One of the most upvoted comments on my #ACR21 plenary was "Why are we still doing observational studies on risks of #steroidsinRA? Given unavoidable confounding, shouldn't we just focus on clinical trials?" I responded in-platform, but thought the topic called for a🧵also. 1/
First of all, clinical trials on this topic are very important, and, when well-designed, will almost always yield results more reliable than those of observational studies when powered to look at common, short-term steroid side effects. 2/
However, many steroid-associated AEs are low-prevalence & require long-term followup to detect. MACE, major fractures, sepsis, VTE, etc. It would be very difficult to power an RCT to reliably assess effect of steroid use (esp. low-dose) on these outcomes. 3/
This is particularly relevant where absolute risk of SAEs may be small but highly relevant at a population level, i.e. avoidable steroid bursts in young, otherwise healthy people. Observational studies can detect these effects; RCTs can't. (see acpjournals.org/doi/abs/10.732…) 4/
People with #RA also use steroids in many ways -- 80% use bursts, 30% long-term low-dose, 30-50% bridge w/treatment escalation. You'd have to do many RCTs to evaluate each exposure type. Well-designed observational studies can account for multiple exposure types. 5/
People who enroll in RCTs of steroid use, esp a study designed to look at harms, are going to have different characteristics than overall population of steroid users, limiting generalizability. 6/
Most RCTs of steroid use have primary outcome = efficacy, with safety secondary. This compounds the power problem & can make steroids look safer than they are due to small # of outcomes. 7/
Tempting to make a composite safety outcome in such studies, but problematic when pooling v. dissimilar AEs (especially when some are short-term/minor and others long-term/major, and study is not powered to either). 8/
Observational studies certainly have major pitfalls re evaluating steroid safety (residual confounding in particular, esp confounding by indication), RCTs aren't perfect either. 9/
As always, we should always design a study to answer a specific question, and should explicitly recognize the limitations of the design we choose. /end
• • •
Missing some Tweet in this thread? You can try to
force a refresh
