adjusted length of stay

[Muhammad Rashid]

Hello,
Can anybody help to understand how can I calculate the risk-adjusted length of stay (LOS), I have los which is a continuous variable counted in number of days, I want to calculate adjusted Los of stay between group A and group B. the idea is that group B is sicker cohort than A that's why group B has higher in-patient stay in terms of numbers of days? however, is it possible that once adjusted for differences in characteristics of the two groups, the Los is similar between both groups? I have tried using predict estimation post Poisson model but this has failed to yield my desired results? any help would be appreciated I am happy to share the code or data if that's essential. but I feel this more sort of a general question instead of a coding issue.

[Weiwen Ng]

I am assuming your dependent variable is LOS. You have a bunch of demographic and clinical variables for use as risk adjusters. You have a group variable (A and B).

Assuming you entered the group variable in the regression, then it strikes me that the exponentiated coefficient for group B already tells you some of what you want to know, e.g. the incidence rate ratio of hospital days is 1.35 (made up number).

If you want things in actual numbers, you can use the margins command and instruct Stats to hold each group at the overall mean, e.g.

margins group, atmeans

[Muhammad Rashid]

Thanks, Weiwen, you are a life saviour, yes this is exactly what I wanted but I want the LOS in median and IQR instead of mean and SD...as my LOS is skewed.

[Weiwen Ng]

Hmm. I sense some possible confusion. What do you mean by the quote below?

... I want the LOS in median and IQR ...

When I read this, I would normally think you are basically asking for descriptive statistics - perhaps you meant LOS by group, or just LOS for the study overall. There's no adjustment involved here. You could get this with the summarize command.

You did a Poisson regression. First, you commented about your LOS being skewed. Length of stay can't go below zero. That skews the data by itself. Good news: Poisson or negative binomial regression may inherently handles that issue (i.e. the DV is a count, and it can't go below 0).

Second, most regressions get you an average treatment effect, conditional on whatever other covariates are in the model. In this context, what's the average effect of being in group B, relative to group A, and conditional on the other covariates you threw into the model.

As far as I'm aware, standard Poisson or negative binomial regression does not estimate the median treatment effect, or the 25th percentile treatment effect. In general, quantile regression would estimate the median, 25th percentile, 75th percentile, etc. But I'm not familiar with that regression, and I believe it may only handle continuous data like linear regression would; so it might or might not be a good fit for a count outcome. Anyway, was that what you were asking about?

That's not related to the question you asked originally, by the way. If you were just looking to show this:

is it possible that once adjusted for differences in characteristics of the two groups, the Los is similar between both groups?

Then I'd argue you've already done so. You just need to present the coefficient for group. You might then get into an argument about the adequacy of your risk adjustment, but that's another matter.

[Weiwen Ng]

Actually, there's a side issue I forgot to mention. When you get hospital data and you look at length of stay, you are looking at people who stayed in the hospital. Assuming you have data on inpatients (i.e. they were admitted to the hospital for an overnight stay, versus just being seen in the emergency room and then released the same day), you probably have data where the dependent variable has a support of 1 to infinity. Recall that, as I alluded to in post #4, the support of the Poisson distribution is actually 0 to infinity. The zero-truncated Poisson or negative binomial models might be more appropriate for your use case.

[Muhammad Rashid]

Dear Weiwen, Thank you for a detailed reply, you are right LOS is not below zero hence I have used Poisson to adjust for differences and compare across the two groups. I think I got it wrong when I thought as I have presented the descriptive LOS as median and IQR and assumed that it should be the same in the model as well however as you rightly pointed that's not the case and hence I should just be presenting the means using margins command.
Thank you so much once again for your help.
you are a star