The relationship between body mass index (BMI) and all-cause mortality has been studied in a number of giant pooled analyses and systematic reviews, mostly showing a "J" shape with mortality increasing both at the lower and upper ends of the BMI scale. We were surprised to see much less research on cause-specific mortality, so we decided in our recent study (https://www.ericas.org/reviews/view/59) to work on a comprehensive analysis of the relationships between BMI and different causes of death, using BMI and other data from a large electronic primary care database (the Clinical Practice Research Datalink, or CPRD), linked to ONS mortality data.
Carrying out this analysis of 3.6 million patient records, and nearly 50 cause-specific mortality outcomes brought many challenges. First, how should we classify causes of death? We considered ICD chapters but these are not always clinically coherent, so we decided to borrow from the Global Burden of Diseases classifications, which usefully operate at several levels of granularity (e.g. non-communicable disease->cardiovascular disease->myocardial infacrtion), allowing for "broad brush" as well as more focussed analyses. Second, what to do about missing data? Around 25% of otherwise eligible patients had no BMI recorded. We did not feel comfortable with using multiple imputation because recording of BMI in a GP context is very likely to be influenced by the BMI itself, a direct violation of the so-called "missing at random" assumption necessary for multiple imputation. So, we instead excluded those with no BMI data, and used a number of sensitivity analyses to reassure ourselves that this did not compromise our conclusions. Third, how to overcome computing challenges? We had data on 3.6 million patients multiplied by sometimes hundreds of rows of data per patient, and needed to run through several different regression models for each of nearly 50 outcomes. The data were ruthlessly reduced to the minimum needed in the initial stages, then we used "high performance cluster" computing to run the many models in parallel, turning days or weeks of analysis into a few hours' processing.
One of the most striking things about our results was that BMI was related in some way to almost every cause of death category, the only exception being transport-related accidents. Excess weight was associated with a higher risk of deaths in every other category except for mental-health related and neurological deaths, while low body weight was associated with deaths from every category except for liver cirrhosis. One question that arises is: to what extent did the striking increases in suicide and neurological deaths at low BMIs represent cause and effect? We took several steps to try and exclude reverse causality, but this is a challenge because mental health and neurological disorders could plausibly cause appetite suppression and thus BMI reduction over many years. We would like to look into this more, perhaps employing a Mendelian randomisation design to further distil the true causal effects.
Dr Krishnan Bhaskaran, Associate Professor, London School of Hygiene and Tropical Medicine, London, UK