topic Interpreting Proc Survey Logistic Output in Statistical Procedures

Interpreting Proc Survey Logistic Output

amng — Tue, 21 Jul 2020 02:11:17 GMT

Hi all,

I'm trying to run an analysis looking at health literacy (1 = low health literacy...4 = high health literacy) predicting self-management of diabetes (0 = low self-management...6 = high self-management). I'm using the following code and just want to make sure everything is in order, since I am getting an OR of 0.743 (showing that individuals with high health literacy are less likely to adequately self-manage, which is the opposite of what I would expect). Any feedback or tips would be appreciated. Code is below:

PROC SURVEYLOGISTIC DATA=test;
WEIGHT _LLCPWT;
CLUSTER _PSU;
STRATA _STRWT;
MODEL MANAGE (event='6')= healthliteracy;
RUN;

Re: Interpreting Proc Survey Logistic Output

SteveDenham — Tue, 21 Jul 2020 12:17:11 GMT

I think by using the 'event=' option you have compressed your response variable in to a binomial. If you want to treat this as a cumulative logistic, I think you should change your model statement to:

MODEL MANAGE = healthliteracy;

You will need a CLASS statement that includes healthliteracy.

If you want to treat this a generalized logit, so that all ORs are compared to a reference level, this would change to:

MODEL manage(reference='6') = healthliteracy/link = glogit;

You still must have the CLASS statement.

I hope this helps some.

SteveDenham

Re: Interpreting Proc Survey Logistic Output

PaigeMiller — Tue, 21 Jul 2020 12:36:09 GMT

According to the documentation

The EVENT= option has no effect when there are more than two response categories.

But I agree that reference='6' is needed.

Re: Interpreting Proc Survey Logistic Output

StatDave — Tue, 21 Jul 2020 14:10:13 GMT

Always check the log. With your code, there will be a message that EVENT= is ignored since this is not a binary variable (as stated in a previous response). There is also a message in the log (and in the results) indicating that you are modeling cumulative logits, meaning that you are treating the response as ordinal, which is probably what you want rather than an generalized logit (LINK=GLOGIT) model which ignores the response level ordering. But as also indicated in that log note (and in the results), you are modeling the probabilities of LOWER ordered levels of the response. The Response Profile table shows that the lower Ordered Values are associated with lower values of your actual response. So, the interpretation of that odds ratio is actually that increasing the predictor decreases the odds of a low response (low management), meaning that it increases the odds of a high response (high management).

Re: Interpreting Proc Survey Logistic Output

StatDave — Tue, 21 Jul 2020 14:17:34 GMT

The results will be more intuitive if you specify the DESCENDING response option (rather than EVENT= or REFERENCE=).

model manage (desc) = healthliteracy;

Also, see this note on interpreting odds ratios in ordinal models.