topic Re: Multivariate regression analysis for skewed medical cost in Statistical Procedures

Multivariate regression analysis for skewed medical cost

petax004 — Sat, 21 Nov 2020 02:01:30 GMT

Hey,

Like most medical cost data, I've right skewed distribution (note: 99% observations are >$0) and my goal is to perform multivariate regression to compare the cost outcome between two groups while controlling for other covariates. I did initial web search to find any code/tutorial that would help me, found some that talked about two part model and would like your help in writing the code and interpreting it.

From my quick research I've written the following code and results are in log values, can I just retransform the values? should I be adjusting for smearing bias?

	proc genmod data=outcomes_stats4;
		class exposure_new (ref="0") male HDHP_Flag Invokanaa_16 uncntrldb15_all invok_adh_q_new_p invok_adh_16_p sglt_adh_16_p;
		model mr_allowed_16 = exposure_new male HDHP_Flag Invokanaa_16 uncntrldb15_all invok_adh_q_new_p invok_adh_16_p sglt_adh_16_p age riskscore eff_sglt2_grp / dist=gamma link=log;
	run;

Re: Multivariate regression analysis for skewed medical cost

StatDave — Sun, 22 Nov 2020 17:08:29 GMT

There is really no need to do any transformation. If you want estimated adjusted means on the response variable scale, use the LSMEANS statement with the ILINK option. If you want to do multiple comparisons among those means, add the DIFF option in the LSMEANS statement. You can get predicted values on the response scale using the PRED= option in the OUTPUT and confidence limits are available with the UPPER= and LOWER= options.

Re: Multivariate regression analysis for skewed medical cost

PaigeMiller — Sun, 22 Nov 2020 18:45:28 GMT

Regression does not require normally distributed Y variables, so sayeth @Rick_SAS and many others (including me). So no need to transform the data because the Y are not normal. (It does require the assumption of normally distributed errors.)