In medical research, common prediction models still predominantly make use of composite endpoints such as progression- or event-free survival. However, these time-to-first-event outcomes do not incorporate important aspects of the individual disease pathway and therapy sequences. In the era of precision medicine with increasing molecular information, the use of a multi-state model is essential to more accurately capture pathogenic disease processes and underlying etiologies.

Especially the availability of big data with a large number of covariates presents several statistical challenges for model building. Effective data-driven model selection strategies for multi-state models are essential to determine an optimal, ideally parsimonious model based on high-dimensional data. Established methods incorporate regularization in the fitting process in order to perform variable selection. A useful technique to reduce model complexity is to combine homogeneous covariate effects for distinct transitions based on a reparametrized model formulation. We integrate this approach to data-driven variable selection by extended regularization methods for model selection within multi-state model building. We propose the sparse group fused lasso penalized Cox-type regression in the framework of multi-state models combining the penalization concepts of pairwise differences of covariate effects along with transition grouping.

This raises the following challenges: First, multiple heterogeneous transitions have to be considered for consecutive treatment phases within the multi-state model. Furthermore, the number of transitions with fewer observations increases during the course of the sequential event history. Finally, model selection procedures have to be efficiently implemented in large-scale multi-state settings.

Thus, model selection strategies for multi-state endpoints are substantial for a more precise understanding and interpretation of individual disease pathways, specific oncological entities along with their precision therapies as well as improved personalized prognoses.

Aging processes in, e.g., the skin are driven by genetic factors (intrinsic aging) as well as environmental factors (extrinsic aging). To assess if intrinsic and extrinsic skin aging develop independently of each other or if these two types of aging influence each other in their development, the secretome or proteome of, on the one hand, intrinsically, and on the other hand, also extrinsically aged skin fibroblasts of the same individuals can be considered and compared.

A problem of the analysis of the extrinsic aging, however, is that the entire human skin ages intrinsically so that it is only possible to measure the combination of intrinsic and extrinsic skin aging in skin areas that are exposed to extrinsic influencing factors and thus to extrinsic aging processes. To study the effects of intrinsic and extrinsic aging in these combinations, it is, therefore, necessary to determine the pure extrinsic signal from the combined intrinsic and extrinsic signal.

For this purpose, we propose a statistical deconvolution method that enables the extraction of signals from mixed data whose values are composed of multiple sources. This nonparametric deconvolution procedures makes essential use of the Fourier transform inversion theorem and piecewise polynomial splines that are fitted to the proteomic data, on the one hand, from the intrinsically aged fibroblasts, and on the other hand, from the combined intrinsically and extrinsically aged fibroblasts. Using numerical integration, the Fourier transforms of the predicted functions are computed, which can then be used to estimate the pure extrinsic signal by considering the quotient of the obtained Fourier transforms.

In a simulation study, we evaluate the performance of the proposed nonparametric deconvolution method. This application to simulated data shows that the proposed procedure is able to extract the signal of a non-measurable source from a mixture of signals, if the signal of the other components of the mixture can be measured. Moreover, we apply the proposed deconvolution method to secretomic and proteomic data from the Gerontosys study that were measured in intrinsically as well as intrinsically and extrinsically aged fibroblasts of individuals belonging to different age groups to extract the pure extrinsic signal from the combined signal and compare the intrinsic and the pure extrinsic signal.

We present a modification of the case-only (CO) approach to time-to-event data for testing multiplicative interactions between binary risk factors from prospective cohort and case-cohort studies. Motivated by a real data example of a cohort study on survival after cardiovascular surgery, we use the event time information to select only patients who died early (i.e. before a pre-specified time point) and modified the CO approach to time-to-event data. The relevant key assumptions of the CO design, rare events and independence between the factors in the general population, were fulfilled. In a simulation study, we investigated the CO approach in the cohort and case-cohort design with time-to-event outcome and compared results from both designs to the classical Cox proportional hazard and logistic regression (LR). For the LR approach, the same cases as in the CO approach were used and censored observations were considered as ‘controls’ in a restricted follow-up scheme in the cohort design and a random subsample in the case-cohort design. In our conducted scenarios with varied event rates and main effects, the applied CO approach was consistently valid in the cohort settings and had a similar power as the alternative analyses. In the case-cohort design, the CO approach was distinctly more powerful than standard LR or Cox regression but in the presence of main effects the estimators are biased and consequently the type I error rate slightly inflated. In summary, under a variety of specific circumstances, the CO approach is as powerful for time-to-event data as the standard models in the cohort framework and even more powerful in the case-cohort framework.

Validated prediction models to estimate the patients’ risk of developing a particular disease or a condition in the future can be a useful decision tool for individualized treatment decisions. Prediction models are usually developed and validated based on data from observational studies, where missing values resulting from different underlying missingness generating mechanisms are a typical problem. In a simulation study conducted by Marshall et al. [2010], the influence of different imputation methods in a Cox proportional hazards framework were investigated. The authors emphasized that further simulation studies in different clinical contexts and underlying distributions of the predictors are required to assess the generalizability of the results. In the clinical context of long COVID, we study the influence of commonly used imputation methods, including multivariate imputation by chained equations (mice), multiple imputation using additive regression, bootstrapping, and predictive mean matching (aregImpute), and non-parametric missing value imputation using random forest methodology (missForest), to handle missing predictor data on model performance. For the set-up of the simulation study, the recommendations of Morris et al. [2019] and Burton et al. [2006] are followed. The data underlying the simulation study is generated based on a Swiss multicenter prospective cohort study [Deforth et al., 2022] to reproduce a “real-world setting”. Several scenarios with different underlying data missingness generating mechanisms, percentage of missing values, stronger and weaker regression coefficients, and varying sample sizes are investigated in a traditional statistical framework. Excluding observations with missing data, complete case, is set as a benchmark. Model performance is assessed in external validation data (without missing values) based on model discrimination ability, calibration-in-the-large, calibration slope and scaled Brier score. The results of the simulation study will be presented in the framework of a neutral comparison study.

References

A Marshall, D G Altman, P Royston, and R L Holder. Comparison of techniques for handling missing covariate data within prognostic modelling studies: a simulation study. BMC Medical Research Methodology, 10(1):7, 2010. doi: 10.1186/1471-2288-10-7.

T P Morris, I R White, and M J Crowther. Using simulation studies to evaluate statistical methods. Statistics in Medicine, 38:2074–2102, 2019. doi: 10.1002/sim.8086.

A Burton, D G Altman, P Royston, and R L Holder. The design of simulation studies in medical statistics. Statistics in Medicine, 25:4279–4292, 2006. doi: 10.1002/sim.2673.

M Deforth, C E Gebhard, S Bengs, K P Buehler, R Schuepbach, A Zinkernagel, S D Brugger, C T Acevedo, D Patriki, B Wiggli, R Twerenbold, G Kuster, H Pargger, J Schefold, T Spinetti, P Wendel-Garcia, D Hofmaenner, B Gysi, M Siegemund, G Heinze, V Regitz-Zagrosek, C Gebhard, and U Held. Development and validation of a prognostic model for the early identification of COVID-19 patients at risk of developing common long COVID symptoms. Diagnostic and Prognostic Research, 6(1):22, 2022. doi: 10.1186/s41512-022-00135-9.

There is an increasing number of potential biomarkers that could allow for early assessment of treatment response or disease progression. However, measurements of quantitative biomarkers are subject to random variability. Hence, differences of a biomarker in longitudinal measurements do not necessarily represent real change but might be caused by this random measurement variability. Before utilizing a quantitative biomarker in longitudinal studies, it is therefore essential to assess the measurement repeatability. Measurement repeatability obtained from test-retest studies can be quantified by the repeatability coffiecient (RC), which is then used in the subsequent longitudinal study to determine if a measured difference represents real change or is within the range of expected random measurement variability. The quality of the point estimate of RC therefore directly governs the assessment quality of the longitudinal study.

RC estimation accuracy depends on the case number in the test-retest study, but despite its pivotal role, no comprehensive framework for sample size calculation of test-retest studies exists. To address this issue, we have established such a framework, which allows for flexible sample size calculation of test-retest studies, based upon newly introduced criteria concerning assessment quality in the longitudinal study. This also permits retrospective assessment of prior test-retest studies.

Ultraviolet (UV) B radiation (290-315 nm) and UVA (315 – 400 nm) are complete carcinogens and both are well known to contribute to the development of skin cancer in humans. Under physiological conditions, human skin is exposed to a mixture of UVB and UVA radiation from natural sunlight. Previous research, however, has primarily focused only on the effects of each type of radiation separately whereas the knowledge of the UVA-UVB-interaction under simultaneous exposure and its impact on human skin is quite unclear. In a previous photocarcinogenesis study we have found that simultaneous exposure of murine skin to UVB and non-carcinogenic doses of UVA increased UVB-induced photocarcinogenesis. These results indicate that detrimental effects caused by UVB and UVA radiation, if applied simultaneously, can enhance each other, even when the UVA dose per se does not cause significant skin damage. In the present study we would like to further analyze the interaction of UVB and UVA radiation, if applied simultaneously, by analyzing gene expression responses (by bulk RNAseq) in human skin cells. We are particularly interested in identifying the minimum UVA dose which is required to enhance UVB-induced skin damage.

The dose-response relationships involved will be examined more closely by using already existing RNAseq data sets of skin cells. Keratinocytes were irradiated either with only UVB or UVA, or with a combination (simultaneous irradiation) of both. The RNA of the cells was isolated at two different time points after irradiation (incubation time). Additional data was acquired for further incubation times and different physiologically relevant dosages and different ratios of UVA and UVB radiation. These settings were chosen to best support the statistical analysis.

When modelling univariate dose-response relationships separately for each incubation time, the information shared across the incubation times is not considered. However, it is critical to identify certain features of the relationship. We will consider data of all incubation times at once by modelling a two-dimensional surface with the help of tensor product B-splines. The identification of the minimum UVA dose, at which the combination irradiation shows a significantly different effect compared to UVB irradiation alone, will ultimately also be investigated and described with the help of spline regression models.

Clinical trials often use a time-to-event endpoints to measure clinical outcome, with event status potentially changing with follow-up time. Biomarkers that are potentially associated with the clinical outcome are also often assessed longitudinally over time. Naturally, the association between the biomarker and the time-to-event endpoint is of special interest. Further interest is to evaluate whether the biomarker can serve as a surrogate endpoint.

In this context, the values of the biomarker observed at a specific timepoint are frequently used as a fixed covariate in a Cox proportional hazards model (Model 1) to analyse its impact on the time-to-event endpoint. Such a model ignores a large amount of biomarker information collected in the longitudinal process. A potentially more appropriate model is to incorporate the biomarker information as a time-varying covariate in the Cox model (Model 2) that implicitly assumes a step function for the biomarker trajectory due to lack of continuously monitored biomarker data. The biomarker value at a scheduled assessment timepoint is assumed constant up to the next assessment timepoint, i.e. the last-observation-carried-forward (LOCF) approach is applied. The resulting step function might not provide a good approximation of the true biomarker trajectory and could result in biased estimate towards zero (Arisido, 2019).

Joint modelling of the longitudinal biomarker data and time-to-event data with shared random effects could be a preferred approach (Model 3), which also allows for the inference on the association between the biomarkers longitudinal process and the hazard of the event. Previous studies showed that such models provide an unbiased estimate, showing improvement over the time-varying Cox model (Arisido, 2019). Moreover, such models can also be used to assess criteria laid out for a surrogate endpoint (Prentice 1989) and compute the proportion of treatment effect explained by a biomarker (Freedman 1992).

As an illustrative example, we applied different models to the data of a randomized double-blind, placebo-controlled phase 3 study. The biomarker of interest was measured at week 4, 8, 16, 26 and 52. The clinical outcome of interest was time to disease progression. Results were compared between models with respect to the association between the biomarker and the hazard of disease progression as well as the potential surrogacy of the biomarker for disease progression.

Crizanlizumab is a humanized monoclonal antibody against P-selectin for the prevention of vaso-occlusive crises in sickle cell disease (SCD). P-selectin (Psel)-mediated multi-cellular adhesion is a key factor in the pathogenesis of vaso-occlusion and vaso-occlusive crisis (VOC) as Psel, that is expressed on the surface of the endothelium, is thought to mediate abnormal rolling and static adhesion of sickle erythrocytes to the vessel surface.

Shedding generates a soluble form of P-selectin that is absent of a transmembrane domain.
We evaluated the relationship between crizanlizumab exposure from 5 mg/kg and 7.5 mg/kg dose and free, unbound soluble P-selectin levels, an exploratory biomarker, in sickle cell disease patients. A sigmoid Imax model with inter-individual variability on baseline free sPsel concentration and IC50 parameter together with a proportional error term was found to fit data best. Covariate search did not reveal any relevant covariate for baseline free sPsel and IC50. As Psel plays a key role in the pathogenesis of vaso-occlusion and vaso-occlusive crisis it is believed that the level of
free P-selectin concentrations may influence the frequency of VOCs in SCD patients.

There is an ongoing debate about deficiencies in the reporting of findings and reproducibility across different areas of medical research [1, 2]. One reason for non-reproducibility is the lack of statistical code to accompany the published research [3, 4]. At the Department of Biostatistics at the University of Zurich, the statistical programming language R is taught to medical students and senior medical researchers of the medical faculty since 2019.

In short introductory R courses of seven hours, participants learn to prepare data for analysis, to compute descriptive statistics, to perform simple statistical tests and to create graphics, using data examples published in the medical field. Upon completion of the course, participants can book code clinic sessions where project-specific questions are solved in one-to-one sessions.

We further teach dynamic reporting with R Markdown to medical students, in a semester course, again with applications of statistical methods in clinical research. Students learn to compile R Markdown reports under a predefined structure related to manuscripts in medical research, including tables and figures generated within R, references, and R package versions. The students appreciate the improved organization of material for their master or dissertation projects, because all important information is stored within one document.

Feedback of the participants is routinely recorded immediately at the end of the courses and in a follow-up survey that will be conducted in April 2023.

Between August 2020 and January 2023, 196 participants of the medical faculty completed the short introductory R course, of whom 174 (89%) filled in a feedback form right after the course. 64 (37%) graded the course as “very good” and 94 (54%) as “good”. Most of the participants (152, 87%) stated they would use R in their next research project, 11 (6%) said maybe they would do so. We will present the survey answers and show details on sustainability of the course, and whether R was used for the analysis in subsequent research by the participants.

For early-stage researchers, knowledge of R in combination with R Markdown is becoming increasingly important. Teaching R and R Markdown to medical researchers is boosting transparency, and therefore reproducibility. This is an important step for better credibility and validity in medical research.

1. Niven, D.J., et al., Reproducibility of clinical research in critical care: a scoping review. BMC Med, 2018. 16(1): p. 26.

2. Wang, S.V., et al., Reproducibility of real-world evidence studies using clinical practice data to inform regulatory and coverage decisions. Nat Commun, 2022. 13(1): p. 5126.

3. DeBlanc, J., et al., Availability of Statistical Code From Studies Using Medicare Data in General Medical Journals. JAMA Intern Med, 2020. 180(6): p. 905-907.

4. Localio, A.R., et al., Statistical Code to Support the Scientific Story. Ann Intern Med, 2018. 168(11): p. 828-829.

ICH adopted the ICH E9-R1 guidance in November 2019. Since the guideline was released for public consultation in 2017, its implementation, especially in superiority trials, has been discussed and described in the literature. However, the implementation of estimands in the context of equivalence trials is not covered in detail by the guideline itself, and less has been published in this context.

This poster will illustrate how the ICH E9-R1 estimand guideline can be interpreted and implemented in equivalence studies. We will present how the guideline was applied to the definition of estimands for a biosimilar equivalence study conducted to demonstrate the therapeutic equivalence between a proposed biosimilar and its corresponding reference product. As a general principle, the assumptions related to the mechanism of missingness, as well as the strategies proposed to deal with the pre-identified intercurrent events were selected with the objective to define sensitive estimands, allowing to detect treatment difference if any.

In particular, the “per-protocol analysis set” analyses were revisited to allow for the Estimand framework. A “Hypothetical continuing per Protocol Estimand” was defined, where data possibly impacted by an intercurrent event were censored and imputed as if the subjects would have continued to follow the protocol.

Definition, analysis, and some results of estimands constructed with an equivalence objective will be presented and discussed.

Introduction: Inflammatory bowel disease (IBD) is characterized by a dysbiosis of the gut microbiome that results from the interaction of the constituting taxa with one another, and with the host. At the same time, host genetic variation is associated with both IBD risk and microbiome composition.

Methods: In the present study, we defined quantitative traits (QTs) from modules identified in microbial co-occurrence networks to measure the inter-individual consistency of microbial abundance and subjected these QTs to a genome-wide quantitative trait locus (QTL) linkage analysis.

Results: Four microbial network modules were consistently identified in two cohorts of healthy individuals, but three of the corresponding QTs differed significantly between IBD patients and unaffected individuals. The QTL linkage analysis was performed in a sub-sample of the Kiel IBD family cohort (IBD-KC), an ongoing study of 256 German families comprising 455 IBD patients and 575 first- and second-degree, non-affected relatives. The analysis revealed five chromosomal regions linked to one of three microbial module QTs, namely on chromosomes 3 (spanning 10.79 cM) and 11 (6.69 cM) for the first module, chr9 (0.13 cM) and chr16 (1.20 cM) for the second module, and chr13 (19.98 cM) for the third module. None of these loci have been implicated in a microbial phenotype before.

Discussion: Our study illustrates the benefit of combining network and family-based linkage analysis to identify novel genetic drivers of microbiome composition in a specific disease context.

Prediction of future fractures and identification of individuals with a higher risk of refracture is essential for the prevention and treatment of osteoporosis [1]. With survival analysis, the time to a fracture event can be assessed. However, a distinction needs to be made regarding study participants for whom the event of interest is not observed either due to classic censoring or due to a competing event such as death. We, therefore, compared three survival analysis models, the Cox Proportional Hazard (PH), the Cause-Specific Cox (CSC), and the Fine-Gray, for estimating time to osteoporotic hip fracture while considering death as a competing risk.

Using a subset of the Study of Osteoporotic (SOF) dataset, we focused on three key variables: age, BMI, and previous fracture (PFX). These three variables were then used to fit two Cox PH models for the events of incident hip fracture and death. Older age and history of PFX, are associated with a decreased time to incident hip fracture. In contrast, a larger BMI seems to confer a slightly protective effect. In the Cox PH model for the event of death, all three variables are positively related to earlier deaths. When comparing the output of the Cox models to the CSC model, we see very minor differences between the hazard ratios calculated from both methods. However, with FG for incident hip fracture, we see a decrease in the hazard ratios for age and previous fracture and a slight increase in the hazard ratio for BMI.

While our results indicate some differences between the models, these differences are mostly minor. Currently, we are extending this study to include more variables, the full SOF dataset, and the use of other machine learning methods with the goal of refining and utilizing these models for clinically applicable risk prediction.

Reference
Morin, S. N., Lix, L. M., & Leslie, W. D. (2014). The importance of previous fracture sites on osteoporosis diagnosis and incident fractures in women. Journal of Bone and Mineral Research, 29(7), 1675-1680.

Encountering missing values is almost inevitable when working with clinical data. Among the variety of methods available for dealing with them, experts often favour multiple imputation over simpler techniques [1]. However, recent literature reviews reveal that clinical risk prediction models show poor handling of missing data and reporting thereof: Most authors either delete missing values or deploy mean imputation, or they rather omit reporting on missingness in their data completely [2]. Why is the problem of missing data widely discussed in theory but seldom tackled in practice? Reasons might be that complex imputation models like MICE are more error-prone and time-consuming and may require consultation with clinical experts. This work aims to answer whether the additional effort of including medical knowledge for imputation results in improved prediction models.

The study uses routine data from ten thousand patients undergoing elective surgery from 2016 to 2022 at the university hospital Charité in Berlin. The dataset includes demographics, vital signs, clinical scores, laboratory values, comorbidities, and surgery information as well as indicators for eight common perioperative complications (e.g., pneumonia, bleeding, or death). For enabling a comparison of missing data handling methods, we developed prediction models for the eight complications by training the methods crosswise with different Machine Learning algorithms. Standard techniques for deletion of missing values, simple imputation, and multiple imputation were compared amongst each other and to new procedures in which each feature receives its own imputation model based on clinician’s appraisals. For this purpose, we performed interviews with clinicians from two different sites. In the analysis, we assessed how much time was needed to specify and train the imputation models. We measured it against the imputation’s effects on prediction performance and the clinician’s trust in the predictions, knowing how missingness is handled in new data.

In our cohort, missing data occurred in all patients and a quarter of the features, with an overall missing ratio of around 15%. The prediction models performed similarly for all missing data handling methods. At the cost of a longer time to specify the imputation models, the procedures involving clinical experts increased trust in the final predictions compared to other methods. We will validate our findings further and assess the effect of altering the imputation methodology on regional and temporal generalizability with data from an ongoing prospective study including four hospitals. We expect that standard methods for handling missing data may overemphasize cohort-specific characteristics of the data, whereas embedding expert knowledge may help to better translate to the broader patient population.

References

1. Jakobsen, J.C., Gluud, C., Wetterslev, J. et al. (2017). When and how should multiple imputation be used for handling missing data in randomised clinical trials – a practical guide with flowcharts. BMC Med Res Methodol, 17, 162. https://doi.org/10.1186/s12874-017-0442-1
2. Nijman, S.W.J., Leeuwenberg, A.M., Beekers, I. et al. (2022). Missing data is poorly handled and reported in prediction model studies using machine learning: a literature review. JCE, 142, 218-229. https://doi.org/10.1016/j.jclinepi.2021.11.023.

Transepidermal water loss (TEWL) describes the loss of water through the epidermis via diffusion or evaporation and is important for the evaluation of the skin barrier function. TEWL is known be altered, for example, in physically damaged or eczema-affected skin. In addition, it was shown to be influenced by age as well as environmental factors such as temperature and humidity. Recent studies also suggested short-term effects of air pollutants on TEWL.

In the longitudinal Study on the influence of Air pollution on Lung function, Inflammation and Aging (SALIA) we investigated the effects of short-term exposure to particulate matter with a diameter of 2.5 micrometers or less (PM_2.5) on TEWL. During a follow-up investigation in 2018/2019 including 224 women aged 75-89 years the TEWL was measured in eight locations across the body. Exposure to PM_2.5 at the participant’s home address on the day of the investigation was assessed via chemical transport models of the German Environment Agency UBA.

Since levels of TEWL differ between different body parts we used a hierarchical Bayesian regression model, which allows for possibly different effects of PM_2.5 among the eight locations while at the same time estimating an overall effect of PM_2.5 on TEWL. The model includes random participant effects to account for the repeated measurements and it was adjusted for ambient temperature, humidity and personal confounder variables. Estimates were derived as percentage change with respect to an IQR increase in PM_2.5.

We found higher levels of TEWL on the forehead (posterior median, 95% credible interval: 1.49, [1.25, 1.90]) and on the upper side of the hand (1.22, [1.02, 1.54]) compared to the overall mean. In addition, TEWL was increased after short-term exposure to PM_2.5 on the wrist (1.25 [1.12, 1.39]) and the cheek (1.11, [1.002, 1.22]). The overall effect of PM_2.5 on TEWL was estimated as 1.07 [0.98, 1.19].

Reference intervals and standard deviation scores (‘z scores’) are widely used as diagnostic tools in various biomedical fields. They are applied to laboratory parameters in clinical chemistry, psychometric tests in neurology, or parameters of children’s growth in pediatrics. Usually, samples from a ‘normal’ population form the data basis for the estimation of reference distributions. This target population may or may not include unhealthy individuals depending on whether the reference should reflect the healthy or the general population. Health status, however, is not collected in some types of reference studies, e.g. in pediatrics.

Parametric estimation of references distributions is complicated by extreme values that may be outliers relative to the assumed (working) distribution or model, even for the unconditional case (i.e. without dependence on covariables such as age). If sample size is moderate, genuine outliers may by chance be over-represented in the sample used for estimating the reference distribution and impair the selection of a suitable model. Second, if the target population is to include unhealthy individuals with their representative share, the reference distribution to be estimated is a mixture of a major ‘healthy’ part plus a ‘contamination’. Finally, the sample may be contaminated by observations that are not members of the target population but remain undetected.

Often in practice, the origin of the extreme values at hand is unknown, but the interest is in one or both of the extreme tails of the distribution. However, most ad-hoc methods proposed in the literature for outlier removal or correction are only designed for one of the cases. In this contribution, the simple approaches of ignoring, deleting or winsorizing outliers are compared to two ex-post correction methods, one of which is newly proposed. Correction methods remove outliers before estimating the reference distribution, thus allowing to select a less complex model; for new observations from individuals to be diagnosed, percentiles are calculated based on the estimated reference distribution and then corrected for the removal of outliers by appropriate rescaling and shifting.

Simulation studies with normal, skewed and heavy-tailed distributions, varying sample size and varying degree of contamination show the strengths and limitations of the considered methods. The simple ad-hoc methods highly depend on their respective assumptions. While the first correction method over-corrects the deletion of outliers, the new correction method adequately corrects for genuine outliers being removed and attenuates the effect of contamination. Data on body mass index and body proportions from a large Austrian pediatric study are used for demonstration.

Diagnostic tests are commonly evaluated by estimating the area under the receiver operating characteristic curve (AUC), as well as sensitivity and specificity at given diagnostic cut-offs. One difficulty with diagnostic trials is that many of them use factorial designs. This means that different combinations of readers and methods may be used to diagnose a patient. In addition, diagnostic studies may generate clustered data by repeated measurements over time or several lesions. See [1, Lange] for a mathematical framework to deal with both of these difficulties.

Additionally, it may be of interest to correct the estimation procedure of the above mentioned accuracy measures for covariates. For example, it may be the case that age, weight or height influence the diagnostic accuracy of a test. In [2, Zapf] a methodological approach is presented to adjust the AUC for such covariates, while also allowing for factorial designs. We developed a modification of the adjustment approach to guarantee unbiased estimators also for sensitivity and specificity.

In this talk, we present the new approach and give a short overview of the corresponding R package (under development).

1) Lange, K. (2011, March 4). Nichtparametrische analyse diagnostischer Gütemaße bei Clusterdaten. Retrieved February 27, 2023, from http://dx.doi.org/10.53846/goediss-3538

2) Zapf, A. (2009, October 23). Multivariates nichtparametrisches Behrens-Fisher-problem MIT Kovariablen. Retrieved February 27, 2023, from http://dx.doi.org/10.53846/goediss-2488

3) Lange, K., & Brunner, E. (2012). Sensitivity, specificity and ROC-curves in multiple reader diagnostic trials—a unified, nonparametric approach. Statistical Methodology, 9(4), 490–500. https://doi.org/10.1016/j.stamet.2011.12.002

Directed acyclic graphs (DAG) have become a well established tool to investigate confounding and bias in observational studies and real-world data. Simulating data according to the causal structures of a given DAG could potentially be useful for teaching, methods research, deciding on appropriate analytical approaches, and even during study design. Convenient simulation functionalities nonetheless remain limited in published software.

The R package dagR, which can also be used to identify (minimal) sufficient adjustment sets, initially included the possibility to simulate data for a given DAG including any combination of binary and continuous nodes/variables, with dependencies conforming to a logistic or linear model, respectively [1,2]. Functionalities to simulate binary variables based on a risk difference specification were added more recently [3].

Some examples are presented to demonstrate potential applications of the package in teaching and research.

Work in progress includes the implementation of time-to-event and repeated measurements simulations, which could pave the way to much wider application in fields of current interest, such as emulated target trial methodology.

1. Breitling LP, Duan C, Dragomir AD, Luta G. Using dagR to identify minimally sufficient adjustment sets and to simulate data based on directed acyclic graphs. Int J Epidemiol 2021;50(6):1772–1777
2. Duan C, Luta G, Dragomir AD, Breitling LP. Reflection on modern methods: understanding bias and data analytical strategies through DAG-based data simulations. Int J Epidemiol 2021;50(6):2091–2097
3. dagR 1.2.1 (2022-10-09), https://cran.r-project.org/package=dagR

In the context of multiple sclerosis (MS), relapse rates are seen as a key indicator of disease activity. Due to its gynecotropia and the peak of initial manifestation in child-bearing years¹, MS-disease management during pregnancy is of critical importance. A study by Confavreux et al.² found that pregnancy reduced relapse rates in mainly untreated MS patients. However, the use of modern disease-modifying treatments (DMTs) adds complexity to pregnancy planning. Most modern biologics are not licensed for use during pregnancy and must be discontinued prior to conception or in early pregnancy.

However, upon discontinuation of DMTs, the issue of rebound relapses arises. Natalizumab, for example, has been found to increase relapse rates in 3 months post-cessation exceeding pre-medication levels ³. It is yet to be determined, whether the protective effect of pregnancy also applies to rebound relapses. Current observational studies remain largely inconclusive^4,5, partly due to the complexity of the temporal relationship between pregnancy duration and the cessation of the disease-modifying therapy (DMT) relative to conception.

To further explore the clinical question of whether pregnancy reduces the intensity of rebound relapses, we must consider the interaction between these two time-dependent covariates. Therefore, in this study we use simulated data to examine which analysis methods are best suited to capture this pregnancy*rebound interaction, in terms of power, bias, and communicability.

We conducted a discrete-time simulation⁶ [DTS], informed by clinical data, to generate 1000 repeat samples of differing parameter sets with varying sample sizes of patients treated with a generic rebound-inducing DMT. We modeled the time-varying effects of the DMT and pregnancy as discrete and continuous effects.

To test the interaction hypothesis, we used forms of recurrent event Cox regression and Poisson regression models. The models are compared based on bias and power of the interaction term only. We will present results and discuss the implications for further studies in this field.

1. Holstiege J, Steffen A, Goffrier B, Bätzing J. Epidemiologie der Multiplen Sklerose – Eine populationsbasierte deutschlandweite Studie. 2017 Dec 7;

2. Confavreux C, Hutchinson M, Hours MM, Cortinovis-Tourniaire P, Moreau T. Rate of Pregnancy-Related Relapse in Multiple Sclerosis. N Engl J Med. 1998 Jul 30;339(5):285–91.

3. Papeix C, Vukusic S, Casey R, Debard N, Stankoff B, Mrejen S, et al. Risk of relapse after natalizumab withdrawal: Results from the French TYSEDMUS cohort. Neurol Neuroimmunol Neuroinflamm. 2016 Dec;3(6):e297.

4. Hellwig K, Tokic M, Thiel S, Esters N, Spicher C, Timmesfeld N, et al. Multiple Sclerosis Disease Activity and Disability Following Discontinuation of Natalizumab for Pregnancy. JAMA Network Open. 2022 Jan 24;5(1):e2144750.

5. Portaccio E, Moiola L, Martinelli V, Annovazzi P, Ghezzi A, Zaffaroni M, et al. Pregnancy decision-making in women with multiple sclerosis treated with natalizumab: II: Maternal risks. Neurology. 2018 Mar 6;90(10):e832–9.

6. Tang J, Leu G, Abbass HA. Discrete Time Simulation. In Simulation and Computational Red Teaming for Problem Solving (eds J. Tang, G. Leu and H.A. Abbass). 1st ed. Wiley; 2019 . https://doi.org/10.1002/9781119527183.ch8

Collider stratification bias is routinely brought up as a potential explanation when the association of obesity with the development of a disease is of different magnitude than the association of obesity with disease outcome. The extreme case, where obesity is a risk factor for the disease, but is associated with improved survival in the diseased, as observed for example for cardiovascular disease, diabetes, chronic kidney disease, and several types of cancer, is known as the “obesity paradox”. Prerequisites for collider bias to occur are an effect of obesity on the development of disease, and an unadjusted confounder of the disease development and outcome relationship, such as for example (unknown) genetic risk factors. Simulating multiple scenarios with different choices of input parameters (i.e. effect of obesity on disease development; strength of the unadjusted disease development and outcome confounder; disease and outcome incidence), we investigated under which scenarios the danger of collider bias is real, and under which it is negligible. A main finding was that the potential impact of collider bias is relatively unaffected by the specific prevalences, but sensitive on the magnitude of the effect of obesity on disease development. As long as this effect is not too large (hazard ratio (HR) between 0.67 and 1.5), even assuming a strong background confounder, collider bias is relatively small, meaning that the true, unbiased HR is distorted by less than 5% on the HR scale (i.e. less than 0.05 on the absolute HR scale if the HR is not too big). For comparison, in large observational studies with a couple of hundred thousand participants, 95% confidence intervals for mortality outcomes for a rare, but not too deadly, disease often cover a range of 0.2 or more on the HR scale, and thus the random variability in HR estimates outweighs the likely magnitude of collider bias. Only in case of small causal effects on the disease outcome, collider bias is able to fully explain the obesity paradox. Classical confounding, detection bias, heterogeneity of disease bias, or model misspecification bear a much higher potential of substantially distorting observed associations than collider stratification in many typical scenarios.

High-throughput amplicon and metagenomics sequencing data have become an invaluable resource for assessing broad statistical patterns of associations between microbial communities and their environment. Host information and a priori expert knowledge represented on a tree are often provided along with the high-dimensional and compositional microbial data. Sparse and compositional aware predictive methods incorporating the hierarchical information and other (non-compositional and non-hierarchical) host-related features are therefore desirable.

Here, we propose a two-stage classification framework encompassing and extending prior hierarchical expert knowledge to associate compositional microbial data with the host's trait, allowing for non-compositional covariates. The first stage extends the sparse log-contrast model to classification cases imposing a zero-sum constraint for compositional awareness and an l1 penalty for sparsity to the convex loss. Another ingredient is the tree aggregation idea allowing for flexible aggregation and selection along the hierarchical tree. The second stage simplifies the model by running another sparse method on log-ratio pairs transformed aggregated variables defined by the support of the first stage.

The applications used the taxonomic tree annotation of the microbes. The method was benchmarked against state-of-the-art black-box models on various datasets, revealing comparable predictive performance while being sparser. The analysis of two real-world datasets revealed the more fine-grained log-ratio Firmicutes (Phylum level) / Bacteroidales (Order level) as possible microbial biomarkers for irritable bowel syndrome instead of the often used fixed-level ratio Firmicutes / Bacteroidetes on the Phylum level.

The scalable and interpretable modeling framework for high-dimensional compositional microbiome data allows for selecting log-ratios pairs predictive of a host's trait. The peculiarity of the log-ratios is to allow for aggregation but to restrict those to a prior defined hierarchical structure. This framework contributes to the growing toolbox for high-dimensional compositional data analysis for microbial data.

Temporally ordered interactions between actors constitute a complex temporal network where past configurations may be partly responsible for future ones. These interactions are sometimes referred to as relational events. The relational event model (REM) seeks to capture the underlying dynamics driving these interactions. Nevertheless, assessing the goodness of fit (GOF) of these models is still mostly an open field of research, particularly for REMs with time-varying and random effects.

Our proposal relies on a cumulative martingale-residual process evaluated for a smooth mixed-effect REM estimated using case-control sampling. We particularly focus on a test of the Kolmogorov-Smirnov type, intended to determine if the covariates are correctly modeled.

In an empirical application, we fit a smooth case-control REM to sequences of alien species invasions. A first record, defined as the first year in which a species is detected as alien in a region where it was not native, is our relational event of interest. To test the GOF of our proposed model, we consider weights being equal to the covariates considered in the model.

Martingale residual stratification may be done in a variety of ways. Our approach can easily be extended to determine whether any other network dynamics features have been adequately incorporated into the model.

As statisticians we have been making many efforts to render our work more accessible for non-statisticians. One aspect however which has been neglected so far is the topic of Digital Accessibility. What can we statisticians do in order to make our work also accessible for people with permanent or temporary disabilities? How can we change the way we present data in order to be inclusive?

Digital Accessibility (DA) is an integral element in creating an inclusive (digital) world and is closely linked to the topic of the conference “From Data to Knowledge. Advancing Life Sciences.”: DA ensures information can be reached, used and understood by everyone..

An estimated 1.3 billion people – or 1 in 6 people worldwide – experience significant disability (WHO, 2022). DA is crucial so that these people but also many of us who do not consider ourselves being disabled can access digital content. Disabilities can be visible, but many disabilities are invisible. The most common categories are

• Visual disabilities (e.g., blindness, low vision, color blindness)

• Auditory (e.g., deaf and hard of hearing)

• Motor (e.g., inability to use a mouse, slow response time, limited motor control)

• Cognitive (e.g., inability in memory, attention/focus, text processing)

In order to make content accessible, it should be easy to see, easy to hear, easy to interact with and easy to understand.

Ultimately, DA means good design and good usability and many other people without permanent disabilities will also benefit: people not fluent in a language, older people, people with “temporary disabilities” due to accident or illness, people with “situational limitations” such as weather-based conditions (bright sunlight, loud environment etc.) as well as people using a slow internet connection or legacy browsers or unable

In our working context DA is very relevant in interactions with our stakeholders (patients, physicians, researchers and students, existing and future employees, investors). It applies to almost everything we do: from choosing fonts and colors to creating tables and graphs which can be used for reports, publications, presentations, websites, social media etc.

In the session I would like to discuss concrete examples what can we do differently as statisticians in order to make our work digitally accessible, e.g.,

• Provide sufficient contrast

• Conscious use of colors (do not use colors alone to convey information)

• Ensure that interactive elements are easy to identify

• Provide clear and consistent navigation options

• Provide easily identifiable feedback

• Use headings and spacing to group related content

People should leave with tangible ideas on easy things to change in order to make our work more accessible and thus easier to understand and to spread.

Sources:

• WHO 2022. Fact Sheet Disability. Available: https://www.who.int/news-room/fact-sheets/detail/disability-and-health [2023, February 28]