0
$\begingroup$

My question is, in logistic regression ,why a variable is significant with a p-value less than 0.05 but no level is significant.

In R

glm(response~var1+var2+var3, family=binomial(link='logit'),data=data)

drop1(model, test = 'Chisq')

all variable with less than ***, except var2 with * but in

summary(model)

the coefficients associated with var2 are not significant.

Thanks for any help!!!

Improve this question
$\endgroup$
5
  • $\begingroup$ var1 and var3 are continuos (discrete) $\endgroup$
    – CJ.Learns.Math
    Oct 24, 2019 at 1:08
  • $\begingroup$ Two reasons. First drop1 is doing likelihood ratio tests while summary is doing Wald tests. These are known to give different results for logistic regression when the fitted probabilities are close to 0 or 1, see for example stats.stackexchange.com/questions/312137 The likelihood ratio tests are more reliable. $\endgroup$
    – Gordon Smyth
    Oct 24, 2019 at 2:05
  • $\begingroup$ Second, drop1 is testing for differences between all the var2 levels at once, whereas summary is testing all the levels individually. The fact that the overall test is significant does not guarantee that any individual level will be significant. That's partly we do the overall test. $\endgroup$
    – Gordon Smyth
    Oct 24, 2019 at 2:10
  • $\begingroup$ @GordonSmyth Could you turn your comments into an answer? $\endgroup$
    – Frans Rodenburg
    Oct 24, 2019 at 2:40
  • $\begingroup$ @FransRodenburg No, I've spent as much time as I have to give. Feel free to repeat my material if you would like to write an answer yourself. $\endgroup$
    – Gordon Smyth
    Oct 24, 2019 at 2:59

1 Answer 1

Reset to default
1
$\begingroup$

By Gordon Smyth,

Two reasons. First drop1 is doing likelihood ratio tests while summary is doing Wald tests. These are known to give different results for logistic regression when the fitted probabilities are close to 0 or 1, see for example link to the example The likelihood ratio tests are more reliable. – Gordon Smyth 2 days ago

Second, drop1 is testing for differences between all the var2 levels at once, whereas summary is testing all the levels individually. The fact that the overall test is significant does not guarantee that any individual level will be significant. That's partly we do the overall test. – Gordon Smyth 2 days ago

Thank you very much for the help Gordon Smyth.

Improve this answer
$\endgroup$
1
  • $\begingroup$ Would you know if there is any way to get a summary coefficient table in R with significance for deviation of coefficients from zero being tested in glm's using likelihood ratio tests instead of using Wald tests? $\endgroup$
    – Tom Wenseleers
    Oct 13, 2021 at 9:48

Your Answer

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

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