Probit or Logit? Which is the better model to predict the longevity of seeds?

Abstract

AbstractThe prediction of seed longevity (P50) is traditionally performed by the use of the Probit model. However, due to the fact that the survival data are of binary origin (0,1), the fit of the model can be compromised by the non-normality of the residues. Consequently, this leads to prediction losses, despite the data being partially smoothed by Probit and Logit models. A possibility to reduce the effect of non-normality of the data would be to apply the principles of the central limit theorem, which states that non-normal residues tend to be normal as the n sample is increased. The Logit and Probit models differ in their normal and logistic distribution. Therefore, we developed a new estimation procedure by using a small increase of the n sample and tested it in the Probit and Logit functions to improve the prediction of P50. The results showed that the calculation of P50 by increasing the n samples from 4 to 6 replicates improved the index of correctness of the prediction. The Logit model presented better performance when compared with the Probit model, indicating that the estimation of P50 is more adequate when the adjustment of the data is performed by the Logit function.