The value of conditional prediction: Can retained species help predict unknown discards in commercial fisheries?

Abstract Species distribution models often predict each species independently based on environmental conditions. In contrast, conditional prediction is an approach that uses information on the known presence or abundance of some species to help predict the distribution of other species in the same assemblage. Despite its potential, conditional prediction has had limited evaluation in real‐world applications. An ideal application is predicting unobserved discarded catch in commercial fisheries, where the discarded and unrecorded portion might be better predicted by incorporating information on the retained and recorded portion. We evaluated conditional prediction for estimating discarding in a trawl fishery using two approaches: stacked species distribution models (SSDMs; using random forests) to compare marginal predictions (based only on environmental variables) with conditional predictions (including known species as predictors), and joint species distribution models (JSDMs; using HMSC and GJAM) to compare joint predictions (modelling species simultaneously) with conditional joint predictions (additionally conditioning on known species occurrences). SSDMs include known species information through additional covariates, while JSDMs capture this information through residual correlations among species. Although marginal and joint predictions showed reasonable performance, conditional and conditional joint predictions added meaningful value only when models had fewer environmental and spatio‐temporal covariates. JSDMs were especially unreliable, with their conditional joint predictions of discarded biomass often less accurate than their joint predictions. This likely resulted from noisy data leading to spurious residual correlations among some species and inflated predictions due to retained species with large abundances contrary to the training data. When predicting a discard time series—such as might be useful for species assessment—the random forest SSDM was most accurate, with marginal predictions having on average 42% less error than a basic ratio estimator, demonstrating the benefit of model‐based discard estimators. Synthesis and applications . Due to their higher performance and reliability, we consider SSDMs valuable for modelling unobserved discards and advise caution when using conditional joint prediction for continuous response variables such as biomass, especially when overdispersion is likely. Our study indicates that identifying and including key environmental variables is more important than incorporating co‐occurring taxa in these kinds of predictive models.