The importance of model choice on pH inferences from scaled chrysophyte assemblages in North America

Abstract

The selection of a reliable inference model is a crucial step in developing ecologically sound reconstructions of environmental variables in the past. We compared intra- and inter-regional regression-based models, and an inter-regional Modern Analogue Technique (MAT) model in their ability to infer lakewater pH from scaled chrysophyte assemblages. The performance of each model was assessed by examining cross-validated coefficients of determination and prediction errors, and through reconstructing the pH of ≈50 modern and fossil samples in south-central Ontario, Canada. Using the intra- and inter-regional data sets, we found little difference in the ability of the regression-based models to infer present-day pH. Partial Least Squares (PLS) regression, Weighted Averaging (WA), and Weighted Averaging Partial Least Squares (WA-PLS) inference models showed similar values for jack-knifed coefficients of determination (r2jack), root mean squared errors of prediction (RMSEPjack), and mean and maximum biases. Based on an analogue matching approach, the inferred values from 48 fossil sediment samples suggested that the intra-regional model did not provide reliable reconstructions for approximately half of the fossil samples. However, inferences from the inter-regional MAT and regression-based models were found to have appropriate analogues and thus considered to be more reliable.