Ecological Archives E088-160-A2

Hélène Fréville, Kevin McConway, Mike Dodd, and Jonathan Silvertown. 2007. Prediction of extinction in plants: interaction of extrinsic threats and life history traits. Ecology 88:2662–2672.

Appendix B. Stepwise modeling procedure.

In all multivariate models, only main effects and two-way interactions between pairs of explanatory (life history or habitat) variables were considered. When a two-way interaction involving a categorical variable was included, the main effect of the variable with which the categorical variable interacted was also included in the model (even if not significant), to satisfy marginality restrictions. By contrast, non-significant main effects of variables that interacted (or appeared as product terms) only with continuous variables were not included, as marginality considerations do not require their inclusion. The modeling process was performed as follows for each time period separately. An initial model involving main effects only was fitted. Product terms, i.e., interactions between continuous variables, and two-way interactions involving categorical variables were added successively to the model until no further improvement of fit could be observed. At this stage, optimal models were modified in three ways. First, any nonsignificant main effects that were not forced into the model because of their presence in significant interactions were removed. Second, we investigated whether adding the squares of the continuous variables in the model would affect the significance of product terms involving these variables. This would indeed be expected if product terms or correlated explanatory variables were acting as proxies for the squares of one or both of the variables included in the product. Finally, random effects whose variance components were not significantly different from zero were removed from the model, in order to improve the accuracy of coefficient estimations for the remaining random and fixed effects. Following this procedure, we obtained different models for the two time periods. We then fitted the whole data set to a third model that included as fixed effects the time period and all the terms that were significant in one or both of the separate models fitted for each period. Time period was included in two-way and three-way interactions with environmental variables and life history traits. Random effects for plots were allowed to differ between time periods, in order to take into account temporal changes in the plot characteristics which may not have been recorded by our threat variables. Nonsignificant terms were then removed one at a time from the overall model until all remaining terms were significant. At this stage, we investigated the relevance of adding squares of continuous variables in the model, and we removed random effects whose variance components were not significantly different from zero.

[Back to E088-160]