Comparison of methods for imputing ordinal data using multivariate normal imputation: a case study of non‐linear effects in a large cohort study