OHDSI Home | Forums | Wiki | Github

What are time bound era covariates?

Hello Everyone,

I was watching tutorials on Atlas for PLP and PLE packages.

I understand covariates are nothing but the characteristics of patient and we try to understand their influence in predicting the outcome. For ex: Age can be used to know whether it has any influence on patient becoming sick or not.

Can you please help me understand what are “Time bound covariates” ex: 365d long term,180d Medium term. with a simple example and explanation please ?

I am not able to get these terms and your response would be helpful


These are the features of the population that appear within the time window ending at the index date (the cohort start date) and starting at either 365d (long), 180d (medium) or 30d (short) prior to index. For example, a person may have a MI 325d prior to index, and that feature ‘condition occurrence of MI’ will show up as a long term feature (because it is within 365d) but not medium or short term (because it is not within 180d or 30d).

Other type of time bound covarate is ‘overlapping’ which is used in the drug domain set of covariates, where the exposure starts before the index and ends after the index.

Hope this helps.


Hi @Chris_Knoll,

Thanks for the response. But one question.

Usually a person is brought into a cohort once he fulfils the cohort entry conditions and stays in cohort until he doesn’t violate cohort end criteria

So when I have identified a cohort of 200 patients. Now I would only be looking at their features / characteristics between cohort start date and cohort end date.

Can you help me understand as to why we might need his past information(prior to being qualified of the cohort) ?

I also understand that it isn’t mandatory to have these features but trying to learn and understand the usefulness of these features


Why are you making this assertion?

In both PLE and PLP, you’re going to be considering the patient characteristics at a ‘baseline’, which is typically indexed on the cohort start date, since that date represents the beginning of something in the population (the initiation of treatment for example).

In Patient-Level Prediction, we’d use these baseline features to construct a predictive model for an outcome happening sometime after the cohort_start_date.

In Population-Level Estimation, we’d use these baseline features to construct a propensity score model to create comparable populations for estimating relative efficacy of treatment.

While it may be useful to understand the population characteristics of ‘in-cohort’ time, I would not say this is the “only time you’d be looking at their features/characteristics”.


Hi Chris,

Exactly. So we look for patient characteristics at baseline which is at the index date and use them in PLP and PLE. As we are talking about cohort start date, why do we have to look at his past 365d,180d etc from index date?

I am assuming whatever feature/characteristic that we find on the index date, they are the baseline features.

But to provide more information to the model (to help us get better results), we extract patient’s past information (365d,180d) etc and code it as 1 and 0 as condition present or not. Is that why extract info from his/her patient?

Am I right to understand that based on our research question, we may have to look at patient’s history of conditions as well because that might influence the outcome as well?

But we may not always need this 365d,180d,90d history covariates?

According to this article, baseline covariates are defined as:

A baseline covaraite is usually defined as a qualitative factor or a quantitative variable measured or observed before a subject starts taking study medication (usually before randomization) and expected to influence the primary variable to be analysed.

So, it is not necessarily things that appeared on the index date. But definitely not things that appear post-index (ie after cohort_start_date). This is the reason why I said you would use pre-index time for covariates and not ‘in-cohort’ time as you mentioned initially.

Yes and Yes.

In research I’ve preformed, things that have been observed up to a year prior to cohort start have been considered ‘baseline covariates’. It’s up to you to decide if that is appropriate for your research question. I would think that you would almost always want to use at least the short-term (30d) prior covariates, but I think it would be common to use 365d (long term) covaraites as well.

1 Like

Thanks for patiently answering my questions. Understood now