Feature request: summary of cause-specific censoring for CohortMethod module

[jmb01]

It would be helpful for interpreting CohortMethod results if there was information available on the number of patients in the target and comparator groups censored for various reasons at various time-points during follow-up. Ideally, this could take the form of a table with one row per day on which a censoring event occurs, with the number of patients censored for each reason in the columns. This can be visualized in a similar manner to a Kaplan-Meier. There seem to be at least seven types of censoring, including:

1. end of observation period
2. death (if available)
3. discontinuation of treatment
4. custom criteria specified in the cohort definition
5. switching to or adding comparator treatment (i.e., the type done by CohortMethod)
6. end of the TAR window
7. outcome event occurrence

However, this seems like a complex task as different types of censoring take place in different modules, and one type may be superseded by another at a later step in the analysis. For example, a patient may be censored 100 days after cohort start due to a custom rule in the cohort definition, then censored further in a CohortMethod analysis where the TAR is set to [cohort start + 1, cohort start + 90]. Furthermore, causes 1-4 above seem to be all rolled into the cohort end date, with no easy way to retrieve them from current output.

[jmb01]

Just adding that one example visualization for this is here. In this case, all types of censoring are treated independently.

[anthonysena]

@jmb01 - I'm going to move this over to the CohortMethod repo as I think this functionality will be specific to that package.

[schuemie]

These could be easily identified inside of createStudyPopulation():

1. end of observation period
2. end of cohort
3. switching to or adding comparator treatment (i.e., the type done by CohortMethod)
4. end of the TAR window
5. outcome event occurrence

Where

• end of observation currently could imply death, but we cannot distinguish that, and
• end of cohort could imply discontinuation or some custom criteria specified in the cohort definition, but we wouldn't know which one

How important do you think it would be to untangle the various reasons for end of observation and end of cohort?

It seems the way to handle one even superseding another is the same way we currently construct our KM curves for the outcome of interest: The y-axis of the plot shouldn't represent the absolute number, but the fraction of those still at risk. The table should therefore probably mimic our cm_kaplan_meier_dist table definitions.