OHDSI Home | Forums | Wiki | Github

Relative risk ratios / odds ratios in ATLAS

Nice to meet everyone; I’m a clinical informaticist who is new to OHDSI tools. My team is trying to wrangle data from a large patient registry that uses the OMOP Common Data Model, and I’ve run into some walls with using ATLAS.

Our goal is to investigate the relationships between several risk factors (e.g., presence of diabetes mellitus) and several outcomes (e.g., limb amputation).

Our organization offers us only ATLAS for accessing data; we do not have direct Postgres access, so we cannot use HADES. And we’ve run into some trouble with ATLAS when trying to determine incidence rates.

ATLAS seems to be able to define patient cohorts defined by the presence of a condition or event, such as when an amputation procedure occurs or when an observation for diabetes mellitus occurs, and we can use the Incidence Rates tab to get incidence rates on these cohorts (e.g., amputations in people with diabetes mellitus).

However, we also need to get the incidence rates for control cohorts (e.g., amputations in people who do not have diabetes mellitus; people with diabetes mellitus who have not had amupations), and we cannot figure out a way to define these control cohorts or get incidence rates for them. Is ATLAS capable of doing this?

Have you thought about using Observation Period entry events in your cohort definitions to establish a ‘common’ starting point for your population (say, 2017-01-01) and in one population, you look for those with prior diabetes, in another you look for no prior diabetes, and then calculate rate of incidence of amputation among those two cohorts.

But there’s various methodological issues with this when trying to produce risk ratios or odds ratios: is it it simply enough to say ‘with or without diabetes’ and do a fair comparison? What about baseline drugs? Age groups? Etc. I’m talking about the perils of comparing ‘crude’ incidence rates in non-comparable populations. The CohortMethod R package will help you account for some of these methodological issues by propensity score matching. I imagine that if you were to characterize the two populations above (as of 2017-01-01 with diabetes vs as of 2017-01-01 without diabetes) I think you would find many differences in baseline characteristics: (diabetes probably have higher proportion of hypertension and obesity, vs. the population that you are explicitly excluding diabetes). People in 2017 that didn’t have diabetes may develop it later, so how would you handle that? Many considerations should be addressed before you you generate statistics…and there’s a lot of resources in the community and EHDEN Academy that can help.


Thanks for the reply. You certainly have a good point about the peril of using odds ratios in populations with probably different covariates. ATLAS by itself unfortunately seems to be unable to do propensity score matching, and we’re still working on getting direct SQL access so we can use HADES packages like CohortMethod.

I’m realizing now that it may be best to shift our study’s focus from risk ratios to simple incidence rates, e.g., 5- or 10-year mortality rates or amputation rates following an infection. ATLAS’s Incidence Rates mode probably can get at least the incidence rates. However, ATLAS without HADES still probably wouldn’t be able to do multivariate analysis, e.g., hazard ratios for diabetes mellitus on amputation/death, or survival curves.

The takeaway I’m getting is that ATLAS is good for exploration but its analytical abilities are limited without HADES.