Under this approach, a random sample is taken from the original data with replacement and the model is applied to predict (impute) the original missing values, given the observed data [32].
We did not replace missing values, given that the attrition rate was modest.
The complementary analyses with imputation for missing values gave support to this assumption.
Missing values were given the value "0" (null), provided that information was given on at least two of the three variables.
The numbers of missing values were given in 'no data' categories.
Furthermore, the on-line questionnaires are divided in several sections and a warning against missing values is given to the respondent every time a new section is submitted.
Missing values were given for children with unknown dates of birth and for outlier z-scores: WAZ > 5 or WAZ < −6; HAZ > 6 or HAZ < −6; WHZ > 5 or WHZ < −5 SD.
The Tree based model is more efficient in dealing with missing values and given that the two approaches identify a nearly identical subset of useful explanatory variables, the possibility that the missing values are having an undue influence on the final regression model is remote.
The regression models used to estimate the missing values are therefore giving the exact values that one would expect and do not allow for random variation in the realised values.
In the case of missing values for a given covariate, we created a separate missing category and included this in the models.
