For predictor variables, thin-plate splines were used, initiated with a dimensional basis of two, thus ensuring simple response curves.
For predictor pairs with a correlation coefficient of ≥|0.8|, we retained for modeling only those variables that influenced the known physiological constraints on the species' distributions based on previous studies [30], [33], [38], [40], [58], [59].
For predictor pairs with Pearson r ≥0.9, we only retained one of the variables for the modeling [48] by selecting the variable with the strongest biological interpretability and the smallest correlation to the other predictor variables (Tables 1, 2).
For predictor variables that were categorical in nature (geographic location and setting of the study, zinc salt used, co-intervention used, and adequacy of blinding procedures) we used subgroup meta-analyses.
For predictor variables, odds ratios (ORs) and 95% confidence intervals (CIs) were calculated.
For predictor variables with more than two levels (e.g., clone), significance tests were based upon assessment of improvement in model fit using a χ2 distribution.
For predictor variables in the indoor endotoxin models, we found insufficient variability across the 12 homes for the more refined categories used in the personal models (small cells).
For predictor sets, we created 10 data sets each containing one randomly selected spot sample per child (i.e., 25 observations total per data set).
Table 1 Descriptive statistics for predictor variables Mean Std.
Similar to this, the Mann–Whitney test (U) was used for predictor variables with two categories.
These latter studies all employed active sensor (radar or ALS) imagery as the single source for predictor variables.
