10-14-2014 01:10 PM
I am trying to analyze the effects of a new drug on the costs of healthcare.
In my randomized control trial, the treatment group were the patients who were treated by the new drug and the control group were the one who received placebo.
The cost of healthcare is measured by the number of subsequent hospital visits after taking the drug.
I run two different sets of analyses: The first one is a logit regression which predicts the chance of at least one hospital visit and the second one is a negative binomial regression which counts the number of subsequent hospital visits.
The following will be my regression results:
for logit regression:
Logit(HospitalVisist)=beta*Drug + Gamma1*Controls
for negative binomial regression:
Log(# of Hospital visists)= alpha*Drug + Gamma2*Controls
I also know the average cost of a hospital visit and the estimates of alpha and beta are both negative which show that the new drug reduces both the log odds and the log number of hospital visits.
Given this information how do you estimate the cost savings associated with this drug?
Shall I estimate the number AND the probability of hospital visits for each patient through both of these models and then multiply them together to have an expected value which will then be multiplied by the average cost of hospital visits?
or shall I just focus on the negative binomial results and simply multiply the number of estimated hospital visits by the average costs?
12-05-2014 05:47 PM
Have you tried fitting a zero-inflated Poisson or a zero-inflated negative-binomial to your counts? I would think that having a single model would simplify your problem.
Why would the difference in average total hospital cost per patient not provide you a fair estimate of the savings?
12-05-2014 07:06 PM
Questions that require stat advice as compared to general SAS programming always garner less response and take longer. There are less users who have the expertise to answer these questions.
stats.stackexchange.com is an alternative location to post these types of questions.