluster standard errors at the state level.

4.2

Price response using SIPP

The second research design measures the impact of child care tax credit generosity on child care prices. Simply regressing prices on child care tax credits received would produce biased estimates, however, due to the fact that the credit received is a function of child care expenses. A family that paid for more expensive child care would receive a larger credit than an identical family that paid for the same quantity of less expensive child care. To address this endogeneity, I offer multiple approaches.19 My basic specification, Strategy 1, with simplified credit notation is: Pi = α + φCi (Emax ) + Xi0 β + i

(4)

where P is the price of child care paid for by household i. X is a vector of controls including the number of children in child care, parental education, a dummy for being a single parent, a dummy for living in an urban area, income bins, and state and year fixed effects. The coefficient of interest, φ, is on the total (state plus federal) household child care credits adjusted for tax liability. I replace the actual child care expenses with the maximum amount that can be claimed in a given year, Emax . The credit is therefore not a function of household child care spending decisions, and I explicitly control for other factors that affect the credit amount, namely income and number of children (see Section 3). I estimate this regression by ordinary least squares (OLS) and cluster standard errors at the state level. Figure 7 offers a visual correlation of this strategy, demonstrating the positive relationship between maximum credit generosity and child care prices. Alternatively, I instrument for the actual child care credits received. The two-stage least squares 19

See Turner (2011, 2012) for application of these empirical strategies in the context of tax based aid for college. His results show that institutions offset increases in aid by increasing their prices.

12

(2SLS) framework is: First stage: Ci (Ei ) = ζ + ρCi (Ez ) + Xi0 γ + νi

(5)

Second stage: Pi = α + φCˆi + Xi0 β + i

(6)

I use exogenous spending, Ez , as an input for the credit calculation on the right-hand side of the first stage. This instrument credit is then used to predict actual credits, Ci (Ei ), that are a function of actual household spending recorded in the data. Strategy 2A is to simply use maximum spending as Ez . This approach expands the reduced form of Equation 4 so that the actual credits are instrumented for via the credits based on maximum spending amounts. Strategy 2B is a modified simulated instruments approach that holds the distribution of spending fixed in order to focus on exogenous policy variation. Using data from the IRS Statistics of Income (SOI), I back out average child care spending in a base year for every state and income range. I then adjust this spending using the Consumer Price Index (CPI) so that I have average spending by state-income bin for every year in the sample, Esim . Using this spending amount as Ez in Equation 5, I calculate the expected credit a household would receive if spending behavior (for their income range and state) was the same as in the base year. Finally, this credit is used as an instrument for actual tax credits. Again, the instrument is not a function of household child care spending decisions. I cluster standard errors at the state level in both IV strategies. The 2SLS estimation requires two assumptions. The first is the exclusion restriction, or simply that Cov(C(Ez ), ) = 0. This illustrates why it is so important to directly control for the other inputs that affect the credit amount, namely income, family size, state of residence, and year. The extensive set of control variables make it unlikely that any omitted variables are correlated with my instruments.20 Secondly, it is imperative that my instrumented credits are correlated with the actual credits, as quantified by a large F-statistic in the first stage. This requirement should be satisfied due to the primarily policy-induced variation in credit size. The formula is identical for both the instrument and the endogenous credit, the only difference being the change in spending input. In summary, my identification strategy in the SIPP analysis is to exploit exogenous policy variation that affects child care credit generosity, thereby allowing me to estimate the causal impact 20

Alternative instruments that investigate the validity of this assumption are presented in Appendix A.2.

13

of the credits on child care prices.

5

Results

5.1

Wage response: OES results

The DD results using the OES data can be found in Table 4. The coefficient of interest is on the interaction between an indicator for the post-period after the expansion of the CDCC and an indicator for working in the child care industry. This estimate captures the difference in hourly wage for workers in the child care industry, who may have been affected by the CDCC expansion, relative to the control group of similar occupations, who should not have been affected by the CDCC expansion. Each column estimates a statistically significant increase in wages for occupations that may have been affected by the CDCC expansion. Each specification includes state and year fixed effects, with later columns including state controls and job fixed effects rather than a treated industry fixed effect. In every case, the hourly wages of the treated group increase by between $0.21 and $0.24. The DDD regression from Equation 3 is shown in Column (4). The main coefficient drops to 0.18 but remains strongly significant. The DDD coefficient is positive and appears to be about half as large as the main coefficient, though unsurprisingly insignificant given the more demanding specification. This result suggests that child care workers in states with larger total credit increases experienced larger wage increases due to the policy change relative to their counterparts in states with smaller credit increases. As an additional representation of this research design, I plot the coefficients from a slightly modified DD specification in Figure 8, also presented in Table 5. I simply replaced the post-period dummy from the interaction 1(j ∈ CC industry)1(t ≥ 2003) with a vector of year dummies. This allows the occupational wage difference to vary over time rather than constraining it to be the same for all pre- and post- expansion years. The figure shows that the coefficients on the interaction between the child care industry and the year dummies increase after 2002, exactly when the CDCC was expanded. While every coefficient in Figure 8 except for 1999 is zero leading up to the CDCC expansion, some readers may suspect that this pattern is cyclical and that I may merely be picking up macroeconomic trends that differentially affect child care workers. This concern is plausible as lower unemployment may increase demand for child care and thus raise child care worker wages, 14

but the pattern and estimates are unaffected when I interact the child care industry dummy with macroeconomic variables such as state median income, unemployment rate, mother’s labor force participation, or GDP.21 I think the most convincing evidence is shown in Figure 9, however. I separately estimated the interaction coefficients for states in the top and bottom terciles of credit generosity. The wage response is larger in the more generous states, consistent with the DDD regression results, which should only be the case if credits are the cause. I cannot rule out the possibility that macroeconomic trends are affecting this pattern but these numerous checks make that explanation less likely.22 Overall, the value of these plots is to show that even with a more demanding DD specification there appears to be a significant wage increase for child care workers that occurs after the CDCC expansion.

5.2

Price response: SIPP results

The SIPP results are presented in Table 6. The coefficient on all tax credit variables should be interpreted as the change in hourly price due to an increase in child care tax credits claimed by the household after adjusting for tax liability where the units have been converted to a $1 hourly subsidy.23 Put more simply: the change in price per dollar of child care credits. Column (1) corresponds to the regression of prices on actual child care credits. As expected, this estimate is large and likely biased by unobservables correlated with actual child care spending and price. This endogeneity-plagued estimate, showing an unreasonable pass-through rate of over 2, motivates the alternative identification strategies. Column (2) includes the results of Strategy 1, the basic reduced form regression of price on the credit calculated based on maximum child care expenses. It appears that an extra $1 hourly subsidy via child care tax credits results in a $0.51 increase in the hourly price of care. This is shy of statistical significance (p = 0.15), representing a 14% increase from the mean. This regression further demonstrates the likely bias of Column (1). 21

See Section 8 for details. I was unable to directly check on overall cyclicality of this industry due to data limitations. The OES wage data only goes back to 1997, when major welfare reform was still being implemented, and the ACS started in 2000. Investigating cyclicality with the Current Population Survey, which is smaller but allowed me to look back to 1980, did not suggest that my results are driven by macroeconomic trends. Using more recent data do not change the results, as the coefficient terms level out around 0.2. 23 The credit is claimed annually but the pass-through relationship is more intuitive using the hourly subsidy representation. To be clear, I divide the total annual credits by the mean number of hours used in the SIPP data over a year: 32 hours for 50 weeks. Assuming two weeks of vacation time is the conservative approach and errors on the side of lower pass-through. See Section 6 for more details. 22

15

The instrumental variable approaches estimate larger magnitudes for the impact of credit on price. Strategy 2A, which instruments for actual child care credits with the credit based on maximum expenses, has a strong first stage with an F-statistic of 254 shown in Column (3). The IV estimate from Column (4) suggests that for every $1 hourly subsidy in child care credits the price increases by $0.74, with a p-value of 0.15. Strategy 2B, which uses a pseudo-simulated instrument based on child care expenses in a base year, has a strong first stage in Column (5) and generates the largest estimate at $0.89 in Column (6), significant at the 10% level. Taking the point estimates at face value, it appears that well over half of every dollar in child care credits is passed through to providers in the form of higher prices.

6

Incidence Discussion

Linking the empirical results to the incidence framework depends on the research design. Referring to the IV estimates from my SIPP results, that of Strategy 2A and 2B, the price of child care increased between $0.74-0.89 for an annual credit increase equivalent to a $1 hourly subsidy. It is important to recognize that this pass-through interpretation is sensitive to how many average hours and weeks are assumed when converting from an annual credit amount to the hourly subsidy. For example, more hours per week results in a smaller hourly subsidy (denominator) and therefore a larger pass-through rate. Using numbers based on average full-time child care usage, however, it appears that over three-fourths of the child care credit is passed-through to the supplier. Calculating pass-through with the wage estimates is slightly different due to the DD research design and the fact that child care workers serve more than one child at a time. Taking the estimate from Column (3) of Table 4, we interpret this coefficient as a $0.24 increase in the hourly wage of child care workers after the expansion of the CDCC. One way to quantify the discrete interaction term that generated this estimate is to calculate the increased credit amount that a typical family would have received in the post period. Using the lowest reimbursement rate (0.2) to be conservative, the expense limit increase alone would have generated a credit increase of $120 per child per year.24 Taking the midpoint child-to-staff ratio for child care workers (5:1),25 this is 24

Using the IRS Statistics of Income Public Use Microdata Sample to calculate the average change for a family in 2002 under both policies yields a remarkably close estimate of $123. 25 The Child Development Council recommendations range from 3:1 for infants to 9:1 for five year olds.

16

an annual increase in child care credits of $600 per worker, or $0.30 per hour for a full-time worker. In other words, we have a subsidy of $0.30 in place of ∆T and a wage increase of $0.24 in place of ∆P , resulting in a pass-through rate of 0.8.26 Hence, both research designs suggest substantial pass-through of child care tax credits.27 The possibility that most of the CDCC is passed-through to suppliers means that the supply side of this market is relatively less elastic than the demand side. This is key: relative rather than absolute elasticities are what determine tax incidence. Prima facie, working parents may seem to be the less elastic side of this market. After all, they need child care to maintain their jobs. This perspective does not factor in the many informal options available to parents, to include friends, under-the-table nannies, and relative care. The supply side of this market, on the other hand, is highly regulated. Strict licensing, child-to-staff ratio, zoning, and facility requirements make providers of child care far less elastic than the completely responsive case of just adding another kid to the classroom. The end result of these regulations is embodied by the long waiting lists that are a staple of child care in the United States.28 Again, even if demand is somewhat inelastic to price, the incidence calculation is based on relative elasticities.29 Thus, it is entirely possible that the supply side of the child care market is less elastic and captures the bulk of child care tax credits. There are a few critical elements that must still be discussed, however. Implicit in the above analysis is interchangeable child care services, such that the increased price of child care resulting from the tax credits does not change the type of child care paid for by parents. Even a cursory search of local providers will reveal the diverse child care options facing any parent. If child care credits incentivize parents to select a better quality child care provider, the price increase need not represent pass-through. Unfortunately, child care quality is notoriously difficult to measure due to 26

Even though labor is the main cost of child care, a puzzle is why these low-skilled workers and not merely the owners benefit from these credits. First, child care providers likely have more bargaining power than their low-skilled counterparts in other service industries, such as waitresses or cashiers. This is due to the fact that parents value steady relationships between their child and the caregiver. Second, many child care providers are non-profit organizations which are theoretically predicted to create rents for their workers. Finally, analysis with the ACS data suggests that self-employed child care workers experienced an even larger wage increase, which is consistent with the idea that owners would benefit from high pass-through. 27 Adjusting the assumptions used in these pass-through calculations generates a range of estimates between 0.4 and 1.4. 28 Anecdotal discussions with child care providers revealed that prices increase and do shorten waiting lists but that norms for allocating placement for siblings of current customers as well as the desire of suppliers to match to families along unobservable dimensions partially explain why waiting lists remain commonplace in this industry. 29 Another explanation is that suppliers are able to claim more of the credit due to market power, possibly generated by this industry’s emphasis on reputation or the high entry costs created by regulation. See Weyl and Fabinger (2013) and Cabral, Geruso, and Mahoney (2014).

17

intangible factors, explaining the limited work on this dimension of child care.30 There is no way to quantify how much of the price response is due to facility improvements or parents “upgrading” their child care, but any shift would make the three-fourths pass-through estimate an upper-bound. What evidence I do have suggests that quality improvements are not driving these results, however. Two of the most commonly discussed inputs to child care quality are staff education and experience.31 I ran the DD regression using education as the outcome variable and found that child care worker education decreased a statistically significant but unremarkable 0.07 years in the post period relative to the control group. Using the Current Population Survey’s Job Tenure Supplement, I compared tenure of child care workers to the control group over the same timeframe. Tenure increased slightly in every position, child care workers not standing out as having substantially reduced job turnover after the CDCC expansion. If education and tenure are rough proxies for worker quality, as has been suggested in the literature, my results are not being driven by quality upgrades in the employee dimension, the primary input of child care services. Furthermore, parents overestimate the quality of their child care provider and choose providers primarily based on location and convenience (Helburn and Howes, 1996). These facts combined with the presence of waiting lists make it unlikely that large numbers of parents responded to this credit increase by switching child care providers. Still, interpretation of the pass-through rates would be affected by quality changes during the sample timeframe. A final point is that child care prices are usually structured as part-time or full-time, with the latter often offering a substantially reduced hourly price. If the tax credits motivate parents to switch from part-time to full-time, the price estimates may be a lower bound. This bias would work in the opposite direction of the quality upgrades. The takeaway is that there appears to be a large percentage of the CDCC going to providers but that I cannot quantify the relative importance of changing quality and pricing tiers. The following section delves more deeply into another aspect of these markets that is important for interpreting my estimates. 30

Blau (2007) estimates that increased regulation does not increase quality but is borne by lower worker pay. Blau and Mocan (2002) suggests, on the other hand, that subsidies may increase the quality of child care centers which may be elastic to input prices. Hotz and Xiao (2011) estimate the heterogeneous impact of regulation, where centers serving low-income parents are likely to shut down but more expensive centers may improve quality. 31 Cross-sectional correlations suggest that recent early childhood training may be a better predictor of care quality (Blau, 2001).

18

7

Heterogeneous Pass-Through

It is possible that the mean pass-through estimates calculated above conceal important heterogeneity between segments of the child care market. As has been documented in the literature (Herbst, 2015; Blau and Mocan, 2002), consumers of inexpensive child care are typically low-income families and quality of care weakly rises with price. Is it possible that the pass-through rate depends on the market segment under consideration? To investigate this question, additional DD regressions are performed, only this time focusing on the 25th and 75th percentiles instead of the median. The results, found in Table 7, still show positive and significant wage increases for child care workers relative to the control group, but the wage increases at the 75th percentile are larger than those at the median or 25th percentile. The heterogeneous response remains when using the more demanding year-dummy interaction specification, the coefficients of which are plotted in Figure 10. These results are suggestive, but the wide standard errors in the figure and a failed test of whether the 75th percentile response is statistically different from the 25th percentile response caution against making too much of these additional results. Having said that, it is worth briefly discussing this pattern in order to better understand pass-through in the child care market. There are two scenarios that could generate this increasing pattern. First, it is possible that the more expensive segment of the child care market has a higher pass-through rate than the less expensive segment of the market. This would be the case if the ratio of demand and supply elasticities was increasing with income (Equation 1). This is plausible for a few reasons and has support in the literature. Blau and Currie (2004) provide a survey of the numerous labor supply elasticities with respect to the price of child care of mothers of young children. It is difficult to draw a strong conclusion from these estimates, however, since they range from 0 to -1, the bulk clustered between -0.2 and -0.6. Moreover, some studies show that married women are more sensitive to the price of child care while others show that single women are more sensitive to the price of child care. Having said that, the safest generalization is that the extensive margin (labor force) elasticity with respect to the price of child care decreases with skill.32 Regardless of potential heterogeneity of demand elasticities, it may be that that child care providers become increasingly inelastic as they 32 It is important to note that most of the estimates of LFP sensitivity to child care prices are before welfare reform and increases in the EITC. These large reforms were specifically designed to increase labor force participation of single parents and reduce welfare usage. Hence, it is possible that mothers have more homogeneous sensitivity to child care prices since many of these studies were published.

19

move up the cost-quality scale. The best child care facilities may not be easily able to expand and maintain the same standards, either in staff ratios or quality, and therefore raise prices rather than accommodate additional children. Indeed, Blau (2001) estimates a large labor supply elasticity for child care workers (1.1), and notes that child care worker wages did not increase substantially throughout the 1980s and 1990s despite rising labor force participation of women. He suggests this is due to availability of low-skilled workers that can switch to child care from other industries when demand increases. It is far easier to open a new family day care home than develop a reputation as a high quality child care center, not to mention the lower regulatory hurdles of inexpensive child care provision. Thus, the literature supports the first explanation of heterogeneous wage responses, that child care has increasing pass-through rates as you consider market segments with higher costs and quality. The other possibility is that heterogeneous wage responses could be due to different effective tax credit expansions. As already discussed, the non-refundability of the CDCC alters the generosity pattern from what would be expected based strictly on the statutory reimbursement rates. If the credit expansion differentially altered generosity along the income distribution, Equation 1 would have different denominators for different market segments. It would therefore be possible that pass-through rates are identical in all parts of the child care market. Put another way, the wage response could be smaller for the bottom quartile of the wage distribution because the effective tax credit is smaller. Using the 2002 IRS Public Use Microdata Sample (PUMS), I took every household who filled out the child care Form 2441 and calculated their maximum credit using TAXSIM. Then, after adjusting the inputs for inflation, I calculated their credits under the 2003 policy. Referring to Figure 11, the credits were still the most generous for the lowest income families even after accounting for tax liability.33 The fact that low-income families did not receive a smaller credit increase even after accounting for tax liability weakens the case that the heterogeneous wage response is consistent with a homogeneous pass-through rate. Using the estimates from Columns (3) and (6) in Table 7 and Column (3) in Table 4, passthrough rates are 0.59, 0.79, and 0.87 for the 25th percentile, the median, and the 75th percentile, respectively. Changing the denominators based on the TAXSIM estimates discussed above, the pass-through rates for the same quantiles become 0.46, 0.83, and 1.04, respectively. While the 33

This pattern is very similar when referring to the aggregate SOI statistics.

20

exact point estimates vary depending on how the denominator is constructed, it appears that child care providers at the top of the price distribution may have increased wages more than providers at the bottom of the distribution after the CDCC expansion. Heterogeneous pass-through will play a part in the distributional implications discussed in Section 9, though it is worth reiterating that this pattern is suggestive and that there is no statistical evidence of different pass-through rates.

8

Robustness

The estimates of both research designs hold up under numerous robustness checks. With regards to the DD strategy, I adjusted the group definitions and found no evidence that a single control occupation was driving the results. I ran the main DD regressions dropping one control occupation at a time. None of these group changes substantially affect the estimates which remain significant and span 0.22 - 0.27. Using more demanding state-by-year fixed effects does not affect the results either. The result of a synthetic control technique is shown in Figure 12. The average demographic information (sex, race, education, age, and marital status) from the ACS at the state-year-occupation level was merged to the OES data. The synth command in Stata, written by J. Hainmuller, A. Abadie, and A. Diamond, matches the outcome variable and selected control variables of the control group to those of the treated group during the pre-intervention period. Child care industry wages matched closely in the pre-period to a weighted combination of teaching assistants, cashiers, and kindergarten teachers. The wages in both groups follow a similar trend until the expansion of the CDCC, after which the child care industry wages continue in a positive direction while the synthetic control group wages decline. Table 8 includes the remaining DD checks. Columns (1) - (4) control for interactions between macroeconomic variables and the child care industry dummy to address possible cyclical drivers of wage differentials; the coefficient of interest actually increases. Columns (5) is a falsification test where I set the policy change two years prior to the actual policy change, resulting in a negative and insignificant coefficient of interest. The supply of child care workers as measured by the log of total employment does not appear to change relative to the control group (Column (6)), and the small size and positive direction makes it unlikely that the quantity of child care workers caused the

21

wage increase.34 The combination of numerous specifications and the fact that the results are very similar when using a completely different dataset (see Appendix A.1) demonstrates the robustness of the DD analysis. Identical and additional robustness checks can be found in Appendix A.1, most notably a placebo test shown in Figure A3. The SIPP estimates, ranging between 0.51 and 0.9 with different levels of statistical significance, are not altered substantially under numerous sensitivity tests, shown in Table 9. Each odd-numbered column uses the maximum credit IV while each even-numbered column uses the simulated IV. For my main SIPP results I drop married households where only one parent works, as they are ineligible for the credit, though including them in the sample does not greatly affect the results (Columns (1)-(2)). I drop families with zero child care credits in Columns (3)-(4) and find no evidence that those observations are driving the results. A few families show up multiple times in my sample due to the panel nature of the SIPP. Using different weights does not affect the results, shown in Columns (5)-(6).35 The main concern with the price response identification strategy is that I do not adequately control for other factors that affect the tax credits. Put another way, the coefficients on the credits could be picking up some unobservable relationship with income or tax liability. My main approach is to flexibly control for income with income bins ($10k bins up to $50k a year, then $25k bins up to $200k a year). Alternative methods using various splines of income and tax liability are included in Columns (7)-(12) of Table 9. These more demanding specifications increase the range of point estimates to 0.42 - 0.77 and widen the standard errors. Similarly, in case there is some endogenous income adjustment taking place I generated the tax credits using a household’s prior year income; the results are unaffected by this modification. Alternative instruments are discussed in more detail in Appendix A.2. Lastly, a non-equivalent dependent variable that may be influenced by the same unobservable relationship with income that should not be directly affected by child care tax credits is the price of sports activities. Running the same regressions with the price of sports played by children in the sample as the outcome variable, the coefficient on child care credits is always insignificant and negative. These checks give me some confidence that the tax credit is reflecting 34

If more child care workers reduced the child-to-staff ratios this could be seen as an increase in quality, however. Turner (2011) divides the first observed weight by the number of times the family appears in the sample. This is my main weighting strategy, though using the recommended weight (un-adjusted for multiple observations) does not affect my estimates. 35

22

the true impact of policy-generated credit generosity and not some underlying relationship with income or tax liability.

9

Distributional Implications

The combination of substantial pass-through and non-refundability generates clear predictions regarding the consequence of increasing child care credit generosity. Recall that the non-refundable nature of this credit means that any increase in generosity is tempered by the household’s remaining tax liability. Thus, as the credit becomes more generous more households hit that limit and cease receiving additional benefit. All households, however, bear the burden of increased child care prices. Consider a simple example with three households: one with zero tax liability, one with $500 of tax liability, and the final with $1000 of tax liability. An increase in the child care tax credit of $1000 will generate tax reductions for these households of $0, $500, and $1000, respectively. With a pass-through rate of 0.75, each family has increased child care expenses of $750. The net change in after-tax, after-child-care (expenses and credits) income for each of these households is -$750, -$250, and $250. In other words, increasing a non-refundable credit with a high pass-through actually redistributes money from households with low tax liability to households with high tax liability, or from poor households to rich households. Figure 13 plots the after-tax, after-child-care income distribution before and after an increase in the CDCC.36 It is clear that low-income families are not benefitting from this credit increase, as they bear the burden of increased child care expense without any offsetting tax benefit. Alternatively, Figure 14 shows the net income distribution assuming the new credit was made refundable. Families across the distribution, in this situation, benefit from the 25% of the credit that stays with the consumer. While still an expensive way to reduce the burden of child care expenses, refundability eliminates the pattern of redistribution from poor to rich households. Another way to make this point is to plot after-tax, after-child-care income changes against various pass-through rates. Figure 15 breaks the income distribution of the SIPP sample into 36

For the chart to be clear I increased the credit by $5,000, which is a little under the mean annual expenses for the SIPP sample. The qualitative result is the same with smaller credits despite being much harder to see.

23

terciles and then increases the CDCC by $1000.37 With the exception of pass-through rates below 0.35, which is below even the most conservative calculation from Section 7, the poorest families are made worse off by an increase in the CDCC. Their limited tax liability does not allow them to claim much of the non-refundable credit yet they still pay the resulting higher prices. The middle and upper terciles benefit regardless of the pass-through rate due to their higher tax liabilities. However, it is remarkable that for pass-through rates at my upper estimate of 0.90 the change in total income turns negative. Not only does this program redistribute from poor to rich families but it is very close to having a negative impact on consumers of child care in total. To avoid overstating this, it is safer to conclude that even for pass-through rates much lower than my empirical estimates an increase in the CDCC disproportionately hurts low income families. The possibility of heterogeneous pass-through rates is explored further in Figure 16. The change in child care expenses net of child care credits resulting from a $1,000 increase in the CDCC is plotted for my main SIPP income distribution. Each line represents a different pass-through scenario. When everyone faces the same pass-through rate, expenses increase for the lowest part of the distribution and decrease for the middle and upper parts of the distribution. A lower passthrough rate for the bottom tercile, either three-fourths or half the top rate, tempers this result, as households shopping in that part of the market experience a smaller price increase. Maintaining a mean pass-through of 0.75 in these scenarios necessitates the increase in pass-through for the upper tercile, thereby reducing their tax benefit. These rates correspond very closely to the results in Section 7. Heterogeneous pass-through limits the negative impact of a credit increase on the lower part of the income distribution and reduces the benefit to the upper part of the distribution. In any case, the lowest income households do not have enough tax liability to offset a price increase resulting from any positive pass-through rate. In light of these findings, policymakers may want to explore alternatives to increases in the non-refundable CDCC as a way to provide tax relief to working families.38 37

I performed all of these simulations using the more detailed and larger IRS SOI PUMS datasets. The conclusions are identical so I used the SIPP for consistency. 38 For example, eliminating the CDCC and increasing the Additional Child Tax Credit or the EITC, which is already indexed based on number of children, would simplify the tax code and still target families with children.

24

10

Conclusion

Child care tax credits are at the center of many important strands of economic research, yet the incidence of these credits has yet to be estimated empirically. This paper fills that gap by providing evidence that three-fourths of state and federal child care tax credits are passed-through to child care suppliers in the form of higher prices and wages. Heterogeneity analysis suggests that passthrough increases along the wage distribution, demonstrating variation in the relative elasticities of the different segments of the child care market. Calls for increased child care subsidies must be considered in light of the possibility that most of that funding will go to child care providers, doing much less to relieve the burden of child care costs than the size of the expenditures would suggest. Moreover, the combination of high pass-through and non-refundable child care credits effectively redistributes income from poor to rich households. Future research may be able to generate more precise estimates, ideally factoring in changing quality and price-tiers, but this paper’s results show that increasing child care credit generosity is a costly policy that may actually hurt the families most in need of child care assistance.

25

References [1] A. Abadie, A. Diamond, and J. Hainmueller. 2010. “Synthetic Control Methods for Comparative Case Studies: Estimating the Effects of California’s Tobacco Control Program.” Journal of the American Statistical Association, 105(490): 493-505. [2] R. Altshuler and A. E. Schwartz. 1996. “On the Progressivity of the Child Care Tax Credit: Snapshot Versus Time-Exposure Incidence.” National Tax Journal, 49(1): 55-71. [3] S. Averett, H. Peters, and M. Waldman. 1997. “Tax Credits, Labor Supply, and Child Care.” The Review of Economics and Statistics, 79(1): 15-135. [4] M. Baker, J. Gruber, and K. Milligan. 2008. “Universal Child Care, Maternal Labor Supply, and Family Well-Being.” Journal of Political Economy, 116(4): 709-745. [5] M. Berger and D. Black. 1992. “Child Care Subsidies, Quality of Care, and the Labor Supply of LowIncome, Single Mothers.” The Review of Economics and Statistics, 74(4): 635-642. [6] Blau, David M. 2001. The Child Care Problem: An Economic Analysis. The Russell Sage Foundation, New York. [7] Blau, David M. 2007. “Unintended consequences of child care regulations.” Labour Economics, 14: 513-538. [8] D. Blau and J. Currie. 2006. "Pre-School, Day Care, and After-School Care: Who’s Minding the Kids?," Handbook of the Economics of Education, Elsevier. [9] D. Blau and H. N. Mocan. 2002. “The Supply of Quality in Child Care Centers.” The Review of Economics and Statistics, 84(3): 483-496. [10] D. Blau and E. Tekin. 2007. “The determinants and consequences of child care subsidies for single mothers in the USA.” Journal of Population Economics, 20: 719-741. [11] M. Busse, J. Silva-Risso, and F. Zettelmeyer. 2006. “$1,000 Cash Back: The Pass-Through of Auto Manufacturer Promotions.” American Economic Review, 96(4): 1253-1270. [12] M. Cabral, M. Geruso, and N. Mahoney. 2014. “Does Privatized Health Insurance Benefit Patients or Producers? Evidence from Medicare Advantage.” NBER Working Paper Series, No. 20470. [13] J. Compton and R. A. Pollak. 2014. “Family proximity, childcare, and women’s labor force attachment.” Journal of Urban Economics, 79: 72-90.

26

[14] G. B. Dahl and L. Lochner. 2012. “The Impact of Family Income on Child Achievement: Evidence from the Earned Income Tax Credit.” American Economic Review, 102(5): 1927-1956. [15] S. Dickert-Conlin, K. Fitzpatrick, and A. Hanson. 2005. “Utilization of Income Tax Credits by LowIncome Individuals.” National Tax Journal, 58(4): 743-785. [16] J. R. Eiler and W. B. Hrung. 2003. “Federal Tax Benefits for Child Care: Progressivity Over Time, EGTRRA, and the Alternative Minimum Tax.” Topics in Economic Analysis and Policy, 3(1), Article 16: 1-12. [17] S. Elango, J. Garcia, J. Heckman, and A. Hojman. 2015. “Early Childhood Education.” NBER Working Paper Series, No. 21766. [18] D. Feenberg and E. Coutts. 1993. "An Introduction to the TAXSIM Model." Journal of Policy Analysis and Management 12(1): 189-194. [19] Gelbach, Jonah B. 2002. “Public Schooling for Young Children and Maternal Labor Supply.” The American Economic Review, 92(1): 307-322. [20] W. M. Gentry and A. P. Hagy. 1996. “The Distributional Effects of the Tax Treatment of Child Care Expenses,” in Empirical Foundations of Household Taxation, M. Feldstein and J. Poterba, eds. (University of Chicago Press), 99-134. [21] G. Goda, M. Farid, J. Bhattacharya. 2016. “The Incidence of Mandated Health Insurance: Evidence from the Affordable Care Act Dependent Care Mandate.” NBER Working Paper Series, No. 21846. [22] Gruber, Jonathan. 1997. “The Incidence of Payroll Taxation: Evidence from Chile.” Journal of Labor Economics, 15(3): 72-101. [23] C. Haeck, P. Lefebvre, and P. Merrigan. 2015 “Canadian evidence on ten years of universal preschool policies: The good and the bad.” Labour Economics, 36: 137-157. [24] Hanson, Andrew. 2012. “The Incidence of the Mortgage Interest Deduction: Evidence from the Market for Home Purchase Loans.” Public Finance Review, 40(3): 339-359. [25] T. Havnes and M. Mogstad. 2011. “No Child Left Behind: Subsidized Child Care and Children’s LongRun Outcomes.” American Economic Journal: Economic Policy, 3: 97-129. [26] T. Havnes and M. Mogstad. 2015. “Is universal child care leveling the playing field?” Journal of Public Economics, 127: 100-114.

27

[27] S. W. Helburn and C. Howes. 1996. “Child Care Cost and Quality.” The Future of Children, 6(2): 62-82. [28] Herbst, Chris. 2010. “The labor supply effects of child care costs and wages in the presence of subsidies and the earned income tax credit.” Review of Economics of the Household, 8: 199-230. [29] Herbst, Chris. 2013. “The impact of non-parental child care on child development: Evidence from the summer participation ‘dip.”’ Journal of Public Economics, 105: 86-105. [30] Herbst, Chris. 2015. "The Rising Cost of Child Care in the United States: A Reassessment of the Evidence." IZA Discussion Paper Series, No. 9072. [31] C. Herbst and E. Tekin. 2010. “Child care subsidies and child development.” Economics of Education Review, 29: 618 - 638. [32] C. Herbst and E. Tekin. 2011. “Child care subsidies and childhood obesity.” Review of Economics of the Household, 9: 349-378. [33] V. J. Hotz and M. Xiao. 2011. “The Impact of Regulations on the Supply and Quality of Care in Child Care Markets.” American Economic Review, 101: 1775 - 1805. [34] Kimmel, Jean. 1995. “The Effectiveness of Child-Care Subsidies in Encouraging the Welfare-to-Work Transition of Low-Income Single Mothers.” The American Economic Review, 85(2): 271-275. [35] Kimmel, Jean. 1998. “Child Care Costs as a Barrier to Employment for Single and Married Mothers.” The Review of Economics and Statistics, 80(2): 287-299. [36] L. Kotlikoff and L. Summers. 1987. “Tax Incidence,” in Handbook of Public Economics Vol. 2, A. Auerbach and M. Feldstein, eds. (Elsevier Press, North-Holland), 1043-1092. [37] M. Kottelenberg and S. Lehrer. 2014. “Do the Perils of Universal Childcare Depend on the Child’s Age?” CESifo Economic Studies, 60(2): 338-365. [38] M. Kottelenberg and S. Lehrer. 2016. “Targeted or Universal Coverage? Assessing Heterogeneity in the Effects of Universal Child Care." NBER Working Paper Series, No. 22126. [39] Leigh, Andrew. 2010. “Who Benefits from the Earned Income Tax Credit? Incidence among Recipients, Coworkers and Firms.” The B.E. Journal of Economic Analysis and Policy, 10(1), Article 45: 1-41. [40] B. Miller and K. Mumford. 2015. “The Salience of Complex Tax Changes: Evidence from the Child and Dependent Care Credit Expansion.” National Tax Journal, 68(3): 477 - 510.

28

[41] H. N. Mocan and E. Tekin. 2003. “Nonprofit Sector and Part-Time Work: An Analysis of EmployerEmployee Matched Data on Child Care Workers.” The Review of Economics and Statistics, 85(1): 38-50. [42] M. Prada, G. Rucci, and S. Urzúa. 2015. “The Effect of Mandated Child Care on Female Wages in Chile.” NBER Working Paper Series, No. 21080. [43] Rohacek, Monica. 2012. “A Summary of Research on How CCDF Policies Affect Providers.” Urban Institute, Washington, DC, http://www.urban.org/research/publication/summary-research-how-ccdfpolicies-affect-providers. [44] Rothstein, Jesse. 2010. “Is the EITC as Good as an NIT? Conditional Cash Transfers and Tax Incidence.” American Economic Journal: Economic Policy, 2(1): 177-208. [45] Sallee, James M. 2011. “The Surprising Incidence of Tax Credits for the Toyota Prius.” American Economic Journal: Economic Policy, 3: 189-219. [46] Turner, Nicholas. 2011. “The Effect of Tax-Based Federal Student Aid on College Enrollment.” National Tax Journal, 64(3): 839-862. [47] Turner, Nicholas. 2012. “Who benefits from student aid? The economic incidence of tax-based federal student aid.” Economics of Education Review, 31: 463-481. [48] E. G. Weyl and M. Fabinger. 2013. “Pass-Through as an Economic Tool: Principles of Incidence under Imperfect Competition.” Journal of Political Economy, 121(3): 528-583.

29

Figure 1: Child and Dependent Care Credit - Maximum Amount

The Child and Dependent Care Credit was increased in 2003. Many states offer additional state credits that are a percentage of the federal credit, so the total credits available (state plus federal) varies between states much more substantially after the expansion. This example shows the baseline federal amount against one of the most generous states (New York).

30

Figure 2: Combined (State + Federal) Child Care Credit Variation

The total child care credits (state plus federal) varies substantially between states and over time. The bars are maximum credit amounts, before adjusting for tax liability, by state and period (before and after the federal expansion). States without a state child care credit are included in the US group, as their residents only receive the federal credit.

31

Figure 3: Geographic Child Care Credit Variation

A geographic view of the change in the maximum credit available for one child in each state before adjusting for tax liability. Total credits are a combination of state and federal amounts. Alaska does not have a state credit. Hawaii’s credit is independent of the federal credit. New Mexico’s credit is reduced by the size of the federal credit.

32

Figure 4: Credit Incidence

The child care credit, represented by subsidy c, will shift the demand curve out and result in a new equilibrium price. The pass-through to the suppliers depends on the relative elasticities of both sides of the market and D can be calculated by ∆P ∆T = S −D . I’ve drawn this chart to reflect my results, which suggest a relatively less elastic supply curve.

33

Figure 5: Median Hourly Wage Trends - OES

Visual of OES data. Parallel trends in the pre-period appear plausible and were tested empirically. See Section 3.1 for details.

34

Figure 6: Mean Adjusted Child Care Credits by State

Based on the main SIPP sample used in the regression analysis, this chart collapses child care credits (adjusted for tax liability and a function of maximum spending limits) into state means, all in 2002 dollars. See Section 4 for details.

35

Figure 7: Maximum Adjusted Child Care Credits and Price

Collapsing the credit amounts into centiles and plotting the mean prices against the mean credits generates a positive correlation. Only two-earner families and single parents; adjusted to 2002 dollars. See Section 4 for details on calculating the credit based on maximum expenses.

36

Figure 8: Hourly Wage Change due to Child Care Credit Expansion - OES

Coefficients on the interaction terms between year dummies and the child care worker dummy from the difference-in-differences regression using the OES data. These results use a less restrictive specification to highlight how the timing of the increase in child care worker wages coincides with the CDCC expansion. Including an interaction term between the child care workers and the state GDP, median income, labor force participation of mothers with children under age 5, or unemployment does not affect the results. See Figure 9 for another way to tackle the macroeconomic trend concern.

37

Figure 9: Differential Hourly Wage Change due to Child Care Credit Expansion - OES

Coefficients on the interaction terms between year dummies and the child care worker dummy from the difference-in-differences regression using OES data. These results were generated by separately running the regression for the states in the top tercile of credit generosity and then again for the bottom tercile of states. The wage response is larger for the more generous states, as would be expected if credits caused the wage response. This evidence addresses concerns that the pattern of increased credits was driven by macroeconomic trends that differentially affected child care workers.

38

Figure 10: Heterogeneous Hourly Wage Change due to Child Care Credit Expansion - OES

Coefficients on the interaction terms between year dummies and the child care worker dummy from the difference-in-differences regression using OES data. These results were generated by separately running the regression using the 25th percentile, the median, and the 75th percentile of wages. The wage response increases with the wage percentile, which suggests increasing pass-through rates for different segments of the child care market, though the standard errors make them statistically indistinguishable. The pattern remains when interacting the treated group with various macroeconomic variables, as discussed in Section 5.1.

39

Figure 11: Maximum Credit Increase by Income

Using the IRS Statistics of Income (SOI) Public Use Microdata Sample (PUMS) for 2002, I run households who filled out the Form 2441 to claim the Child and Dependent Care Tax Credit through TAXSIM to calculate their maximum credit in 2002. Then, adjusting the inputs for inflation, I run them through TAXSIM again to find their maximum credit in 2003. These bins are chosen to match the SOI aggregate statistics, which show very similar changes by bin despite being affected by behavioral responses and selection issues.

40

Figure 12: Synthetic Controls

Synthetic control technique comparing the wages of the child care industry to those in the weighted control group. The average demographic information (sex, race, education, age, and marital status) from the American Community Survey at the state-year-occupation level was merged to the Occupation Employment Survey data. The synth command in Stata, written by J. Hainmuller, A. Abadie, and A. Diamond, matches the outcome variable and selected control variables of the control group to those of the treated group during the pre-intervention period. Child care industry wages matched closely in the pre-period to a weighted combination of teaching assistants, cashiers, and kindergarten teachers. The wages in both groups follow a similar trend until the expansion of the CDCC, after which the child care industry wages continue in a positive direction while the synthetic control group wages decline.

41

Figure 13: Distributional Effects of Increasing Non-Refundable Credit

Distribution of adjusted gross income after taxes and child care (expenses and credits) from the main SIPP sample. Increasing the CDCC with a pass-through rate of 0.75 reduces this measure of income for low income households. This is due to the high pass-through and non-refundability of this credit. Put another way, increasing these credits in their current form effectively transfers money from poor to rich households. This visual uses a $5000 credit increase in order to highlight the changes, but the qualitative result is the same for smaller credit increases. The light outlines use +/− one standard error of my pass-through estimates.

42

Figure 14: Distributional Effects of Increasing Refundable Credit

Distribution of adjusted gross income after taxes and child care (expenses and credits) from the main SIPP sample. The green line demonstrates that increasing the CDCC by the same amount as in Figure 13 yields a very different result when the credit is made refundable. In this case, families across the distribution are made better off, abstracting from the costs of this program. The light outlines use +/− one standard error of my pass-through estimates.

43

Figure 15: Variation in Pass-Through Rates

Change in child care expenses, net of credits, after an increase in the credit of $1000. Considering the full range of pass-through rates it is clear that, with the exception of very low rates that are unlikely given the supply challenges discussed in Section 6, increasing the CDCC disproportionately hurts lower income families. Moreover, the total change in this income measure turns negative for pass-through rates near my upper estimates of 0.90. This demonstrates that, with a high enough pass-through, increasing the CDCC not only redistributes from poor to rich but makes the total group worse off. To be clear, a refundable credit would follow the path of the top tercile, as this is just the difference between the full credit amount and the amount captured in the form of higher prices.

44

Figure 16: Heterogeneous Pass-Through Rates

Change in child care expenses, net of credits, after an increase in the credit of $1000, allowing for heterogeneous pass-through rates. Starting with the SIPP income distribution and a baseline pass-through of 0.75, increasing the credit increases the burden of child care for the bottom of the distribution. With segmented markets where the bottom tercile experiences a pass-through rate 3/4 or half of the top tercile’s pass-through, the pattern is less pronounced. With lower pass-through there are some households that have enough tax liability to offset the smaller increase in price, but the households with the lowest income will still end up paying more than the size of their claimable credits. Maintaining a mean pass-through of 0.75 necessitates the increase in the top tercile’s pass-through, thereby reducing their tax benefit in the latter two scenarios.

45

Table 1: Form 2441 - Reimbursement Rates Before 2003 IF your adjusted gross income is: Over: But not over: $0 — 10,000 10,000 — 12,000 12,000 — 14,000 14,000 — 16,000 16,000 — 18,000 18,000 — 20,000 18,000 — 22,000 22,000 — 24,000 24,000 — 26,000 26,000 — 28,000 28,000 — No limit

2003 - present THEN the percentage is: .30 .29 .28 .27 .26 .25 .24 .23 .22 .21 .20

IF your adjusted gross income is: Over: But not over: $0 — 15,000 15,000 — 17,000 17,000 — 19,000 19,000 — 21,000 21,000 — 23,000 23,000 — 25,000 25,000 — 27,000 27,000 — 29,000 29,000 — 31,000 31,000 — 33,000 33,000 — 35,000 35,000 — 37,000 37,000 — 39,000 39,000 — 41,000 41,000 — 43,000 43,000 — No Limit

THEN the percentage is: .35 .34 .33 .32 .31 .30 .29 .28 .27 .26 .25 .24 .23 .22 .21 .20

Note: Reimbursement rates taken from the IRS Form 2441. A household multiplies their annual child care expenses (or the capped limit which is currently $3,000, previously $2,400) by the appropriate reimbursement rate based on their adjusted gross income. This potential credit is then adjusted to account for tax liability due to the non-refundable nature of the credit.

Table 2: OES Summary Statistics Control Child Care Industry mean sd mean sd Median Hourly Wage 10.13 4.44 8.40 1.43 State minimum wage 5.42 0.60 5.43 0.60 Fraction of population under age 5 0.07 0.01 0.07 0.01 Median State Income 42953 6454 42962 6450 State unemployment rate 4.72 1.14 4.72 1.14 Log CCDF funding 17.34 1.04 17.33 1.04 LFP mothers with kids under 5 0.66 0.07 0.66 0.07 Observations 2278 915 Note: 1999-2007 Bureau of Labor Statistics Occupational Employment Statistiscs (BLS OES) data adjusted to 2002 dollars. State control variables from the Census, Department of Health and Human Services, Tax Policy Center, and BLS. Control group made up of teaching assistants, kindergarten teachers, cashiers, waiters, and residential advisors. The child care industry is child care workers and pre-school teachers. 46

Table 3: SIPP Summary Statistics mean sd Total CC credit, C(Ei ) 577 316 Total CC credit, C(Emax ) 645 327 Total CC credit, C(Esim ) 575 295 Hourly price of formal CC 3.60 3.40 Annual Income 74667 54611 Weekly (total) hours in formal CC 38.6 22.9 Average (per kid) weekly hours in formal CC 31.9 14.0 Weekly cost of formal CC 116 93 Children under 6 years old 1.27 0.48 Total kids in formal CC 1.21 0.43 Number of infants (0-1 years) 0.28 0.46 Number of toddlers (2-3 years) 0.49 0.54 Number of children (4-5 years) 0.51 0.53 Parent has high school degree or some college 0.44 0.50 Parent has college degree or higher 0.54 0.50 Single parent 0.19 0.40 Lives in urban / metro area 0.80 0.40 Black 0.11 0.31 Asian 0.04 0.19 Hispanic 0.08 0.27 Observations 3616 Note: 1999-2011 Survey of Income and Participation (SIPP) data adjusted to 2002 dollars. Sample restricted to single-parent and two-earner married families with children under age 6 who paid for formal child care without outside assistance. Simulated credit based on state data from IRS Statistics of Income.

47

Table 4: OES Results

(job ∈ CC industry) X (t ≥ 2003)

Dependent Variable: Median Hourly Wage (1) (2) (3) (4) 0.212*** 0.237*** 0.238*** 0.180*** (0.067) (0.058) (0.058) (0.067)

(job ∈ CC) X (t ≥ 2003) X (generous)

0.119 (0.117)

State minimum wage

0.047 (0.080)

0.048 (0.077)

Fraction of population under age 5

-11.398 (20.744)

-10.580 (19.724)

Log of state median annual income

0.550 (0.508)

0.431 (0.490)

State unemployment rate

0.022 (0.037)

0.014 (0.037)

Log CCDF funding

-1.220** (0.485)

-1.144** (0.471)

LFP mothers with kids under 5

0.600* (0.308)

0.551* (0.294)

Year FE

X

X

X

X

State FE

X

X

X

X

Industry FE

X

Job FE X X X Observations 3193 3193 3193 3193 R2 0.111 0.904 0.904 0.904 Mean of Dependent Variable 9.64 9.64 9.64 9.64 Note: 1999-2007 Bureau of Labor Statistics Occupational Employment Statistiscs (BLS OES) data adjusted to 2002 dollars. Control group made up of teaching assistants, kindergarten teachers, cashiers, waiters, and residential advisors. The child care industry is child care workers and pre-school teachers. Standard errors clustered at state level: *** p < 0.01, ** p < 0.05, * p < 0.1

48

Table 5: Year Interaction Results - OES

(job ∈ CC industry)X(t = 1999)

Dependent Variable: Median Hourly Wage (1) (2) (3) All States Least Generous Most Generous 0.186** -0.022 0.155 (0.084) (0.118) (0.122)

(job ∈ CC industry)X(t = 2000)

-0.018 (0.070)

-0.086 (0.090)

-0.071 (0.170)

(job ∈ CC industry)X(t = 2001)

-0.015 (0.029)

-0.068 (0.043)

-0.015 (0.064)

(job ∈ CC industry)X(t = 2003)

0.054 (0.035)

0.044 (0.070)

0.082 (0.080)

(job ∈ CC industry)X(t = 2004)

0.267*** (0.071)

0.362*** (0.122)

0.280** (0.131)

(job ∈ CC industry)X(t = 2005)

0.375*** (0.095)

0.390*** (0.129)

0.544** (0.214)

(job ∈ CC industry)X(t = 2006)

0.422*** (0.093)

0.447*** (0.110)

0.563** (0.226)

(job ∈ CC industry)X(t = 2007)

0.261*** (0.087)

0.362*** (0.079)

0.362** (0.170)

Year FE

X

X

X

State FE

X

X

X

Job FE

X

X

X

State Controls X X X Observations 3193 1132 1055 R2 0.904 0.905 0.896 Mean of Dependent Variable 9.64 9.28 10.01 Note: 1999-2007 Bureau of Labor Statistics Occupational Employment Statistiscs (BLS OES) data adjusted to 2002 dollars. Control group made up of teaching assistants, kindergarten teachers, cashiers, waiters, and residential advisors. The child care industry is child care workers and pre-school teachers. Column (1) is the main regression, Column (2) is the regression using only the states in the bottom tercile of child care credit generosity, and Column (3) uses only the states in the top tercile of credit generosity. Standard errors clustered at state level: *** p < 0.01, ** p < 0.05, * p < 0.1

49

Table 6: SIPP Results Dependent Variable:

Total CC credit, C(Ei )

Price

Price

C(Ei )

Price

C(Ei )

Price

(1) OLS (Endogenous) 2.054*** (0.283)

(2) Strategy 1

(3) Strategy 2A 1st Stage

(4) 2A

(5) Strategy 2B 1st Stage

(6) 2B

Total CC credit, C(Emax )

0.740 (0.513) 0.512 (0.358)

0.890* (0.528)

0.692*** (0.043)

Total CC credit, C(Esim )

0.761*** (0.053)

High school grad

0.256 (0.268)

0.323 (0.256)

0.018 (0.016)

0.310 (0.259)

0.016 (0.016)

0.303 (0.257)

College grad

0.599** (0.289)

0.674** (0.284)

0.022 (0.017)

0.658** (0.286)

0.020 (0.017)

0.651** (0.285)

Single parent

-0.595*** (0.189)

-0.480** (0.198)

0.034*** (0.008)

-0.505** (0.204)

0.042*** (0.009)

-0.515** (0.207)

Total kids in formal CC

-0.727*** (0.124)

-0.390*** (0.148)

0.057*** (0.011)

-0.432*** (0.168)

0.053*** (0.014)

-0.466*** (0.158)

Lives in urban / metro area

0.782*** (0.143)

0.801*** (0.141)

0.010** (0.005)

0.793*** (0.140)

0.009* (0.005)

0.792*** (0.140)

Year FE

X

X

X

X

X

X

State FE

X

X

X

X

X

X

Income bins X X X X X X Observations 3616 3616 3616 3616 3616 3616 R2 0.146 0.139 0.736 0.143 0.717 0.143 Mean of Dep Var 3.6 3.6 0.4 3.6 0.4 3.6 F-stat 253.8 207.1 Note: 1999-2011 Survey of Income and Participation (SIPP) data adjusted to 2002 dollars. Sample restricted to single-parent and two-earner married families with children under age 6 who paid for formal child care without outside assistance. Child care credits were converted to units of a $1 hourly subsidy. Simulated credit generated from IRS Statistics of Income data. Standard errors clustered at the state level: *** p < 0.01, ** p < 0.05, * p < 0.1

50

51

0.001 (0.026)

State unemployment rate

X

X X

State FE

Industry FE

X

X

0.146 (0.190)

X

X

X

(4) 0.236*** (0.081)

X

X

(5) 0.261*** (0.081)

X

X

0.798 (0.490)

-0.977 (0.629)

0.028 (0.045)

-0.068 (0.542)

-32.830 (23.947)

0.063 (0.092)

(6) 0.262*** (0.081)

DV: 75th Percentile Wages

Job FE X X X X Observations 2742 2742 2742 2742 2742 2742 R2 0.314 0.728 0.733 0.257 0.756 0.757 Mean of Dependent Variable 7.10 7.10 7.10 9.95 9.95 9.95 Note: 1999-2007 Bureau of Labor Statistics Occupational Employment Statistiscs (BLS OES) data adjusted to 2002 dollars. Additional variables from the Census, Department of Health and Human Services, and Tax Policy Center. Control group made up of teaching assistants (K-12), cashiers, waiters, and residential advisors. The child care industry is child care workers and pre-school teachers. Standard errors clustered at state level: *** p < 0.01, ** p < 0.05, * p < 0.1

X

X

Year FE

LFP mothers with kids under 5

-0.956*** (0.303)

0.012 (0.342)

Log of state median annual income

Log CCDF funding

-18.277 (13.284)

(3) 0.178*** (0.038)

Fraction of population under age 5

(2) 0.177*** (0.038) 0.207*** (0.050)

(1) 0.165*** (0.038)

State minimum wage

(job ∈ CC industry)X(t ≥ 2003)

DV: 25th Percentile Wages

Table 7: Heterogeneous OES Results

52 X X

X X

State FE

Job FE

X

X

X

-0.438 (0.637)

(3) 0.236*** (0.057)

X

X

X

-0.678 (0.637)

1.387 (1.054)

-0.115 (0.097)

(4) 0.317*** (0.073)

X

X

X

-0.052 (0.052)

(5)

X

X

X

DV: Log of total employment (6) 0.037 (0.026)

State Controls X X X X X X Observations 3193 3193 3193 3193 1415 3153 R2 0.904 0.904 0.904 0.904 0.898 0.953 Mean of Dependent Variable 9.64 9.64 9.64 9.64 9.55 8.76 Note: 1999-2007 Bureau of Labor Statistics Occupational Employment Statistiscs (BLS OES) data adjusted to 2002 dollars. Control group made up of teaching assistants, kindergarten teachers, cashiers, waiters, and residential advisors. The child care industry is child care workers and pre-school teachers. Columns (1) - (4) include interaction terms between the child care industry dummy and various macroeconomic variables to address concerns that the wage response is driven by other trends. Column (5) is a falsification check where CDCC expansion is set to occur one year earlier. Column (6) shows the change in total employment between the two groups after the policy change. Standard errors clustered at state level: *** p < 0.01, ** p < 0.05, * p < 0.1

X

1.844 (1.127)

X

-0.119 (0.094)

(2) 0.282*** (0.065)

Year FE

Placebo treatment: (job ∈ CC industry)X(t ≥ 2001)

(job ∈ CC industry)X(log state median income)

(job ∈ CC industry)X(LFP mothers with kids under 5)

(job ∈ CC industry)X(unemployment rate)

(job ∈ CC industry)X(t ≥ 2003)

(1) 0.290*** (0.073)

Dependent Variable: Median Hourly Income

Table 8: Additional Results - OES Robustness

53 X X

X X

State FE

Income bins

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

Tax liability spline (cubic) X X Observations 4717 4717 3383 3383 3616 3616 3616 3616 3616 3616 3616 3616 R2 0.144 0.145 0.140 0.141 0.142 0.143 0.136 0.137 0.137 0.138 0.143 0.144 Mean of Dependent Variable 3.81 3.81 3.67 3.67 3.60 3.60 3.60 3.60 3.60 3.60 3.60 3.60 F-Stat 438.30 344.84 163.93 136.34 270.59 228.33 271.49 209.80 275.14 216.72 185.59 154.98 Note: 1999-2011 Survey of Income and Participation (SIPP) data adjusted to 2002 dollars. Sample restricted to single-parent and two-earner married families with children under age 6 who paid for formal child care without outside assistance, though Columns (1)-(2) include married couples with one earner. Child care credits were converted to units of a $1 hourly subsidy. Simulated credit generated from IRS Statistics of Income data. Odd columns are second stage results using maximum credits as an instrument; even columns use the simulated credit. Columns (3)-(4) drop families with zero child care credits. Columns (5)-(6) use given survey weights. Columns (7)-(12) control for various income and tax liability splines. Standard errors clustered at the state level: *** p < 0.01, ** p < 0.05, * p < 0.1

Income spline (linear)

Income spline (cubic)

X

X

Year FE

-0.619*** (0.206)

0.652** (0.280)

0.291 (0.262)

0.793*** (0.140)

0.296 (0.263)

Lives in urban / metro area 0.888*** 0.888*** 0.800*** 0.799*** 0.796*** 0.795*** 0.784*** 0.782*** 0.788*** 0.787*** 0.795*** (0.138) (0.138) (0.149) (0.148) (0.147) (0.147) (0.140) (0.140) (0.141) (0.141) (0.140)

0.236 (0.257)

-0.395** (0.157)

0.244 (0.259)

-0.473*** -0.493*** -0.392** -0.424** -0.447*** -0.490*** -0.359** -0.399** -0.403** -0.445*** -0.349** (0.108) (0.103) (0.190) (0.185) (0.161) (0.147) (0.170) (0.158) (0.169) (0.157) (0.165)

0.234 (0.258)

Total kids in formal CC

0.241 (0.259)

-0.416** -0.413** -0.483** -0.491** -0.517** -0.529** -0.580*** -0.595*** -0.586*** -0.602*** -0.613*** (0.163) (0.162) (0.222) (0.225) (0.208) (0.209) (0.209) (0.213) (0.212) (0.216) (0.205)

0.404 (0.256)

Single parent

0.410 (0.256)

0.678*** 0.668*** 0.316 0.313 0.757*** 0.750*** 0.611** 0.603** 0.611** 0.602** 0.658** (0.226) (0.229) (0.377) (0.377) (0.283) (0.282) (0.279) (0.278) (0.281) (0.281) (0.280)

0.073 0.071 (0.327) (0.326)

(12) 0.627 (0.579)

College grad

0.249 (0.228)

(11) 0.422 (0.555)

0.258 (0.227)

(10) 0.769 (0.494)

High school grad

Total CC credit, C(Ei )

Dependent Variable: Hourly Price of Child Care (1) (2) (3) (4) (5) (6) (7) (8) (9) 0.662** 0.783** 0.694 0.822 0.757 0.947* 0.417 0.593 0.583 (0.331) (0.371) (0.716) (0.759) (0.467) (0.486) (0.463) (0.465) (0.490)

Table 9: SIPP Robustness Checks

A

Appendix

A.1

American Community Survey Difference-in-Differences

The difference-in-differences (DD) analysis presented above using the Occupational Employment Statistics (OES) was replicated using the American Community Survey (ACS). Since the ACS data is at the individual level it is possible to include race, education, age, and sex as additional demographic controls. The empirical strategy is identical to the one presented in Section 4.1 and the regression in Equation 2 is modified only by the addition of individual controls. State, year, and job fixed effects are included, and I once again use OLS and cluster at the state level. Summary statistics for the ACS data are shown in Table A1. The treated group is once again child care workers and pre-school teachers since they may have been affected by the CDCC expansion. The control group, which should not have been affected by the CDCC expansion, is again teaching assistants, kindergarten teachers, cashiers, waiters, and residential advisors. As before, the treated group earned slightly less than the control group, likely due to the presence of kindergarten teachers in the latter. Now that individual controls are available, it is clear that there are less male employees in child care and that child care workers are slightly older. These differences highlight the need to control for these observable characteristics in the regression framework. Child care workers are also much more likely to be self-employed, likely due to family day care homes and nannies. This variable presents another opportunity for comparison to the main analysis, since the OES does not sample self-employed workers. Figure A1 shows the wage trends between the child care industry and the control group. Once again, it appears that before 2003 the two groups exhibit parallel trends and that after the CDCC expansion the child care workers’ wages do not decline nearly as much as the control group. Different pre-trends were tested empirically with p-values never dropping below 0.39. The DD results using the ACS are shown in Table A2. The coefficient of interest in the first row is again consistently positive and statistically significant. Column (1) starts with the most parsimonious specification and each column thereafter adds other control variables. Since these data are at the individual level I can control for demographic characteristics, effectively reducing the magnitude of the point estimates in Columns (3) and (4) to be in line with the OES results. The range of ACS estimates suggest that expansion of the CDCC increased child care worker wages 54

between $0.27 and $0.36, a wider range than the OES but still positive and significant. Alternate specifications described in Section 8 proved similarly robust when using the ACS data and, while I omit them to conserve space, Figure A2 plots the coefficients on the interaction terms between the year dummies and the child care industry dummy. While not as clear as the the OES version in Figure 8, the pattern remains that the child care worker wage increase appears to be timed with the CDCC expansion. Lastly, I ran a placebo test with the ACS data to investigate the likelihood of estimating this wage response by chance. After dropping child care workers and randomly assigning the same ratio of treatment to workers from the control group and the larger Personal Care and Service Occupations group, I estimated the same DD regressions 500 times. As shown in Figure A3, my main estimate is larger than most of the placebo distribution (p-value = 0.072), further supporting the interpretation that the estimated wage increase of child care workers is the result of child care credit generosity.

A.2

Additional Instruments: SIPP

Two additional methods for generating an instrument for the SIPP analysis are included in this appendix. Both methods are intended to demonstrate that the main instruments described in Section 4.2 were not biased by the use of household income in calculating the instrumental credit. The concern is that the decision to work and possibly how much to earn (via which job to take) is endogenous to the price of child care; hence, I would not be able to make the assumption that the instrument is uncorrelated with unobservables in the second stage. Both additional methods demonstrate that high pass-through is likely even when using alternative inputs for TAXSIM. The first approach is to use average credits for a given family type as an instrument for actual credits. I took every family with children under the age of 6 before 2003 in the SIPP sample and ran them through TAXSIM using maximum expenses. I recoded every family to be from each state and adjusted inputs for inflation in order to do this for every year in the sample. Thus, the same pre-expansion distribution of families was used to calculate total credits in every state-year combination. I then took the average credit for each family type based on race, age, marital status, education, family size, state, and year bins. It is important to mention that this sample includes zeros generated from married households where only one parent works. Thus, the instrument is 55

now the average credit a family of that type, regardless of their endogenous work decisions, would have received under the given state-year policy. The second approach is to use predicted income based on exogenous variables as an input to the TAXSIM calculator, once again circumventing the endogenous labor force decision. I regressed wages on age, age squared, and dummies for race, education, and state of residence using the pre2003 SIPP households with children under the age of 6. The predicted wages, based strictly on these controls, were used as inputs to the TAXSIM calculator. The instrument in this case is the maximum credit a household would receive under a given state-year policy assuming they earned income predicted by their household characteristics. These results are shown in Table A3. The left panel includes the results of the averaged IV and the right panel includes the simulated income IV. Reduced form estimates in Columns (1) and (4) show that for every dollar in IV credits the price of child care increases between $0.66-0.8, very similar to the main results in Table 6. The first stage results are shown in Columns (2) and (3), both of which have large F-statistics though, as to be expected, are not quite as large as the main estimates. The first stage coefficients are smaller and therefore cause the second stage results, in Columns (3) and (6), to be much larger than the main estimates. My interpretation of these alternative IV results is that my main results are not driven by endogenous household decisions related to income and that my main SIPP pass-through estimates may be conservative.

A.3

Child Care Market Rate Survey Data

My decisions to use the SIPP, OES, and ACS datasets were primarily due to the absence of reliable child care price data. In addition to the above analysis, I also generated a first-of-its kind dataset in an attempt to estimate the incidence of child care tax credits using better price data. In order to receive Child Care and Development Fund (CCDF) block grant funds, states are required to conduct a market rate survey of child care facilities at least every two years. It has been previously noted that there is substantial variance in the way these surveys are conducted (Rohacek, 2012), yet they remain the most detailed and standard measurement of child care prices that pre-dates the 2003 CDCC expansion. I was able to gather the summary reports of these surveys from 22 states, something that has yet to be done due to the idiosyncrasies of various state

56

record keeping and public information policies.39 Some states produce much more detailed and informative reports while others are very brief, occasionally including nothing more than a table of prices. In any case, I focus on the 75th percentile of child care prices at formal child care centers in the available states.40 Thus, abstracting from the caveats about state survey differences, I have compiled what may be the best data on child care prices across states over this timeframe. The next challenge was to quantify the generosity of child care credits appropriate for this analysis. My approach addresses both endogeneity of spending decisions and non-refundability. To start, I calculate the maximum predicted federal credit available in each year, which is identical across states before and after the federal credit expansion. Next, I adjust for tax liability using the aggregate state IRS SOI data which I used in the simulated instruments context. The IRS data allow me to calculate the average credit in a given state-year-income range. The ratio of this average credit over the maximum credit can be thought of as the average credit after adjusting for tax liability, which addresses the non-refundability issue. The second modification uses the same fraction of actual over maximum credits, but propagates that fraction across years. In other words, the after-adjustment ratio in a base year is multiplied by the maximum credit in all other years. This results in credits adjusted for tax liability and spending decision changes. Finally, I take a weighted average of the adjusted tax credits based on how many people received federal credits in each state-year-income range. Thus, I have the average credit that would have been claimed in any given state-year, adjusted for tax liability and primarily only a function of state and federal child care credit policies. Due to the small sample size, unbalanced panel, and alternating survey years, regression analysis is not feasible at this point, though I continue to work with the remaining states in hopes of expanding my dataset. However, a plot of prices against credit generosity shown in Figure A4 yields the same pattern: higher credits are correlated with higher prices. Despite my attempts to sterilize my credit variables of the aforementioned endogeneity, this pattern could still be affected by reverse causation. In any case, this additional plot complements the earlier analysis with a novel dataset. 39

For example, some states had these reports destroyed after a certain period of time, while others were lost in accidental fires or hard drive failures. Ten additional states provided what reports they had but were not usable due to a lack of pre-CDCC expansion observations. 40 I use the 75th percentile because it is required by the CCDF rules and is therefore the most commonly reported price. I use child care centers because there is far less missing data for these facilities.

57

Figure A1: Hourly Wage Trends - ACS

Visual of ACS data. Parallel trends in the pre-period appear plausible and were tested empirically. See Section A.1 for details.

58

Figure A2: Hourly Wage Change due to Child Care Credit Expansion - ACS

Coefficients on the interaction terms between year dummies and the child care worker dummy from the difference-in-differences regression using the ACS data. The control group comprises teaching assistants (K-12), kindergarten teachers, cashiers, waiters, and residential advisors. These results use a less restrictive specification to highlight how the timing of the increase in child care worker wages coincides with the CDCC expansion. While not as clear as the OES version in Figure 8, the visual supports the regression results.

59

Figure A3: Placebo Test

Distribution of DD coefficients generated from placebo treatments using ACS data. After dropping child care workers and randomly assigning the same ratio of treatment to the control group and the larger Personal Care and Service Occupations group, the same DD regression was estimated 500 times. The red line indicates the most conservative estimate from my main ACS estimates in Table A2.

60

Figure A4: Market Rate Survey Prices Against Credit Generosity

Infant child care prices and credit generosity for years 1999-2008, in 2002 dollars. Prices compiled from state-level Child Care Market Rate Surveys. States include Arizona, California, Georgia, Idaho, Illinois, Indiana, Iowa, Louisiana, Maine, Maryland, Massachusetts, Missouri, Nebraska, New Jersey, New York, North Carolina, North Dakota, Ohio, Oregon, Texas, Washington, and Wyoming. Tax credits generated using IRS SOI data.

61

Table A1: ACS Summary Statistics Control Child Care Workers mean sd mean sd Hourly wage 8.64 5.66 8.16 5.58 Male 0.21 0.41 0.04 0.21 Black 0.10 0.29 0.12 0.33 Hispanic 0.12 0.33 0.13 0.33 Asian 0.05 0.21 0.02 0.15 Other race 0.01 0.10 0.01 0.10 High school graduate 0.68 0.47 0.70 0.46 College degree or higher 0.10 0.30 0.16 0.37 Age 32.2 13.9 36.6 13.3 Married 0.34 0.47 0.52 0.50 Self-employed 0.01 0.10 0.31 0.46 State minimum wage 5.41 0.66 5.43 0.67 Fraction of population under age 5 0.07 0.01 0.07 0.01 Median state income 43259 5807 43549 5833 State unemployment rate 4.96 0.96 4.96 0.98 Log CCDF funding 18.17 0.95 18.15 0.96 LFP mothers with kids under 5 0.63 0.05 0.63 0.05 Observations 240977 70276 Note: 2000-2007 American Community Survey data adjusted to 2002 dollars. State control variables from the Census, Department of Health and Human Services, Tax Policy Center, and Bureau of Labor Statistics. Control group made up of teaching assistants (K-12), kindergarten teachers, cashiers, waiters, and residential advisors.

62

Table A2: ACS Results

(job ∈ CC industry)X(t ≥ 2003)

Dependent Variable: Hourly Wage (1) (2) (3) (4) 0.356*** 0.283*** 0.270*** 0.275*** (0.081) (0.078) (0.078) (0.079)

High school graduate

0.896*** (0.081)

0.896*** (0.081)

College degree or higher

2.901*** (0.121)

2.901*** (0.121)

Age

0.071*** (0.002)

0.071*** (0.002)

Married

0.237*** (0.057)

0.237*** (0.057)

Male

1.146*** (0.060)

1.146*** (0.060)

Black

0.522*** (0.115)

0.522*** (0.115)

-0.247 (0.167)

-0.247 (0.167)

-0.285*** (0.095)

-0.285*** (0.095)

0.146 (0.167)

0.146 (0.166)

Hispanic

Asian

Other race

Year FE

X

X

X

X

State FE

X

X

X

X

Industry FE

X X

X

X

Job FE

State controls X Observations 311253 311253 311253 311253 R2 0.029 0.074 0.128 0.128 Mean of Dependent Variable 8.68 8.68 8.68 8.68 Note: 2000-2007 American Community Survey (ACS) data adjusted to 2002 dollars. Control group made up of teaching assistants (K-12), kindergarten teachers, cashiers, waiters, and residential advisors. The child care industry is made up of child care workers and pre-school teachers. State controls include minimum wage, log state median income, unemployment rate, log of CCDF funding, fraction of population under 5, and labor force participation of mothers with children under age 5. Standard errors clustered at state level: *** p < 0.01, ** p < 0.05, * p < 0.1

63

Table A3: Additional IV Results Dependent Variable:

Total CC credit (averaged IV)

Price

Actual Credit

(1) Reduced Form 0.792* (0.456)

(2) 1st Stage 0.442*** (0.045)

Total CC credit (actual)

Price

Price

(3) (4) 2nd Stage Reduced Form

Actual Credit

Price

(5) 1st Stage

(6) 2nd Stage

1.752* (1.043)

Total CC credit (sim. income IV)

2.002 (1.484) 0.658 (0.504)

0.329*** (0.027)

Year FE

X

X

X

X

X

X

State FE

X

X

X

X

X

X

Income bins

X

X

X

X

X

X

Demographic controls X X X X X X Observations 3611 3605 3605 3616 3616 3616 0.146 0.590 0.152 R2 0.147 0.563 0.153 Mean of Dep Var 3.6 0.4 3.6 3.6 0.4 3.6 F-stat 95.2 149.5 Note: 1999-2011 Survey of Income and Participation (SIPP) data adjusted to 2002 dollars. Sample restricted to single-parent and two-earner married families with children under age 6 who paid for formal child care without outside assistance. Child care credits were converted to units of a $1 hourly subsidy. Instrument in columns (1)-(3) is averaged maximum credit amount by race, age, marital status, education, family size, state, and year bin. Instrument in columns (4)-(6) is maximum credit using predicted income as input to TAXSIM. Standard errors clustered at the state level: *** p < 0.01, ** p < 0.05, * p < 0.1

64