Data analysis
All analyses were conducted in R 4.2.2 (R Core development team, 2022). Continuous variables were centred prior to analysis, and we used the equivalent sum to zero contrasts approach for categorical variables (Schielzeth, 2010). Centering variables reduces problems that otherwise arise with model averaging when interaction terms are included as predictors (Schielzeth, 2010; Cade, 2015; Tyre, 2017). We modelled relative species richness and Sørensen index of i) trees/shrubs, ii) forbs, iii) graminoids, and iv) climbers using linear mixed effects methods with study ID as a random effect, using the lme4 package (Bates et al., 2015).
In all cases models had Variance Inflation Factors (VIF) less than 10 indicating that results are not markedly impacted by collinearity between predictors (Hair et al., 1992, Craney & Surles, 2002). We also checked for linearity of responses by including square terms and comparing the model fit to equivalent models that only included a linear term. The fit of all models was also checked using model diagnostic plots.
We constructed all possible ecologically realistic models (n = 32;Appendix 5, Table S4 ) given our suite of predictor variables, i.e., time since fire (years; ln transformed), fire type (fixed factor: prescribed/non-prescribed burns), biomes (fixed factor: TSMBF, TSBDF, TSCF, TSGSS & FGS), and protection status (fixed factor: protected/non-protected). We included interaction terms between each of our two fire metrics (time since fire, and fire type) and i) biomes, and ii) protection status to test whether biome type or protected area status moderated the relationships between each fire metric and our outcome variables.
We used D2 as a measure of explanatory capacity; D2 = (ND − RD)/ND where ND is the null deviance and RD is the residual deviance, which cannot be explained by the model, thus ‘ND–RD’ is the explained deviance. D2 varies between zero and one and equals one when the deviance can be explained completely by the model (Guisan & Zimmermann, 2000).
We used an information-theoretic criterion approach to obtain a set of models whose Δ AICc values were within two points of the best performing model, i.e., that with the lowest AICc value, and then conducted model averaging (Burnham & Anderson, 2004).