Comparison of information-yield from different experimental designs used in algal toxicity testing of chemical mixtures

In controlled laboratory toxicity studies one often encounters situations where several variables might influence the biological response(s). Well-designed multivariate experiments allow for exploration of the effects of several variables simultaneously, and produce maximum information with maximum precision for a given experimental effort, as compared to the one-variable-at-a-time (univariate) approach. Thus, multivariate experimentation usually is the most cost-efficient but is rarely reported in the ecotoxicological literature. The aim of the present study has been to investigate the information-yield and predictive performance of models based on different statistical experimental designs for a given problem. Petroleum refinery wastewaters are complex mixtures. Previous work identified aromatic hydrocarbons and cyanide as the main contributors of a specific refinery effluent to in vivo chlorophyll a fluorescence inhibition of the marine diatom Skeletonema costatum. In the present article the effects of three variables (the aromatic hydrocarbon fraction [A], cyanide [C] and exposure time [T]) were investigated using four different duplicated experimental designs of various resource requirements (two-level factorial [FF(2)], central composite face-centred [CCF], Box-Behnken [BB] and three-level factorial [FF(3)] designs). Partial least squares (PLS) regression and analysis of variance (ANOVA) were applied in the data analysis. The prediction performances of the four models were evaluated by cross-validation and by predictions for new observations. The results show that the multivariate approach gathered all the information relevant to the stated problem in one mathematical model, and that considerable experimental effort can be saved by adopting a sequential modelling approach.