Error message: system is computationally singular

General
#1

Hello,

during my 2SLS-Regression I have a problem with an error popping up that says:

Error in solve.default(vcov.hyp) :
system is computationally singular: reciprocal condition number = 2.85793e-17

It appears, when I call the summary. The regression itself works actually and the results are saved in the environment. Any ideas, what is the problem?

What does the error message mean actually? I found that the error it has to do with singular matrices and that it might appear, or if there are too many variables. Is that correct? In fact I have 8 exogenous variables and 13 instruments and I don't get the error in a standard linear regression. But anyways I have a very similar analysis with 8 variables and 35 instruments that works fine. As a solution I found that I might change the tolerance in the "solve" function, but not sure how to do that in the summary function. So help or a bit of explanation is highly appreciated. Thanks in advance.

Best regards

1 Like
#2

Hi mihefra,

I'm running into the exact same problem and none of the solutions posted for resolving this in other contexts than 2SLS seem to work.
If you find a solution, please be so kind to post it here.

Best

#7

Since I couldn't fix the error I have other questions.

  1. What alternatives are there to the summary() function? I need the p-values of the t-test and adjusted the R-squared
  2. I would like to change the tolerance in solve(), but summary() uses solve.default(). Is there an option to change it inside summary()?
    I can post my code, if that helps. Thank you in advance, I appreciate it a lot.
#8

This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.

#3

Alright, will do. By the way, the error also occurs using stargazer() instead of summary().

#4

Anyone here knows, what "solve.default(vcov.hyp)" is? Looks like something about covariance. Maybe it is possible to isolate the fault and find a solution.

#6

I can't be certain without seeing your data, the code that you ran and the model output, however, this sounds like a situation where two or more of the independent variables in your regression are highly correlated or perhaps even linearly dependent. This means that there isn't a unique solution for the parameters of your regression or that whatever solution you're getting has huge standard errors.