Estimating the relative importance of habitat loss (or amount) and fragmentation is difficult because the two processes are highly correlated but it is necessary so that management actions make efficient use of limited conservation resources. We hypothesized that the wide variety of statistical methods used in previous studies may have influenced estimates of their relative importance. After surveying the literature we found that most studies used one of the following methods to estimate relative importance of habitat amount and fragmentation: residual regression, model or variable selection, averaged coefficients from all supported models (ΔAIC < 4), summed Akaike weights from models that include each predictor, classical variance partitioning, or hierarchical variance partitioning. Using simulations and predictors with a realistic covariance structure, we compared these techniques under identical conditions, against each other and against a multiple regression model using all predictors with no adjustments for multicollinearity.

Results/Conclusions

We found that different techniques generated different rankings of the predictors and that some metrics were biased in unexpected ways. Residual regression and variance partitioning approaches were highly biased by the correlations among predictors and also by the direction of a predictor’s effect (i.e., masking or suppressor effects). Overall, our simulations suggest that many efforts to deal with the correlation between amount and fragmentation may have done more harm than good. In an effort to avoid the high variability due to multicollinearity these efforts substitute highly-biased estimates of relative effect strength, which represent the correlation structure of the predictors more than any underlying mechanisms. When extraneous confounding effects are controlled and adequate thought is given to the ecological mechanisms behind modeled predictors, then partial regression coefficients are unbiased estimates of the relative importance of amount and fragmentation, even when predictors are highly correlated.