Assessing Systematic Differences in Industry-Award Rates of Social Security Disability Insurance

Transcription

1 Assessing Systematic Differences in Industry-Award Rates of Social Security Disability Insurance Till von Wachter * University of California Los Angeles and NBER Abstract: Although a large body of literature has examined the role of worker incentives and characteristics for explaining high and rising Social Security Disability Insurance (SSDI) rolls, little is known about the role of firms and industries. Yet, firms face potentially important incentives to exploit social insurance systems, as explicitly recognized in the design of Unemployment Insurance, and may have substantially different work environments. In this project, we characterize differences in SSDI award rates by industry using the March Current Population Survey (CPS). The CPS allows us to construct multiple measures of SSDI receipt, use a longitudinal match to construct new SSDI awards, and to adjust industryspecific SSDI award rates for differences in worker characteristics, such as educational background, family status and occupation. Our findings indicate substantial differences in the average rates of new SSDI awards across industries that are robust to important differences in worker composition. * tvwachter@econ.ucla.edu. I thank Joe Song for excellent research assistance. This research was supported by the U.S. Social Security Administration through grant RRC to the National Bureau of Economic Research as part of the SSA Retirement Research Consortium. The findings and conclusions expressed are solely those of the author(s) and do not represent the views of SSA, any agency of the Federal Government, or the NBER. 1

2 1. Introduction The high and rising number of Social Security Disability Insurance (SSDI) beneficiaries has been an important subject of concern and debate among policy makers and academics alike. The overwhelming focus in explaining the high and increasing incidence of applications to SSDI in the academic literature has been on the characteristics and incentives of individuals (e.g., Bound and Burkhauser 1999, Duggan and Imberman 2006). For example, it has been shown that application rates to SSDI, and to a lesser degree, SSDI rolls expand in recessions (e.g., Rupp and Stapleton 1995), and that this effect has risen over time (e.g., Autor and Duggan 2003). However, the number of SSDI beneficiaries has been rising in between recessionary periods as well, and are now at historically high levels. Part of the secular trend is due to changes in the characteristics of the insured population. For example, Duggan and Imberman (2006) show that part of the increase can be explained by changes in the age structure and fraction female among workers insured by SSDI. While much of the literature in economics has studied the effect of SSDI on the behavior of insured workers, workers employers also face important incentives to influence and potentially to raise SSDI application and potentially award rates of their employees. Among others, these incentives can derive from an effort to lower costs from medical insurance, to ease the transition out of the labor force of less healthy and potentially less productive workers, or to reduce the work force of older, potentially more expensive workers. The incentives of firms to exploit a social insurance system whose costs are borne by society are explicitly recognized in the Unemployment Insurance (UI) system. In the UI system, firms experiencing a high inflow rate into UI are partially penalized by an increase in their 2

3 portion of UI taxes for the cost their high award rate imposes on the system. Existing research has shown that UI award rates at the firm and industry level respond to these incentives (e.g., Anderson and Meyer 1994). While the incentives faced by firms in the case of SSDI are potentially of similar nature, there is little known about systematic differences in award rates by firms or industries. Even if firms do not engage in explicit efforts to manipulate SSDI claims of their workers, it might still be that there are persistent differences in disability rates by firm, industry, or occupation arising from differences in work environments. Part of these differences may reflect the differences in the inherent riskiness of the job, or they may result from differences in workplace safety. Again, since contributions to Social Security are not experience rated, it is important to know to what extent the SSDI program implicitly subsidizes the employment health risk of workers and firms in certain industries. If it does, the policy question arises whether some of the costs of such risk would be better internalized by workers and firms. A large literature in labor economics has recognized that there are important systematic firm and industry level differences in other outcomes such as mean earnings, worker turnover rates, productivity, even after accounting for worker characteristics. Among others, these findings suggest there are differences in human resource policies among firms and industries, something that could also affect SSDI award rates. Another strand of literature has shown that SSDI award rates tend to change with the economic conditions in the industry (e.g., Black, Kermit, and Sanders 2002). Firms in difficult economic conditions may find it easier to reduce their workforce if they can minimize the economic costs to workers by raising award rates. This 3

4 may lead to persistently higher SSDI award rates in industries with long term declines in employment. Despite the growing empirical literature on the role of industries, occupations, and firms in the labor market, very little is known about industry, occupation, or firm specific rates of SSDI receipt. This makes it difficult to assess whether industries or firms are a relevant dimension when trying to explain and potentially prevent further growth in SSDI rolls. To begin addressing this gap in the literature, in this project we characterize differences in SSDI award rates by industry and by occupation using publicly available survey data from the March Current Population Survey (CPS). As discussed in detail below, to measure the incidence of SSDI awards the paper uses both the fraction of individuals receiving SSDI in a given year, and the flow of new SSDI awards between two years. The march CPS is an extremely useful data source for this purpose, because it combines large sample sizes with information on industry and occupation status, various measures of receipt of SSDI, and detailed demographic information on workers. Moreover, since individuals can be matched across surveys, the March CPS allows us to construct actual transition rates into SSDI. This combination of features is important for the analysis and not available in alternative survey or administrative data. 1 We begin our analysis by analyzing assessing the ability of the March CPS to accurately measure the stock and flow of SSDI awards in the population. To do so, we work both with cross sectional data, as well as a version of the CPS in which we merge information of individuals from two consecutive years. Given the CPS only captures industry and occupation 1 Other survey data sources have either too small sample sizes to measure industry specific SSDI award rates or do not contain sufficient information to measure SSDI receipt. Administrative data from the Social Security Administration have very large samples and good measures of SSDI receipt, but lacks important demographic information such as educational attainment. 4

5 status for workers with recent labor force attachment, the main analysis focuses on the rate of new SSDI awards. Key results of the paper are obtained by analyzing differences in the average SSDI award rates by major industries and occupations. Note that we cannot analyze difference in application rates at the industry or occupation level using the CPS, since the survey does not contain information on application to SSDI. Hence, all empirical results based on the CPS refer exclusively to SSDI receipt, not applications to SSDI. One potential concern with the analysis is that average industry level differences in award rates could partly arise due to differences in worker characteristics. An advantage of using survey data is that it allows us to control for differences in SSDI award rates arising from differences in worker characteristics, such as educational background, family status or occupation. Once we control for such characteristics, the remaining differences more closely represent the true differences in average industry award rates, which among others may be driven by differences in human resource policies or safety of the work environment. Using this approach, we obtain the following main findings: The CPS based measures of the rate of incidence and new awards of SSDI in the working age population we construct match well comparable measures of the incidence of SSDI from administrative data. The CPS based measures of the rate of incidence and new awards of SSDI vary with age, education, race, and marital status as expected. SSDI award rates differ significantly across major industries, irrespective of what measure of SSDI receipt we use; the standard deviation of industry award rates is about half the average SSDI award rate. 5

6 The industry differences in new SSDI award rates we document are robust to controlling for individual characteristics. Overall, these findings suggest that new award rates of SSDI do not only differ widely across demographic groups, but also vary significantly by industry. Thus, further analyzing the pattern of differences in SSDI receipt by industry, occupation, or firms, and to what extent they are related to different health risk, strategic behavior, or other factors, is a worthwhile pursuit for future research. The next section of the paper describes how we measure SSDI receipt in the CPS and what sample we use. Section 3 compares our alternative measures and benchmarks them with similar measures based on published data from the Social Security Administration. Section 4 contains the main results relating to industry level differences in SSDI award rates, including a regression analysis incorporating worker characteristics. Section 5 summarizes and concludes. 2. Measurement of SSDI Receipt in CPS and Sample 2.1 Definitions of SSDI Receipt in Survey Data To analyze how the incidence of SSDI receipt varies by industry, we use information from the March Demographic Supplement to the Current Population Survey (CPS). The CPS allows us to construct measures of both the stock and flow of SSDI recipients in the US population. We use three distinct but related definitions based on different CPS questions available in the March Demographic Supplement. The first definition counts a person as receiving SSDI if the reason for receiving Social Security Benefits is disability and the person is 64 years old. This measure is based on CPS 6

7 questions that aim directly at measuring whether a person is receiving SSDI, and hence will be our preferred measure below. The relevant question was introduced in The second definitions counts a person as receiving SSDI if he or she is receiving any Social Security benefits, is under 62 years old, and in addition reports a work limiting disability. Given the inability to distinguish disabled early claimants of Social Security old age pensions (Old Age Survivor Insurance) from SSDI recipients below the federal retirement age, this measure requires individuals to be below age 62. Since many SSDI recipients are older, this is the most restrictive of the three measures. An advantage of this measure is that the CPS questions on which this measure is based are available before 2003 as well. The third and final definition counts a person as receiving SSDI if he or she receives Medicare and is under 65 years old (and does not receive railroad retirement disability insurance). Since the only way a person below age 65 can receive Medicare is to receive SSDI, this measure should accurately capture SSDI receipt. It is available since Due to the different nature of the survey questions, the three definitions differ somewhat in who is counted as SSDI recipient. For example, by construction the second definition will tend to undercount the overall incidence of SSDI by excluding workers age 62 to 64. On the other hand, we suspect the third definition over counts SSDI receipt among workers close to retirement age. Moreover, since each definition is based on questions regarding potentially difficult concepts (regarding the nature of Social Security benefits, the source of health insurance, and the presence of work limiting disabilities), they all may be affected by measurement error. Hence, we will present results for all three definitions in our empirical analysis below. 7

8 2.2 Stock and Flow Measures of SSDI Receipt in the CPS Based on these three definitions, the March CPS readily yields measures of the total number of SSDI recipients in March of a given year. Using this measure of the stock of SSDI recipients, we define the fraction of SSDI recipients in the working age population as, Below, we will compare this measure of the rate of SSDI coverage with the ratio of the total number of SSDI recipients divided by the total population in the same age range published yearly by the Social Security Administration. These concepts are not expected to be equal, since the SSA figures for the number of SSDI recipients generated by its own administrative records and the population estimates differ as well. However, they should be at least approximately similar. While the stock of SSDI recipients is informative, for analyzing the role of industry patterns, the rate of new SSDI awards is more relevant. The majority of individuals receiving SSDI at any given point in time has been on the program a long time. The CPS does not record an industry affiliation for these individuals. The CPS only asks information on industry for workers who are in the labor force, or if they worked within the last 12 months. The rate of new awards is also conceptually more appealing measure. For policy purposes, one is presumably first interested in which industries are currently most contributing to the growth in SSDI rolls. While the industry composition of the total stock could also be interesting for example as a 8

9 measure of the total legacy of an industry to the SSDI program it cannot be analyzed with the CPS and we leave it for future work. 2 For these reasons, we construct several measures of new SSDI award rates based on the March CPS data. The definition of the aggregate rate of new SSDI awards among individuals insured by SSDI used here is Pr 1. For analyzing new award rates by industry, this measure further requires that all individuals were affiliated with a given industry in March t 1. Using the March CPS, two possible approximations of this ideal measure of the new award rate can be constructed. One measure is simply the estimate of the fraction of total SSDI recipients, from the cross section, but restricted to individuals with a valid industry code. Given the way the CPS collects industry information, these are either workers who are currently employed, unemployed, or were employed in the last 12 months. 3 This is a reasonable measure of the rate of new SSDI awards because the numerator will only include SSDI recipients that had recent labor force attachment. Since SSDI is a program that insures working individuals against a permanent, work limiting disability, most new awards should be for people coming from the labor force. 2 Administrative data from SSA contains information on the industry of individuals employers since the early 1980s, and hence can be used to construct industry affiliation for at least part of the stock of individuals currently receiving SSDI. For a description and use of this industry data to analyze incidence of SSDI awards by industry, see von Wachter, Song, and Manchester (2010). 3 As per the questionnaire, the CPS collects industry information for the main job a worker is currently employed or at which the worker last worked. As per the interviewer instructions, the question is only asked for workers who held a job within the last 12 months (see page 29 of Chapter 4 at While this is not explicitly mentioned, this restriction is likely to apply only to workers out of the labor force, not workers searching for jobs and hence declared unemployed. 9

10 To better see what this cross sectional approximation of the new award rate captures and why it is a sensible measure, it can be written as,, The numerator captures (1) SSDI recipients in March of year that are not working but that had worked within the last 12 months and (2) SSDI recipients that are currently working. Since few individuals receiving SSDI tend to work, the numerator should be a close approximation to the actual number of new awards, which by the scope of the program should mainly consist of those permanently transiting out of their previous line of work because of a disability. The denominator includes all working individuals plus non working individuals with recent labor force attachment, whether they receive SSDI or not. This should be a reasonable approximation of the pool of workers in the labor force or with recent labor force attachment at risk of applying to SSDI. Clearly, the cross sectional can be no more than an approximation. Among others, by the nature of the program which requires a waiting period without substantial gainful activity and then often includes some time before a claim is adjudicated typically SSDI awards occur with a lag. Hence, the denominator of the cross sectional measure is not the actual at risk group for new SSDI recipients in the immediately preceding months, and is sensible only in so far as the size of the relevant population (and later the industry distribution) evolves sufficiently slowly. 10

11 Hence, we pursue a second strategy to estimate the rate of new SSDI awards based on a true longitudinal match between individuals appearing in consecutive March Current Population Surveys. With two observations on the same individual, we can obtain an estimate of the new award rate that is conceptually closer to the ideal measure. The second measure based on a sample of matched consecutive March CPS can be written as,, The advantage of this measure is that it is closer to the actual concept we want to measure, described above by. The numerator captures those individuals transiting into the SSDI program between two points in time that at the beginning of that period were in the labor force (an approximation of being insured by SSDI). The denominator captures the correct at risk group of all individuals that were in the labor force in working age at the beginning of the period. As a result,,, more closely approximates the ideal definition of the new award rate for a cohort of individuals than the cross sectional measure does. The disadvantage of using the second measure based on longitudinally merged consecutive CPS is that it is based on much smaller sample sizes (as further discussed below). However, since entry into SSDI is not a frequent event, to measure differences in the rate of entry into SSDI among single sectors with sufficient precision large sample sizes are required. Hence, we will present results for both measures in the empirical section. 2.3 Construction of CPS Sample and Longitudinal Match 11

12 The main goal of this paper is to measure SSDI award rates by industry status. The CPS records industry codes in a consistent fashion from 2003 onwards. 4 Hence, since we are also interested in merging consecutive years of the March CPS, we restrict our main analysis to the March surveys of the years 2004 to (We use earlier data for years in our aggregate analysis to assess longer term trends.) To construct our working sample, we take the full sample of individuals in the CPS in working age (age 25 to 64) and exclude the small number of individuals employed in the military. For this sample, we construct our three definition of SSDI receipt discussed in Section 3.1, as well as several variables relevant for our analysis, as further described in the Data Appendix. Column 1 of Table 1 provides basic descriptive statistics for our sample combining all eight years in the sample. All statistics are weighted by the CPS sampling weights. The average annual number of observations in our sample of working age individuals per year is about 100,000. The resulting average estimated number of individuals in working age per year is about 170 million, which is close to what Census Bureau reports based on the 2010 Census. The gender, education, and age distributions are all as expected. Based on this sample, individuals appearing in consecutive March CPS years were matched based on household and person identifiers, as well as demographic characteristics. 4 In the section Historical data and comparability and availability of the document Occupational and industry classification systems used in the Current Population Survey, the Bureau of Labor Statistics states: In January 2003, the CPS adopted the 2002 Census occupational and industry classification systems; they replaced the 1990 Census occupational and industry classifications. The introduction of these classification systems created a complete break in comparability with existing data series at all levels of occupation and industry aggregation. The composition of detailed occupations and industries changed substantially in the 2002 systems compared with the 1990 systems, as did the structure for aggregating them into major groups. Hence, any comparisons of data on the different classifications are not possible without major adjustments. ( On the same webpage, the Bureau of Labor Statistics provides a cross walk to earlier years that could potentially be used to extend our analysis of SSDI award rates by industry backwards if desired. 12

13 The Data Appendix describes the procedure of the match and robustness checks in detail. Based on the outgoing rotation group structure of the CPS, 50% of individuals in any given month can be matched to the same month in the previous year. The match rate for this group is about 50 60%, depending on the month and year, which is typical. Column 2 of Table 1 displays characteristics of individuals for whom a valid match in the previous March survey is found (the matched sample). The sample size of the matched sample is about a third of the main crosssectional sample, something that will become relevant below. Overall, the distribution of individuals characteristics in the matched sample is quite similar to that in the cross section. As expected, given the CPS does not follow people who move, the matched sample has fewer younger and non white individuals, slightly more married and more older individuals. As is discussed next, the rate of SSDI receipt in the population and the rate of new SSDI awards are similar in the cross sectional and matched samples as well. 3. Benchmark Results Based on CPS Measures of SSDI Receipt 3.1 CPS Based Estimates of the Fraction of SSDI Receipt in the Population The first row of Table 1 shows the average proportion of individuals receiving SSDI in for our three alternative measures of SSDI receipt. The three measures put the average fraction of SSDI recipients among the population age broadly at about 3%. As discussed in Section 2, the first measure uses CPS questions explicitly focused on measuring SSDI receipt, and hence is our preferred measure. This measure suggests the average stock of SSDI recipients in was 3.21% of the population age As expected in Section 2, the second measure based on reports of work limiting disabilities implies a smaller fraction of 13

14 only 2.61%, since it excludes workers Finally, the third measure based on Medicare receipt yields an estimate of the fraction of SSDI recipients of 3.73%. The remainder of the table shows the average proportion of individuals receiving SSDI in different demographic groups. The resulting patterns are as expected based on the previous literature (e.g., Bound and Burkhauser 1999, Duggan and Imberman 2006); while the proportion is similar for men and women, less educated workers are four to six times more likely to be receiving SSDI benefits than college graduates, and the oldest age group (55 64) is much more likely to receive SSDI than the youngest age group (25 34). The demographic patterns among SSDI award rates appear to confirm findings in the foregoing literature that the CPS yields reliable measures of SSDI receipt in the population. To further benchmark the rate of SSDI receipt in our sample, Figure 1 compares the rate implied by our three measures by major age groups to similar proportions calculated using data published by the Social Security Administration (SSA). 5 The SSA publishes the total number of SSDI for each year, as well as the number of individuals in the population by age group. The ratio of these two concepts is used in Figure 1 as the SSA based measure of SSDI receipt in the population. It should be noted that this ratio is not strictly comparable to the CPS based measure, since the source is administrative data unaffected by measurement problems typically affecting survey based measures of disability. Hence, we would expect the CPS measure to understate somewhat the rate of SSDI receipt. The patterns in Figure 1 are as expected. First, the figure confirms that SSDI receipt rises rapidly with age. Second, it suggests that all three measures broadly capture the same order of 5 The age groups correspond to those published by SSA and hence differ from those in Table 1. 14

15 magnitude and growth of SSDI receipt by age. Third, it confirms that the three CPS based measure mostly differ among workers that are above 60 years old, and are very similar for younger workers. Finally, it shows that the SSA based measure is higher for all age groups. The one exception is the third measure based on Medicare that rises rapidly for the year olds, suggesting that perhaps the incidence of Medicare may be overstated by the CPS in that age group. Overall, the results in Table 1 and Figure 1 suggests that the CPS does a reasonable job in capturing the overall level and age gradient of SSDI receipt in the population, but that it tends to understate the full extent of coverage. 3.2 CPS Based Estimates of the Rate of New SSDI Awards As discussed in Section 2, to gauge the contribution of industries to the growth in SSDI rolls we will turn to an analysis of the rate of new awards. Table 2 displays our first approximation of the rate of new awards based on single CPS cross sections. This measure is the proportion of workers with valid industry codes (and hence with current or recent labor force attachment) receiving SSDI, divided by all individuals with valid industry codes (see Section 2.2). As shown in the first row of the table, the average annual rate of new awards for this group ranges from 0.3% (measure 2) to 0.6% (measure 3). As in the case of the overall stock of SSDI recipients, the rate of inflow differs by demographic characteristics as expected; it is about 3 times higher for high school dropouts than for college graduates and at least double for year olds than year olds. One can again use numbers published by SSA to benchmark these values. Based on data published by SSA, we calculated the ratio of new awards for SSDI to the insured population for year olds for a given year, This ratio is 0.3%, very close to the number shown in 15

16 Table 1; given this number excludes very young workers, who have low inflow rates (but also a lower proportion among the insured population), the comparable number is likely to be somewhat smaller. Nevertheless, the measure is likely to be in the ballpark of the figures in Table 2. Table 3 shows the conceptually cleaner measure of the rate of new SSDI awards based on the longitudinally matched March CPS discussed in Section 2.2. The range of new award rate for the three measures is again between 0.3% and 0.5%, very similar to the purely crosssectional measure in Table 2, and somewhat larger than what results from SSA data. The new award rates in Table 3 also display very similar patterns across education, age, and race as Table 2. The broad similarity between the two CPS measures suggests they are capturing the same underlying concept, and gives us confidence in the purely cross sectional measure for which we can exploit larger sample sizes. As an additional check of the quality of the longitudinal match, we have also compared the fraction of the stock of SSDI recipients in the population (defined as in Table 1) in each available year in the purely cross sectional and in the matched sample. The results are shown in Figure 2A to Figure 2C for our three definition of SSDI receipt, respectively. All three figures show the increasing trend in the fraction of the population receiving SSDI. Importantly for our purposes, both level and trend are similar for the original and the matched sample, giving comfort in the quality of the matched despite the fact that we can only match 60% of those individuals that could in principle be linked. An advantaged of the longitudinal link is that we can decompose the new award rate into flows coming from different labor force status. For each available year, Figures 2A to 2C 16

17 also show the new award rate for workers coming from employment and from unemployment. Interestingly, there appears to be a clear upward trend in new SSDI awards coming from employment. Although the figures are noisier due to the smaller number of unemployed workers, there appears to be no upward trend in the rate of new SSDI awards going to workers that are unemployed. This in itself constitutes an interesting finding, but clearly deserves further scrutiny with data that allows more precise measurement. 4. Main Results From Industry Specific Rates of SSDI Receipt in CPS 4.1 Differences in Average New Award Rates by Major Industry The average rate of new SSDI awards by major industry for our three definitions of SSDI receipt is shown in Table 4. For ease of comparison, Figure 3 displays the industry specific award rates for our three definitions of SSDI receipt. The upper panels of the table and figure, respectively, show the rate approximated using only CPS cross sections, based on much larger samples; the respective lower panels show results based on the longitudinally matched sample. There is clearly substantial variance in SSDI award rates across industries. For example, using the cross sectional measure and our preferred definition, the upper panel Table 4 shows the inflow rate from manufacturing is 0.25%, whereas that from wholesale and retail trade and leisure and hospitality is around 0.5%, respectively. Professional and business services and leisure and hospitality, two large industries which together account for 30% of employment lie at around 0.4%, roughly equal to the population average for that measure. The upper panel of Figure 3 suggests that these patterns are independent from the particular definition of SSDI receipt adopted. 17

18 The longitudinal new award rates in the lower panels of the table and the figure also suggest and important degree in dispersion by industry. The patterns bear broad similarity to the cross sectional measure. However, there are also some differences. When comparing the two sets of numbers one should keep in mind that the longitudinally matched data has much smaller sample sizes. The resulting higher variance may lead to discrepancies that are simply due to sampling error. 6 Overall, we view the patterns in Table 4 and Figure 3 as confirmation that there are important industry specific pattern of inflows into SSDI the SSDI program. To gauge the potential effect of such fluctuations, column 1 of Table 4 also displays the overall industry distribution in the population. It is remarkable how different the new award rates for the largest economic sectors are. This suggests that the variance in industry specific award rates we find is likely to play a significant economic role. Hence, it might be worth at least considering further whether to monitor and potentially reduce new SSDI awards focusing on particular industries is a promising avenue. 4.2 Robustness to Differences in Worker Composition It is well known that industries differ in the demographic structure of their work force. For example, some industries have workers that are on average more educated or older. Since we have seen in Section 3.2 that new SSDI award rates differ significantly by these and other characteristics, differences in worker composition could explain part of the documented industry differences in new award rates. To assess the importance of worker composition, we estimated regression models that hold constant differences in some key observable 6 We calculated the standard errors for the matched sample, and for many industries the differences are within plus or minus two standard errors. 18

19 demographic worker characteristics. We regressed an indicator of whether a worker with recent labor force attachment received SSDI in a given year (for the cross sectional measure) and whether a worker received SSDI in a given year that had not received SSDI in the previous year (for the longitudinal measure) on industry indicators, and a range of worker characteristics. The regression model we estimated is where is an indicator variable indicating the respective SSDI status of individual i in year t, are twelve dummies for major industry, and is a vector of observable characteristics including year dummies, age dummies, education dummies, race dummies, a gender dummy, and a dummy for marital status. Since the dependent variable is an indicator variable, the resulting model is a linear probability model. If one was to regress on alone (excluding a constant), the resulting coefficients (, one would simply reproduce the mean industry specific award rates shown in Table 4. Once is added to the regression model, the coefficients on the industry dummies represent the industry specific award rates that cannot be explained by differences in the worker characteristics contained in. Consider first the results based on the CPS cross section in Table 5 for our preferred definition for SSDI receipt (the regression tables for definitions 2 and 3 can be found in the appendix), which for convenience are also displayed graphically in Figure 4. These are the findings that are most reliable given the substantial sample sizes of the unmatched CPS. Clearly, not surprisingly worker characteristics do matter (coefficients not shown), and alter the size of the industry differences in the approximate new award rates. However, overall the coefficients 19

20 on the industry dummies remain highly statistically significantly different from zero. This is underscored by the results of an F test of joint insignificance shown at the bottom of Table 5. The coefficients also remain economically significant, and Figure 4 underscores that much of the differences is unaffected by the added controls. Among the worker characteristics, it is dummies for completed education and occupation dummies that matter most. The inclusion of education tends to reduce interindustry differences, suggesting that those industries with higher new award rates also had a higher fraction of low educated workers. Occupation controls, on the other hand, tend to have the opposite effect and raise differences across industries. I.e., those industries with higher award rates tended to have more workers in occupations with lower award rates. This could be for example if, say, managers have lower SSDI award rates, and there are proportionately more managers in, say, manufacturing industries. Interestingly, occupation controls tend to work in the opposite effect than education in several cases. Table 6 shows the corresponding regressions for the award rates based on the longitudinally matched CPS for our preferred definition (the regression tables for definitions 2 and 3 can be found in the appendix). Overall, the results support a similar message as for the cross sectional award rates. This is underscored in the bottom panel of Figure 5 differences in composition matter, but the importance of industry characteristics remain. However, once additional characteristics are included, the coefficients on most industry dummies in Table 6 are not statistically significant from zero (with the exception of public sector and information). This is likely to be chiefly due to the lower sample sizes of the matched sample. The sample in the cross section is about three times larger than that of the matched 20

21 sample. As a result, the standard errors in Table 6 should be about 1.73 larger than those in Table 5. In fact, this comes close to the actually observed difference in standard errors. Hence, while in principle the matched measure is conceptually more appealing, the smaller sample sizes make it less useful for studying award rates in smaller groups, even after pooling several years of data. Yet, fortunately for the present study the findings in Tables 5 and 6 suggests that the cross sectional measures yield very similar and more robust results. Overall, the regression results support the interpretation that the mean industry differences in Table 4 represent to a large extent true differences in new award rates across sectors, and do not mainly reflect differences in demographic characteristics of the work force. 5. Summary and Discussion This project has produced industry specific award rates for SSDI based on cross sectional and longitudinal data from the March Current Population Survey (CPS). The large sample sizes, detailed demographic characteristics, multiple SSDI measures, and longitudinal component make the CPS an ideal candidate to study SSDI receipt by smaller units such as industry or occupation. In a first step, the project confirms the feasibility of using the CPS as a reliable measure of SSDI receipt. Using multiple definitions of SSDI receipt and multiple measures of new SSDI awards, the main finding based on the CPS is that award rates can differ substantially by major industries. The standard deviation of industry specific award rates is about half of the average new award rate. In a key step, the project has shown that these differences are largely robust for adjustment of differences in worker composition. Studying industry differences in SSDI receipt is important, since firms may differ in the rate to which their workers enter SSDI, and these differences may vary at the industry level. As 21

22 mentioned at the outset, firms that have higher SSDI award rates do not face higher costs from the program, and hence they may face an implicit subsidy of unsafe work environments or risky jobs. They may also face a potential incentive to help more costly or less productive eligible workers transition into SSDI. The differences in industry level SSDI award rates we document here suggest that it is worth studying such potential differences and incentives further. However, it is worth emphasizing that the results in this paper do not provide information about the sources of industry differences in award rates, other than the role of worker characteristics. Hence, no inference is possible on the potential sources of industry or firmspecific differences in SSDI receipt mentioned above, or other explanations, without additional research. 22

23 References Anderson, Patricia Linear Adjustment Costs and Seasonal Labor Demand: Evidence from Retail Trade Firms. Quarterly Journal of Economics 108(4): Anderson, Patricia and Bruce Meyer (1994). The Extent and Consequences of Job Turnover. Brookings Papers on Economic Activity 1994:1. Autor, D. H. and M. G. Duggan The Rise in The Disability Rolls And The Decline In Unemployment. Quarterly Journal of Economics 118(1): Black, Dan, Kermit Daniel, and Seth Sanders The Impact of Economic Conditions on Participation in Disability Programs: Evidence from the Coal Boom and Bust. American Economic Review 92(1): Bound, John and Richard Burkhauser Economic Analysis of Transfer Programs Targeted on People with Disabilities. In Handbook of Labor Economics, Vol. 3C, ed. O. Ashenfelter and D. Card, Amsterdam: Elsevier North Holland. Duggan, Mark and Scott Imberman Why Are the Disability Rolls Skyrocketing? In Health in Older Ages: The Causes and Consequences of Declining Disability among the Elderly, ed. David Cutler and David Wise. Chicago: University of Chicago Press. Rupp, Kalman and David Stapleton Determinants of the Growth in the Social Security Administration s Disability Programs: An Overview. Social Security Bulletin 58(4):

24 Table 1: Sample Characteristics and Fraction SSDI Receipt in Population, March Current Population Survey (CPS) for Years March CPS Longitudinally Matched March CPS SSDI Measure 1 SSDI Measure 2 SSDI Measure 3 Total Obs (Weighted) Total Proportion Receiving SSDI in Population 1,389,579, ,306, % 2.61% 3.73% Percentage Among Category Proportion Receiving SSDI Among Category Men 48.98% 48.59% 3.30% 2.68% 3.66% Women 51.02% 51.41% 3.13% 2.54% 3.79% Education Less Than High School 11.79% 10.70% 6.69% 5.26% 7.80% Education Equal High School 30.35% 30.49% 4.17% 3.40% 4.77% Education Some College 27.17% 27.11% 3.05% 2.52% 3.34% Education College or Beyond 30.69% 31.71% 1.08% 0.88% 1.47% Married 64.33% 68.68% 2.17% 1.68% 2.58% Unmarried 35.67% 31.32% 5.10% 4.28% 5.79% Age % 20.99% 1.04% 0.96% 1.23% Age % 27.08% 1.89% 1.76% 1.97% Age % 29.55% 3.71% 3.41% 3.79% Age % 22.38% 7.19% 4.81% 9.33% White 79.77% 82.67% 3.04% 2.47% 3.47% Non-white 20.23% 17.33% 3.90% 3.15% 4.75% Note: The size of the unmatched March CPS sample of individuals age 25 to 64 is 860,432, the size of the longitudinally matched sample in the same age range is 215,717. Columns 3 to 5 are based on the unmatched CPS sample. All proportions are based on counts using CPS person weights.

25 Table 2: Sample Characteristics and Fraction SSDI Receipt in Population with Valid Industry Information, March Current Population Survey (CPS) for Years March CPS Longitudinally Matched March CPS SSDI Measure 1 SSDI Measure 2 SSDI Measure 3 Total Obs with Valid Industry Code (Weighted) Total Proportion Receiving SSDI in Population with Valid Industry Code 1,105,700, ,094, % 0.28% 0.56% Percentage with Valid Industry Code Among Category Proportion with Valid Industry Code Receiving SSDI Among Category Men 53.43% 52.87% 0.34% 0.25% 0.51% Women 46.57% 47.13% 0.43% 0.31% 0.61% Education Less Than High School 9.47% 8.38% 0.59% 0.44% 0.92% Education Equal High School 29.36% 29.52% 0.51% 0.38% 0.70% Education Some College 27.83% 27.75% 0.41% 0.29% 0.56% Education College or Beyond 33.33% 34.35% 0.19% 0.13% 0.32% Married 64.74% 69.11% 0.22% 0.14% 0.33% Unmarried 35.26% 30.89% 0.67% 0.54% 0.97% Age % 21.96% 0.25% 0.18% 0.40% Age % 28.76% 0.29% 0.22% 0.40% Age % 30.86% 0.42% 0.34% 0.51% Age % 18.42% 0.69% 0.43% 1.18% White 80.49% 83.29% 0.37% 0.28% 0.52% Non-white 19.51% 16.71% 0.42% 0.27% 0.70% Note: The size of the unmatched March CPS sample of individuals age 25 to 64 is 860,432, the size of the longitudinally matched sample in the same age range is 215,717. Columns 3 to 5 are based on the unmatched CPS sample. All proportions are based on counts using CPS person weights.

26 Table 3: Sample Characteristics and Fraction and Rate of New SSDI Awards in Population, Current Population Survey (CPS) for Years Longitudinally Matched March CPS SSDI Measure 1 SSDI Measure 2 Total Observations (Weighted) Total Proportion Newly Re in Population 369,094, % 0.29% Proportion Newly Recei Among Categor Men 52.87% 0.39% 0.29% Women 47.13% 0.40% 0.30% Education Less Than High School 8.38% 0.79% 0.57% Education Equal High School 29.52% 0.57% 0.45% Education Some College 27.75% 0.37% 0.27% Education College or Beyond 34.35% 0.17% 0.12% Married 69.11% 0.29% 0.18% Unmarried 30.89% 0.64% 0.54% Age % 0.24% 0.19% Age % 0.25% 0.20% Age % 0.41% 0.35% Age % 0.77% 0.47% White 83.29% 0.35% 0.26% Non-white 16.71% 0.59% 0.45% Note: The size of the unmatched March CPS sample is , the size of the Longitudinal sample is Columns 3 to 5 are based on the unmatched CPS sample. All proportions on counts using CPS person weights.

27 , March SSDI Measure 3 eceiving SSDI 0.51% iving SSDI ry 0.47% 0.56% Averge Annual Sample Size Total in 2010, Census 46,136, ,623,999 THIS EXCLUDES THOSE NIU (TABLE 3A INCLUDES THEM) 0.94% 0.73% 0.46% 0.26% 0.37% 0.83% 0.24% 0.29% 0.46% 1.27% 0.45% 0.79% lly matched s are based

28 Table 4: Sample Characteristics and Fraction SSDI Receipt in Population by Industry, March Current Population Survey (CPS) for Years Percentage Among Category SSDI Measure 1 SSDI Measure 2 SSDI Measure 3 Proportion Receiving SSDI Among Category Cross-Section March CPS Agriculture 1.35% 0.56% 0.42% 0.75% Mining 0.51% 0.23% 0.18% 0.27% Construction 8.15% 0.16% 0.13% 0.33% Manufacturing 12.13% 0.25% 0.18% 0.36% Wholesale and Retail 13.08% 0.55% 0.41% 0.76% Transportation and Utilities 5.55% 0.23% 0.13% 0.36% Information 2.44% 0.33% 0.28% 0.50% Financial Activity 7.27% 0.27% 0.16% 0.40% Professional and Business 11.00% 0.41% 0.30% 0.51% Education and Health Service 22.13% 0.44% 0.33% 0.63% Leisure and Hospitality 6.61% 0.53% 0.40% 0.84% Other Services 4.68% 0.64% 0.45% 0.99% Public Administration 4.68% 0.20% 0.13% 0.32% Longitudinally Matched March CPS Agriculture 1.34% 0.53% 0.44% 0.70% Mining 0.51% 0.54% 0.33% 0.42% Construction 7.73% 0.37% 0.31% 0.46% Manufacturing 12.81% 0.35% 0.29% 0.40% Wholesale and Retail 13.02% 0.49% 0.36% 0.65% Transportation and Utilities 5.56% 0.43% 0.32% 0.50% Information 2.54% 0.16% 0.10% 0.34% Financial Activity 7.43% 0.28% 0.15% 0.40% Professional and Business 10.75% 0.40% 0.26% 0.44% Education and Health Service 22.60% 0.36% 0.29% 0.47% Leisure and Hospitality 5.74% 0.62% 0.50% 0.85% Other Services 4.57% 0.55% 0.33% 0.78% Public Administration 4.57% 0.24% 0.19% 0.43% Note: The size of the unmatched March CPS sample for individuals age 25 to 64 is 860,432, the size of the Longitudinally matched sample in the same age range is 215,717. All proportions are based on counts using CPS person weights.

29 Table 5: Estimates of Fraction SSDI Receipt in Population with Valid Industry Information by Major Industry Controlling for Worker Characteristics, Linear Probability Model, Definition 1 of SSDI Receipt, March CPS , Cross Section Sample (1) (2) (3) (4) (5) Mining * ** * * ** (0.0013) (0.0016) (0.0016) (0.0016) (0.0017) Construction *** *** *** *** *** (0.0002) (0.0009) (0.0009) (0.0009) (0.0011) Manufacturing *** *** *** *** *** (0.0002) (0.0009) (0.0009) (0.0009) (0.0011) Wholesale and Retail Trade *** * (0.0003) (0.0009) (0.0009) (0.0009) (0.0011) Transportation and Utilities *** *** *** *** *** (0.0003) (0.0009) (0.0009) (0.0009) (0.0011) Information Services *** ** ** ** (0.0006) (0.0011) (0.0011) (0.0011) (0.0012) Financial Activity *** *** *** *** (0.0003) (0.0009) (0.0009) (0.0010) (0.0011) Professional and Business Services *** * * ** (0.0003) (0.0009) (0.0009) (0.0009) (0.0011) Education and Health Service *** (0.0002) (0.0009) (0.0009) (0.0009) (0.0011) Leisure and Hospitality Services *** ** (0.0004) (0.0010) (0.0010) (0.0010) (0.0012) Other Services *** (0.0006) (0.0011) (0.0011) (0.0011) (0.0012) Public Administration *** *** *** *** *** (0.0003) (0.0009) (0.0009) (0.0010) (0.0011) Results for Joint Test that Industry Dummies are Equal to Zero F-Statistic P-Value Note: The size of the unmatched March CPS sample for individuals age 25 to 64 is 860,432. All regression models use CPS person weights. Excluded industry is agriculture. Model (1) includes Industry Dummies; model (2) adds year dummies; model (3) adds sex, marital status, age and race dummies; model (4) adds education dummies; model (5) adds occupation dummies. Robust standard errors in parentheses; *** p<0.01, ** p<0.05, * p<0.1

30 Table 6: Estimates of Fraction of New SSDI Receipt in Population with Valid Industry Information by Major Industry Controlling for Worker Characteristics, Linear Probability Model, Definition 1 of SSDI Receipt, March CPS , Cross Section Sample (1) (2) (3) (4) (5) Mining * (0.0029) (0.0033) (0.0033) (0.0033) (0.0035) Construction *** (0.0006) (0.0017) (0.0017) (0.0017) (0.0020) Manufacturing *** (0.0004) (0.0017) (0.0016) (0.0016) (0.0020) Wholesale and Retail Trade *** (0.0005) (0.0017) (0.0017) (0.0017) (0.0019) Transportation and Utilities *** (0.0008) (0.0018) (0.0018) (0.0018) (0.0020) Information Services ** ** ** (0.0007) (0.0017) (0.0017) (0.0017) (0.0020) Financial Activity *** (0.0005) (0.0017) (0.0017) (0.0017) (0.0019) Professional and Business Services *** (0.0006) (0.0017) (0.0017) (0.0017) (0.0020) Education and Health Service *** (0.0004) (0.0016) (0.0016) (0.0016) (0.0020) Leisure and Hospitality Services *** (0.0009) (0.0018) (0.0018) (0.0018) (0.0021) Other Services *** (0.0009) (0.0018) (0.0018) (0.0018) (0.0021) Public Administration *** * ** (0.0006) (0.0017) (0.0017) (0.0017) (0.0020) Results for Joint Test that Industry Dummies are Equal to Zero F-Statistic P-Value Note: The size of the longitudinally matched March CPS sample for individuals age 25 to 64 is 215,717. All regression models use CPS person weights. Excluded industry is agriculture. Model (1) includes Industry Dummies; model (2) adds year dummies; model (3) adds sex, marital status, age and race dummies; model (4) adds education dummies; model (5) adds occupation dummies. Robust standard errors in parentheses; *** p<0.01, ** p<0.05, * p<0.1

31 Figure 1: Fraction of Working Age Population Receiving SSDI by Age Group, Different Measures Based on March CPS and on SSA Data for SSDI Measure 1 SSDI Measure 2 SSDI Measure 3 SSA Data

32 Figure 2A: Annual Fraction of Working Age Population Receiving SSDI (Stock) and Rate of New SSDI Awards by Labor Force Status (Flow), SSDI Measure 1, Based on Cross-Section and Matched March CPS Stock Match Stock Match Flow - Employed Match Flow - Unemployed

33 Figure 2B: Annual Fraction of Working Age Population Receiving SSDI (Stock) and Rate of New SSDI Awards by Labor Force Status (Flow), SSDI Measure 2, Based on Cross-Section and Matched March CPS Stock Match Stock Match Flow - Employed Match Flow - Unemployed

34 Figure 2C: Annual Fraction of Working Age Population Receiving SSDI (Stock) and Rate of New SSDI Awards by Labor Force Status (Flow), SSDI Measure 3, Based on Cross-Section and Matched March CPS Stock Match Stock Match Flow - Employed Match Flow - Unemployed

35 Figure 3: Alternative Approximations of New SSDI Award Rate by Major Industry and Different SSDI Measures, March CPS Panel A: Fraction SSDI Receipt in Population by Industry Based on Cross Section March CPS 0.01 SSDI Measure 1 SSDI Measure 2 SSDI Measure Panel B: New SSDI Award Rate by Industry Based on Matched March CPS SSDI Measure 1 SSDI Measure 2 SSDI Measure