Adding previous winter rain to a linear model increased mean squared error in 12 plant count predictions on two Portal control plots

We tested whether rain from the previous winter improved a model of annual plant counts in winter censuses at Portal, Arizona. We used 24 censuses from 1991 to 2025, with the same 32 quadrats on two plots that have always been controls. The base linear model used current rain, census date and calendar year. The added previous winter log-rain coefficient was 0.027, with a nominal 95% interval of -0.500 to +0.553. We fit models only to earlier censuses and tested 12 later censuses from 2008 to 2025. The added term increased mean squared error from 2.610 to 3.125 on the log count-density scale. The error ratio was 1.20, with a three-year calendar block interval of 0.97 to 1.53. One test year supplied 61% of the added error. Sample and model checks also gave ratios above one. This fixed rain term did not improve these later predictions. The small sample and wide coefficient interval do not rule out ecological memory. The package preserves source data, sample exclusions, all results, figures, code and independent implementation checks.

desert annual plants · ecological memory · precipitation · Portal Project · temporal prediction · sampled zeros · open data