Replication of proc logistic in SAS to sklearn.linear_model Logistic Regression has issues

[Bond007]

Hi,

I have researched on how to replicate proc logistic from SAS to Logistic Regression in Python and come up with the following observations:

1. SAS uses unpenalized regression and python uses penalty=l2 by default. So I have changed it to penalty=none.

2. SAS has default convergence criteria GCONV=1E-8. Python has default value for convergence criteria with tol= 1e-4. I have updated it to 1e-8.

3. I have used solver=lbfgs (default value in Python).

According to the articles that I have found online, the parameter changes 1 and 2 should get the intercept and coefficient outputs for all model variables similar upto few decimal points in SAS and Python implementations.

SAS code:

proc logistic DESCENDING data=dataset_name;

model target_var=&modelvariables/ selection = none CLPARM=WALD lackfit RSQ STB ;

output out=result p=pred;

run;

Python code:

clf=LogisticRegression(penalty='none',solver='lbfgs',tol=0.00000001)

clf.fit(X, y)

When I compare the SAS and Python outputs, I am having an issue as the intercepts have difference of 2 and one of the variable has the coefficient with opposite sign.

Example: 

Intercept: SAS (-7.03), Python (-5.21)

Coefficient for one of the variable: SAS (2.56), Python (-2.45)

The remaining variable coefficients match if I round them up to one decimal point.

Can anyone suggest if any other parameter changes are required in Python?

Thanks in advance.

[Reeza]

https://communities.sas.com/t5/Statistical-Procedures/Understanding-the-difference-between-Logistic-...

Try the C option instead?

[Bond007]

I have tried the C=1e9 option, instead of penalty=none and there are no changes with the output. The issue is not resolved.

[Reeza]

STB displays the standardized estimate. If you remove the STB from SAS do your answers match?

[Bond007]

The SAS logistic regression output gives both the estimate and standard estimate columns, along with few other columns (Error, Chi-Square, Pr > ChiSq).
I am attempting to accurately replicate the SAS model using Python. So for this case, SAS model is non-editable.

[Ksharp]

You got different sign from " SAS (2.56), Python (-2.45)" ,
that means you model different level of Y between SAS and Python,
Try remove "DESCENDING " option from PROC LOGISTIC ,
or try "target_var(event='0')" or "target_var(event='1')"

[Bond007 · Accepted solution]

I have used the following code and the results from SAS and Python Logistic Regression are a match.

clf=LogisticRegression(penalty='none',solver='newton-cg',tol=0.00000001)

clf.fit(X, y)