21,429 reputation
13080
bio website biostat.mc.vanderbilt.edu/…
location Nashville, TN
age
visits member for 3 years, 6 months
seen 40 mins ago

I am Professor of Biostatistics and Chairman of the Department of Biostatistics at Vanderbilt University School of Medicine, Nashville TN USA. I am Associate Editor of Statistics in Medicine, a member of the Faculty of 1000 Medicine, and a member of the policy advisory board for the Journal of Clinical Epidemiology. I am a Fellow of the American Statistical Association. I am author of Regression Modeling Strategies (Springer, 2001). My specialties are development and validation of predictive models, clinical trials, observational clinical research, cardiovascular research, technology evaluation, clinical epidemiology, medical diagnostic accuracy, biomarker research, pharmaceutical safety, Bayesian methods, quantifying predictive accuracy, missing data imputation, and statistical graphics and reporting. I am a long-time user of R. In August 2014 I was given the WJ Dixon Award for Excellence in Statistical Consulting by the American Statistical Association. Among many other things, Dr Dixon was the lead developer of the first general-purpose statistical software package, BMD.


1h
revised Logistic Regression using fractional polynomials in R
fixed incorrect work in title (factorial -> fractional)
1h
answered Logistic Regression using fractional polynomials in R
5h
comment Propensity score to match on exposure thats not a treatment
Yes. If you don't have a problem with non-overlap, a regression spline in the logit of PS is a very flexible bias-removing high-precision approach. We occasionally assume linearity in logit(PS) but the regression spline does away with the need for that assumption while solving the problem of residual confounding.
5h
comment How can I get more precise regression tree?
If the precision of a prediction is not adequate, the precision of a classification is even more problematic. Classification is a minimum information/maximum variance approach to the problem. Think of binary classification as providing one bit of information, which is the minimum amount of information you can have other than no information at all. You never help with a problem by discarding information; you only make the problem seem to be simpler.
6h
comment How can I get more precise regression tree?
Random forests will almost always beat single trees, but you lose all interpretability advantages of a tree. Flexible regression (possible with penalization) is likely to be very competitive with RF and more interpretable depending on dimensionality. But it is not the case that smooth methods perform worse with noisy data.
6h
comment Propensity score to match on exposure thats not a treatment
Subclasses will result in residual confounding due to binning continuous variables. Weighting on PS will increase variance. Covariate adjustment, after checking for non-overlap PS regions, is likely to work well. Using the terms "case" and "control" confuses the issues to me and implies retrospective sampling. Try to state the hypothesis as a model, e.g. identify a parameter with family history and ask whether median total cholesterol or mean log chol. is associated with fam hx adjusting for all covariates measured concurrently with cholesterol (a slightly strange H0).
6h
comment Propensity score to match on exposure thats not a treatment
You are implying in your last sentence that you will be deleting observations. This is not necessarily a scientific approach to the problem. But more to the point, what is your real hypothesis/goal? You already know from the literature how many of your variables relate to myocardial infarction. History of MI needs to be further defined to take age/time into account.
6h
comment How can I get more precise regression tree?
Please try to understand my answers, and respond with single comment entries. For a method to be competitive, it must perform well in simple situations where we know the truth, such as linear models. Trees don't perform well in that situation. Then consider smooth relaxation of the linearity assumption using for example regression splines. Regression methods can flexibly fit nonlinear relationships with high precision (low variance); single trees cannot. And the answer from @user3969377 has to do with bias but not with variance; minimize bias by using a tiny bucket size but ruin precision
7h
comment When normalization is counter-productive
The context is too broad, and you did not state why you would want to optimize an improper accuracy scoring rule such as classification accuracy. You also need to state what problem normalization would attempt to solve. In general normalization gets in the way of understanding more than it helps, but occasionally normalization reduces the dimensionality of the model in a helpful way.
7h
comment Propensity score to match on exposure thats not a treatment
PS (best done using covariate adjustment on regression splines of logit PS) is done to correct for measured potential confounders when assessing the association between an exposure (often a treatment) and $Y$. It is only logical to develop a PS model to predict the exposure (treatment) of interest. It would help if you were to state the goal of the modeling exercise, along with the sample size and distribution of $Y$ and total number of potential confounders. Depending on the effective sample size and dimensionality of the problem you may not need PS at all.
7h
answered How can I get more precise regression tree?
23h
comment Marginal effects of a logistic model, and their standard errors
Thanks that are functions of covariates are conditional, not marginal, if I understand you. I stand by my statement that marginal effects are in general not transportable outside the current sample.
1d
comment Any other non-parametric alternative to Kruskal-Wallis?
No this has been warned against at other places on this site. Pairwise Wilcoxon tests (using the mean rank for group A in the context of groups A and not A) is not consistent with a K-W test that distinguishes B and C among the not-A observations.
1d
comment Stratified sampling for creating test/training sets when there are continous and categorical variables to consider?
Imbalance is the least of your worries. The sample size is very inadequate to split the data, and perhaps even for fitting a model on the whole dataset.
1d
revised Stratified sampling for creating test/training sets when there are continous and categorical variables to consider?
expanded related to using whole sample for all purposes
1d
answered Any other non-parametric alternative to Kruskal-Wallis?
1d
answered Stratified sampling for creating test/training sets when there are continous and categorical variables to consider?
2d
comment Interactions in Propensity Score Models
I don't recommend the use of matching when it discards any observations. Also please state why ordinary covariate adjustment doesn't work in your case. Propensity score analysis (preferably through covariate adjustment by a nonlinear function of the logit propensity after checking non-overlap regions) is usually reserved for the case when the number of potential confounders is too great for their coefficients against $Y$ to be reliably estimated.
2d
comment Ranking two models based on ROC-AUC and PR-AUC
It's best if the two model outputs can be transformed to the same scale; then there are more options. Otherwise consider the "is one model concordant for a pair of observations when the other model isn't" approach of the R Hmisc package rcorrp.cens function.
2d
comment Ranking two models based on ROC-AUC and PR-AUC
Note that ROC area is too insensitive a measure to be used to compare two models. It is useful for summarizing predictive discrimination for a single model, in my opinion.