I'm working with panel data with evidence of heteroskedasticity. I am looking at returns to college major based on gender, so fixed effects is not an option. I thought quantile regression with clustered standard errors (Santos Silva) would be a good option, but I am now second guessing because I do not think I have enough observations to get reliable results. I have about ~2,800 observations. When I ran OLS, the results for all of my other explanatory variables made sense (phds, master's, and BA degrees earned more than associate degrees ect.). However, when I run qreg2 at the default .50, associate degrees are predicted to earn more than other education levels, and other explanatory variables do not make sense. I would really appreciate any suggestions for the best regression technique to use. Thank you!
Code:
qreg2 log of income gender race marital_status highest_degree occupation hours_worked math_score college_major majorXfemale
(all respondents are approximately the same age)