Experimental design matters for statistical analysis: how to handle blocking

Nowadays, evaluation of the effects of pesticides often relies on experimental designs that involve multiple concentrations of the pesticide of interest or multiple pesticides at specific comparable concentrations and, possibly, secondary factors of interest. Unfortunately, the experimental design is often more or less neglected when analysing data. Two data examples were analysed using different modelling strategies. First, in a randomized complete block design, mean heights of maize treated with a herbicide and one of several adjuvants were compared. Second, translocation of an insecticide applied to maize as a seed treatment was evaluated using incomplete data from an unbalanced design with several layers of hierarchical sampling. Extensive simulations were carried out to further substantiate the effects of different modelling strategies.

It was shown that results from suboptimal approaches (two-sample t-tests and ordinary ANOVA assuming independent observations) may be both quantitatively and qualitatively different from the results obtained using an appropriate linear mixed model. The simulations demonstrated that the different approaches may lead to differences in coverage percentages of confidence intervals and type 1 error rates, confirming that misleading conclusions can easily happen when an inappropriate statistical approach is chosen.

To ensure that experimental data are summarized appropriately, avoiding misleading conclusions, the experimental design should duly be reflected in the choice of statistical approaches and models. We recommend that author guidelines should explicitly point out that authors need to indicate how the statistical analysis reflects the experimental design. © 2017 Society of Chemical Industry.