1

I've fitted a binomial GLM, which returns the summary table shown below. What could be the reason that the estimated value of the predictor spikeCountj2m4 has such a large standard error, compared to the other predictors? (highlighted by a red circle in the table below).

See also the contingency table of the response spikeCountj1 and the predictor spikeCountj2m4 below. Is it possible that the standard error in question is inflated because there are no observations where the predictor is different than zero when the response is equal to one? If so, how can I remedy this?

enter image description here

enter image description here

harisf
  • 149
  • 11

1 Answers1

3

"Is it possible that the standard error in question is inflated because there are no observations where the predictor is different than zero when the response is equal to one?"

Exactly right.

"If so, how can I remedy this?"

Either collect more data to allow the cells of your contingency table to fill in, or acknowledge that this predictor adds no predictive or explanatory information about this dependent variable.

In the future you'll want to check coefficient tables for any such "astronomical" values either of a coefficient or of its standard error.

It may be an interesting exercise to note the degree to which other results from the regression differ between versions that include and exclude the problematic predictor.

rolando2
  • 11,645
  • 1
  • 39
  • 60
  • thank you! Do you know of any references that discuss the role of such contingency tables and the regression results? Or is the argument based on "common sense" (i.e. little data --> large uncertainty)? – harisf Feb 03 '17 at 20:33
  • Your situation is related to one of "complete separation" as discussed at http://support.sas.com/kb/22/599.html – rolando2 Feb 04 '17 at 13:23