Missing BMI

[eribul]

I would like to know the level of missing BMI data especially in the NJR as historically, BMI data has been poorly completed and up to 50% of BMI data can be missing. If so, has imputation modelling been used to mitigate against such missing data in the statistical modelling.

[eribul]

Review: The reason why we reported missing numbers for BMI in Sweden/SHAR but not in England and Wales/NJR (figure 1), was that BMI was not included in the statistical prediction model (although considered before variable selection in the derivation cohort). Hence, the proportion of missing BMI in SHAR was 3,096 of 57,923 patients (5,3 %) and 55 patients with BMI > 50 were also excluded (0.08 %), since those measures were considered suspiciously high. We understand, however, that the difference between obesity due to reported BMI above 30, or diagnosed obesity due to the ICD-10-code E66, could be emphasized further. We have therefore made sure that the condition used in the statistical model is always referred to as "diagnosed obesity" (line X and tables).

Imputation is a good alternative assuming that data is "missing (completely) at random". This might not be the case for BMI in SHAR, however, which is sometimes measured exactly at the hospital, sometimes estimated by the physician, and sometimes self reported by the patient (different settings in different hospitals). We have anecdotal evidence that patients with high BMI tend to under-report their values if asked to. Also low BMI might be assumed less relevant to record precisely, due to perceived less relevance for the planned procedure. Taken together, we can not preclude association between the missing values and the probability to be missing. We thus have data "missing not at random", which, unfortunately, disqualifies the use of imputation.