were lowest for series 3 models, which used weather data
on the day of disease assessment and the previous 24 h
period; both continent-wise and cross-continent ANN
models showed this consistent trend (Fig. 3). Prediction
errors were generally similar for series 1 and 2 models that
used either average weather during the week prior to disease
assessment (26·6–81·8% for ANN; 29·4–81·8% for
REG models) or weather for the seventh day before
assessment (26·6–72·7% for ANN; 36–72·7% for REG
models). Due to some missing weather data, the number
of days with disease and weather data varied slightly for
the three series of models. However, these differences were
relatively small and may not have greatly influenced
prediction errors (Fig. 3). All further work therefore, concentrated
on series 3 models.