Losing a species from a community can cause further extinctions, a process also known as coextinction. The risk of being extirpated with an interaction partner is commonly inferred from a species' host-breadth, derived from observing interactions between species. But observational data suffers from imperfect detection, making coextinction estimates highly unreliable. To address this issue and to account for data uncertainty, we fit a hierarchical N-mixture model to individual-level interaction data from a mutualistic network. We predict (1) with how many interaction partners each species interacts (to indicate their coextinction risk) and (2) how completely the community was sampled. We fit the model to simulated interactions to investigate how variation in sampling effort, interaction probability, and animal abundances influence model accuracy and apply it to an empirical dataset of flowering plants and their insect visitors. The model performed well in predicting the number of interaction partners for scenarios with high abundances, but indicated high parameter uncertainty for networks with many rare species. Yet, model predictions were generally closer to the true value than the observations. Our mutualistic plant-insect community most closely resembled the scenario of high interaction rates with low abundances. Median estimates of interaction partners were frequently much higher than the empirical data indicate, but uncertainty was high. Our analysis suggested that we only detected 14-59% of the flower-visiting insect species, indicating that our study design, which is common for pollinator studies, was inadequate to detect many species. Imperfect detection strongly affects the inferences from observed interaction networks and hence, host specificity, specialisation estimates and network metrics from observational data may be highly misleading for assessing a species' coextinction risks. Our study shows how models can help to estimate coextinction risk, but also indicates the need for better data (i.e., intensified sampling and individual-level observations) to reduce uncertainty.