To protect blood donors from anemia, donors with low hemoglobin levels are generally deferred from donating. In this thesis we consider various statistical techniques to gain insight in the trajectory of hemoglobin and to predict future values, which both can be useful to prevent deferrals. We examine the longitudinal association of hemoglobin (Hb) and zinc protoporphyrin (ZPP), a biomarker which is believed to be predictive for future Hb levels. We apply a multivariate autoregressive mixed-effects model, and find that our data suggest that there is not a time-dependent association, but rather a correlation of individual specific average values of Hb and ZPP. In the context of out-of-sample predictions, the usefulness of ZPP as a predictor for future Hb levels seems very limited. In order to successfully predict future Hb levels we examine a variety of methods that can be used for longitudinal forecasting. We propose a hierarchical specification of the mean-reverting Ornstein-Uhlenbeck process model, which matches well with the theoretical properties of the trajectory of Hb levels, and we provide a way to generate dynamic predictions with this model. Furthermore, we implement a Bayesian variable selection technique to incorporate an additional relatively high-dimensional set of blood levels, and we consider two decision tree ensemble methods to capture possible non-linearities in our data. We focus on out-of-sample forecasting and find that (i) the hierarchical OrnsteinUhlenbeck model performs slightly better than traditional mixed-effects models, (ii) the tree based methods seem to be most successful in determining eligibility for donation and (iii) incorporating additional blood levels can improve predictions.

, , , , , , ,
hdl.handle.net/2105/45039
Econometrie
Erasmus School of Economics

Fokkinga, J.J., & Paap, R. (2019, January 3). Modelling hemoglobin levels of blood donors. Econometrie. Retrieved from http://hdl.handle.net/2105/45039