Fossil Chironomidae (Insecta: Diptera) as quantitative indicators of past salinity in African lakes

We surveyed sub-fossil chironomid assemblages in surface sediments of 73 low- to mid-elevation lakes in tropical East Africa (Uganda, Kenya, Tanzania, Ethiopia) to develop inference models for quantitative paleosalinity reconstruction. Using a calibration data set of 67 lakes with surface-water conductivity between 34 and 68,800 mu S/cm, trial models based on partial least squares (PLS), weighted-averaging (WA), weighted-averaging partial least squares (WA-PLS), maximum likelihood (ML), and the weighted modern analogue technique (WMAT) produced jack-knifed coefficients of determination (r(2)) between 0.83 and 0.87, and root-mean-squared errors of prediction (RMSEP) between 0.27 and 0.31 log(10) conductivity units, values indicating that fossil assemblages of African Chironomidae can be valuable indicators of past salinity change. The new inference models improve on previous models, which were calibrated with presence-absence data from live collections, by the much greater information content of the calibration data set, and greater probability of finding good modern analogues for fossil assemblages. However, inferences still suffered to a greater (WA, WMAT) or lesser (WA-PLS, PLS and ML) extent from weak correlation between chironomid species distribution and salinity in a broad range of fresh waters, and apparent threshold response of African chironomid communities to salinity change near 3000 mu S/cm. To improve model sensitivity in freshwater lakes we expanded the calibration data set with 11 dilute (6-61 mu S/cm) high-elevation lakes on Mt. Kenya (Kenya) and the Ruwenzori Nits. (Uganda). This did not appreciably improve models' error statistics, in part because it introduced a secondary environmental gradient to the faunal data, probably temperature.To evaluate whether a chironomid-based salinity inference model calibrated in East African lakes could be meaningfully used for environmental reconstruction elsewhere on the continent, we expanded the calibration data set with 8 fresh (15-168 mu S/cm) lakes in Cameroon, West Africa, and one hypersaline desert lake in Chad. This experiment yielded poorer error statistics, primarily because the need to amalgamate East and West African sister taxa reduced overall taxonomic resolution and increased the mean tolerance range of retained taxa. However, the merged data set constrained better the salinity optimum of several freshwater taxa, and further increased the probability of finding good modern analogues.We then used chironomid stratigraphic data and independent proxy reconstructions from two fluctuating lakes in Kenya to compare the performance of new and previous African salinity-inference models. This analysis revealed significant differences between the various numerical techniques in reconstructed salinity trends through time, due to their different sensitivity to the presence or relative abundance of certain key taxa, combined with the above-mentioned threshold faunal response to salinity change. Simple WA and WMAT produced ecologically sensible reconstructions because their step-like change in inferred conductivity near 3000 mu S/cm mirrors the relatively rapid transitions between fresh and saline lake phases associated with climate-driven lake-level change in shallow tropical closed-basin lakes. Statistical camouflaging of this threshold faunal response in WA-PLS and ML models resulted in less trustworthy reconstructions of past salinity in lakes crossing the freshwater-saline boundary. We conclude that selection of a particular inference model should not only be based on statistical performance measures, but consider chironomid community ecology in the study region, and the amplitude of reconstructed environmental change relative to the modern environmental gradient represented in the calibration data set.