Directorates
·	Director General's Office
·	Data Development
·	Strategic Evaluations
·	Partnership and Continuous Evaluations
Publications
·	Evaluation Reports
·	Evaluation Briefs
·	Labour Market Development Agreement Series
·	Lessons Learned Series
·	Evaluation Tool Kit Series
·	Evaluation Forum Newsletter
·	From Theory To Practice

In addition to investigation of the determinants of unemployment duration and the role played by UI changes such as those introduced by C-17, there is also some interest in the related modelling of the determinants of UI recipiency durations themselves. These durations — here termed benefit durations for brevity — are important for several reasons. First, assessment of the budgetary effect of a policy change, such as C-17, must require assessment of the effects on the level and duration of such benefit durations. Second, benefit durations, as opposed to unemployment durations, may fit more closely with the theoretical notion of the insurance vs. moral hazard tradeoff in models of optimal program design and size. Third, much past Canadian work on UI, including some of the most influential work (e.g., Ham and Rea 1987), has been forced to use benefit durations simply because the administrative data on which the work relied had no information beyond the end of the period of UI receipt. This said, it should be noted that benefit durations are clearly distinct from unemployment or jobless durations and that, as an empirical matter, these differences may be large. Lévesque (1987, 1989) compares LFS measured unemployment and the behaviour of UI recipiency data, and his work has been extended, with particular focus on cyclical issues, by Barnes & Picard (1992) and Roy (1994).¹³

In presenting this analysis of benefit durations, I have chosen to concentrate on results for the full sample. As with the mass of results in the previous section on unemployment durations, there were few differences in the main effect of C-17 by separation reason. Moreover, since the presentwork is necessarily restricted to those eligible for UI, and further restricted to those who actually initiated a claim, the sample size for the VQ/Dis group becomes quite small in many cases. As above, the analysis proceeds by examining a series of duration models, investigating alternativee specifications, alternativedistributional assumptions, and alternative sets of controls as supplement to the Before/After quasi-experimental effect. I again begin with the Cox partial likelihood model, and then investigate two alternative specifications.

Tables 19 and 20 present the results of estimating Cox models of the determinants of benefit durations, restricted to the subsample who were both eligible for UI and who initiated a claim. In Table 19, Model 1 gives the pure quasi-experimental effect which, unsurprisingly, is large, and is much bigger than the analogous effect for unemployment durations from Table 1. Converted to hazard ratios, the point estimate of 1.241 in Table 19 Model 1 implies a proportional upward shift of the hazard of almost 3.5. Moreover, this large effect holds up when a control for local unemployment rates is added in Model 2, and when the standard broad set of demographic and other controls are added, without or with local unemployment rates, in Models 3 and 4 respectively.

It is interesting that, unlike the earlier unemployment duration results, these benefit duration models exhibit a significantly negative effect from local labour market conditions, both without other controls (Model 2) and when regional and demographics are addressed (Model 4). Also, in the final two models of Table 19, the data display significant effects for sex (men have a higher hazard and hence shorter expected durations), age (with a negative leading term), and some significant regional effects (even after controlling for local unemployment rates). Moreover, having held a full-time job prior to the benefit claim acts to lower the hazard significantly: such persons have a harder time finding a suitable new job and are more likely to remain on claim.

Table 20 then presents results for three related specifications. Model 1 studies the subsample not expecting return to the former job while Model 2 studies the complement who did report expecting such a return. There are some differences in the results between these two columns and relative to the overall sample from the previous Table, but these are not large. The coep effect is large and positive in both models, and other significant explanatory variables include sex and, for model 2, some provincial variables. The local unemployment rate point estimate stays about constant across the two columns and differs little from the Table 19 figures, but it loses statistical significance, perhaps as a consequence of the smaller sample sizes. Finally, the effect of having held a full-time position in the former job is only significant in model 1, for the subsample not expecting return, although the model 2 estimate is also negative, but insignificant. The final column of Table 20 augments the set of explanatory variables with a measure of the number of weeks of eligibility: this variable has a significantly negative effect, tending to lower the hazard and increase durations. However, one might prefer to model such UI eligibility and potential UI exhaustion issues in the context of time-varying covariates. This will be done in a joint model of benefit and unemployment durations in the next section.

For comparability with the earlier results and assessment of robustness, I also briefly present results from estimating two parametric models of the hazard out of benefit receipt. Tables 21 and 22 give the constant hazard results for the exponential specification, while Tables 23 and 24 give the analogous Weibull estimates.

The exponential models yield results that are qualitatively similar to the Cox specification, with significantly positive coep effects, some role for local labour market conditions, and some significant effect of sex, age, some provincial dummies and a role for full-time status on the past job. The coep effects are numerically smaller than in the Cox model but are still large in terms of their effects on the hazard: Table 21 Model 4, for example, implies a hazard 1.8 times higher for COEP95 than for COEP93.

The Weibull models in Tables 23 and 24 also give a similar set of results, with depressing effects of local unemployment rates on the hazard in Model 2 of Table 23 being swamped by the inclusion of the full set of controls in Model 4 of the same Table. Whereas the estimated shape parameter for the Weibull models of unemployment durations was negative, implying a declining hazard out of unemployment, the estimates of ln q here are all significantly positive, implying a hazard for benefit spells that rises as such spells lengthen. Again, one suspects that such findings may be proxying for effects related to the potential exhaustion of UI benefits, so that longer benefit durations mean moving closer to the end of benefit entitlement. We turn to this in the next Section. Before doing so, however, it is worth remarking that, again, the coep variable has sizeable and positive estimated coefficients in all these Weibull models. This estimate of the effect of C-17, given the other controls, appears once again to be quite well-determined in these data.

Note: Standard errors in parentheses with p<0.05 = *, p<0.01 = **. Based on weighted sample from COEP93 and COEP95.

Note: Standard errors in parentheses with p<0.05 = *, p<0.01 = **. Based on weighted sample from COEP93 and COEP95. Model 1 is for sample reporting that they do not expect return to the reference job; model 2 is for the sample reporting that they expect return to the reference job; and model 3 studies the effects of weeks of UI eligibility for the full sample.

Notes Standard errors in parentheses with p<0.05 = *, p<0.01 = **.Based on weighted sample from COEP93 and COEP95. All models include a constant term.

Notes Standard errors in parentheses with p<0.05 = *, p<0.01 = **. Based on weighted sample from COEP93 and COEP95. Model 1 is for sample reporting that they do not expect return to the reference job; model 2 is for the sample reporting that they expect return to the reference job; and model 3 studies the effects of weeks of UI eligibility for the full sample. All models include a constant term.

Notes Standard errors in parentheses with p<0.05 = *, p<0.01 = **. Based on weighted sample from COEP93 and COEP95. The ln q parameter is the estimated shape of the baseline hazard for the Weibull distribution, with ln q < 0 implying a decreasing hazard.

Notes Standard errors in parentheses with p<0.05 = *, p<0.01 = **. Based on weighted sample from COEP93 and COEP95. Model 1 is for sample reporting that they do not expect return to the reference job; model 2 is for the sample reporting that they expect return to the reference job; and model 3 studies the effects of weeks of UI eligibility for the full sample. The ln q parameter is the estimated shape of the baseline hazard for the Weibull distribution, with ln q < 0 implying a decreasing hazard.

Footnotes

13	See also the discussion of unemployment and UI spells in Corak and Jones 1995, pp.560-1.