topic Re: Exact Conditional logistic in Statistical Procedures

Exact Conditional logistic

iuri_leite — Wed, 14 Sep 2022 20:39:03 GMT

Dear All,

We are carrying out a study in which patients are followed-up during a year with quarterly visits. In each visit they are teste for many diseases. One of them is hepatitis B. We are running a logistic model to measure the difference in proportions of participants with hepatitis B (HEPAT_B) in week 0 and week 52.

As the measures are from the same individuals, we used a conditional logistic. Generally, I use proc genmod with the identification of individuals (copid).

PROC GENMOD DESCENDING;

CLASS COPID Week (REF='0');

MODEL HEPAT_B=week/ DIST=BIN LINK=LOGIT;

REPEATED SUBJECT=COPID/TYPE=UN;

ESTIMATE 'TEMPO' week 1 -1/EXP;

RUN;

It runs perfectly.

But because there in no event at week 52, I tried to include some exact test.

I used the following command:

PROC LOGISTIC DATA=TOTAL exactonly;

CLASS week;

STRATA COPID;

MODEL HEPAT_B=week;

EXACT week;

RUN;

It did not work.

However, if I used only the exact logistic without the option strata, the model works fine.

proc logistic data=TOTAL;

class week/ param=ref;

model HEPAT_B=week;

exact week;

run;

I would like to know how to can I used a conditional logistic model with a option exact.

Thank you in advance.

Regards,

Iuri

SAS Output:

NOTE: ODS Graphics are not available when the EXACTONLY option is specified.

WARNING: There are no valid observations since all strata are noninformative.

NOTE: The SAS System stopped processing this step because of errors.

NOTE: There were 4473 observations read from the data set WORK.TOTAL.

NOTE: PROCEDURE LOGISTIC used (Total process time):

real time 0.04 seconds

cpu time 0.01 seconds

Re: Exact Conditional logistic

StatDave — Wed, 14 Sep 2022 20:52:40 GMT

First, note that the model you fit in PROC GENMOD is not a conditional logistic model. It is a Generalized Estimating Equations (GEE) model, though that population-averaged type of model might be fine for your purposes. Also, the model in PROC LOGISTIC without the STRATA statement is an ordinary (unconditional) logistic model. As noted in the description of the STRATA statement in PROC LOGISTIC, the conditional method only uses the data in strata that are informative - that is, strata in which there is some variation in the responses. You said that you have no events in any of your pairs (strata), so none of the strata are informative.

Re: Exact Conditional logistic

iuri_leite — Wed, 14 Sep 2022 22:15:45 GMT

Dear Dave,

thanks for correcting me.

What I really want is to get a p-value to see if there is difference across the categories of variable week taking into account that the observations are from the same individuals when measured at week 0 and at week 52.

When there are events in both weeks the GEE model works fine. But in this specific case, where there is no event at week 52 I do not get the models estimates.

If I ignore the fact there is correlation between observations, using exact logistic regression I get the estimates.

What should I do?

Is it possible to get a p-value considering a GEE model that give me the results when there is no event at week 52?

Thank you in advance.

Best regards,

Iuri

Re: Exact Conditional logistic

StatDave — Wed, 14 Sep 2022 22:44:30 GMT

Exact estimation is not possible in GEE models. However, your data can be thought of as a set of stratified 2x2 tables. As a result, you can take a non-modeling approach by using the CMH option in PROC FREQ in order to get an analysis not using exact methods. An exact approach can be done using PROC MULTTEST. Both are shown in this note.

Re: Exact Conditional logistic

iuri_leite — Fri, 16 Sep 2022 11:10:37 GMT

Dear Dave,

many thanks.

regards