Proc Glimmix: How do I test whether an intervention effect varied by another variable?

[blmorgan]

Hello,

This is my first time posting to the SAS community board, so please bear with me. I'm analyzing the effect of an intervention that was rolled out across several healthcare clinics on antibiotic prescribing. We have a year of pre-intervention data, a year of the intervention roll-out, and two years of post-intervention data. I'm following the process/code in this paper (specifically, scenario #3): https://www.lexjansen.com/wuss/2014/74_Final_Paper_PDF.pdf

I had no issues with the first analysis looking at the effect of the intervention organization-wide and within the clinics. However, we now have added information regarding whether the visit was in the office or through telehealth and we'd like to know if the intervention effect varied depending on the type of visit. I was originally thinking I would just include an interaction term (intervention*visit type) but I'm unsure if that is sufficient? Any recommendations on how this is usually done would be greatly appreciated; this is my first time using proc glimmix.

I didn't include any data or my code because it's more a theoretical question than me having trouble with coding, but I can if it would be helpful (it's basically the same coding as scenario #3 in the paper but tailored to my data)

[SteveDenham · Accepted solution]

The interaction effect is precisely what you need to identify whether intervention is different by type.  While I generally believe in GLIMMIX, this analysis sounds much like a moderator/mediator effect. Additionally, your dataset is much more likely to be observational than from a designed experiment/survey.  This sounds to me like a good candidate for PROC CAUSALMED.  Check the documentation and especially the examples found at https://documentation.sas.com/doc/en/statug/15.2/statug_causalmed_overview01.htm 

The https://www.lexjansen.com/wuss/2014/74_Final_Paper_PDF.pdf  deals with the interaction by working with random effects, implying that the mediation affects the observed variability, rather than causing a location shift.

SteveDenham

[blmorgan]

Thank you for your response! I appreciate your time and insight. I chose to use proc glimmix (and the linked paper) based on my data having nested hierarchies, repeated measures, clustering, clinic-wide and organization-wide pre-, post-, and intervention effects, and the intervention being phased across clinics. It wasn't designed as a true experimental study but as a quasi-experimental study. I'm hesitant to start the entire analysis over with a new procedure, but I'll definitely look into proc causalmed and see if it would be a better fit! 

Thank you again!

Sincerely,

Brittany