OHDSI Home | Forums | Wiki | Github

Negative control outcomes implementation

Hi everyone,

Just need some clarification on how the negative controls outcome analysis is implemented and if this affects the numbers of the population used. I am doing first user comparative studies using a selection of different drugs with dementia outcomes. I have my target and comparator populations and I excluded people who had a dementia outcome before drug as my hypothesis is that certain drugs are associated with a reduction in dementia risk. I want to apply negative control outcomes which I now have a list of >50 outcomes that are not associated with the exposure drugs so I would expect the HR to be 1. My issue is understanding is how you can use the exact populations (target, comparator) from my initial study when trying to use different outcomes as there are going to be cases with negative outcomes where the events only occur before the drug use so these people should be excluded? Therefore for each negative control outcome analysis for each of the 50 outcomes, there could be a different number of people from my original populations. Or does the implementation only class people with a negative outcome only if it occurs after the drug start and ignores it if it occurs before? Hopefully this makes sense but let me know and happy to discuss further :smiley: Thanks Danielle

Hi Danielle!

To make the analyses for the negative control outcomes as similar to the analysis for the outcome of interest I recommend applying the same rules to all. So if the rule for your outcome of interest (dementia) is that people who had the outcome prior should be removed, then you should do the same for your negative control outcomes. That does mean that each analysis will use a slightly different set of people.

This is the default behavior implemented in the CohortMethod package. To save time, the default approach the package takes is to apply all exclusion criteria except removing people with the outcome prior, then fit the propensity model, then remove the people with the outcome prior, and then perform matching / stratification / weighting. This means you’ll only have to compute the propensity model once (which is often quite expensive since we fit very large propensity models) for all outcomes.

Thank you for the quick response and explanation - really helpful! :slight_smile: