I'm trying to do a probit regression with some categorical and continuous variables but Stata keeps omitting certain variables and even claiming that some can't be used to to collinearity problems (I checked if there's a correlation between my DV and said PV but it isn't perfectly collinear or anything). When I also try to find the ML (mfx), it doesn't work.

Here's what I put into the command-box:

probit Vote i.BioDadEdu i.BioMomEdu i.Age i.Region i.Citizenship i.EnrollStat HGC FamilyIncome i.HHSize i.HDE i.MarStat i.UrbanRural i.Income IncomeVal i.SpouseIncome SpouseIncomeVal i.GovtInt if RaceEth==1

Please help! This is for my senior thesis and I'm so lost.

  • 1
    $\begingroup$ Should we guesss the difference between SpouseIncome, SpouseIncomeVal, FamilyIncome and Income IncomeVal or are you going to provide a definition of the variables you intend to use? If you want a guess, I guess you have income of mum and dad plus family income which is the issue. $\endgroup$
    – AKdemy
    Commented Oct 28, 2022 at 4:48
  • $\begingroup$ SpouseIncome & Income are binaries to show that an individual or the spouse does have an income while the other 3 are continuous values $\endgroup$ Commented Oct 28, 2022 at 5:40
  • $\begingroup$ OK, so my guess was right. $\endgroup$
    – AKdemy
    Commented Oct 28, 2022 at 6:41

2 Answers 2


Stata is correct, you must have perfect colinearity.

If stata is just dropping some controls and you’re not interested in those coefficients its fine. Ignore it and just focus on your variable of interest, since the dropped controls would not add anything to the model anyway (due to perfect colinearity).

If you don’t like the variables stata is dropping, then drop some yourself.

In your case, I suspect some of the income variables might be causing the colinearity.

  • $\begingroup$ I do not recommend to simply ignore it and move on. If you do not understand the basics, the entire task is just creating numbers out of thin air. $\endgroup$
    – AKdemy
    Commented Oct 28, 2022 at 6:42

Note there might not be perfect collinearity between variables $x_i$ and $a_i$ as $cor(x_i,a_i)\neq 1 \lor -1$, but still perfect collinearity between $x_i$ and $a_i$ conditional $z_i$ i.e. it is possible $cor(x_i,a_i)|z=1 \lor-1$.

Additionally, note that regression might not be run on your whole sample if there are missing observations. If you regress $y_i$ on, $x_i$, $a_i$, $b_i$ and $c_i$ regression, even without any 'if' condition, will only be run over observations where none of the variable misses an observation. So if you have 300 observations, $y_i$ is available for all 300, $x_i$ has some N/A so it is only available for 250 observations, $a_i$ for 260, $b_i$ for 207 and $c_i$ for 265, the regression will only be run over sample composed of observations where for observation $i$ only if $y_i,x_i, a_i, b_i, c_i \neq N/A$ at the same time.

In addition to the explanation in first paragraph again $cor(x_i,a_i) \neq 1 \lor -1$ but at the same time $cor(x_i,a_i) |y_i,x_i, a_i, b_i, c_i \neq N/A= 1 \lor -1$.

As to what to do about this it depends on what is actually causing the problem which you have to investigate first.

Is it the fact that you want to run regression only if RaceEth=1? Then remove that condition or one of the variables with perfect multicollinearity. If variables are perfectly collinear you don't get any more information from regression by controlling for the additional variable since the coefficient would be either exactly same or inverse which is something you can check via correlation between the variables.

If its issues with observations try to collect larger sample.

If you can't get more data or drop variables try to pivot your topic. I hope you didn't fall into common trap of many students writing nearly whole thesis without first checking if the regressions can be run given the data you have. If you didn't made this mistake you should be able to talk to your thesis advisor about pivoting the topic due to insufficient data to answer the current research question. Most professors will not mind at the early stages of thesis writing process because they expect some ideas might have data issues that make them infeasible.

If you made the mistake of already writing everything out and you are at a late stage of thesis writing process you should still talk to your supervisor. Your supervisor might be able to accept output that is subpar taking into account data issues but even if accepted it will most likely be still reflected in grading.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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