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 the present analysis of the EI changes using these two indicators, I draw upon two related approaches and research methodologies. The first of these is quasi-experimental and the second is a more traditional structural analysis. Although it is not feasible to review the extensive, detailed and technical literature on the benefits and potential pitfalls associated with each approach, it is nonetheless appropriate to give a brief summary of the issues. In either case, "durations" may mean either of the two main indicators discussed above.

The standard structural approach seeks to model the determinants of durations, either through simple regression techniques or by the use of duration modelling methods such as estimation of the hazard function. The hazard is the conditional probability that a given spell will end in a certain period, conditional on its not having ended prior to that point. The regression model is simple to implement but suffers from econometric problems if the true hazard varies with the duration of the spell (the case of "duration dependence") or if some spells are still in progress at the survey date and are hence right-censored (so that the end date is not observed). In addition, the regression model does not naturally fit into a choice-theoretic framework (since an individual may not choose at the outset how long a duration to experience), whereas the on-going and sequential nature of a jobless duration (based on a sequence of decisions as to whether or not to accept a job offer or to continue with further job search) corresponds more naturally to the underlying econometric structure of the hazard specification.

Statistical and econometric methods used to estimate the hazard function can allow for such "duration dependence" and can naturally accommodate right-censoring. In either case, the structural approach seeks to control for other determinants of the duration — factors such as education, sex, marital status, and regional characteristics — and then to assess the role played by the variable of interest (e.g., the UI/EI replacement rate).

The key problem in the standard approach is the identification of the effects of variation in the central variable. Typically this problem is "solved" by assuming that its effect is not completely captured by the other controls. Whether this truly solves the problem is an open issue, however, and is one that depends on the credibility of the identifying assumption. An example of this approach is the use of the legislative maximum on insurable earnings under UI/EI as a means of generating sample groups with different ex post replacement rates. In each case, the structural approach relies on the identifying assumption that some variation in the key variable can be used to separate its own effect from the effects of other controls.

The chief problem with such structural modelling arises from the observation that there is in fact little exogenous variation in the parameters of UI/EI programs in many countries, including Canada. If the level of weekly UI/EI benefits depends on past earnings — both the level of these earnings (subject potentially to legislative minima and maxima) and the duration of the insured employment in the qualifying period — then it may be very difficult to separate out the direct effects of UI on behaviour and the indirect effects of all of the factors that influence past earnings. Estimates of benefit effects on unemployment durations, say, may then be biased, although in principle it is hard to determine the direction of this bias. Furthermore, since Canadian UI/EI program parameters are largely national in scope (with the exception of regional aspects of benefit eligibility and the variable entrance requirement to qualify for UI/EI), something like the state-level variation exploited in some U.S. research, for example, is not available. If researchers or policy practitioners have misgivings about the ability of non-experimental data to distinguish between direct UI effects and the influence of all of the other variables that affect UI eligibility and receipt, conclusions from structural studies must be treated with caution.

The alternative approach is a quasi-experimental (or natural experimental) technique that exploits some variation in program parameters that can be thought of as exogenous to the individuals involved. Of course, such an approach requires appropriate data that in some way brackets a legislative change (or some other form of quasi-experimental exogenous variation in the program design). In the present case, such data are available by use of the various cohorts of the Canadian Out Of Employment Panel (referred to as the COEP96 to differentiate it from two earlier surveys by HRDC, the COEP93 and the COEP95). Essentially, the COEP96 data from the initial four cohorts can serve to provide the "before" group under the old UI system, while data from the subsequent cohorts of COEP96 provide the "after" group subject to various elements of the provisions of EI. Specifically, cohorts 5 and 6 provide data on the phase-in period of EI, when some provisions were operative but others were not, while cohorts 7-10 can serve for assessment of the initial effects of EI when fully operational. As can be seen in research on the labour market effects of Bill C-17 (e.g., Jones 1997), a reform of UI earlier in the 1990s, comparison of the durations for these "before" and "after" groups enables a quasi-experimental assessment of the overall effects of the changes. Such quasi or natural "experimental" methods have the advantage of using exogenous variation in UI/EI program parameters to identify key effects, something that is typically not possible with non-experimental data.