After identifying significant risk factors a summary risk index was constructed for planned and unplanned suicide attempts separately for developed and developing countries. We began by combining the predictors that had ORs over 2 or under 0.5 in the earlier bivariate models into new multivariate models. These models included statistical controls for country, age, and age squared in order to ensure that effects were not explained by these factors. The predictors that continued to have ORs above 2 or below 0.5 in the multivariate models were used to create the index by constructing a predicted probability of the outcome from a version of the multivariate model that contained only these predictors. This distribution of predicted probabilities was then inspected and divided into strata using the logic of stratum-specific likelihood ratios (SSLRs).46 That is, we divided the distribution into strata large enough to find statistically significant monotonic relationships between all contiguous pairs of strata. Three measures of screening scale performance were then calculated for each stratum: positive predictive value (PPV; the proportion of the respondents in the stratum who made an
