Binary Data with Count Dependent Variable?

Question

Howdy internet strangers!

I've put a sample of my data below, but basically I have the number of accidental falls in a particular set of patients vs. drugs that the patients were taking. The drugs they were taking are coded in a binary manner - 1 if they were on the drug, 0 if they were not. Any ideas on how to analyze this data?

The question I would like answered is "Do patients on drug X have an increased risk of falling?" And an risk ratio for each drug would also be great.

I have an engineering background and can run scripts in R but I am by no means a statistic expert. I was doing GLM with Poisson regression (in R) for each drug, but I have an inkling this is the wrong test to be doing. I would like to use some sort of multiple comparisons (that takes into account all of the drugs that each individual patient was on) but I don't know what test is most appropriate. Thank you in advance for your help!

Answer 1

Gabe, welcome to the site!

I'll try to give you some answers and some clarifications so that you can dig deeper.

• First of all, your dependent variable (# of falls) is not continuous it is discrete, despite having infinitely many values that it can take.

• Second (and most importantly) of all, Poisson regression was not a bad idea at all and might be the best way to go. One import consideration is how long each patient has been on the drugs. The parameter being estimated in Poisson Regression is a rate, how many times we expect an event to happen over a fixed period of time. Poisson regression can take this into account with something called exposure, which I suggest you look into.

• You talked about comparing an increase in fall risk, which makes it sound like a binary event (fall or no fall) versus a count event (how many times will they fall). Do you want a risk instead of a rate? If so you might want to consider dichotomizing the result and trying logistic regression, though personally I think the Poisson regression is preferable.

• Lastly, if you don't have a lot of data, you might want to consider Bayesian methods and/or reducing the number of questions you're trying to answer.

Hope this was helpful and gave you enough to keep digging. Let me know if you have follow-ups.