 |
 |
Just a caveat. I train in EBM, and many newbies get the impression (usually from the research articles themselves) that as long as the study used a "good" method of imputing data, the amount of attrition doesn't matter. I'd like to see it made it clearer that high attrition is a serious weakness apart from how the values are imputed, because we can't automatically assume randomness of the missing data-- even if we control for all the variables we can think of, there may be unknown variables that caused the missingness to be not random. (Steve, thanks for mentioning the 30% attrition rate in this study as a weakness.) For the sake of newbies, it would be great if thought leaders and authors would routinely present that caveat in the introduction to any discussion of imputation methods, just to set the stage; e.g., "high attrition presents a serious risk of bias that can never be cancelled out by statistical gymnastics; but if researchers decide to go ahead with ITT analysis in spite of a high number of drop-outs, multiple imputation is one of the best methods for filling in the missing values." That would help the newbies to understand the distinction, and not jump to their wrong conclusion (that a good imputation method obviates any risk of attrition bias). Thanks! Teresa Benson, MA, LP Clinical Lead, Evidence-Based Medicine McKesson Health Solutions [log in to unmask] -----Original Message----- From: Evidence based health (EBH) [mailto:[log in to unmask]] On Behalf Of Steve Simon, P.Mean Consulting Sent: Thursday, May 16, 2013 5:07 PM To: [log in to unmask] Subject: Re: Multiple imputation (branch of "Why did the CMAJ publish this study") Obviously we come from different perspectives. But to clarify, multiple imputation is not designed to provide narrower confidence intervals. It is designed to prevent the biases caused by simpler flawed methods such as complete case analysis and last observation carried forward. The solution to the problem is for every paper to include a copy of the data set, appropriately de-identified. Then we can recreate the original analysis and try other reasonable alternatives. Steve Simon, [log in to unmask], Standard Disclaimer. Sign up for the Monthly Mean, the newsletter that dares to call itself average at www.pmean.com/news