Logistic regression -accounting for clustering/repeated values

[derbygun]

Hi all,

I have a dataset with a binary outcome (The prescription of drug A Yes/No). My dataset is at patient level, meaning there are unique patients in the dataset. We wanted to consider how neighborhoods could affect the use of drug A, so we merged our data by census tracts to neighborhood-level factors (proportion in the census tract living below the poverty level, proportion with a high school degree). The dataset is now set up such that patients in the same census tracts have the same neighborhood level values (see dataset below).  I want to run a logistic regression to predict the use of drug A, but I would like to account for the repeated values as a result of the census tract. How do I do this? 

PatientID	age	sex	diabetes	arthritis	Drug A	census_tract	prop_below_poverty	prop_with_highschool
1	47	male	0	1	1	47157002000	15.0	47.0
2	51	female	0	0	1	47157002000	15.0	47.0
3	34	female	1	1	0	47157002000	15.0	47.0
4	65	male	1	0	0	47157008500	8.6	75.0
5	27	male	1	0	1	47157008500	8.6	75.0
6	34	male	0	0	0	47157008500	8.6	75.0
7	70	female	1	1	1	47157008500	8.6	75.0
8	62	male	1	0	1	47157021136	12.1	62.0

Drug A = dependent variable /outcome. (Was determined at patient level)

Diabetes ( 0 = no diabetes, 1 = has diabetes)

arthritis (0 = no arthritis, 1 has arthritis)

prop_below_poverty  and prop_with_highschool (continuous variables calculated as percentages)

Thank you. 

[derbygun]

proc logistic data =  A descending;
Class  diabetes  (ref = '0')/ param = ref; 
Class  arthritis (ref = '0')/ param = ref; 
class sex (ref = '0')/ param = ref; 
model DrugA (EVENT = '1') =  age sex diabetes arthritis prop_below_poverty prop_with_highschool;
run;

The above SAS codes show the simple logistic regression I did without accounting for the repeated values of "prop_below_poverty" and "prop_with_highschool"

[StatDave · Accepted solution]

You can use the REPEATED statement in PROC GEE to fit the logistic model accounting for the correlation within neighborhood clusters. For example, 

proc gee;
class census_tract sex;
model drug_a(event="1") = age sex arthritis prop_below_poverty prob_with_highschool / dist=bin;
repeated subject=census_tract / type=exch;
run;