Hosmet and Lemeshow Test Goodness Fit Test

[aranganayagi]

Hi, what is the ideal Pvalue which tells the model is in good fit in H-L Goodness fit test. My understanding is lower the pr>chi sq, better the model. How low it should be.

[PGStats]

Note: the test is called Hosmer-Lemeshow.

It's the other way around. Read the doc.

Large values of [Chisquare] (and small p-values) indicate a lack of fit of the model.

i.e. p-values  greater than 0.05 indicate that the test doesn't detect a signigficant lack of fit.

[aranganayagi]

Thanks PG stats. So Pvalue > 0.05 indicates good fit of model?

[Ksharp]

Yes. The bigger Pvalue , the better your model .

i.e.

H0 : model fit data well 

P=0.9 stand for your model fit the train data very well .( H0 is true )

BTW, if you have big table, H-L is not good test due to it will always refuse H0,

under this scenario you could try GOF option in MODEL statement.

or @Rick_SAS  blog plot the predicted / real Y by proc sgplot .

https://blogs.sas.com/content/iml/2019/02/20/easier-calibration-plot-sas.html

https://blogs.sas.com/content/iml/2018/05/14/calibration-plots-in-sas.html

[aranganayagi · Accepted solution]

In other places, lesser p value means higher significance. Why in H-L test alone, higher p value indicates better model. Can you please help to understand.

[Ksharp]

Did you see my H0 hypethesis ?

P<0.05 stand for refusjing H0 , the higher P value the more close to H0 .

[aranganayagi]

Yes... I understood now. Thanks for explaining. While parameter estimation, p should be lesser, lesser p means it rejects null hypothesis. Whereas here in goodness fit test, h0 is hypothesis built by the model and p should be larger to say that built h0 is true.

[Ksharp]

Yeah. You nailed it .

For " parameter estimation" , the H0 is  the parameter coefficient = 0