I would suggest you focus on full-time workers in your analysis, for example, those who work for 35 hours or more per week. Part-time workers and full-time workers exhibit very different labour market behaviours.

You can also work with "hours-weighted average wages." To do this you would create a new weight variable = survey weight * hours worked. Then the average wages that you are analyzing pertain to the average hour worked, not the average worker. It is as if you are analyzing a universe of person-hours rather than a universe of people. This is a valid way to think about things, and it is not at all unprecedented. For example, if one calculates average hourly wages by summing all earnings in a period and dividing by all hours worked in that period, that average is implicitly hours-weighted: jobs with longer hours worked will contribute more heavily to the average. This could lead to a higher average wage if it is the case that better paying jobs are associated with longer hours worked in a week or year.

Thank you very much for your answers. Hon-Kwong, even if I analyze part-time and full-time workers separately, I still have large differences in workload especially among part-time workers. Thomas, I like your suggestion. I am only concerned about the fact that my dependent variable (hourly wage) is computed by dividing annual wage by hours worked for each individual. Therefore, hours worked would be used twice, to compute both the dependent variable and the weighting variable. Could this be a problem ?

That's a good question. I have never heard anyone raise that objection before, however, and I don't think you need to worry about it -- certainly plenty of authors have ignored the problem! But it is interesting: Suppose a given case had an erroneously high reported number of hours. That would result in (a) an understatement of their hourly wage, and (b) an increase in the weight attached to that observation. That would tend to bias the weighted average hourly wage downwards for the whole sample. But as long as that bias is roughly similar across subgroups of the population and over time it should not have much effect on the Oaxaca decomposition.

Here's another random Oaxaca tip. If you use Stata you have the option of specifying "pooled", which specifies that the coefficients used to evaluate the "explained effects" of differences in X variables between the two periods are the coefficients from a regression that pools the data, and includes a dummy variable to distinguish the two periods. The thing I like about that specification is that it does not rely on the period 1 or the period 2 coefficients, but rather on a kind of average of the two which seems prudent. Also, the total "unexplained effect" will come out exactly the same as the coefficient on the dummy variable in the pooled regression: it is the part that cannot be explained by the differences in X.

Best of luck,

Tom