To monitor epidemics, different indicators such as incidences, hospitalisation, and deaths play an important rôle. The corresponding data form a multivariate time series whose development over time is governed by several effects: the non-linear reproduction equation describing the spread of the disease, delays due to reporting and the patient-dependent progression of the disease, as well as under-reporting. We show that state space models offer a flexible framework to incorporate these effects. They can easily be fitted even to incomplete data either using an extended Kalman filter or following the likelihood-based approach proposed by Durbin & Koopman (2001). Using publicly available data on COVID-19, we demonstrate how the predicted evolution of states can be used both to monitor the epidemic as well as to generate short-term forecasts.

Pertussis or whooping cough is a highly contagious vaccine preventable disease. Incidence of pertussis has known a steady decline after the introduction of pertussis vaccination in the first half of the previous century, nevertheless, pertussis incidence increased over the past two decades in many countries. The analysis of serial serological survey data can improve our understanding about the dynamics of pertussis.

However, the development of assays for the detection of IgG antibodies in sera entails that various assays have been used for different survey years and thus the antibody titre measurements are not directly comparable. Comparable sero-epidemiological results would enable exploring statistical and mathematical models to estimate time-varying epidemiological parameters, i.e. ‘standardized’ assay results are required. The long-term goal is to investigate the consequences of the uncertainty related to the standardization of pertussis toxin IgG antibodies results from three serological surveys conducted in Belgium (2002, 2006, 2013). In the current project, we focus on finding the standardization model for each survey.

In each survey, 150 samples were selected such that the range of the original values for IgG antibodies against pertussis toxin was as best as possible covered. All 450 samples were then tested using a magnetic bead-based multiplex immunoassay (MIA). We explored different models for standardization in a Bayesian framework using log-transformed titer values. We considered several strategies for dealing with the censored data that occur due to the limit of detection in assays. i.e. in the MIA assay (dependent parameter) as well as in the original assays (independent parameter). Models and strategies regarding censoring are contrasted using fitted curves with confidence bands as well as posterior predictions.

The choice of model for standardization depends on the strategy used for censored data, which can lead to substantial differences in predictions. The uncertainty in the standardization of antibody titres needs to be reflected in models aimed at estimating time-varying epidemiological parameters from serial serological survey data.

Description: Although randomized studies previously assessed short-term effects of behavioral factors on overweight/obesity (OW/OB), they could not assess the effects under real life conditions nor over long time spans. Using methods of causal inference, we therefore aimed to study the long-term effects of hypothetical behavioral interventions on OW/OB from childhood to adolescence, i.e. to answer questions such as “What would happen to the incidence of OW/OB if all children continuously adhered to screen time recommendations over a period of 13 years”. Our sample comprised 10 877 children aged 2 to <10 years at baseline who participated in the well-phenotyped IDEFICS/I.Family cohort¹. Children were followed over 13 years from 2007/2008 to 2020/21. The risk of developing OW/OB was estimated under various single and joint hypothetical behavioral interventions using the parametric g-formula².

The 13-year risk of developing OW/OB was found to be 30.7 [28.4;32.7] percent under no intervention and 25.4 [22.1;27.7] percent when multiple interventions were imposed jointly, corresponding to a risk reduction of 17%. The most effective interventions were to meet screen time recommendations and to meet moderate-to-vigorous physical activity recommendations which could reduce the incidence of OW/OB by -2.2 [-4.4;-0.7] and -2.1 [-3.7;-0.8] percentage points (risk difference [confidence interval]), respectively. Meeting sleep recommendations (-0.6 [-1.1;-0.3]) showed a similar intervention effect as compared to increasing sleep duration by 30 minutes/day (-0.6 [-0.9;-0.3]). If all children were members in a sports club, the incidence of OW/OB could be reduced by -1.6 [-2.7;-0.4] percentage points over a 13-year period. Unexpectedly, reducing the consumption of sugar-sweetened beverages by 1 drink/day increased the risk by 0.73 [0.16;1.4] percentage points; however, this effect disappeared in a sensitivity analysis where exposures were redefined so as to precede the outcome by several years.

Our analysis is one of only few practical applications of the parametric g-formula to cohort data. We will discuss the plausibility of the underlying assumptions and point to practical challenges in its implementation in the context of our specific research question: These address, for instance, the irregular and fairly long time intervals between waves and the modelling of time-varying covariates.

To conclude, we evaluate the utility of our approach for estimating intervention effects based on observational data and highlight potential sources of bias. We make suggestions for strengthening confidence in any results obtained from observational data by following the principle of target trial emulation.

References

1. Ahrens W, Siani A, Adan R, et al. Cohort Profile: The transition from childhood to adolescence in European children-how I.Family extends the IDEFICS cohort. Int J Epidemiol 2017;46(5):1394-1395j.

2. Naimi AI, Cole SR, Kennedy EH. An introduction to g methods. Int J Epidemiol 2017;46(2):756-762.

Population-based cohort studies using longitudinal electronic health records (EHR) data are commonly used to explore adverse health outcomes for specific conditions (e.g., is eczema associated with subsequent development of fractures/cancer/cardiovascular disease/etc...). These studies aim to answer causal questions to provide actionable evidence for decision makers, however the status quo of conducting individual “one exposure – one outcome” studies is problematic, in its inefficiency, lack of transparency, lack of comparability between studies, and due to concerns about publication biases. Applying generic and adaptable approaches consistently to multiple research questions, while preserving the ability to incorporate the required expert knowledge and critical thinking, will allow generating evidence on important health related questions more quickly, make results more comparable, reporting more transparent, and studies more easily updateable with components such as more detailed disease phenotypes.

We describe how "hypothesis-testing" EHR studies are currently conducted and delivered as a "one exposure - one outcome" package, and problems with this approach. We then explore how a "one exposure - many outcomes" approach could address problems with current approaches and propose a framework to implement such an approach. We will discuss 3 main themes: 1. The role of heterogeneity between studies, i.e. which parts of studies can remain the same for research questions on different outcomes, and which parts should be different; 2. The role of clinical expertise and for each outcome selecting which analyses should be considered as main analyses and which as sensitivity analyses; 3. The organisation of outcomes into categories, based on universally applicable hierarchies, and exposure-specific categories.

We will demonstrate the approach using the applied example of adverse health outcomes in adults with (atopic) eczema. Eczema may be related to several adverse health outcomes, however for most previously explored outcomes there is only low or moderate certainty evidence, and some outcomes may be unknown. This uncertainty should be addressed in order to improve patient management including implementing screening and preventive measures where appropriate. We will describe a framework for a research pipeline, with generic methods to set up cohorts, estimate exposure-outcome associations within these cohorts, and account for confounding. Using this pipeline, we will explore adverse health outcomes associated with eczema, replicating previously published cohort studies, and investigating associations with outcomes that have not or not adequately been researched in the existing literature. The choice of outcomes will be guided by findings from a recent large-scale review of the previous evidence. Results for the applied example will be produced shortly, pending data access.

In summary, we will create a high-quality and comprehensive evidence source on the topic of adverse health outcomes associated with eczema while describing a framework adaptable to other skin diseases and other areas of research.

Background: The pathophysiological causes of chronic kidney disease development are not fully understood. Genome-wide association studies (GWAS) can be utilized as a foundation for causal inference to identify molecular pathways involved in the pathogenesis. One of the key challenges is the translation of genetic variants associated with disease to the respective physiological relevant units – the genes. While transcriptome-wide association studies rooted in genetic instruments have advanced our capabilities for causal inference using gene expression (e.g. PrediXcan and FUSION approaches) the joint analysis of related transcriptomic and proteomic features is an unmet statistical challenge.

Results: We applied a Mendelian Randomization technique that focused on instrumental genetic variables within the respective gene region (cis-variants) and allows us to conduct a genome-wide screen in a two-sample design. We combined tissue-specific weights derived by elastic net regression in the GTEx project for transcriptome-wide association studies (TWAS) in relevant tissues (kidney cortex, kidney tubule, liver, and whole blood), plasma-specific weights derived for proteome-wide association studies (PWAS), and the most recent genome-wide association study (GWAS) summary statistics for three markers of kidney filtration (glomerular filtration rate (GFR) estimated with serum creatinine, GFR estimated by serum cystatin C, and blood urea nitrogen) and one of kidney damage (albuminuria). This allowed us to assess the effects of 12,893 genes and 1,342 proteins on each of the kidney markers. We found 1,561 significant associations (Bonferroni adjusted) distributed among 260 genomic regions that were supported by TWAS and/or PWAS as putatively causal. We then prioritized 153 of these genomic regions using additional Bayesian colocalization analyses (posterior probability >80%). These findings intertwined with local co-regulation of neighboring genes left us with an unmet statistical challenge of the joint analysis of the different tissues and neighboring genes. To integrate the genetic influence on the transcriptome and the proteome, we used regression with a summary statistics approach to incorporate the cis-regulated genetic correlation of the different models, performing conditional causal inference to identify the underlying tissue of an association as well as to evaluate whether multiple independent signals were contained within the same genomic region. Our findings were supported by existing knowledge (e.g., animal models for MANBA, DACH1, SH3YL1, INHBB), exceeded the underlying GWAS signals (28 region-trait combinations without significant GWAS hit), were confirmed by experimental follow up in a clinical cohort (INHBC kidney disease progression hazard ratio=1.86, CI=1.37-2.52) and differentiated markers of kidney filtration from those with roles in creatinine and cystatin C metabolism.

Conclusion: In summary, this study combined multimodal, genome-wide association studies to generate a catalog of putatively causal target genes and proteins relevant to disease. We extended the causal inference approach to allow for conditional analysis to prioritize tissues, and demonstrated the application based on the example of kidney function and damage which can guide follow-up studies in physiology, basic science, and clinical medicine.