Hosmer and Lemeshow Goodness-of-Fit on Validation Sample, Help...Many thanks

[Kanyange]

Hi,

I would like to perform the Hosmer and Lemeshow test on my Validation set. I have managed to use it on the training that I have used to build the model, but really I want to run it on the validation sample (out of sample) to see if my model generalizes well....This below was my process for the Training and Output results....But I would like to know how to apply it to validation once I have scored it using the model created on the training. Your help will be much appreciated. Thank you..

proc logistic data=Xsell_File DESCENDING plots(only)=roc;

class VAR1

      VAR2

      VAR3

      VAR4

      VAR5

;

model Click_Flag = VAR1

      VAR2

      VAR3

      VAR4

      VAR5

/ selection=stepwise lackfit;

SCORE DATA=Validation OUT=Validation_Scores (RENAME=(P_1=p));

run ;

Partition for the Hosmer and Lemeshow Test
Click_Flag = 1		Click_Flag = 0
Group	Total	Observed	Expected	Observed	Expected
10	9,236	588	600.52	8,648	8635.48
9	9,664	494	496.83	9,170	9167.17
8	9,668	472	442.13	9,196	9225.87
7	9,656	402	403.19	9,254	9252.81
6	9,716	392	374.06	9,324	9341.94
5	9,711	327	346.88	9,384	9364.12
4	9,612	303	319.56	9,309	9292.44
3	9,665	316	297.07	9,349	9367.93
2	9,665	267	271.16	9,398	9393.84
1	10,047	232	241.61	9,815	9805.39
Hosmer and Lemeshow Goodness-of-Fit
Test
Chi-Square	DF	Pr > ChiSq
7.08	8	0.5281

[jwexler]

Hi, I just posted a similar response to Kanyange.  You would be best served posting this question in SAS STAT Community

I would consider this more of a stat than data mining question

Thanks,

Jonathan

[Kanyange]

Hi Jonathan,

I am quite confused ,  I thought that modelling is part of DataMining??? Also this test helps to validate the model...to see if your actual and predicted are actually similar....as far as I am aware this is datamining...

Thanks

[Reeza]

I think DataMining here refers to Enterprise Miner Software. Logistic regression is a portion of data mining though, and is part of the e-Miner software suite

[Kanyange]

Hi Reeza,

Thanks for your response..I am having troubles to get an answer on this, could you please help? Does EM have this test? Many thanks

[PGStats · Accepted solution]

I think you might be able to pull this off by writing your final model parameter estimates to a dataset with option OUTEST= and then calling PROC LOGISTIC again with your validation data set with DATA=, bringing in your previous model with INEST=, preventing a new fit with MAXITER=0 and requesting H-L test with LACKFIT.

Good luck

PG

[Kanyange]

Thank you very much PG, that's really helpful...