adjustment for multiple comparisons in proc logistic

[Kyra]

I am looking into association of many different job titles with head and neck cancer.

I ran separate logistic regression models for all the job titles adjusting for same confounders .

The control group for job titles is same for all the models.

How should i adjust for multiple comparisons?

[StatDave]

Specify your job title variable in the CLASS statement and use the PARAM=GLM option in that statement. Of course, job title will also appear in your MODEL statement. Then you can also specify it in an LSMEANS statement and use the ADJUST= option to select one of the various adjustment methods available. See the LOGISTIC documentation for details on these statements and options.

[Kyra]

Thank you very much for the prompt reply. I really appreciate it. But I still did not get my answer. The code I used is (the level of cook is 1 vs 0 in below code):

proc logistic data=red.merged;

class studyname ageq sex Education_levelmi1 race region drink_dayq tobaccoyearq cook/param=GLM;

model case= studyname ageq sex Education_levelmi1 race region drink_dayq tobaccoyearq cook / scale=none aggregate;

LSMEANS cook / e diff oddsratio cl adjust=dunnett;

run;

SAS log says

The ADJUST=DUNNETT p-values and confidence limits cannot be computed because the

Dunnett-Hsu approximation did not converge. Try ADJUST=SIMULATE.

when I did adjust=simulate it says

Cannot compute simulated p-values with inestimable LSMEANS.

I am posting part of my codes just to clarify what I did previous to adjustment:

proc logistic data=red.merged;

class studyname ageq(ref="<40") sex(ref="Female") Education_levelmi1(ref="No education") race region drink_dayq(ref="0 (Non-drinkers)") tobaccoyearq(ref="0 (Never smoker)") cook(ref="0")/param=reference;

model case= studyname ageq sex Education_levelmi1 race region drink_dayq tobaccoyearq cook/ scale=none aggregate;

run;

proc logistic data=red.merged;

class studyname ageq(ref="<40") sex(ref="Female") Education_levelmi1(ref="No education") race region drink_dayq(ref="0 (Non-drinkers)") tobaccoyearq(ref="0 (Never smoker)") painter(ref="0")/param=reference;

model case= studyname ageq sex Education_levelmi1 race region drink_dayq tobaccoyearq painter/ scale=none aggregate;

run;

I ran this kind of model for 25 job titles.

[StatDave · Accepted solution]

If the point is to make comparisons among your job titles, then it makes no sense to run a separate model for each title. Instead, fit a single model containing a job title variable that has as many levels as you have titles and use all the data in that single model fit. Then your LSMEANS statement will have multiple comparisons which it would make sense to adjust. In the separate models that you have fit, there is no point in doing any adjustment since each has just a single comparison. 

[Kyra]

Thank you very much for the reply. The goal of our study is to see association of each job title with head and neck cancer irrespective of other jobs.

We do not want to make comparisons among job titles.

I suppose we will not make adjustments then. Thank you once again.

[StatDave]

In that case, then you could adjust the p-values from the individual job title parameters by creating a simple data set with those p-values in a variable and then use PROC MULTTEST to do an adjustment. See the example titled "Inputting Raw p-Values" in the Examples section of the MULTTEST documentation. The Holm adjustment method is generally applicable.

[Kyra]

Thank you very much. I will try this.

[Kyra]

Hi,

Proc multtest with Holm method worked for me.

Thank you very much for your help.