I am working with a large healthcare dataset from 15 states to conduct a difference-in-difference analysis. I have questions related to appropriate treatment of standard errors and interpretation of genmod results. I am attempting to develop hierarchical fixed effect regression models predicting the impact of a policy intervention on treatment frequency (procedures per 100,000 people), adjusting for patient-level confounders. However, I want to ensure that the standard errors appropriately reflect clustering at the state level. The dataset is not a sampling or survey data, but rather complete capture from each state. Reading through the different posts on this site, I have seen some people advocate using the absorb statement in proc glm, but I am unsure if this produces accurate standard errors. Does anyone know if that is the case? Alternatively, others advocate using proc genmod, with example variables and code below: outcome=treatment_freq intervention=policy (1=intervention; 0=no intervention) time-period (pre vs. post-intervention)=period (1=post-intervention, 0=pre-intervention) confounders=conf1, conf2, etc. State variable=state proc genmod data=have; class policy (ref='0') period (ref='0') conf1 (ref='0') conf2 (ref='0') state; model treatment_freq= conf1 conf2 policy period policy*period; repeated subject= state/type=ind; run; This produces the output shown in the attachment. Am I correct in interpreting the policy*period parameter estimate of 0.0434 as reflecting a 4.34% increase in treatment frequency associated with the intervention, with the standard errors accounting for clustering at the state level? Thank you!!
... View more