Moisture content and water activity relations in honey: A Bayesian multilevel meta-analysis

Research objectives were to investigate i) linear moisture content (m) - water activity (a_w) relations in honeys, ii) quantify variability due to geographical origin, iii) prediction of a_w from m using this quantified variability. Bayesian multilevel modeling was used to reach these objectives using thirty literature data sets. The rate of change in a_w as function of m (the slope) was identical per origin, whereas actual values of a_w at the same m (intercepts) differed. Geographic variation characterized by multilevel modeling (partial pooling) was quantified and compared to single-level modeling with all data pooled (complete pooling) and single-level modeling for each origin (no-pooling). Multilevel modeling results in an overall slope and intercept at the population level but also in individual intercepts per origin, hence variation is characterized at two levels. The obtained multilevel population parameters predicted almost exactly the relation between a_w and m of glucose-fructose solutions resembling honey, confirming that a_w-m relations in honeys are determined by glucose/fructose but not the actual a_w-values themselves. Multilevel modelling, a compromise between over- and underfitting, gives the best prediction of a_w of honeys from m including origin variability. The applied procedure is a blueprint to characterize variation in all types of food.