Conference Agenda

Over the past three decades, scholars working in the fields of causal inference and missing data, have developed a variety of techniques to produce estimators with highly desirable robustness and efficiency properties. We apply these techniques, to derive various AIPW (augmented inverse probability weighted) estimators of occupation probabilities of a latent multi-state process under two levels of coarsening; right censoring and baseline exposure, allowing for both time constant and time dependent confounders. The key exchangeability assumption for identification is coarsening at random (CAR). The AIPW estimators are motivated from a different and arguably simpler identification result, than the common product integral representation of occupation probabilities. We investigate the performance of the estimators of occupation probabilities in a simulation experiment under different scenarios.

Predictive algorithms are widely used in hospitals to aid patient management, but when treatment policies shift, they can quickly lose accuracy. While predictive models can be dynamically updated, these updates take time due to the need to gather sufficient new data. For instance, a clinical model’s predictive accuracy may decline when a new treatment is introduced, requiring extensive sample sizes to re-estimate the effect of the predictors.

In this work, we delineate an alternative approach where external information on the effectiveness of the new treatment is used to estimate predictions under the upcoming intervention strategy. Predictions under interventions, which address “what if” scenarios by estimating outcomes under hypothetical interventions, have been used primarily for personalized decision-making. Here, we combine them with model updating strategies to create “interventional updates,” aiming to reduce lag in dynamic model updates when new treatments are introduced in clinical practice.

For each treatment that changes in standard care—whether newly introduced, discontinued, or modified in use—there are two key time points: when the changes in care begin (e.g., when a new treatment starts being used) and when these changes stabilize (new treatment has become the new normal). This defines three distinct phases: the old standard care, a transitional period, and the new standard care. In the old and new standard care phases, standard updating techniques can be used. In the transitional period we propose to use interventional updating which incorporate evidence of treatment effect from external sources. We carefully write down how to extend known methods for dynamic updates for time-to-event prediction models, extending the most commonly known updating strategies: refitting, intercept recalibration and Bayesian updating methods.

We illustrate our methods using electronic health records from 4,064 patients hospitalized with COVID-19 in four Dutch hospitals in 2020 and 2021. A prediction algorithm was trained on first-wave data (February–July 2020) to estimate the 28-day risk of mortality or ICU admission from the time of hospital admission. We then compared the performance of “standard” and “interventional” model updates on second-wave data (August 2020–May 2021), assessing discrimination (c-index and AUCt), calibration (calibration intercept, slope and OE ratio), and overall performance (Brier score). For the “interventional” updates, we focused on the change in dexamethasone use, as this was identified as the most significant treatment shift in 2020. We incorporated evidence on dexamethasone’s effectiveness from the RECOVERY trial.

At Australian universities, PhD candidates are expected to submit their thesis within three to four years of starting their project. Extensions are possible, and were widely granted for students whose projects were affected by COVID. However, funding is tight, and University administrators are keen to know how long candidates are taking to submit their theses, and whether targeted support mechanisms such as statistics advice, intensive writing retreats, and the like make a measurable impact on time to completion.

This presentation will focus on a comprehensive analysis undertaken at the Australian National University to address these pressing questions. We will begin by outlining the structure of the PhD program and the nature of the data available to answer the research questions. This will include a detailed description of the various support mechanisms in place and the metrics used to measure their effectiveness.

While the survival models used to analyse the data will be off-the-shelf, the presentation will highlight the complexities encountered in understanding and interpreting the data. Two proposed solutions will be discussed in depth, revealing the challenges that analysts face when dealing with secondary analysis of administrative data in an educational context. These solutions will be evaluated in terms of their generalisability and scalability, as well as potential limitations.

The presentation will conclude by drawing connections between the research question, proposed solutions, and alternative analyses from within a causal framework. This will involve a critical examination of the assumptions underlying our approach and a discussion of how causal inference methods could potentially enhance our understanding of the factors influencing PhD completion times.

By sharing these insights, I aim to contribute not only to the technical discussion of survival analysis methodology but also to the broader conversation on improving PhD completion rates and the effective allocation of resources in higher education.

It should be well known that in a competing risk setting the use of the Kaplan Meier estimator counting only one type of event is biased for estimation of the cumulative incidence probability (Schuhmacher et al., Schmeller et al.). Several analytical proofs of this exist which confirm examples from real data. Therefore, it is clear that the false-Kaplan-Meier is not an estimator for the cumulative incidence but it is questioned if this estimator still has a meaningful interpretation. The aim of this talk is to show a proof that the Kaplan-Meier estimator has a causal interpretation for the survival probability in a time to combined endpoint analysis that would have been observed if the random censoring had been avoided. However, the talk will show that this cannot be adapted to a competing risk setting and the Kaplan-Meier estimator counting only one type of event and censoring the other cannot be interpreted as the estimator for the cumulative incidence of the event of interest if the competing events had been avoided (the intervention distribution with intervention "no competing event"). The reason is that the occurrence of the competing events are not independent. The simple proof is based on a simple causal graph and a straightforward application of the g-computation rule. It is conceptually related to but technically simpler than the recent argument by Young et al. however, without requiring that one cause of death precedes the other as in Young et al. .

Schuhmacher et al. : Schumacher M, Ohneberg K, Beyersmann J (2016) Competing risk bias was common in a prominent medical journal. J Clin Epidemiol 80:135–136

Schmeller et al. : Schmeller S, Fürst D, Beyersmann J (2023). Konkurrierende Risiken Modelle. In Jan Gertheis, Matthias Schmid, and Martin Spindler, editors, Moderne Verfahren der Angewandten Statistik. Springer Spektrum, Berlin, Heidelberg, 2023.

Young et al.: Young JG, Stensrud MJ, Tchetgen Tchetgen EJ, Hernán MA (2020) A causal framework for classical statistical estimands in failure-time settings with competing events. Stat Med 39(8):1199–1236