How to interpret residual plot of a logistic regression？

[sufeipopo]

Hi all,

I'm working on a default probability model and was modeling the default probability using logistic regression. The pearson residual and deviance residual plots are attached. Can anyone give me some idea about how to interpret these two plots? Do these plots indicate heteroskedasticity? serial correlation? overdispersion? curvature etc.?

Thanks!!

 

[sufeipopo]

Seems I should post this to Biostat sub forum. Can I delete this thread? And how to? Thank you!

 

[hlsmith]

Nice avatar. I think this is fine in Statistics since you did not introduce any biologic context.

 

[noetsi]

When I have seen deviance and pearson used in logistic regression it tests how good the model is with grouped data. Which is of course different than residual plots. According to the link below if the n are high enough then these plots work much like standardized residual plots in linear regression. But since logistic regression does not assume homoscedacity or normality it is of little value for that. It's major use appears to be for detecting outliers and for leverage analysis such as Cook's d.

https://onlinecourses.science.psu.edu/stat504/book/export/html/160

http://support.sas.com/documentatio...efault/viewer.htm#statug_logistic_sect057.htm

 

[sufeipopo]

Hi Noetsi, thank you for your reply!!

You remind me of the different diagnosis for grouped and ungrouped data. This is very important to me!

I have another question, so if my data is ungrouped -- cuz I have many continuous data for this model, which analysis should I use to test the model fit? I found hosmer-lemeshow is the one to use but didn't find many helpful examples for me to get a better idea of its application. Do you happen to know anything that would be helpful for my case? Thank you in advance!

 

[sufeipopo]

Hi Noetsi,

I have a question. I used hosmer-lemeshow test to test the goodness of fit for logistic regression on ungrouped data. I wonder if a goodness-of-fit guarantees a good projection forward as well? Or it just shows the model fit is good but doesn't have anything to do with model validation?

Also, if the hosmer-lemeshow test yields a small p-value, what's the common method to improve the model? It would be great if you can give me some guidance. Really appreciate that!!