Optimizing the scale of spatial aggregation in minirhizotron studies of crop root system distribution

Research into root system distribution often employs analyses in which depth is treated as a categorical variable. It is not presently known to what extent the choice of strata size affects type I and type II error rates in such analyses, or how to maximize statistical power while controlling for false positives. This research addresses those questions using a simulation study: mixed models were fit to each of one thousand simulated data sets, for 400 treatment combinations associated with differing levels of spatial and temporal autocorrelation, different effect sizes, and different degrees of spatial aggregation. The results show that statistical power declined with increasing degrees of aggregation, especially for small effect sizes and in the presence of spatial autocorrelation. Specifically, in the absence of spatial autocorrelation and with a true effect size of 6, aggregating 80 data points into 4, 20 cm depth class reduced statistical power from a very high initial rate of 0.946 (95% Confidence Interval: 0.935 - 0.955) to the still acceptable rate of 0.855 (0.839 – 0.870), but for an effect size of just one, initial power was already lower, at 0.656 (0.635 – 0.677) when no aggregation was performed and fell to just 0.373 (0.352 – 0.395) upon aggregating to 20 cm depth classes. This pattern is even more pronounced in the presence of spatial autocorrelation. Overall, the study recommends that researchers choosing to employ such an analysis for their minirhizotron data use the smallest computationally feasible depth classes.