When considering this type of approach, a natural question arises: Why bother matching the temporary workers and nontemporary workers? Why not simply estimate a regression model that controls for the variables used in matching?
There are several answers to this question. First, use of the matched comparison group brings us (incrementally) closer to a random assignment design, by trying to limit the comparison group to those who match actual temporary workers in their likelihood of taking a temporary job. Second, including persons in a regression analysis with characteristics that indicate that they are very unlikely to be temporary workers adds no additional information to our estimate of the effect of temporary work. Finally, a regression model typically assumes that the relationship between the independent variables and the dependent variable is structurally similar for all members of the sample. Thus, inclusion of many regular workers and non-workers who are dissimilar to temporary workers in the regression could produce spurious results if the relationship between their background characteristics and subsequent outcomes differs from that of those who are similar to temporary workers. Including people dissimilar from the temporary workers in the regression thus may decrease the ability of the regression to accurately estimate how the choice of temporary work affects future employment, for those people for whom this is a reasonable choice.
A good summary of this argument is provided by Dehejia and Wahba (1998) who make the following comments about the methods under consideration:
"[Propensity score methods] reduce the task of controlling for differences in pre-intervention variables between the treatment and the non-experimental comparison groups to controlling for differences in the estimated propensity score (the probability of assignment to treatment, conditional on covariates). It is difficult to control for differences in pre-intervention variables when they are numerous and when the treatment and comparison groups are dissimilar, whereas controlling for the estimated propensity score, a single variable on the unit interval, is a straightforward task. We apply several methods, such as stratification on the propensity score and matching on the propensity score, and show that they result in accurate estimates of the treatment impact." (p.1)