• generalized linear models;
  • rainfall;
  • stochastic models;
  • climate change;
  • Botswana;
  • ENSO;
  • imputation;
  • missing data


The application of spatial-temporal stochastic rainfall models to semi-arid or arid areas is expected to be particularly challenging because of the high variability of rainfall, sparse rain gauge networks with significant periods of missing rainfall and potential data quality issues. In this article, a generalized linear model (GLM) has been fitted to daily rainfall data from the period 1975–1999 for 13 gauges in a 7660 km2 sub-basin of the Limpopo basin in Botswana, with the objective of exploring applicability of the GLM for infilling historic records and for climate change analysis. Several relevant statistics of rainfall space-time variability were used to analyse model performance, including use of an independent validation period and sites that were not used in the fitting. The GLM was considered to simulate rainfall adequately for the purpose of sub-basin-scale water resource studies, although the model uncertainty is high. The main factors affecting rainfall space-time variability were found to be seasonality, autocorrelation of daily rainfall, altitude, latitude and longitude. Addition of large-scale drivers of rainfall (pressure, humidity and temperature) further improved representation of inter-annual variability, and this link to large-scale climate potentially facilitates downscaling of global climate model outputs. Although the model was locally sensitive to data quality issues, there was no evidence that these issues affected sub-basin scale analysis. Copyright © 2011 Royal Meteorological Society