Frank Harrell

Biostatistician/Professor/Consultant/Founding Chair of Biostat, Vanderbilt U. Blog: Statistical Thinking: @f2harrell on

[Frank Harrell @f2harrell]

@EvanThomas84 Yes, and a good way to destroy equipoise (and being wrong about it).

[Frank Harrell @f2harrell]

@Metaanalysis360 Yes. Keep in mind that when there are random effects, Bayesian models typically work better than frequentist ones such as what orm does. See the rmsb package blrm function or more generally the brms package.

[Frank Harrell @f2harrell]

Version 8.2 of the #rstats rms package is now on CRAN. This is a milestone release due to major improvements and extensions of the general-purpose semiparametric modeling function orm, including improved computational accuracy and two types of random effects. Details at https://t.co/XC99HU9TBP #Statistics

[Frank Harrell @f2harrell]

@eric_weinhandl @5_utr @RWJE_BA In line with that, behavioral economists assumed for years that birth month is a valid instrumental variable until it was pointed out that those born in months that make them almost 7 years old when starting elementary school had a developmental advantage lasting > 10y.

[Frank Harrell @f2harrell]

@RSKemp01 @Dr_R_Kurzrock Yes, although RCTs are barely sized to be able to detect even moderate overall average effects. But in observational studies we usually cannot even estimate an average effect (main effect) and an effect in a subgroup (main effect + interaction effect) is hopelessly biased.

[Frank Harrell @f2harrell]

@5_utr @Dr_R_Kurzrock I cringe every time I see a researcher search for something that doesn't exist (since the relationship is smooth). And if you really want a threshold it has to be different for every patient.

[Frank Harrell @f2harrell]

@5_utr @firexgamerx Check out https://t.co/qT3zgSJkSU e.g. a scatterplot of pre vs post-test probability.

[Frank Harrell @f2harrell]

@arajand @DavidKlaveren @Stats_research @BenVanCalster Great, and keep in mind that we often want to adjust for clinical variables when doing that, calling for modeling the biomarker with a restricted cubic spline in a multivariable model such as a cumulative probability semiparametric ordinal regression model https://t.co/j951pZ9LY7

[Frank Harrell @f2harrell]

Spearheaded by @DavidKlaveren & with co-authors @stats_research, Patrick Serruys, @benvancalster our new article in Stat in Med provides a minimal-assumption way to estimate the relationship between survival probability and a continuous variable: https://t.co/J0CYCd2QLz #Statistics

[Frank Harrell @f2harrell]

@arajand @DavidKlaveren @Stats_research @BenVanCalster Perhaps; this is for the case where you have a specific time horizon in mind. For biomarker research in general we relate a marker value to continuous time to event (better: multistate transition probabilities or longitudinal data analysis).

[Frank Harrell @f2harrell]

@Richard_D_Riley Thanks Richard. No I haven't done any research on pseudo observations other than in a simple jackknife setting. What's your prediction on how they would compare? Do you have a compact R function that estimates S(t|x) for specified constant t that I could add to my comparison?

[Frank Harrell @f2harrell]

I've updated simulations of performance of various methods for estimating S(t|x) vs. continuous x with right censoring. The new loess method is a winner. https://t.co/NkKlnpJKyc #Statistics

[Frank Harrell @f2harrell]

Spearheaded by @DavidKlaveren & with co-authors @stats_research, Patrick Serruys, @benvancalster our new article in Stat in Med provides a minimal-assumption way to estimate the relationship between survival probability and a continuous variable: https://t.co/J0CYCd2QLz #Statistics

[Frank Harrell @f2harrell]

@FinkelUriah @DavidKlaveren @Stats_research @BenVanCalster It's in my Zotero bibliographic database but I have not read it recently enough. Excellent paper; thanks for pointing this out.

[Frank Harrell @f2harrell]

@FinkelUriah @DavidKlaveren @Stats_research @BenVanCalster I haven't looked enough into the multistate setting. But no variation of overlapping window smoothed KM estimates was really competitive.

[Frank Harrell @f2harrell]

@RafaelOliveLeit @5_utr @RWJE_BA We've got to find this Harrel and stop him or her. Glad I'm Harrell.

[Frank Harrell @f2harrell]

@5_utr @RWJE_BA Interesting; relates to relative insensitivity of P(treatment effect > 0) with changing priors but sensitivity for P(treatment effect > trivial).

[Frank Harrell @f2harrell]

@VickersBiostats @DavidKlaveren @Stats_research @BenVanCalster loess is something I often teach before introducing linear regression. It's a go-to method. If Y is binary you need to use lowess(x,Y,iter=0).

[Frank Harrell @f2harrell]

@See_Elegance @5_utr There are many occasions where we want to estimate S(t|x) as a function of x without assuming a model. This is especially true when x is S-hat(t | X) (estimated survival probabilities as a function of lots of covariates X), i.e., estimating a calibration curve.

[Frank Harrell @f2harrell]

@5_utr @RWJE_BA I haven't thought of that and haven't run into a situation where I think that is likely to happen. I tend to just think of an effect β that an RCT estimates and an effect β + γ that observational data estimates and put a prior on γ,

[Frank Harrell @f2harrell]

@5_utr @RWJE_BA I'll bet that if a key part of a DAG has a probability of 0.9 of being true, the final conclusion of the analysis is at risk, especially if the evidence for an association is not exceedingly strong.