Another limitation is due to missing values that required imputing the presence or absence of neuropsychiatric symptoms for a specific domain.
Thus, new methods to cope with a greater number of variables and that provide accurate prediction and robustness against missing values are required.
These programs were used to format the pedigree files for BOADICEA, validate the data and to estimate missing values as required (see Appendix).
This particular method was chosen due to its significant advantages: (a) dealing with missing values, (b) requiring data scaling, (c) implying a computationally efficient variant of gradient boosting algorithm [27], (d) providing satisfactory results in ML competitions [28] and was successfully used in other studies and domains (see [29, 30]).
We used multiple imputation to replace those missing values, but this requires the strong assumption that the data are missing at random, given the available covariates.
The k-nearest neighbor method [ 26] was used to estimate any missing values, as ARACNE requires a complete expression matrix.
Only the questions on baseline characteristics were selected as required, since missing values were preferred over terminated questionnaires.
Using standardized methods [ 7, 8], we recalibrated raw scores as required, imputed missing values for individual questionnaire items where possible, and transformed scores for each domain to a scale of 0 to 100.
The second step is the single imputation of missing values in background variables which requires an imputation model.
As GeneANOVA requires no missing values, we implemented the program with the SAM imputed data set in which the missed values has been calculated from the mean of 10 non-missing nearest neighbour.
