Event Description

https://www.dsts.dk/events/todagesm%C3%B8de-efter%C3%A5r-2022-symposium-celebration-kragh-andersens-contributions-statistics

Haderup Auditorium, Panum Institute, Blegdamsvej 3, 2200 Copenhagen N

Tuesday November 15:

14.00-14.05: Welcome

14.05-14.50: Ørnulf Borgan (Oslo)

Supersampled nested case-control and case-cohort data

Nested case-control and case-cohort studies are useful for studying associations between covariates and time-to-event when one covariate is expensive to measure. Information on the expensive covariate is collected for the nested case-control and case-cohort samples only, while cheaply measured covariates are often observed for the full cohort. Standard analysis of such case-control samples ignores any full cohort data. Previous work has shown how all data can be used by (i) maximum likelihood estimation using all available data for the full cohort, or (ii) multiple imputation of the expensive covariate for all individuals who are not in the case-control sample. For large cohorts these approaches are computationally expensive or even infeasible. An alternative is to base the analysis on a supersample, where we add more controls to a nested case-control study and enlarge the subcohort of a case-cohort study. The cheaply measured covariates are observed for the added controls/enlarged subcohort, but not the expensive one. In the talk I will explore methods of analysis for such supersampled nested case-control and case-cohort data.

14.50-15.35: Henrik Ravn (Novo)

The 3½ Habits of Highly Effective Statisticians

15.35-16.00: Coffee break

16.00-16.45: Torben Martinussen (KU)

Predictive accuracy of covariates for survival times.

We study a graphical measure, the generalized positive predictive value function, to quantify the predictive accuracy of covariates for survival times. This measure characterizes the event probabilities over time conditional on a thresholded linear combination of covariates and has direct clinical utility. We construct a fully nonparametric estimator based on the corresponding efficient influence function for this function under right censoring. The efficient influence function depends on complex density functions but estimation of these can be avoided in the estimation process of the target parameter leading to a surprisingly simple estimator. To bypass the estimation of the complex density functions involved in the asymptotic variances, we use  the bootstrap approach and establish its validity.

16.45-17.25: Hein Putter (Leiden)

Smooth estimation of transition hazards and probabilities in general interval-censored Markov multi-state models

Estimation of transition hazards and probabilities in interval-censored Markov multi-state models is a notoriously difficult problem. Researchers have therefore simplified the problem to either restricting to parametric models (most notably piecewise constant hazards as implementend in the msm package in R) or to the illness-death model (the SmoothHazard package in R). In this work we develop an approach to smooth estimation of transition hazards based on P-splines. We divide time in a very fine grid of equally spaced intervals and use an iterative Expectation-Maximization (EM) algorithm. The E-step consists of calculating, for each time bin, conditional expectations of at risk status (exposure) and observed transition status (occurrence) for each subject and transition, based on the subject’s observed data and current transition rates. These are calculated based on solvinfg the Chapman-Kolomogorov differential equations. The M-step is based on Poisson generalized linear modeling. The approach is applicable for arbitrary (Markov) multi-state model structures, and allows for inclusion of covariates in a proportional hazards model.

17.25: Beers and soft drinks.

19.00: Dinner at Madklubben, Østerbro

Wednesday November 16:

9.00-9.45: Thomas Gerds (KU)

The etiology of a misspecified medical risk prediction model

All statistical models are wrong and medical risk prediction models are no exception. The sources of why a specific medical risk prediction model is bound to be wrong(ly interpreted) are the product of inappropriate data, inappropriate statistical tools, and missing thoughts about the actual application of the model.

9.45-10.30: Erik Parner (AU)

Regression analysis for censored event data using pseudo-observations

The pseudo-observation method has become a popular method for performing regression analysis for censored event data. Pseudo-observations are transformations of the event data; once they are computed, they can be treated as observations for regression analysis and often standard statistical software can be used for the analysis. Applications include regression analysis for cumulative risk, restricted means and number of life years lost due to specific causes of death. I discuss under which conditions we may expect the pseudo-observation method to provide unbiased estimates. In particular, I consider regression models for cumulative risk in a cohort with left-truncation. Some variants of pseudo-observations are also discussed.

10.30-10.50: Coffee break

10.50-11.35: Maja Pohar Perme (Ljubljana)

Simplifying survival analysis using pseudo-observations

Pseudo-observations present a general tool that can simplify a wide range of analyses in the survival field. The basic idea is to first handle the issue of censoring by redefining an outcome that is available for each individual at each follow-up time. Having defined pseudo-observations, the basic need for special survival analysis methods is removed and one can use standard approaches available outside the survival field. In this talk, we shall approach the topic from the applied point of view and consider three different problems of survival analysis where pseudo-observations provide simple and flexible solutions: graphical methods for assessing goodness-of-fit for regression models, estimation in relative survival and regression with transition probability in non-Markov  multi-state models as the outcome of interest. We will use these examples to discuss the properties of pseudo-observations, their advantages and their issues. While it shall often turn out that compared to model specific solutions, pseudo-observation approach may be less efficient, their importance becomes clear when considering models for which no method has yet been developed.

11.35-12.20: Richard Gill (Leiden)

The Dutch new herring scandals

(Repeated measurements with unintended feedback)

We analyse data from the final two years of a long-running and influential annual Dutch survey of the quality of Dutch New Herring served in large samples of consumer outlets. The data was compiled and analysed by a university econometrician whose findings were publicized in national and international media. This led to the cessation of the survey amid allegations of bias due to a conflict of interest on the part of the leader of the herring tasting team. The survey organizers responded with accusations of failure of scientific integrity. The econometrician was acquitted of wrong-doing by the Dutch authority, whose inquiry nonetheless concluded that further research was needed. We reconstitute the data and uncover its important features which throw new light on the econometrician's findings, focussing on the issue of correlation versus causality: the sample is definitely not a random sample. Taking account both of newly discovered data features and of the sampling mechanism, we conclude that there is no evidence of biased evaluation, despite the econometrician's renewed insistence on his claim.

12.20: Sandwich to go