Keynote Speakers' Program
Bayesian Methods in Clinical Research
Emmanuel Lesaffre, L-Biostat, KU Leuven, Leuven, Belgium
Bayes theorem is well –known among medical researchers, but probably only as a tool to derive from sensitivity, specificity and prevalence the predictive value of a diagnostic test. The Bayesian methodology is also popular for (medical) decision support systems. However, a search in the current dental literature shows that Bayesian methods are rarely used in dental research outside the dental age assessment in forensic and criminal applications. Nevertheless, the Bayesian approach to statistical inference has become increasingly popular the last 25 years in a great variety of research areas including medical research.
Initially, the Bayesian approach has been primarily applied to model complex data structures, such as complex hierarchical data structures. In the last decade there is, however, a strong interest for this approach in clinical trial research. Indeed, the fact that the Bayesian approach allows the use of prior information into the analysis of the current study opens the possibility to reduce the necessary sample size of a clinical trial.
In this talk we review the basic principles behind the Bayesian approach and contrast them with the classical statistical fundaments. We indicate why it has taken about 250 years for the Bayesian approach to become popular. The usefulness of the Bayesian methodology will be illustrated using a variety of oral health and medical examples. The computational power of the approach will be illustrated via analyses on a complex longitudinal dental study, the Signal Tandmobiel® study.
Beware of the fashionable: keeping it real in oral epidemiological research
WM Thomson, The University of Otago
Fashions in dental research come and go. You can see how it happens. Someone reports a hitherto-undescribed association from a survey and “over-interprets” the findings. That finding gets replicated in other cross-sectional studies; the media gets involved; work gets underway to elucidate putative biological mechanisms; funding bodies get interested; and research centres get set up as the new field burgeons. Eventually, the whole thing gradually loses steam for a number of reasons, chief among which are the failure to confirm a causal relationship, and the use of more appropriate study designs, measures and analyses to investigate the issue. One of the initial problems in such a phenomenon is the early and inappropriate use of the language of causality. This paper will consider the issues involved and make some recommendations for improving the accuracy of scientific writing.
When Harmful Exposures Look Good
Kaufman JS, Department of Epidemiology, Biostatistics and Occupational Health, McGill University, Montreal, Quebec, Canada
Aim: Selection bias is a ubiquitous concern in longitudinal follow-up of health data.
The aim of this talk is to raise awareness and understanding of this and several
other causal inference issues.
Method: The presentation will define selection bias using directed graphs and
counterfactual quantities. Then a simulated case-control study will be used to
illustrate a number of issues including control selection and its impact on
parameter interpretation, and the roles of the causal consistency and positivity
assumptions. Finally and most importantly, the example will be used to
show what happens when the units remaining at risk of the outcome change over
Results: When study denominators change over time, exposures that can only harm an
individual can appear to be protective. This is due to conditioning on remaining
at risk, as is done in Cox Proportional Hazards models and other common statistical
Conclusion: Valid causal inference requires absence of confounding, selection and
misclassification biases. Of these, selection is probably the least well appreciated
in practice, and yet can have profound effects on inference, especially when outcomes are not rare.
Analysis of Zero-inflated Count Data from Longitudinal Oral Health Studies
KyungMann Kim, University of Wisconsin-Madison
In the first half of the talk, I will review statistical methods for analysis of longitudinal data such as linear and generalized linear random-effects models and models based on generalized estimating equations. Statistical models for the analysis of longitudinal data must take into account the association among serial observations with each study unit. Two-stage models naturally lead to random-effects models. Generalized estimating equations approach is semiparametric in that it only assumes specific parametric models for the first two moments, i.e. mean and variance-covariance.
In the second half of the talk, I will discuss how to formulate these models to analyze caries indices with excessive zeros. More specifically, I will review zero-inflated regression models which view data as being generated from a mixture model for a point mass at zero and a non-degenerate distribution. These models have emerged as a popular framework to characterize the dependence of count data with excess zeros on covariates. Applications of these models focus on the latent class formulation where the mean response of the so-called at-risk or susceptible population and the susceptibility probability are both related to covariates, but separately. While this formulation in some instances provides an interesting representation of the data, it often fails to produce easily interpretable covariate effects on the overall mean response which are often of scientific relevance. Recently we developed two new approaches that circumvent this limitation. The first approach consists of estimating the effect of covariates on the overall mean from the assumed latent class models, while the second approach formulates a model that directly relates the overall mean to covariates. I explore longitudinalization of these models motivated by an oral health study on low-income African-American children, in which the overall mean model is used to evaluate the effect of sugar consumption on caries indices.
This is based on a joint work with David Todem of the Michigan State University and Wei-Wen Hsu of the Kansas State University.
Development and Implementation of an Oral Health Data Registration and Evaluation System for the Belgian Population
Dominique Declerck, KU Leuven Department of Oral Health Sciences, Research unit Population Oral Health studies, Leuven, Belgium
Aim:Information regarding health condition, related behaviors and utilization of health care services is essential for the planning and organization of care delivery. Recently, an oral health data registration and evaluation system was developed in Belgium. The aim of this presentation is to describe the set-up and implementation of this system, including a discussion of challenges encountered.
Methods: The Belgian Oral Health Data Registration and Evaluation System (OHDRES) was implemented nationwide in 2009-2010, and a second time in 2013-2014. The system combines a Health Interview Survey (HIS) and a Health Examination Survey (HES). The HIS-part consists of an interview (from 2013-2014 onwards, part of the National Health Interview Survey organized by the Belgian Scientific Institute for Public Health) and the completion of a self-administered questionnaire (focus on oral health related habits, delivered by the dentist-interviewer). The HES consists of a standardized oral examination by a trained dentist-interviewer at the participant's home. The collected data were supplemented with data on utilization of (oral) health care services, registered by the Belgian health insurance funds. All data were transferred to a platform for patient related data acting as a trusted third party, where individual files were linked. Sampling of Belgian residents (> 5 years) consisted of a multi-stage, stratified cluster sampling procedure. Ethical approval was obtained and informed consent collected from all participants.
Results: Results from both data collection periods will be presented, with a strong focus on methodological challenges that were encountered. In households that accepted to participate, clinical and questionnaire data could be obtained in a large amount of the individuals. Dentist-interviewers reported difficulties with the scoring of some clinical variables, eg periodontal condition and enamel defects. Contacting the selected households was the most difficult and time-consuming task for the dentist-interviewers, impacting on retention rate.
Conclusion:The Belgian OHDRES was successfully implemented at two occasions, after long and complex preparations. The integration with the National Health Interview Survey that was realized in the OHDRES-2014 survey resulted in a net enrichment of the data set and created considerable added value.
Causal analyses in observational studies
Odd Aalen, University of Oslo, Oslo, Norway
Causality is a major issue in epidemiological and clinical research. One wants to find out whether certain interventions, like treatments or preventive measures, do have a real effect. Deducing causality from observational statistical data is a tricky issue; nevertheless, this is often what we want to do. Over recent years new and promising approaches to causal inference have been developed. A useful system of causal diagrams have been developed by Judea Pearl at UCLA and his co-workers. Donald Rubin and James Robins at Harvard introduced the ideas of counterfactual causal inference. We shall give and introduction to these fundamental ideas with some applications.