Paying for Disability Insurance?: Firm cost sharing and its employment consequences

Transcription

1 Paying for Disability Insurance?: Firm cost sharing and its employment consequences Amelia Hawkins and Salla Simola JOB MARKET PAPER This version: November 20, 2018 Please read the most recent version found here Abstract This paper offers the first empirical evidence on the employment consequences and tradeoffs of tying a firm s Disability Insurance (DI) tax payments to the DI claims of the firm s own workers. The evidence comes from a reform in Finland that increased employer DI coinsurance rates while decreasing (non-experience rated) payroll taxes. The reform applied to some firms but not others, and to varying degrees. The evidence indicates that among firms bearing higher costs of DI claims by their own employees, (i) DI receipts of their employees declined, (ii) the composition of new hires shifted away from higher-disability-risk groups, and (iii) among current employees the expected gains from increased sustained employment exceed the declines in expected DI receipts. The University of Michigan, 611 Tappan Street, Ann Arbor, MI aaehawki@umich.edu Aalto University, Arkadiankatu 7, Helsinki. salla.simola@aalto.fi 1

2 1 Introduction Disability Insurance (DI) in the US and in other developed countries is a large and expensive social insurance program. For example, in the US in December % of adults aged received Social Security DI benefits which cost a total of $11.3 billion that month (Social Security Administration, 2017). At varying points in time, rising numbers of program recipients have also caused concern. In response to these trends, the academic and policy communities in both the US and around the world have sought ways to improve efficiency in their DI programs, reduce overuse and promote the accommodation of workers with disabilities by tying a firm s costs for the DI system to the DI claims of their own employees (Bound and Burkhauser, ; Autor and Duggan, 2006; Burkhauser et al., 2008; Burkhauser and Daly, 2012; Burkhauser et al., 2014). The federal DI program in the US, as in many other developed countries, is not experience rated but is instead funded by FICA (Federal Insurance Contributions Act) payroll taxes with a 1.8% rate 1. Recent proposals would have firms pay more for the DI costs generated by their own employees, as firms already do in state Workers Compensation and Unemployment Insurance programs through experience rating (Autor and Duggan, 2010; Burkhauser and Daly, 2011; Burton and Guo, 2016). Without experience rating or some other way for firms to bear directly the additional costs when their workers move onto the DI rolls, firms may not have sufficient incentive to internalize the behaviors that lead their employees to seek DI. From this point of view, workers use the DI program too much and firms accommodate or rehabilitate workers with disabilities too little. If a worker could continue working with appropriate accommodation, rehabilitation and return to work services (but firms do not have incentive to provide them), then the costs will be shifted onto the DI system. Presumably, if employers and employees could negotiate freely over the employee s compensation including accommodations, then without government intervention, firms would provide accommodations to their workers so long as the cost of the accommodation to the firm is less than the benefit to the worker. Apart from the moral hazard argument, under 1 Generally speaking, experience rating is when a firm s loss history is used to determine its future premium rates. So, for example, firms whose employees enter DI at rates higher than similar firms would have higher future premiums or higher FICA payroll tax rates. The National Academy of Social Insurance explains the rationale for not experience rating DI: Punishing the employer for an employee s need for non-work related disability... is unfair to the employer. Still, New Jersey s Temporary Disability Insurance program is experience rated and many private disability insurance plans in the US are also experience rated. 2

3 provision of accommodations might arise if firms and employees have imperfect information about accommodation costs and value or if firms that provide accommodations attract higher-disabilityrisk workers creating an adverse selection problem. While tying a firm s DI tax payments to the DI claims of the firm s own workers might encourage firms to accommodate current workers who have become disabled, it might also discourage them from hiring potential workers who are disabled or who they think might make a disability claim in the future, thus contributing to the unemployment of people with disabilities rather than ameliorating it. In testimony to the House Ways and Means Committee Subcommittee on Social Security Lisa D. Ekman, JD, MSW on behalf of the Social Security Task Force Consortium for Citizens with Disabilities articulates this unwanted consequence, We do not believe that experience rating SSDI FICA payroll taxes will increase employment or retention of people with disabilities. In fact, we believe that this proposal would have the opposite effect it would make it less likely that people with disabilities or chronic conditions would be hired in the first place. 2 However, little empirical evidence exists quantifying these desired and undesirable consequences. We provide the first empirical evidence on these trade-offs by leveraging a natural experiment in Finland s DI system where co-insurance directly ties firms DI costs to the DI use of their own employees. That is, some firms owed up front and in lump-sum a portion or all of the expected stream of DI benefits their worker would generate. We exploit a reform in that raised firms co-insurance rates while decreasing their premiums (non-experience rated payroll taxes); this change applied to some firms but not others and to varying degrees. Using an event study design and a rich administrative dataset, we find that among firms with higher DI costs for their own employees Disability Insurance in-flows fell by 13-19%. Our results suggest that a 1 percentage point increase in the co-insurance rate / decrease in overall payroll tax obligation led to a 1.1 to 1.6% fall in the DI inflow rate. We find that the composition of new employees shifts away from higher-disability-risk groups, i.e. less educated workers, older workers and those 2 This concern has been echoed in the Netherlands (e.g. OECD (2008); Koning and Lindeboom (2015); Fultz (2015); Koning (2016)) and in Finland (e.g. (Antolin et al., 2001; OECD, 2004). In a recent survey by the Finnish Center for Pensions, 70% of the surveyed Finnish employers agreed with the sentence: firms don t hire highdisability-risk workers because these workers might result in expenses to the firm [translation is our own] (Llukko et al., 2017) (in Finnish). 3

4 with a history of non-employment. However, at least 74% of the reduction in DI in-flows came from incumbent pre-reform employees rather than from changes in the composition of employees. Finally, we find among current employees the fall in DI receipt is more than offset by an increase in sustained employment without substitution to other social assistance or social insurance programs. Previous research in the Netherlands shows that experience rating reduces DI entry (Koning, 2009; de Groot and Koning, 2016) though the evidence in Finland has been more mixed (Korkeamäki and Kyyrä, 2012; Kyyrä and Tuomala, 2013; Kyyrä and Paukkeri, 2018). Korkeamäki and Kyyrä (2012) study the same reform, though taking a different empirical strategy, and find a substantial reduction in entry to the DI and sickness benefit programs. However, these initial studies of moral hazard look solely at DI entry rates, not other outcomes of interest such as hiring patterns and worker retention. Hullegie and Koning (2018) take a step towards addressing this gap finding that a series of Dutch reforms reduced DI entry, increased employment up to 4 years after a hospitalization, and did not increase the use of other social assistance benefits. However, in the Dutch context many aspects of the DI system changed at once or in close secession, making it difficult to disentangle the contribution of any particular policy (de Jong, 2015). Our study is the first to link the outcomes of those who would have gone on DI, but don t to the increase of firms direct costs for DI per se. There is even less evidence connecting firms DI costs to the hiring of workers with disabilities. And, as far as we are aware, these concerns are based solely on anecdote. 3 Koning and Lindeboom (2015) report descriptive evidence that the decline in Dutch DI inflows has been accompanied by a decrease in the fraction of workers with permanent contracts, those for which firms must cover worker DI costs, as opposed to temporary contracts, which are exempt from cost sharing rules. 4 Our study is the first to empirically examine if firms direct costs for DI reduces hiring of workers most at risk of entering that social insurance program. Our paper is organized as follows: Section 2 briefly characterizes the disability insurance system in Finland, how it is financed, and the rules we leverage for identification. In Section 3 we discuss the data, and in Section 4 the empirical strategy. In Section 5 we present our findings and robustness checks, and Section 6 concludes. 3 Hullegie and Koning (2018) look at overall employment after a hospitalization which would include the net effect of workers retained at their old job and any hiring effects. 4 In response to these trends in 2016 the Netherlands changed their system to include firm cost sharing for temporary contracts. 4

5 2 Background 2.1 Overview of Disability Insurance in Finland Since its inception, The Social Insurance Institution (SSA) has provided federal disability benefits, first paid out in Disability benefits in Finland, as in the U.S., follow a two tiered system: earnings-related Disability Insurance (for which the claimant must have an earnings history and have paid into the system) and national disability benefits (a program for workers with very low earnings or an insufficient work history ). 6 In the s the Finnish disability benefit system was rather generous compared to many OECD countries (OECD, 2008, 2010). In, 303,398 Finnish residents ( 9.4% of the population aged 18-64) received disability benefits. 25,281 new awards were made that year with an 70% award rate (Finnish Center for Pensions, ). In Finland spent 2,935.9 million euros on disability benefits ( 1.3% of GDP that year (Finnish Center for Pensions, 2018; Eurostat, 2018)). Of all disability benefit recipients in, 82% (247,636) received an earnings-related disability insurance payment (Finnish Center for Pensions, 2018). For the remainder of this section we restrict our discussion to the earnings-related DI system covering private sector employees during the s. The Ministry of Social Affairs and Health set uniform rules guiding the application and appeals process, financial and medical eligibility and benefits that did not vary across firms or insurance providers. All firms must enroll their earnings-related-insurance-eligible employees into a disability insurance plan. Employees became earnings-related insurance eligible after just one month of work. Workers did not lose their eligibility for earnings-related disability benefits when they left their employer. There were no other financial requirements of the applicant, such as the Substantial Gainful Activity in the US. The replacement rate, which accrued during work and non-work periods by at most 1.5% per year, was capped at 60%. However, there was no benefit amount cap. The average replacement rate in this time period was 55% (Korkeamäki and Kyyrä, 2012). When a worker turned 65, DI benefits were converted to an Old-Age pension benefit. The DI system in Finland also included work disincentives. Beneficiaries could work up to an earnings threshold before benefits would be suspended. If beneficiaries worked above that threshold 5 For a detailed history of the Finnish social security system in English see Hannikainen and Vauhkonen (2012). 6 Also, as in the U.S., a distinct workers compensation covers income loss from workplace accidents and occupational diseases. Worker s Compensation is the primary pension, meaning that while a worker can receive both a DI payment and a Worker s Compensation payment, the DI payment is reduced dollar for dollar. 5

6 for more than 2 years, then benefits would be terminated. However, few beneficiaries exercised this option. Furthermore, although SSA would reassess the eligibility criteria of time-limited DI beneficiaries, they didn t periodically review indefinite DI beneficiaries. As such, Disability Insurance was an absorbing state. In, less than 1% of DI recipients left for employment the next year and less than 3% worked while receiving benefits (see Figure C.1a). Of course, the cross-sectional rate of leaving DI for work can be quite small compared to a longitudinal rate (Liu and Stapleton, 2011). However, Figure C.1b shows that between even 10 years after DI entry beneficiaries earnings show no signs of the rebound we might expect if a significant portion of beneficiaries eventually become employed. For more details on the Finnish Disability Insurance system see Appendix A. 2.2 How DI was financed A series of pension acts governed DI provision for Finnish workers. The Finnish Center for Pensions determined which pension act a worker s occupation fell under, and that determination set the rules governing the DI provision for that worker. 7 In the private sector ( 65% of the workforce), the Employees Pension Act covered non-seasonal workers ( 85% of private sector workers. 8 ) This paper focuses on the DI system for private sector workers covered by the Employees Pension Act, hereafter called covered workers. In the private sector DI for non-seasonal workers was funded by a system of premiums (nonexperience rated payroll taxes) and co-insurance. 9 Firms owed payroll taxes to their insurance provider each year for their current employees. In addition, for their current and former employees, firms owed co-insurance, a share of the payment made against the claim, or in this case, a share of the stream of future benefits owed in lump-sum at the time the worker entered DI. Firms did not owe co-insurance for all current and former employees, but generally did for (i) workers whose disability onset occurred within 3 years of being employed at the firm, (ii) workers who worked for 7 In the s, seven pension acts governed DI for private sector occupations and three pension acts governed public sector occupations. The rules governing DI varied slightly across pension acts and sectors. Due to occupation switches during their working careers, workers could be covered by different pension acts at different points in time, but whichever pension act the worker was covered under before disability onset would regulate the DI system they entered (Antolin et al., 2001). 8 Authors tabulation from the Finnish Longitudinal Employer-Employee Dataset. 9 Firms remitted both components to the insurance provider. 6

7 the firm for at least a year, and (iii) workers who had not entered long-term unemployment benefits first. 10 In practice, SSA actuaries determined this stream of future Disability Insurance benefits by multiplying the benefit amount by an age specific constant that accounted for the number of years an average person of that age entering Disability Insurance would receive benefits. So, for example, a firm that paid 100% co-insurance for a 50-year old with a 43.5% replacement rate would owe 3.7 years of earnings at the time that worker entered DI. Each firm s number of covered workers in December determined their split between these two components, the payroll tax obligation and the co-insurance rate. Figure 1 shows how the split between these two components varied with firm size before. Firms with fewer than 50 covered workers paid only a non-risk adjusted payroll tax set at the same rate for each employee (similar to FICA in the US). Firms with 50 to 299 covered workers paid only a risk adjusted payroll tax, with age as the only adjustment factor. For firms with covered workers the DI obligation was split. They paid a portion of their risk-adjusted payroll tax obligation and a portion in co-insurance. The degree of this split was determined by their firm size, so that larger firms faced higher co-insurance rates but also had a smaller overall payroll tax obligation. Firms with more than 100 workers paid no payroll taxes for DI, but paid for the entire stream of future benefits. 11 Firms got a rebate for the benefits not paid out if their former employee exited DI for employment, but not if the former employee died (interview with Christina Lindell, retired employee of the Finnish Center for Pensions, Friday, August 4, 2017). As far as we are aware, a secondary insurance market did not exist, though of course nothing prevented firms from saving funds to cover unexpected co-insurance costs. 10 See Ylinen (2010) (in Finnish) for exceptions to these rules. The workers last employer before disability onset was responsible for the co-insurance cost. If the individual worked for more than one employer before disability onset, the coinsurance costs were apportioned to both employers; however, 5% of individuals list more than one firm ID each December during our study period in our dataset. 11 The smaller the firm the lower the co-insurance rate under the assumption that small firms have less ability to pay large sums (interview with Christina Lindell, retired employee of the Finnish Center for Pensions, Friday, August 4, 2017). A trade-off exists between shifting enough of the burden of DI costs onto firms so that they internalize the additional costs of DI, but not shifting so much of the burden as to make it prohibitively expensive for firms, penalizing them for bad luck or worker moral hazard out of their control (Cutler, 2002; Autor, 2015). 7

8 2.3 The reform Figure 1 shows how the split between the co-insurance rate and payroll tax obligation varied with firm size after the reform compared to what it was before the reform, a swivel to the left of the line bifurcating the two components. The reform decreased the overall payroll tax obligation and increased the co-insurance rates for firms with 50 to 999 covered workers, but to varying degrees producing a triangular shape; meanwhile SSA left the rules unchanged for firms with fewer than 50 or more than 1,000 covered workers. 12 The largest changes occurred for firms with 300 covered workers. So, for example, a firm with 300 covered workers before would owe no co-insurance but owe the full risk-adjusted premium rate for each covered employee. After, firms with 300 covered workers had a 26% co-insurance rate. But, at the same time, the total payroll tax obligation for their current covered workers fell by 26%. This reform was in the works for a year. On June 8,, labor union representatives signed off on the drafted regulation. On December 20, the Council of State finalized the new regulation which became effective January 1,. 13 Over the course of the s amid reforms to the Finnish social security system the fraction of working age Finns receiving disability insurance fell (see Figure C.2). 14 As far as we are aware, this was the only reform in the s based on firm size The Finnish Longitudinal Employer-Employee Dataset (FLEED) and sample selection We use the Finnish Longitudinal Employer-Employee Dataset (FLEED); compiled from various administrative and survey sources, this data includes information on the socio-demographic characteristics of the whole Finnish resident population age 16 to 70 from 1988 to FLEED contains information on both working and non-working periods. For workers, the FLEED links 12 The firm size threshold for paying DI co-insurance costs was lowered from 300 to 50 for administrative ease to coincide with rules for the long-term unemployment benefit system (interview with Christina Lindell, retired employee of the Finnish Center for Pensions, Friday, August 4, 2017). 13 See Appendix A.4 for a time-line of events leading to this reform. 14 See e.g. Kyyrä and Wilke (2007), Niemelä and Salminen (2006), Antolin et al. (2001) for more of these reforms. 15 See Appendix A.4 for a discussion of DI reforms during our study period 8

9 employees to the primary and secondary 16 firm and establishment where they worked in the last week of each year. Unfortunately, the FLEED does not contain any information on the disability or health status of Finns nor does the dataset contain information on workplace accommodations, or rehabilitation services such as job demands or tasks, flexible scheduling or even hours worked. We construct our analysis sample focusing on firms most likely to be affected by the reform and a group of comparable firms. First, we restrict the time period from 1988 to because in 2000, another policy that differentially affected firms by firm size was implemented. Second, we exclude public sector workers because the DI system for the public sector has its own set of rules and was not impacted by this reform. We exclude firms with fewer than 10 workers to have a similar set of firms in each year because the FLEED did not release data on the smallest firms in all years. Third, we restrict our sample to firms in the dataset in and calculate each firm s number of covered workers that year to determines whether, and to what degree, firms would be affected by the reform. We construct our treatment status variables using the number of covered workers before the reform announcement because firm size itself might have responded to the policy announcement or implementation. We investigate this possibility empirically by looking at the histogram of firm size before and after the reform announcement in for visual evidence of heaping at the 50 firm size kink point and find none (histograms shown in Figure C.3). Finally, we create weights to ensure the comparison group firms (with fewer than 50 or more than 1000 covered workers) look more like treated group firms (with 50 to 999 covered workers) in terms of industry and region. We created these weights from a logit model with 2-digit industry codes, 19 region codes and full set of interactions. 17 To the extent that firms in our treated and comparison groups are in different industries and regions, in different labor markets, the re-weighting helps us isolate more closely the effect of the reform of interest rather than picking up other factors that differed across industry and region over our study period. The Table C of unweighted and re-weighted summary statistics shows that in firms in our treated and comparison groups differed most notably in industry and region makeup. We collapse the individual level data down to the firm level to arrive at our final sample-an unbalanced panel of 2,094 treated group firms (371,946 employees in, 22% of all Finnish employees economy-wide in ) and 9,272 comparison group firms (617,029 employees in, 16 Few Finns in our study period worked a secondary job 5% 17 We take the modal industry and region of residence among each firm s workforce in to assign each firm an industry and region. 9

10 36% of all employees economy-wide in ). See Table C.2 for how each sample restriction affects our sample size. In Section we check the robustness of our main results to each of our sample selection choices. 4 Empirical Strategy Our primary empirical strategy will be an event study design comparing our treated and comparison group firms in the years leading up to and after the change in DI financing. Our identification strategy relies on the assumption that, if the reform had not taken place, the outcomes would have evolved similarly in the more and less affected firms of different sizes. Therefore, it is important to evaluate whether pre-reform trends in these outcomes of interest were indeed similar across the more and less affected firms in our sample. In our Results Section 5.2, we evaluate this assumption by comparing the trends of our treated group firms that faced higher co-insurance rates and lower payroll tax obligations to the trends of our comparison group firms. We run the following weighted least squares models at the firm level to test this parallel trends assumption: y ft =β 0 + β f1 1(firm f ) + t= [β t1 1(year t ) + β t2 1(year t ) T reat f ] + ɛ ft (1) where y ft is a firm-level outcome of interest for firm f in year t. Year indicators 1(year t ) allow for flexible time effects controlling for time varying factors which affect all firms. Firm fixed effects 1(firm f ) take into account time-invariant firm-level characteristics that differ between our more and less affected firms and could influence the outcomes of interest. T reat f designates our treatment status variable. We construct two variables: (i) a binary treatment status equal to 1 if the firm employed 50 to 999 covered workers in and 0 otherwise, and (ii) a dosage continuous treatment status ranging from 0-100, the triangular shape depicted in Figure 1, indicating the percentage point increase in the co-insurance rate / decrease in the payroll tax obligation. This variable imposes a linear relationship between the treatment dose and the outcomes of interest. In Figure 1 this would mean the two sides of the triangle of equal height are restricted by the model to have the same treatment effect. For example, the model averages the treatment effects of a firm with 145 workers and a firm with 734 workers both of which had a 10 percentage point increase in their co-insurance rate and 10 percentage point 10

11 decrease in their overall payroll tax obligation. Both models, with a binary treatment status or with a continuous treatment status, have advantages. The binary status doesn t impose the linear structure between the treatment dose and the outcomes of interest; it is also more robust to measurement error. Because firm size fluctuates, we probably do a better job assigning the correct treatment status in the binary case than in the continuous case. On the other hand, the dose response models allow us to further test our hypothesis that firms with the largest increased incentive to accommodate or rehabilitate workers see the largest drops in DI entry and those same firms, which also have the largest swings in the relative costs of hiring higher-disability-risk groups are the least likely to do so after the reform. β t2, the interaction terms between each year indicator and firm treatment status are the coefficients of interest. is the omitted year. We plot these coefficients with confidence interval bars in the Results Section 5.2. If more- and less- affected firms have similar trends before the reform, and diverge only after the reform, these figures provide evidence that the regulation caused these changes rather than a pre-existing differential time trend. If indeed the regulation caused the change in outcomes, then the estimated treatment effects, the coefficients on the interaction terms of treatment status variable and the year fixed effects, β t2 should be zero for the years prior to the reform and non-zero after the reform. With each figure we show in a corresponding table β t2 from the following simple difference-indifferences regression: y ft =β 0 + β f1 1(firm f ) + β t1 1(year t ) + β t2 1(P ost) T reat f + ɛ ft (2) which includes a post-reform indicator 1(P ost) instead of individual year dummies. We cluster standard errors at the firm level. We do not population weight the regressions. We know the treatment status and dosage varied by firm size and therefore it is not obvious that some doses should be given more weight than others We have not yet bootstrapped the standard errors to take account of the weights, but this will be added soon. 11

12 5 Results 5.1 How much did firms costs change? Descriptive evidence Ex-ante: Shifts in expected costs across firms in the sector to firms with higher than average disability risk workers The risk-adjusted payroll tax rates were set each year so that, in expectation, each cohorts DI costs would be fully funded by the firms co-insurance payments together with their cohorts payroll tax contributions. 19 Therefore, a firm s expected DI costs for a worker who has an average risk of entering DI given their age would not change but would simply shift from being paid as a payroll tax to being paid as co-insurance. Likewise, a firm with the average mix of workers would have no change in their expected DI costs after the reform. However, the reform would increase the expected DI costs for firms with higher than average disability-risk workers and decrease the expected DI costs for firms with lower than average disability-risk workers. To illustrate how the reform shifted expected costs across firms, we have used the Disability Insurance rules, SSA parameters, and data from the Finnish Longitudinal Employer-Employee Dataset (FLEED) to impute what firms would expect to owe for their current and former employees under the new and old rules. 20 Figure 2a shows for each firm the percent change of expected labor costs for their current employees in. The variation moving from left to right comes from the policy change itself, while the variation moving up and down comes from the interaction of the policy with the existing disability risk structure of workers at each firm. Figure 2a shows that the reform would actually save firms money on average if they did not owe co-insurance for former employees. This apparent sector-wide cost savings results from the fact that employed people generally do not enter the DI system and thus are of lower than average risk of entering DI given their age, Figure 2b shows the percent change in firms expected labor costs for current employees in and for former employees from, and for whom the firm still might owe co-insurance. The average change in expected labor costs sector wide is now close to 0, which would be the case for a revenue neutral and actuarially fair policy. These figures highlight an important point: having higher co-insurance makes it most costly to separate from 19 Figure C.4 depicts these rates by age and over time. This figure shows that the payroll tax rates risk adjusted by age in roughly mirrored the Kaplan-Meter empirical hazard by age. 20 See Appendix B for details of this imputation. 12

13 high-disability-risk workers exactly those workers who might enter DI Ex-ante: Shifts in expected costs within firms to higher-disability-risk workers Regardless of a firm s mix of current workers, this reform also shifts the relative costs of employing a high-disability-risk worker compared to a low-disability-risk worker. Due to data limitations, we won t be able to disentangle whether firms hire different workers for the same job opening than they would have otherwise or if they open different jobs than they would have otherwise due to this reform. So, for example, a firm might have one job posting and decide to hire a lowerdisability-risk applicant over a high-disability-risk applicant. Alternatively, a firm might decide to open a new job posting that typically attracts low-disability-risk applicants over one than typically attracts high-disability-risk applicants. For example, suppose a firm s employees work either on a factory floor or in a call center, and, for whatever reason, workers on the factory floor tend to be higher-disability-risk than workers in the call center; so the reform might induce the firm to open additional jobs in the call center rather than on the factory floor. 22 To get a sense of how much the cost of hiring workers at higher disability risk increased relative to workers at lower disability risk, we calculated the change in relative expected labor costs over the working life for the average high school dropout compared to the average high school graduate within each firm before and after the policy change. 23 Figure 3a plots this value expressed as a percent change for each firm and shows how this value varies across firms of different sizes. The variation going left to right reflects the change in the policy faced by firms of different sizes while the variation going up and down reflects the fact that the degree to which high school dropouts are at higher disability risk than high school graduates varies for different types of firms. Figure 3a reveals that at most the expected cost of the average high school dropout relative to the average high school graduate rose 0.82% with the average increase among affected firms in the size range at 0.22%. We did the same calculations for other (correlated) groups that disproportionately 21 These figures plot the firms expected labor costs for current and former employees in ; these calculations are highly correlated across time using employees from other years. We would expect a highly positive correlation across years if firms employ the same people or the same types of people every year (see Figure C.5). 22 Employers can lawfully ask applicants about their capacity to perform the advertised job and about their general health status. See Appendix A.5 for additional details on discrimination law in Finland. 23 See Appendix B for details on how we did this imputation and Table C.5 for the imputed change in relative expected worker costs for different worker characteristics. 13

14 enter DI: in Figure 3b for workers who were not employed the past year relative to workers who were employed the past year and in Figure 3c for workers over the age of 50 relative to workers under age 50. How well do different worker characteristics, like education, divide the cases into those who will and won t enter DI? The companion Table C.4 gets at this question for a variety of worker characteristics. This table shows the rate at which high school dropouts enter DI and their representation among DI entrants in the private, non-seasonal sector. In, high school dropouts were almost twice as likely to enter DI than the age population at large. For both groups DI entry was a low probability event with less than one percent of the age population entering DI that year and 1.62% of high school dropouts entering DI that year. Almost 60% of DI entrants that year were high school dropouts despite making up 31% of Finns age Taking Figure 3 and Table C.4 together, these calculations suggest that while some personal characteristics are indicative of DI entry, the ex-ante change in expected costs both across firms in sector and within firms for different types of workers are modest Ex-post: How much did the amount of co-insurance increase? Ex-post the amount in co-insurance firms owed could have increased by at most 50 weeks of earnings (that is, for a 50-year-old worker with a 43.5% replacement rate in a 300 covered worker sized firm for which the co-insurance rate increased from 0 to 26.3%). The average annual earnings before DI entry among entrants in our sample period was e25,723 (2017 euros) or $30,896 (2017 USD); therefore, the increase in co-insurance firms owed could have at most increased by e25,122 (2017 euros) or $30,174 (2017 USD). For the average aged worker (aged 39 among our affected firms in ) at the average firm size (168 covered workers among our affected firms in ) the co-insurance rate increased from 0 to 12.3% corresponding to a e8714 (2017 euros) or $10,466 (2017 USD) increase. This increase in co-insurance becomes a cost averted if the employee does not enter DI, effectively decreasing the cost of providing accommodations, assuming that the accommodation works to keep the employee off of DI. For perspective, survey estimates suggest that the median costs for personal-assistance services over one year was $12,033 (2017 USD) ($148,855 over 15 years) and the median cost for non-personal- assistance accommodations over one year was $3,054 14

15 ($11,608 over 15 years) What were the impacts? Regression results For each of our main results we present both a figure and table. The figures graphically illustrate the relationship between an outcome of interest and the treatment by plotting the β t2 coefficients on the interaction terms 1(year t ) T reat f from equation 1 and the 95 percent confidence interval on each point estimate. The vertical lines at and are meant to draw attention to the pre- time period, before the reform announcement, and the post- period, after the reform implementation. The omitted year is expressed as a dot. Below each figure, a table presents the Wald estimator from equation 3 with cluster robust standard errors. We present the binary treatment status results in one panel and the dose response relationship in the other panel Disability Insurance entry rates We first establish that increasing firms co-insurance rates and decreasing their payroll tax obligation led to a decline in the inflow to DI as the regulation intended. We define a firm s DI entry rates as the fraction of workers at the firm in t 1 who enter DI in t. The majority, 64.67%, of covered private sector, non-seasonal, workers who entered DI in entered within one year of last being employed. (We will look at a longer time horizon below.) The time pattern of results shown in Figure 4a suggest the regulation change rather than other factors led to the decline in firm DI entry rates. We find no significant effect of being a sized treated group firm on DI entry rates prior to the reform. The coefficients on the interaction terms 1(year t ) T reat f are small in magnitude and statistically not different from zero. Beginning in, after the new regulations were implemented, we find that being a sized firm predicts a significant reduction in DI entry. The sharp break in supports the attribution of the decline in DI entry rates to the regulation rather than a differential trend in DI entry rates across our two groups of firms; however, we will probe this interpretation further in Section 5.2.7, further 24 Among employers providing personal- assistance services the median annual costs for those services was $9,773 (2017 USD) plus $2,260 for one-time costs and for non- personal- assistance accommodations the median annual cost was $611 plus a $2,443 one-time cost (Solovieva et al., 2009). See Bardos et al. (2015) for a comprehensive cost/benefit analysis of return to work policies. These authors take into account additional employer costs including HR costs, labor compensation and productivity, finding that from the employer s perspective it would cost $201,330 (2017 USD) to re-employ a worker for 17 years after disability onset. 15

16 examining the relationship between firm size and DI entry. Our results in Figure 4a suggest that on average for affected firms the reform led to a decline in firms DI entry rate significant at the one percent level. Relative to the comparison group firms average DI entry rate (0.0059) this change represents a 19 percent drop. Stated another way, this decline represents a 13.8 percent drop relative to the treated group firms mean DI entry rate prior to the increase in experience rating (0.0084). Turning to the dose response model, in Figure 4b we see a similar time pattern in the estimated coefficients on the interaction terms; however, the trend break in is less pronounced in this model, and we find significant effects in the pre-period, specifically in and. The fact that the pre-reform point estimates do not lie as closely to zero in the dose response model compared to the binary model could be due to the regression weights. We create the weights from a logit model with an indicator of being a sized firm as the left-hand-side variable. This approach will up-weight comparison group firms in the same industry and region as the sized firms but not additionally ensure that less affected firms have the same industry and region make up as more affected firms as would be the case with a Generalized Propensity Score. Our results in Figure 4b suggest that every one percentage point increase in the co-insurance rate and decrease in the overall payroll tax obligation led to a decline in firms DI entry rate, significant at the one percent level. This estimate implies that a move to a 100% co-insurance rate (or full experience rating) would lower the claim rate by percentage points. Relative to the average DI entry rate among comparison group firms (0.0059) this change represents a 1.6 percent drop. Alternatively, this decline represents a 1.1 percent drop relative to the treated group firms mean DI entry rate prior to the increase in experience rating (0.0084). The point estimates from the binary model are roughly 12 times the size of those from the dose response model as might be expected seeing as the co-insurance rate increased from 0 to 12.3% for a treated group firm with the average number of covered workers in. These estimates imply for every $845 increase in co-insurance, or put another way for every $845 increase in the cost of not accommodating or rehabilitating a disabled worker, the DI entry rate fell by %. 25 Because DI entry is a relatively rare event, a 1.1% drop amounts to For the average aged worker (39 years old in the treated group of firms in ) with a 27% replacement rate (assuming a 1.5% increase in the accrued replacement rate since age 22 the maximum accrual rate at the time) a firm with full experience rating would owe approximately 2.73 years of earnings in co-insurance at the time of DI entry using SSA program parameters. The average earnings of DI entrants before DI entry was approximately 16

17 fewer covered workers entering DI from an average-sized firm. 26 Taken together, these estimates imply that for every $28,490 (2017 USD) increase in co-insurance owed, 1 less person enters the DI system the next yea. 27 This implied willingness to pay for an averted DI entrant equates to more than twice the median personal-assistance-services cost over two years or what the median non-personal-assistance accommodations cost over 15 years. Although most workers enter DI within one year of their last employment, others have not worked for several years by the time they start receiving benefits. Table C.6 shows the years elapsed between last employment and DI entry for for those workers who worked in the private, non-seasonal sector at least once between 1988 and. Among the workers who entered DI in, 91.05% did so within 3 years of last employment. 28 Therefore, we look next at DI entry rates over a longer 2-year and 3-year period in order to see whether workers with a longer period of detachment from their previous employer show a similarly large decline in DI receipt. The results presented in Figure 5 show that as a result of the reform (in the binary treatment status model) the fraction of workers entering DI within 2 years fell by , and within 3 years it fell by , Figure 5. These point estimates, roughly similar in magnitude to the results for DI entry within one year, constitute an 11.9% and 8.5% drop relative to the comparison group mean (an 8.6% and 6.2% drop relative to the treated group mean in respectively). The dose response model results follow a similar pattern, suggesting a smaller reduction in DI entry rates over the longer 2-year and 3-year windows; however, a pre-trend evident in the dose response model suggests that these estimates are upward biased. Taken together, these results suggest that the longer workers are detached from their previous employer before entering DI the less efficacious it is to tie a firm s costs to the DI entry of its own workers. That is, over the longer windows tying a firm s costs to the DI entry of its own workers does less to reduce DI entry rates. Perhaps this finding is a result of workers who experience disability onset after exiting the firm, workers whom the firm might not have even known became disabled and thus would have had no opportunity to $30,896 (2017 USD) (see Figure C.1b). These values imply that a one percentage point increase in coinsurance for the average aged person is roughly $ The average number of covered workers in a treated group firm in was Using the SSA parameters for a 50-year-old instead, our results imply for every $1146 increase in co-insurance owed, the DI entry rates fall by % or that for every $38,649 increase in co-insurance 1 less person enters DI. 28 Table C.6b shows the same statistics but among all Finns whether or not we see them working in the private, non-seasonal sector before DI entry in. Among all Finns entering DI in, 84.08% did so within 3 years of last employment. 17

18 intervene with accommodation or rehabilitation. What do workers who would have entered DI do instead? To answer this question, we would ideally follow disabled workers who in the absence of the reform might have entered the DI system, but because of the reform go on to do something else instead. Unfortunately, DI eligible workers, not to mention those workers contemplating the take-up of DI benefits, would be a difficult group to identify in any dataset. The program s complex medical criteria and the subjective nature of administrative decisions make this task impossible. Lamentably, data to identify marginal workers whose future might have been altered by the reform such as survey data on work-limiting disabilities or on Activities of Daily Living does not exist and administrative data on hospitalizations is not yet linked to the our dataset (as has been done in the Netherlands Hullegie and Koning (2018)). Instead, we restrict our sample to workers whose characteristics make them at high-disabilityrisk and hence hopefully a marginal worker likely to be affected by the policy change. We then compare their outcomes across our treated and comparison group firms to determine what they do if they do not enter Disability Insurance. Specifically, we identify workers under the age of 60 with a predicted probability of entering DI in the top 75th percentile of the population distribution in. We then collapse their individual level data to the firm level. We restrict the age of workers to those under age 60 to follow them for 10 years (individuals age out of the dataset at age 70). We take the predicted probability of DI entry in thus incorporating at least one year of employment data in order to follow the same group of Finns over time. The predicted probabilities come from the logistic regression seen in Table 1 which controls for age, gender, language, education, marital status, household size, number of children and the previous year s employment. As a falsification check we also follow Finns in the bottom 25th percentile of the predicted probability distribution. If Finns whose characteristics make them the least likely to be DI entrants have a similar pattern of outcomes when we compare the treatment to comparison group firms, we might suspect that a factor other than the DI system rules was driving our results. First, we look at DI entry rates among these marginal workers. Figure 6 shows the binary treatment status on the top panel and the dose response model on the bottom panel. Figure 6a indicates that indeed our DI entry rate results are concentrated among high-disability-risk workers. The reform led to a decline in firms DI entry rates among these workers significant at the one percent level. Relative to the average DI entry rate in the comparison group firms ( ), 18

19 this change represents a large (27.6 percent) drop. 29 The dose response model also suggests a decline in DI entry among the high-disability-risk group (Figure 6d); though these results do not exhibit a clear trend break after the new regulation went into effect in. We turn next to the results for the low-disability-risk group. If being in the bottom 25th percentile of our predicted probability distribution perfectly screened out disabled Finns, we would expect the point estimates in Figure 6b to be zero because nobody in this subgroup would have been DI eligible. The point estimate on the binary treatment status model is significant at the 10 percent level, about a fourth the size of our main point estimate and 1/30th the size of the high-disability-risk group. However, the confidence intervals on the interaction terms plotted in each figure make the time pattern difficult to assess. One might expect such results for both high-disability-risk and low disability-risk workers if one assumes firms intervene to provide accommodations or rehabilitation after the reform regardless of the disabled worker s characteristics so long as the costs averted exceed the cost of the intervention Worker retention Do workers who might have entered DI stay employed instead? Our results suggest, yes they would stay employed. In Figure 7 the outcome is the fraction of workers in each firm t 1 still employed in t among high-disability-risk Finns. We find positive and significant effects of higher co-insurance rates and a lower payroll tax obligation. The binary treatment status results suggest a 2 percent increase in employment on average, and the dose response model suggests that for every one percentage point increase in the co-insurance rate the employment rate rose by 0.13 percent. This rise in employment more than offsets the fall in DI entry. The binary treatment status results imply that for each firm the average increase in employment was heads per firm while on average fewer heads per firm entered DI. 30 As a falsification test, Figures 7b and 7e show the same outcome for the low-disability-risk workers. These figures suggest a small (0.4 percent) decline in employment in the late s though with a significant pre-trend. Together these results indicate that the employment effects are concentrated among current workers with characteristics that make them more likely to enter DI; this is the group most likely to be affected 29 Alternatively, it represents a 20.2 percent drop relative to the treated group firms mean DI entry rate prior to the increase in experience rating (0.0402). 30 In, the average treated group firm had by our measure 24 high-disability-risk workers and 44 low disability risk workers. 19