The impact of the employment insurance repayment policy: nonexperimental approaches

Abstract

This research examines the impact of the strengthened repayment provision under the Employment Insurance (EI) system in Canada. The provision has introduced an increase in benefit repayment rates and a reduction in the repayment threshold. Without a randomized experiment, this article uses various nonexperimental approaches, particularly matching, to estimate the policy effect. Using data from the Survey of Changes in Employment (CIE) and the Survey of Labour and Income Dynamics (SLID), the results suggest that the new repayment policy has reduced the probabilities of filing a claim among workers whose annual income is equal to or greater than $48 750. The estimated decline in the claim rate ranges between 6 and 12 percentage points, depending on the datasets. The results appear to be robust regardless of the methods of estimation.

Notes

¹ Currency in this chapter is in Canadian dollars.

² The sample does not include individuals whose potential income is less than $39 000, because they are unaffected by the new repayment policy. Individuals whose potential income is between $39 000 and $48 750 are also excluded because, whether they are affected by the new policy depends on their claim history, which is not available in either dataset.

³ It should be noted that although the quality of the matches may be improved by imposing the common support restriction. However, it is possible to lost high quality matches at the boundaries of the common support. Besides, the sample may be considerably reduced, so imposing the common support restriction is not necessarily better (Becker and Ichino, 2002).

⁴ Propensity scores and matching estimators are estimated through a user-written Stata program designed by Becker and Ichino (2002).

⁵ In radius matching, the treated unit is matched to all control units with estimated propensity scores falling within a self-defined radius r = 0.001. In kernel matching, a match is constructed for each treated unit using a kernel-weighted average over all individuals in the control group. The weights depend on the distance between each control and treated unit. Control units that are very far away get a very low weight rather than a zero weight (see for example Heckman et al., 1997a, b).

⁶ The first 10 cohorts were sampled during the period from July 1995 to December 1997. Due to administrative reasons, only two samples were collected between January 1998 and June 2000: cohort 13 (July–September 1998) and cohort 17 (July–September 1999). Beginning in July 2000, the CIE restarted its regular quarterly collection (cohort 21) with redesigned weighting methodology. The CIE is still active, and the most recent available data is cohort 25. For consistency reasons, cohorts 21 and onward that used a redesigned weighting scheme were not used in this study. Cohort 1 was also excluded because, it was merely a test, and its target population was different from other cohorts.

⁷ In response to SLID income questions, respondents can grant their permission for Statistics Canada to access their tax file data from Revenue Canada solely for the purposes of the SLID. In 1997, for example, about 81% of the population has used the tax route.

⁸ See Appendix B for detailed variable construction.

⁹ Since most people do not know their actual annual income at the time of job separation, an individual's income from the previous year is used as a proxy for potential annual income. Unfortunately, income from the previous year is not available in the CIE data; therefore, two methods are used to calculate an individual's potential annual income. See Appendix B for details.

¹⁰ For the CIE data, the original sample from cohort 2–10, 13 and 17 consists of 47 964 observations. Sixty-six were dropped from the sample because, they did not live in any of the 10 provinces, 613 were eliminated because of age (≥ 65), and 868 were removed due to self-employment. Individuals whose reasons for separation were other than ‘layoffs’ or ‘contract ended’ were excluded (15 706), and those who were not eligible for UI/EI (as defined in Appendix B) were excluded (2672). The sample included only full-time workers (4233 were removed), and those who had no information about their previous income and education status were excluded (619). Finally, only high-income workers were selected from the sample (an additional 20 202 were dropped), resulting in 2985 observations selected from the CIE data.

¹¹ For the SLID data, there were 57 729 job interruptions that occurred during 1993 and 1998 (33 543 were permanently displaced, and 24 186 were temporarily laid off). Among them, 154 were dropped because of age (≥ 65), and 978 were removed due to self-employment. Individuals whose reasons for separation were other than ‘layoffs’ or ‘contract ended’ were excluded (34 767), and those who were not eligible for UI/EI (as defined in Appendix B) were excluded (4718). The sample included only full-time workers (2373 were removed), and those who had no information about education and firm size were excluded (471). Finally, only high-income workers were selected for the sample (an additional 13 150 were dropped), resulting in 1118 observations from the SLID data.

¹² Construction of this variable is described in Appendix B.

¹³ The balancing hypothesis can be written as T  ⊥  X  |  p(X) (see Becker and Ichino, 2002). To satisfy the balancing property, observations with the same propensity score must have the same distribution of characteristics independent of treatment status. In other words, for each covariate, differences in mean across treated and control units within each interval are not significantly different from zero. Notice that the individuals are divided into intervals according to the estimated propensity score.

¹⁴ That is, in the radius matching the weight for a given control j is 1/N^c (where N^c is total number of controls within the radius for treated i).

¹⁵ Regression on propensity score is estimated in two steps. First, a probit of the treatment on X is estimated, and then a regression of outcomes on the treatment dummy and fitted propensity score is estimated. The estimated propensity score should contain all the information in the covariates that is relevant for estimating the treatment effect. The advantage of using the propensity score as a regressor is that, it rules out any values of X such that p(X) = 1. The drawback, however, is that it ignores the residual error from the first-stage probit estimation.

¹⁶ Predicted number of new claims is evaluated at the mean value of Claimrate_highincome (= 0.7120885).