To ensure that our modelling strategy has accounted for heteroscedasticity due to cluster structure, the residuals from the best model were further analyzed by fitting a residual intercept model (i.e., null model) and a residual random intercept model.
Heterogeneity across studies and dependency between data originating from the same study will be taken into account by fitting a random intercept for each original study.
By fitting a random intercept model for each SNP to the 9 RASav values, it is possible to estimate the variance components attributable to pool construction and the measurement error.
The same estimate is provided by an intercept-only logistic regression model.
Models were fitted in order of forward selection; starting with an intercept-only model that had no predictors, successive predictor variables were added one at a time.
Hence, beforehand, we decided to distinguish a class with change and one without, and we fitted the trend in the stable class with an intercept-only predictor.
For each compound, we tested for heterogeneity by fitting a model with random intercepts and slopes by cohort.
We assessed model fit using a likelihood ratio test (comparing versus an intercept-only model).
For comparison, an intercept-only (or horizontal line) model was also fitted, as this reflects the relationship implied by the traditional approach.
Adding sex to an intercept-only model predicting weight improved model fit (Table A1; B(Male) = 7.32, 95% CI 3.70 10.93).
