I am studying the effect of changes in household state on transport mode changes in people's life trajectories. My model is such that each observation I have is a 2-year period in a person's life. My DV is mode change, and it is binary: has there been a mode change in the period? My two predictors are similarly binary: has the respondent married in the period? has the respondent had a child in the period? I am interested in whether these two events have a significant relationship with the likelihood that one changes transport mode.

Question: I do know that it is impossible that someone who is already married marries again. Thus, I am confused about whether I should filter out observation windows in which the respondent is married at the beginning. My worry is that these observations (which happen most often) may be biasing my overall results in the regression. Does this make sense?

  • $\begingroup$ I think you mean that mode change is your DV not IV? $\endgroup$ Jan 22 '15 at 3:49

Your variable really has 3 levels, not 2:

  • Got married
  • Already married
  • Not married and didn't get married

So you should consider adding in 2 dummies, one for "Already Married" and another for "Not Married", with "Got Married" as the omitted category (or shift around to whatever makes more theoretical sense).

If you restrict your sample to only include individuals who are not married at time=0, then your interpretation of the coefficient on "has child" changes. It becomes the average change in log-odds among those that were not married at time=0. I don't think that is what you are looking for.

Also, keep in mind, you should be careful running the restricted sample model compared to the full model with a logit regression. You cannot necessarily compare logistic regression models across different samples, because of the latent scaling factor/unobservable heterogeneity in a logit model. See this post for a more detailed description.

You can include interaction terms if you believe the effect of having children is different for those that got married vs those that were already married vs those that never married.

  • $\begingroup$ Thanks, Robin, but I cannot quite grasp what do I gain from adding the third level... Can you point me to any reference or explanation that helps with that? Is it because I'd now be able to compare the categories 1 and 3 in your list? $\endgroup$
    – nazareno
    Jan 22 '15 at 14:15
  • 1
    $\begingroup$ Adding the third level lets you explicitly distinguish between the never married and the already married. You stated in your OP that you were concerned that there would bias when comparing the newly married with the already married when looking at the effect of getting married. This let's you test all the different relationships: newly married vs. never married, newly married vs already married, already married vs. never married. You can always test if the coefficients are statistically the same with a Wald or LR test, in which case you don't need all 3 levels. $\endgroup$ Jan 22 '15 at 15:29

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.