The Digital Health and Informatics Network (DHIN) launched their first event with Professor Sally Cripps’ presentation “Bayesian Breakfast: To P, or not to P” discussing the use of Bayesian statistics in health research data. This session has been made publicly available here.
Sally began with a brief introduction to Bayesian theory, revisiting three well-known examples where the lack of coherent probabilistic thinking led to invalid conclusions. She considered Bayesian statistics more intuitive than frequentist statistics, particularly in the interpretation of the p-value (and misuse of!) and hypothesis testing conditional on the available data.
Sally then presented a case study using longitudinal data to identify risk factors associated with disengaged youth. Here, she compared findings from the Bayesian approach to the multiple logistic regression approach. In particular, she highlighted how a model’s predictive ability is unrelated to the model’s posterior probability. That is, the most probable model may not be the best model for predictions. While some risk factors may overlap, the most probable model may identify different risk factors from the model with the greatest predictive ability. In Sally’s study, education, depression and sleep duration were present in the most probable model, while maternal authority, physical neglect and fatigue were present in the model with the highest predictive ability. Both models also considered emotional abuse as a risk factor.
In conclusion, Sally suggested not to P and, if you must, to be well aware of its limitations. While Bayesian Statistics has its merits, one of the biggest barriers to its uptake is its availability and ease of use in current statistical programs. Sally alluded to this, and encouraged researchers to get in contact with a biostatistician familiar with Bayesian statistics. Its low uptake is also likely to be overcome in the not too distant future and, with advocates such as Sally, it will surely be more frequently used in health research.
If you’d like to see the entire presentation and questions, take a look here.
Nicole De La Mata, Biostatistician
Sydney School of Public Health