I hate to be "that guy," but the statement about physiologic "price" is true, but also unsupported by the study that is linked in this tweet...1/x https://twitter.com/loomba_rohit/status/1297588518538051587

The study of vasoactive therapy in observational datasets offers a perfect example of confounding by indication, and few, if any datasets contain the type of data needed to draw causal inference from associations between drug and outcome...2/x

Putting aside the known limitations of the PHIS dataset in assessing pediatric cardiac surgical outcomes, and the fact that the exposure variable of interest is ever/never prescribed a drug (no dose, duration, or combination data), the analyses are not adequately controlled 3/x

for variables that would give any information on a causal association. Further, the methods do not include any mention of controlling for clustering within center, and we know that practices AND outcomes vary across hospitals, so this could impact observed associations...4/x

Methods that could enhance causal inference include propensity matching (wouldn't be possible in this dataset because illness severity and physiologic data needed to match are not available) or instrumental variable analyses..5/x

which could be useful to study individual drugs that are used by some hospitals a lot, and never at some other hospitals. But the real reason I'm responding to this tweet above is because it highlights one of the pitfalls of #medtwitter 6/x

In the paper, which I read today, the authors highlight some of these limitations that I've pointed out. In the tweet there's no nuance. This is where we need to be careful in our #SoMe communication. If we promote our science we have communicate in the same... 7/x

balanced way we do when we submit peer-reviewed material. Final thought: if anyone is interested in trying to learn more about vasoactive use in the #PedsCICU the @pc4quality database is now collecting much more detailed data on postop and medical patients than we used to. 8/x

We would still struggle to deal with some of the limitations I raise above, but there may be opportunities to understand how certain practices impact outcomes. Study ideas are welcome! Final, final thought: understanding the right indications...9/x

for the "right" drug would be very, very, very hard. I'm not sure there is a study design that would really work outside of *maybe* an RCT. This may be one area of #PedsICU practice where first principles and physiology are the best guide. 10/x

If there are other good ideas out there on how you would apply statistical methods to enhance causal inference in observational datasets for this study question please reply or DM me. 11/11