Linearity assumption in Logistic

[Heebyung Koh]

Hi statalist.
I wonder during logistic regression, do I have to check linearity assumption between continuous independent variable & logit of dependent variable in aspect of univariate model or multivariate model?

[Clyde Schechter]

Well, implicitly you do that when check the calibration of the model, for example with -estat gof-.

[Rich Goldstein]

yes, you should check the linearity of this relation; in addition to what Clyde suggested, this can also be done directly with -lowess-; see

help lowess

[Heebyung Koh]

yes, you should check the linearity of this relation; in addition to what Clyde suggested, this can also be done directly with -lowess-; see

help lowess

Thank you Clyde Schechter & Rich Goldstein .
I Knew that -estat gof- ; Hosmer and Lemeshow's goodness-of-fit test is test for goodnees-of-fit, but isn't it for linearity?
I knew there are several way of idendify linearilty in Logistic regression. eg Box-tidwell regression or draw variable - logodds graph.
But I don't know why, maybe number of data is too large (300,000), when I ran boxtid, it just showed that it is in progress and didn't show result.
So I'm searching another way to identify linear assumption in Logistic regression.
And.. Isn't lowess only available in univariate analysis?
I want to check the effect of variable in multivariate model. Isn't it enough to perform linearity test in univariate?

[Rich Goldstein]

-lowess- will give you a good idea (in my opinion) of whether there is anything to worry about re: linearity; your results might change when adding additional predictors but I have never seen it change from non-linear to linear in that situation; -lowess- also gives an idea of what functional form might be appropriate if there non-linearity is shown (or even hinted at); none of this, however, means that the same form will be there after adjusting for other variables; I do note that if there is an interaction between the apparently non-linear predictor and some other predictor that your job might be more complicated (but interactions are more complicated anyway)

[Heebyung Koh]

-lowess- will give you a good idea (in my opinion) of whether there is anything to worry about re: linearity; your results might change when adding additional predictors but I have never seen it change from non-linear to linear in that situation; -lowess- also gives an idea of what functional form might be appropriate if there non-linearity is shown (or even hinted at); none of this, however, means that the same form will be there after adjusting for other variables; I do note that if there is an interaction between the apparently non-linear predictor and some other predictor that your job might be more complicated (but interactions are more complicated anyway)

Thanks you