Patient-reported outcomes (PROs), which refer to patients’ appraisals of their own health and quality of life, are increasingly used to assess the impact of treatment interventions over time. Estimates of intervention effects may be biased by item non-response, which can also result in loss of statistical power. We estimated the bias and mean squared error of the factor scores for full information maximum likelihood, conditional proportional odds multiple imputation and non-negative matrix factorization, which approximately separates a high-dimensional data matrix into two low-dimensional matrices. We used Monte Carlo simulation for longitudinal item response theory to compare ordinal item non-response methods for estimating change in latent factor scores over time. Results were obtained under the assumption of: (1) longitudinal measurement invariance, which implies that the latent variable parameters are constant over time; and (2) different missing data mechanisms with varying proportions of item non-response. The impact of the missingness mechanism and the choice of methods for handling ordinal item non-response on estimated change in latent factor scores will be discussed.