Effect of Cohort Definition on Regression Coefficients

[Nguyen Huong]

I am analyzing the effect of years since immigration (yrsimm) on annual income (incwage) and want to control for cohort effects, where cohorts are constructed based on the year of immigration. I noticed that the coefficient on yrsimm changes depending on how cohorts are defined (e.g., 1-year, 4-year, or 5-year cohorts). In particular, the last cohort initially consisted of only one year. After merging this last year into the previous cohort, the coefficients changed dramatically compared with the original specification (5-year cohorts + last cohort with only 1 year of observations). Could you help me to understand that why it happened and how I should define the cohorts to maintain a more stable results?

[George Ford]

The effect sizes vary by year, and you're getting averages when you define a cohort. one year is likely to look a little different than 5 year averages. Play around with it and try to determine what provides somewhat stable patterns. I'd think there'd be a pattern to the coefficients.

[Nguyen Huong]

I tried using different cohort definitions in the regression, and the coefficients do change, but they vary less when the cohorts have equal numbers of years.
Thank you for the suggestion.