Weed biological control programs require extensive testing prior to release to ensure that imported natural enemy species are unlikely to harm nontarget plant species. Typical screening tests confine individuals on the nontarget plant in a laboratory and measure their ability to oviposit or develop. Currently, there is no widely accepted standard for translating laboratory host specificity test results into field risk to nontarget plants. As a guide for decision-making, we developed a database of historic weed biocontrol projects in North America and relate laboratory host specificity test measures to the probability of field use using logistic regression analysis. Using receiver operator characteristic curves, we also examine the tradeoffs between accurately excluding the agents destined to use nontarget plants in the field and mistakenly excluding agents that would actually be safe. Three key findings are that (1) oviposition tests alone appear to be less effective at predicting field use as compared to larval development or full life cycle reproduction tests; (2) an index that combines different laboratory measures of nontarget host use (e.g. choice oviposition and larval development) appears to offer better accuracy for correctly predicting field use; and (3) common methods of host specificity screening are insufficient to eliminate risk without also mistakenly excluding some acceptable agents that would be host specific in the field. These findings can help guide regulatory and permitting decisions in biological control, and they illustrate how biological control records that are routinely collected can be used to predict and manage risk to nontarget plants.