Therefore, among the 1,055 possible participants, 1,000 for Cooperation, 997 for Self - Control, and 1,026 for Assertion were available because of
missing predictor variables or covariates at 2 years of age.
Not exact matches
Several analyses focused on
missing data.36 To explore
missing data patterns, we coded loss to follow - up as a binary
variable and tested baseline
variables as
predictors using a stepwise logistic regression.
To assess the potential effect of
missing data (ie, ignorable vs informative
missing data), a pattern - mixture analysis was implemented using 2 - tailed tests.51 We defined patterns using a binary completer status
variable, which was entered as a
predictor in the RRM and MMANOVA.
We undertook multiple imputation (MI) of
predictor variables to assess the sensitivity of results to
missing data using the chained regression method of MI to generate five imputed data sets.
Baseline drinking status (ever vs never tried alcohol) did not predict attrition, but to account for attrition bias related to other
variables, estimation was carried out after multiple imputation using the standard
missing at random assumption (ie,
missing data are assumed
missing at random conditional on observed
predictors included in the model).27 The imputation model included all the
predictors in the alcohol models plus a number of auxiliary
variables that were not of direct theoretical interest but were nonetheless predictive of missingness so as to improve the quality of the imputations and make the
missing at random assumption more plausible.28
Multivariable regression models were considered, but the modest sample size and patterns of
missing data did not allow for simultaneous consideration of multiple
predictor variables.