Specific inclusions

[eribul]

Can you delineate the specific inclusions for diseases of the CNS. And cutoff for diagnosis of obesity? (BMI > 30?) If you have BMI data it seems unwise to rely on the diagnosis of obesity, rather than including the specific patient level data either as a dichotomous variable or as a linear BMI variable

[eribul]

Review: The definition of each comorbidity (including CNS disease and obesity) is presented in table 1. We have now expanded the caption with a reference for the individual ICD-10-codes underlying each sub-category.

It is seen from the table that obesity was equivalent to obesity as defined by Elixhauser et al. (codified by Quan et al, 2005). This corresponds to patients with ICD-10 = E66 (as also specified on line 252), thus with "Drug-induced obesity, Morbid (severe) obesity with alveolar hypoventilation, Overweight, Other obesity, Obesity, unspecified. It would be possible to provide such details in the text but if so, similar details should also be provided for at least cancer and kidney disease as well, since those commodities were also used in the final model. Cancer alone (malignancy and metastatic solid tumours according to Charlson or lymphoma, metastatic cancer and solid tumours according to Elixhauser, as defined in table 1), entails more than 1,000 individual ICD-10-CM codes. We fear that descriptions with this level of details would be of limited use to most readers, although we hope that table 1 (with its improved caption) might provide enough details for the ones with a true interest.

We have tried to describe in the text that both BMI, as well as diagnosed obesity, were included as potential predictors in the model. Only the latter, however, was included in the final model after variable selection using LASSO-regression and repeated bootstrap samples. This is an empirical result, which indicates that the inclusion of diagnosed obesity in the model, does improve predictive power of the model, whereas BMI (as a linear variable) did not. This is the reason we include diagnosed obesity and not BMI. A possible confusion regarding this matter is the distinction between traditional inference modeling and prediction modeling. The first setting would value interpret-ability more, and the best understanding of association on group level might very well be provided through the study of BMI rather than diagnosed obesity. This is not the case in prediction modeling, with the goal to filter out the most important variables but where causal relations among predictors and outcome might be less intuitive.