Heller Hurwicz Data Initiative Spring Semester, 2015 Summary

Transcription

1 Heller Hurwicz Data Initiative Spring Semester, 2015 Summary Zachary Mahone September 9, 2015 The Heller Hurwicz Data Initiative is a longitudinal data gathering project housed in the Heller Hurwicz Economics Institute. Driven by undergraduate researchers, it provides educational outreach to the university community and simultaneously generates a valuable resource for economists. The [current] project constructs and maintains a publicly available unemployment insurance (UI) database and associated calculator that can be used in conjunction with panel data of US workers to construct estimates of the UI benefits available to them. Below we detail the progress made by the HHDI in the Spring semester of 2015 and provide some basic descriptions of the data. 1 Background and Summary Unemployment insurance is an important part of the safety net in the United States. It provides weekly income over a fixed period of time to an unemployed member of the labor force who has lost a job through no fault of their own. Economists have long recognized, however, that by replacing lost income UI provides disincentives for job search. Thus, exactly how much income should be insured (referred to as the replacement rate, or the fraction of weekly earnings that UI replaces) and how it should be disbursed (over what time frame, at a constant or varying level etc.) are the subject of debate. In order to quantitatively estimate the costs and benefits of any program then, it is important to have an accurate measure of the option value a worker has if they do not find (or accept) a job. Over the years this has led to the construction of various UI calculators that take data on earnings, year and location and estimate the replacement rate available to the worker (see Gruber (1997), Chetty (2008) and Kuka and East (2014) for examples). The construction of these calculators is time intensive because each state runs their own UI program; i.e. the criteria for determining replacement rates, maximum and minimum payouts and eligibility change from state to state, and this information is not centrally aggregated. The aforementioned papers have each constructed calculators for some period of time, but these efforts are not ongoing. The HHDI takes up the work where Kuka and East (2014) left off, aggregating state level data and annually updating 1

2 the calculator. Valuable advice for this project was provided at the start by Jonathan Gruber and the initial dataset and calculator was provided by Kuka and East. The database and calculator is will be made publicly available by the end of the summer of Progress in Spring 2015 During the Spring semester of 2015, five undergraduate students were employed as research assistants on the HHDI. They were trained in the basics of the Stata language and assigned a set of states for which they were responsible. Both the dataset and code were updated from 2007 to 2014 using annual Department of Labor summary reports on UI laws by state as the primary source. These five RAs (in alphabetical order) are Ryan Dias, Brady Durst, Edward Linds, Minh Tran and Ryan Webster. 2 Description of the Dataset and UI Calculator This project contains two interrelated components. The first is a database which details for each year, state and number of dependents the values of the minimum and maximum weekly benefit amounts (b min, b max ), along with the dates when these values take effect. The second is a Stata-based calculator that can be used with panel data on workers to create an estimate of available benefits. To do so, it uses the information in the dataset along with the benefit formulas as written in state legislation. The complication is that if y are earnings in the past twelve months, the relevant earnings for the benefit formula will be some function of y, I(y) and it will vary across states. If Texas, for example, computes unemployment benefits as x fraction of some measure of earnings in the last twelve months, I(y), with a maximum and minimum benefit of b max, b min, the code includes (x, I( ), b min, b max ). To summarize, in its current form the calculator contains all information on minimum and maximum benefits along with benefit formulas, while it ignores the dates in which laws come into effect (implicitly assuming this is January 1st in all cases). The dataset contains all information on minimum and maximum benefits along with the dates in which these takes effect, while not including benefit formulas. In addition, the calculator covers a shorter time period than the data set ( as compared to ). One are of future work will be to harmonize the information available in the dataset and calculator. 2.1 Terminology UI benefit computation is not terribly complicated but has its own jargon. This is a brief guide. Assume time is discrete and each period one month. Suppose a worker files for unemployment insurance at time T (we will assume at the beginning of the period). The base period for earnings will be the prior twelve months (y T 12, y T 11,..., y T 1 ) where y denotes total monthly income. We divide these into quarters: q 4 = (T 12, T 11, T 10), q 3 = (T 9, T 8, T 7) and so on. Summing earnings over the relevant months for 2

3 each quarter, we obtain quarterly earnings for the base period, y qj j {1,.., 4}. Exactly which statistics are used for computing benefits depends on the state. Let s call this I, as in the example above. Note that I will not generally be the income earned over the base period. Many states for example use I=high quarter wages, which would be max j In this case, if earnings are volatile at the quarterly frequency only the highest earning quarter matters. Other states use I=average of the two highest quarters to determine benefits. States also bound benefits above and below. Sometimes eligibility requirements put a lower bound on benefits. These are usually of the form I I MIN which has an equivalent b min given the benefit formula. Other times a b min is directly written into law. Similarly, a maximum payout is usually defined, such that b(i) b max. Knowledge of base period wages, how I is defined, and the eligibility and maximum benefit rules allow us to compute expected benefits. 2.2 The Dataset The dataset contains observations from 1967 to 2014 for all states in the US. States are identified in the variable statefip using their Federal Information Processing Standards (fips) code (although a separate, sequential, identifier state is provided to make looping easier). Because the formulae for computing benefits may change based on the number of dependents, the state variables are the triple (year, state, # dependents), where the number of dependents varies from one to thirteen. As of 2014, 14 states include consideration of dependents in their computation of UI benefits. For each (year, state, # dependents) triple, the minimum and maximum weekly benefit amounts (wba) are reported, stored in min and max, respectively. The date at which a change in law comes into effect is stored in nndate in the format yyyymmdd. In most cases, this is January 1st of the year in question. In some cases however, nndate may take on another value, such as This value of nndate would indicate that any change in the state s UI calculations for the year 1983 effects only claims made after March 1st. 2.3 The UI Calculator The UI calculator assumes earnings, location (state) and family (number of dependents) data is available. The code is straightforward: first estimate the wba for the worker given information on recent earnings, dependents, year and state of filing, using the formula as given in the state law. Next, check that the computed wba does not violate the upper or lower bounds on the wba as stated in the law. If the bounds are violated, re-set the value of the wba at the relevant bound. Finally, double check that eligibility requirements are satisfied. If not, re-set the value of the wba to zero Earnings The calculator assumes that at least the value of earnings in the three months prior to claiming UI are observable. Panels such as the Survey on Income and Program y qj. 3

4 Participation (SIPP) satisfy this requirement. Others, such as the National Longitudinal Survey of Youth (NLSY) or the Panel Study of Income Dynamics (PSID) would have to construct estimates of these values using data on hourly pay, hours per week and weeks worked. If only three month prior earnings are provided, then the code will impute this value to all other wage variables as appropriate. For example, high quarter wages and the average of wages in the two highest quarters will be given the same value (three month prior earnings). Base period wages (earnings over the period of twelve months immediately prior to filing a claim) will be imputed to be the three-monthly earnings multiplied by four. If a researcher has the data to construct more precise values for high-quarter wages, second-highest-quarter wages etc, the formulas in the code reflect the actual law, so more accurate values can be constructed Caveats and Future Work Because the initial code was designed to work with the SIPP which has some peculiarities regarding state identification, South Dakota, North Dakota, Wyoming, Vermont and Maine were not initially updated, so wba and eligibility estimates can be computed only for the period for these states. Future work should complete these missing years. Some codes are also suspicious although for the moment unchanged. In particular, the existing code for eligibility in Montana is difficult for us to map to the law. This results in non-eligibility for levels of earnings that are anomalous in comparison to all other states, and also means that eligibility is not determined after We are currently also missing eligibility from for states with fips codes 4,5,6,10,11 and 12. These should be updated in future work as well. 3 Descriptive Statistics of UI, In this section we perform a simple what if scenario, computing UI benefits and replacement rates for 12 imaginary earners (six earnings levels with zero or two dependents). The six annual earnings levels we consider are 5K, 14K, 20K, 30K, 50K and 80K. We keep the real level of earnings at these amounts, so equivalent earnings will inflate over time (as will the minima and maxima written into state law). The choices of annual earnings are random aside from the focus on the lower end of earnings since presumably these are the people for whom UI plays an important role. The exercise proceeds as follows. All earnings are assumed evenly divided by fiscal quarters and hours. This is NOT an innocuous assumption since most states have eligibility formulas that require a certain level of consistency in work. For example, many states have a requirement that high quarter wages (hqw) must be no more than two thirds of base period wages (bpw). To see the impact of this constraint, consider a worker who was not working for a period of nine months and then started a job, earning $5K in three months prior to being let go. Their bpw=$5k, and their wages by quarter (over the most recent twelve month period) are (0,0,0,$5K) for quarters 1-4, respectively. Clearly hqw=$5k. In this case, the requirement that bpw 1.5*hqw is not satisfied since the worker earned all of 4

5 their earnings in the high quarter and did not work in any other quarter of the base period, so they are ineligible to claim UI benefits. Our assumption that earnings are evenly distributed over the base period implies that the particular eligibility condition described above will always be satisfied. For those interested in marginal workers who work primarily seasonal jobs or otherwise tend to enter and exit the labor force often, these sorts of rules should be kept in mind. Using the earnings levels above and varying dependents to be zero or two, we then compute eligibility and replacement rates, where the replacement rate is defined as RR = wba (bpw/52). Below we report averages and standard deviations of replacement rates and eligbility to get an idea of both how large are differences across states and how generosity levels have changed over time. Note that these are not population weighted, so Texas and Rhode Island receive the same weight in reported statistics. We also stress that all of the reported values are preliminary, first-run analyses and may change as our database and code are updated in the future Eligibility In Figure 1 in the appendix we report for each year the fraction of states in which an earner of a given level would be eligible to receive benefits (only states with non-imputed eligibility are included here). For all but the lowest earner, our assumptions on the distribution of earnings over the base period guarantees UI eligibility in all states across all time periods. For the lowest level of annual earnings considered ($5K) we observe a modest decreast in elibility from , then reverting back to 100% eligibility by the end of the sample. Notice that this fall is only by a few percentage points, representing a few states only. This graph motivates our imputation of eligibility for all earners in the states and years in which we are currently missing eligibility formulas. Also note that the number of dependents does not effect eligibility in most cases and the figure is identical for two dependents. 3.2 Replacement Rates We now turn our attention to replacement rates, a crucial parameter in labor market models with unemployment benefits. We begin by plotting in Figure 2 (appendix) average replacement rates over time for the six earnings groups and two different dependent numbers (zero and two). We first notice that these two graphs look nearly identical; there is very little consideration given to dependents in most states (for UI at least, other safety net programs may be different). The most striking observation in either graph is the downward trend in average replacement rates for all groups. This is visually clearest for the lowest earners. It would appear that over the last 25 years then there has been a marked decline in the 1 To deal with the data issues mentioned in the section on Caveats, we impute eligibility to states where those formulas are missing. In practice, since all but the lowest earners are eligible across all states and years, we assume all earning levels are eligible for those observations missing eligibility formulas. Those states missing the wba formulas themselves are dropped from the analysis. 5

6 generosity of UI benefits as measured by replacement rates. For earnings groups 1-6, average replacement rates in 2013 expressed as a fraction of replacement rates in 1992 (i.e. RR 2013 RR 1992 ) are 0.57, 0.6, 0.65, 0.68, 0.68 and 0.68, respectively (this is for zero dependents; the numbers are practically identical for two dependents). In percentage terms, generosity has fallen in all groups although the size of the decline is weakly monotonic in earnings, with a ten percentage point difference between the highest and lowest earners we consider. We are also interested in variation across states. It may be that our conclusions about changes in UI generosity are being driven by a few outliers. It is also interesting in its own right to ask how much replacement rates vary across states holding constant ones earnings profile. Because we would expect that the standard deviation would be falling over time in step with the mean (and it is), we instead present scatter plots for the six earnings groups, displaying in each year the replacement rates for all states along with the median observed. These are shown in the appendix in Figure 3. The first thing to note is that the decline in generosity holds for all earnings groups for both the median and the upper and lower bounds observed. This is true for the lowest earnings group even if we exclude the substantially more and less generous states. Further, the variation in generosity across states is non-negligible, particularly for lower earners. For a worker earning $20K in the base period, in 2013 they could receive replacement rates anywhere from %10 to %24 depending on their state of residence. While interesting to note, exactly how large these differences are require theory. Since optimal replacement rates will be a function of wealth and other characteristics that may also vary across states (see Lentz (2009), for example) the variation above may not be that dramatic. Alternatively, in understanding the cyclical properties of unemployment there is some debate about the importance of replacement rates compared to other opportunity costs of employment (see Karabarbounis and Chodorow-Reich (2015)), which might mitigate the relative importance of these changes. 4 Conclusion While we continue to emphasize the preliminary nature of the results that we present above, the main conclusions we draw from our work is that the generosity of UI levels have fallen over the last 25 years in the United States. This is consistent (thankfully) with what Kuka and East (2014) report. Our focus on similar earners over time also allows us to see that the decline has been somewhat larger among low earners. This decline can be seen in average replacement rates, medians and the maximum and minimum rates across states and so appears to be a fairly robust result. 6

7 5 Bibliography Chetty, Raj. Moral Hazard vs. Liquidity and Optimal Unemployment Insurance, Journal of Political Economy 116(2): , Chodorow-Reich, Gabriel and Karabarbounis, Loukas. The Cyclicality of the Opportunity Cost of Employment, forthcoming Journal of Political Economy. East, Chloe and Kuka, Elira. Reexamining the Consumption Smoothing Benefits of Unemployment Insurance, forthcoming Journal of Public Economics, Gruber, Jonathan. The Consumption Smoothing Benefits of Unemployment Insurance, American Economic Review, 87(1), March 1997, p Lentz, Rasmus. Optimal Unemployment Insurance in an Estimated Job Search Model with Savings, Review of Economic Dynamics. January 2009, vol. 12(1), pp

8 6 Appendix Figure 1: Non-population weighted probability of eligibility Figure 2: Average Replace Rates Over Time 8

9 Figure 3: Replacement Rates Across States 9