Amusing Friday thoughts: I've been reading Stuart Pocock's 1983 book Clinical Trials: A Practical Approach (do not concern yourself with the reason).
There is a passage on "Statistical Computing" in Chapter 11 of the book which one might have expected would age poorly, but is in fact remarkable for how well several of the statements have held up.
"I would like to refer briefly to the frequent misuse of statistical packages. Since they make each analysis task so easy to perform, there is a real danger that the user requests a whole range of analyses without any clear conception of what he is looking for."
"In my experience it pays to be economical in computer analysis. Only perform analyses you are really interested in and allow adequate time after each analysis to interpret the findings properly"
"The user of a statistical package should aim to suppress superfluous information from the computer output and certainly ignore it when it cannot be omitted"
(there are some other passages which are, ah, just kind of funny to read today, though that is no fault of the author in a book that's nearly 40 years old...but the feeling behind all of these remains today, IMO)
• • •
Missing some Tweet in this thread? You can try to
force a refresh
Fun thread using some simulations modeled on the ARREST trial design (presented @CritCareReviews a few months ago) to talk through some potential features you might see when we talk about “adaptive” trials
DISCLAIMER: this is not just a “frequentist” versus “Bayesian” thread. Yes, this trial used a Bayesian statistical approach, but there are frequentist options for interim analyses & adaptive features, and that’s a longer debate for another day.
DISCLAIMER 2: this is just a taste using one motivational example for discussion; please don’t draw total sweeping generalizations about “what adaptive trials do” from this thread, as the utility of each “feature” must always be carefully considered in that specific context
Here is a little intro thread on how to do simulations of randomized controlled trials.
This thread will take awhile to get all the way through & posted, so please be patient. Maybe wait a few minutes and then come back to it.
This can be quite useful if you’re trying to understand the operating characteristics (power, type I error probability, potential biases introduced by early stopping rules) of a particular trial design.
I will use R for this thread. It is free. I am not interested in debates about your favorite stats program at this time.
If you want to do it in something else, the *process* can still be educational; you’ll just have to learn to mimic this process in your preferred program.
Here’s a brief follow-up thread answering a sidebar question to the last 2 weeks’ threads on interim analyses in RCT’s and stopping when an efficacy threshold is crossed
The “TL;DR” summary of the previous lesson(s): yes, an RCT that stops early based on an efficacy threshold will tend to overestimate the treatment effect a bit, but that doesn’t actually mean the “trial is more likely to be a false positive result”
(Also, it seems that this is generally true for both frequentist and Bayesian analyses, though the prior may mitigate the degree to which this occurs in a Bayesian analysis)
As promised last week, here is a thread to explore and explain some beliefs about interim analyses and efficacy stopping in randomized controlled trials.
Brief explanation of motivation for this thread: many people learn (correctly) that randomized trials which stop early *for efficacy reasons* will tend to overestimate the magnitude of a treatment effect.
This sometimes gets mistakenly extended to believing that trials which stopped early for efficacy are more likely to be “false-positive” results, e.g. treatments that don’t actually work but just got lucky at an early interim analysis.
Having one of those mornings where you realize that it's sometimes a lot more work to be a good scientist/analyst than a bad one.
(Explanation coming...)
Processing some source data that could just be tabulated and summarized with no one the wiser, thereby including some obviously impossible data points, e.g. dates that occurred before study began, double-entries, things of that nature.
Not exactly an original observation here, but when we talk about issues with stats/data analysis done by non-experts, this is often just as big of an issue (or a bigger issue) than whether they used one of those dumb flow diagrams to pick which analysis to do.
OK. The culmination of a year-plus, um, argument-like thing is finally here, and it's clearly going to get discussed on Twitter, so I'll post a thread on the affair for posterity & future links about my stance on this entire thing.
A long time ago, in a galaxy far away, before any of us had heard of COVID19, some surgeons (and, it must be noted for accuracy, a PhD quantitative person...) wrote some papers about the concept of post-hoc power.
I was perturbed, as were others. This went back and forth over multiple papers they wrote in two different journals, drawing quite a bit of Twitter discussion *and* a number of formal replies to both journals.