Mediation with bootstrapped confidence intervals, multi-categorical predictor, and count outcome

[confooseddesi89]

Hello,

I am interested in examining whether city Size (mediator; continuous) mediates the association between Race (predictor; 5 categories) and PA_days (outcome; count variable). Covariates are Sex and Depression (both dichotomous). Below are my data:

Data Race_PA;
Input ID Race Size PA_days Sex Depression;
datalines;
8	4	169.425	5	1	0
20	2	172.3	1	0	0
21	2	164.3	0	1	1
22	4	163.8	4	0	0
28	2	174.7444	0	1	0
35	3	147.8	0	1	1
36	3	164.675	1	1	0
37	2	167.55	5	0	0
50	4	179.675	3	0	0
52	2	171.55	1	0	1
57	3	173.925	3	0	0
58	3	128.8	5	0	0
60	5	178.4111	1	1	0
71	2	164.05	2	1	0
77	2	160.9667	7	0	0
80	4	175.425	0	1	0
83	5	159.05	0	0	1
88	1	155.8	3	1	0
91	2	158.3	4	0	1
107	4	172.64	3	1	1
118	2	166.55	4	0	1
124	1	160.3	4	0	0
128	4	176.925	2	0	1
147	3	176.55	3	0	0
164	3	172.425	7	1	1
165	2	170.17	7	1	0
169	3	173.8	6	1	1
172	1	173.46	2	0	1
174	2	174.87	3	1	0
184	5	165.925	4	1	1
187	2	172.35	0	1	1
191	1	175.05	0	0	1
197	3	164.8	5	0	1
198	1	173.425	4	1	1
203	5	153.675	0	0	0
204	3	178.3	7	0	1
205	2	180.05	3	1	0
206	3	163.925	5	1	0
207	1	172.55	6	1	1
229	3	172.8	6	1	1
233	3	164.33	2	1	1
247	4	174.925	1	1	0
258	2	176.05	5	0	0
268	2	176.9	5	0	1
272	2	153.3	3	1	0
287	1	168.55	5	1	0
290	5	170.1	4	0	1
292	1	169.175	3	0	1
295	2	155.3	4	1	1
299	5	170.925	0	0	1
309	2	172.925	3	0	0
316	3	171.05	4	1	0
317	2	156.425	6	1	0
336	2	167.925	6	1	0
340	1	171.3	5	0	0
346	1	170.175	1	1	1
362	3	163.175	4	0	0
369	4	162.175	0	1	0
374	1	168.3	3	0	1
378	3	148.05	4	0	1
383	1	170.05	6	0	0
385	4	171.34	5	0	0
388	1	167.5	4	1	1
406	2	172.05	3	1	0
407	1	175.55	2	1	0
413	3	166.175	2	1	1
416	2	177.425	4	1	0
418	3	175.4111	4	0	0
419	2	179.8	2	0	0
424	3	168.82	4	0	1
433	2	164.07	7	0	1
435	4	156.42	2	0	1
436	2	174.8	0	0	0
439	3	175.8	3	1	1
451	4	164.64	4	1	1
456	4	145.425	5	0	1
459	4	163.675	4	1	1
466	5	159.55	0	1	0
;
run;

I would like to use bootstrapped confidence intervals for the assessment of the indirect effect of mediation (i.e., the Andrew Hayes method). I know PROC CAUSALMED can handle count outcomes but not multi-categorical predictors, and Hayes' PROCESS macro can handle multi-categorical predictors but not count outcomes. Is there a way to assess mediation using bootstrapped CIs with a multi-categorical predictor and a count outcome in SAS? Thanks.

[sbxkoenk]

I moved your topic to the "Statistical Procedures" - board.

Maybe @SAS_Rob can help?

(he's getting notified by me at-mentioning him)

Koen

[StatDave]

Just to note something you might consider - if you have a count of some event recorded in cities and you have the population sizes, then typically the interest is in modeling the rate, count/size. Size would be used as an offset in the count model. That can be easily done by creating a variable as log(size) and then using that in the OFFSET= option in a log-linked Poisson model. For example, 

proc genmod;
model pa_days=race sex depression / d=p offset=logsize;
run;

[SAS_Rob]

From within SAS Viya there is a BOOTSTRAP option in CAUSALMED that uses bootstrap resampling to compute standard errors and confidence intervals for causal mediation effects and decompositions.  Below is a link to the documentation.

SAS Help Center: Bootstrap Methods

[confooseddesi89]

As I mentioned in my original message, PROC CAUSALMED does not allow for multi-categorical predictors. I have a multi-categorical predictor.

[SAS_Rob]

If you mean that it does not allow treatment effects that are more than 2 levels, then that is correct.  It does however allow multi-level predictors in the COVAR statement.  

All that being said, if you are looking for a way in SAS to fit a causal mediation model with a treatment variable with more than 2 levels, then I am afraid that is not possible.

[confooseddesi89]

Yes, I meant the "treatment" and not covariates in this case.