Dienstag, 24.10.2023, 12-13 Uhr in W9-109

Jan-Ole Koslik
Universität Bielefeld

Mitigating consequences of the Markov property

Statistical methods are essential for understanding complex real-world phenomena. So-called hidden Markov models HMMs) are a powerful instrument for modelling time series data with underlying sequential dependencies. While HMMs have gained popularity in various fields, they have also faced criticism for their reliance on the Markov assumption, suggesting that the present can entirely describe future events without additional consideration of the past. Traditionally, HMMs have assumed homogeneity in the underlying process, where the most probable time spent in a hidden state is fixed at one. However, recent years have seen a growing interest in inhomogeneous models, allowing for time-varying state-transition dynamics, such as seasonality. In this talk we investigate whether the common criticism of HMMs as overly simplistic in capturing real-world processes remains valid for more complex inhomogeneous models by deriving important properties of inhomogeneous Markov chains. Furthermore, we explore hidden semi-Markov models (HSMMs) as a flexible and promising extension of HMMs, enabling explicit modelling of state dwell times and, consequently, relaxing the Markov assumption. Our objective is to enhance the model formulation by integrating covariates into the state dwell-time distributions, thereby augmenting the capabilities of HSMMs in modelling real processes. Two case studies reveal that both approaches hold significant potential in mitigating the unrealistic consequences of the Markov assumption, each offering unique advantages and limitations. 

Dienstag, 07.11.2023, 12-13 Uhr per zoom

Dr. Joachim Schnurbus
Lehreinheit für Computergestützte Statistik und Mathematik der Universität Passau

Heeding the Call of Science: What Leads PhD Graduates to Pursue an Academic Career?

Building on self-determination theory and intelligent career theory, we investigate what leads PhD graduates to pursue an academic career. Using panel data from a cohort of PhD students who completed PhDs in business, economics, law, or social sciences in the 2014 academic year in Germany, we find that, in general, intrinsically motivated PhD students are more likely to choose an academic career. However, they need to be embedded in a specific social context, i.e., a specific set of social relationships, for the intrinsic motivation to have its positive effect. In particular, our results show that the positive relationship between the PhD students’ intrinsic motivation and their probability of pursuing an academic career is strengthened when they collaborate more with others. Further, we find that social relationships such as mentoring aimed at developing the PhD students’ career identity are favorable, whereas mentoring targeted at increasing their knowledge and skills or exposure to networks becomes irrelevant, or even dysfunctional, when their collaborative working environment is considered. By disentangling the interplay of intrapersonal factors and social context in the academic career choice, our findings contribute to a more comprehensive, multilevel view of the factors that lead to the choice of such a career.

 

Dienstag, 21.11.2023, 12-13 Uhr in W9-109

Prof. Dr. Matej Demetrescu
Ökonometrie und Statistik, TU Dortmund

Tests of No Cross-Sectional Error Dependence in Panel Quantile Regressions

This paper argues that cross-sectional dependence (CSD) is an indicator of misspecification in panel quantile regression (QR) rather than just a nuisance that may be accounted for with panel-robust standard errors. This motivates the development of a novel test for panel QR misspecification based on detecting CSD. The test possesses a standard normal limiting distribution under joint N,T asymptotics with restrictions on the relative rate at which N and T go to infinity. A finite-sample correction improves the applicability of the test for panels with larger N. An empirical application to housing markets illustrates the use of the proposed cross-sectional dependence test.


Dienstag, 05.12.2023, 12-13 Uhr in W9-109

Prof. Dr. Dominik Liebl
Institute of Finance and Statistics, University of Bonn

Fast and Fair Simultaneous Confidence Bands for Functional Parameters

Quantifying uncertainty using confidence regions is a central goal of statistical inference. Despite this, methodologies for confidence bands in Functional Data Analysis are still underdeveloped compared to estimation and hypothesis testing. In this work, we present a new methodology for constructing simultaneous confidence bands for functional parameter estimates. Our bands possess a number of positive qualities: (1) they are not based on resampling and thus are fast to compute, (2) they are constructed under the fairness constraint of balanced false positive rates across partitions of the bands' domain which facilitates the typical global, but also novel local interpretations, and (3) they do not require an estimate of the full covariance function and thus can be used in the case of fragmentary functional data. Simulations show the excellent finite-sample behavior of our bands in comparison to existing alternatives. The practical use of our bands is demonstrated in two case studies on sports biomechanics and fragmentary growth curves.

Dienstag, 19.12.2023, 12-13 Uhr in W9-109 - ausgefallen

Julia Dyck
Universität Bielefeld

Signal detection of adverse drug reactions: The Bayesian Power generalized Weibull Shape Parameter test

After release of a drug on the market, pharmacovigilance monitors the occurrence and changes in known adverse drug reactions (ADRs) as well as detects new ADRs in the population. This is done to keep a drug‘s harm profile updated and can potentially result in adjustments of the prescription labeling or even a recall of the product from the market. In recent years the interest in the use of longitudinal electronic health records for pharmacovigilance increased. Sauzet and Cornelius (2022) provided a test based on the power generalized Weibull (PgW) distribution shape parameters (PgWSP). If both shape parameters of the PgW distribution are equal to one, the distribution reduces to an exponential distribution with constant hazard over time. This is interpreted as no temporal association between a drug and an adverse event. Signal detection can be improved by incorporating existing knowledge about the ADR profile of drugs from the same family or based on expert knowledge about the drug mechanism. Therefore, we propose the development of a Bayesian PgWSP test. The test compares a region of practical equivalence (ROPE) around one reflecting the null hypothesis with the estimated credibility intervals (Krushke, 2015). Depending on the single parameters’ outcome and the chosen combination rule, the tool raises a signal. We performed a simulation study to tune the optimal ROPE and credibility interval for signal detection using a Bayesian PgWSP approach. Samples are generated under varying conditions regarding sample size, prevalence, and proportion of adverse events. Prior assumptions considered are no ADR, ADR at the beginning, middle, or end of the observation period. A range of ROPE and credibility interval types as well as combination rules are considered. The optimal ROPE and credibility interval tuning parameters are determined based on the area under the curve. 

Dienstag, 16.01.2024, 12-13 Uhr in W9-109

Dr. Denny Kerkhoff
Psychologische Methodenlehre und Evaluation, Abteilung Psychologie an der ​Universität Bielefeld

Extending a-priori simulation practice to assess and maximize replicability in psychological research

With the awareness that attempts to replicate psychological research findings are prone to fail, research on improving replicability has stressed the necessity for methodologically sound primary studies with sufficient power, unbiased estimates and correctly specified models. A prominently suitable method to assess these criteria for a planned study is the Monte Carlo simulation. Adjacently, multi-model approaches such as the multiverse analysis provide an efficient conceptualization to identify practice-relevant simulation conditions as well as methods to visualize and report complex results. Fostered by the importance of producing replicable results, a-priori simulation analyses to assess estimation quality and required sample sizes for a planned research project have become common in psychological research. However, in practice, such analyses often boil down to a-priori power analyses to determine the minimum sample size to reliably detect an effect in the target statistical model, such that extensive simulation capabilities and multiverse analysis considerations remain underutilized. I introduce a planned research project with the goal to integrate simulation and multiverse capabilities into a-priori simulation practice. Four challenges to achieve this goal are presented: First, current a-priori simulation practices must be systematically described and evaluated. Second, the range of simulation inputs and outputs must be succinctly defined and interpreted in terms of replicability. Third, multiverse analysis must be adapted for a-priori usage. Fourth, a workflow must be developed to efficiently identify relevant simulation inputs, identify patterns in results, and interpret, visualize, and report results in terms of replicability. For each challenge, I present the current state of research and methods to address each challenge. 

Dienstag, 30.01.2024, 12-13 Uhr in W9-109

Alexandra Strobel
Institut für Medizinische Epidemiologie, Biometrie und Informatik der Martin-Luther-Universität Halle-Wittenberg

Uncovering hidden influences: Impact of omitted covariates on treatment estimates in randomized and propensity score matched studies

Hazard Ratios (HRs) are the most common treatment effect measure in clinical studies focusing on time-to-event outcome. However, recent years have witnessed an increasing critique of HRs, particularly regarding their non-collapsibility or with respect to their (causal) interpretation. This work addresses another critical aspect related to unobserved or omitted covariates that may impact the survival outcome and/or the treatment allocation. Misspecification of the Cox model due to unobserved or omitted covariates could result in heavily biased treatment estimates, affecting both randomized and propensity score matched trials. Researchers frequently lack clarity on whether (and, if so, which) covariates might induce this bias. Therefore, we present a methodological approach called Dynamic Landmarking, providing a visual indication of biased treatment effect estimates and identifying omitted covariates that could induce this bias. By simulations we show that Dynamic Landmarking indeed serves as an effective visual tool for detecting biased treatment estimates in both study designs. Furthermore, we use the method to assess the built-in selection bias in individual patient data from 27 large randomized controlled trials (RCTs). We found no empirical evidence of this bias in these trials and conclude that this kind of bias is of limited practical relevance in most cases. This is mainly due to small treatment effects and homogeneous patient populations resulting from inclusion and exclusion criteria. In summary, Dynamic Landmarking can be used to verify if estimated treatment effects from a Cox model are biased and whether omitted covariates induce this bias. The empirical investigation of RCTs suggests that HRs are not materially affected by the built-in selection bias and can therefore be safely used, at least concerning this aspect.