Re: Biomedical Informatics Approaches to Identifying Drug–Drug Interactions

    loading  Checking for direct PDF access through Ovid


We read Han et al’s “Biomedical Informatics Approaches to Identifying Drug–Drug Interactions: Application to Insulin Secretagogues” with interest.1 We have concerns whether the underpinning assumptions of the self-controlled case series method, whose novel use we support, are satisfied.
Self-controlled case series conditions on follow-up period, and assumes the outcome event does not affect the start and end of follow-up. In this study, discontinuation of an insulin secretagogue defined follow-up end. It is likely that after the outcome event of hypoglycemia, patients are more likely to be switched to an alternative that causes less hypoglycemia and, thereby, end follow-up. This could bias the analysis in either direction. We recommend choosing a completely independent observation period, such as entry and exit from the database.
Violation of this assumption can be investigated with a histogram of the intervals between outcome and follow-up end to look for an excess of short positive intervals. Alternatively, a sensitivity analysis could be performed removing subjects whose follow-up ended shortly after the outcome event, although this could be prone to bias if it selectively removed more severe hypoglycemic events. Farrington et al2 describe a self-controlled case series extension to deal with event-dependent censoring in a study of the association between antipsychotics and stroke.
Another issue concerns drug pairs for which there is a known potential drug–drug interaction. Self-controlled case series assumes the outcome does not affect the likelihood of subsequent exposure. However, patients diagnosed with hypoglycemia would thereafter be unlikely to receive a drug pair with that known risk. The bias here would be upwards because follow-up time after the outcome will tend to be unexposed. Again, this can be investigated graphically. A histogram of the intervals between outcome and exposure would show a paucity of exposures after an outcome event. If this is temporary, it is possible to address this violation by including a pre-exposure risk window removing this time from baseline. However, if it is permanent, the standard self-controlled case series method is inappropriate. Douglas et al3 describe an example of violation of this assumption in the context of orlistat and liver injury where liver injury is associated with an increased short-term likelihood of subsequent orlistat exposure.
In summary, the self-controlled case series has many strengths in observational pharmacoepidemiology, but analyses require careful consideration of whether the method’s assumptions are realistic.
    loading  Loading Related Articles