@DongNguyeb I'm glad you read the book. I really like it. It emphasizes how Bayesian reasoning is more actionable, e.g., if trying to win money you'll take more profitable actions with Bayesian methods. This translates to scientific decision making also.

@DongNguyeb @5_utr Read Nate Silver’s book The Signal and the Noise to get a good sense of how Bayesian methods are highly advantageous in a decision making framework.

@trumanfrancis @DongNguyeb @5_utr Great question (but nonparametric data -> semiparametric test). covers a good deal of this. A Wilcoxon test gives you a unitless concordance probability and an odds ratio, both of which can be interpretable. But translate those to means/quantiles.

@DongNguyeb @5_utr A lot of conventions are silly and wrong. You have provided no defense of this convention, and I'm glad the convention is not followed by most decision makers.

@DongNguyeb @5_utr Ronald Fisher would disagree with you.

@DongNguyeb @5_utr Using cutoffs for P-values is widely criticized, though accepted by many who don't want to delve too deeply into its defects.

@PavelRoshanov Wonderful displays! And the paper helps us fight the notion that change in renal function is what’s important. Absolute function is almost always the dominant variable.

The R rms package, now 34 years old, has a massive update. A key new #Statistics capability is profile likelihood confidence intervals and LR tests 4 general contrasts. And for ordinal regression the number of model intercepts is now unlimited. #RStats

@Daniel_R_Kowal I don't agree with the disadvantages you listed. Semiparametric models are essentially adjusted ECDF (empirical cumulative distribution functions) and there are not degeneracy problems with ECDFs. Transformation models seldom allow for discontinuities, bimodal distributions, ...

@Daniel_R_Kowal Not at all. Efficiency is 3/pi compared with parametric methods WHEN normality holds, and prediction is easy on any scale: mean, quantiles, exceedance probabilities, ...

@flscientist @5_utr My resources related to this are at with the bottom line being the use of the highest resolution measure that is clinically or physiologically relevant or relevant to patients.

@5_utr @flscientist @MayoUrology @MayoRadOnc @vandy_biostat I’m surprised anyone is still trying to solve for cutpoints by minimizing p-values. Really questionable statistical practice, and will allow you to always find a cutpoint even when it doesn’t exist.

@Daniel_R_Kowal One reason is that semiparametric ordinal models are ready to handle any distributional problem including a discontinuity.

@ADAlthousePhD @drjgauthier @5_utr One of the primary estimands of the longitudinal model is mean time in a good state, which is like mean restricted survival time. It would be good to know the RMST estimates from the study.

@selcukorkmaz This is related to

@ajordannafa (1) Fortran 2018 which is lightning fast and easier to code in than C or C++; (2) Stan, which will open you up to amazing statistical modeling capabilities.

#Statistics thought of the day: Choosing & constructing a high-resolution outcome variable that captures what's important to patients is all-important in therapeutic studies. Avoid time to first event and WIN ratio/odds. Embrace longitudinal Y.

@mendel_random And any time a reporter ignores what DS says they are really blowing it. @d_spiegel

@TopfHNS @TassoneHNS @sindhurasridhar Only for assessing some mechanism, never for judging treatment effectiveness. Even more secondary non-intent-to-treat mechanism analysis I'd probably use a multistate model, allowing effect of baseline variables to differ by state being predicted.