Abstract
Aim Combining camera trap data from multiple surveys has the potential to address large‐scale ecological and conservation questions. But individual surveys can differ substantially in study design and, at the large scale, these surveys typically do not constitute a spatially representative sample. Here, we use camera trap data from Snapshot USA and Snapshot Europe, two near‐continental collaborative initiatives, to explore how large‐scale sampling bias and variation in sampling design affect ecological inference. Location Continental USA and Europe. Time Period 2019–2023. Major Taxa Studied Terrestrial mammals > 300 g. Methods We analysed how environmental variables affected selection of sampled locations using logistic regression; and how they affected aspects of study design (number and spacing of sampling locations, proportion of locations along roads/trails, survey duration) using generalised linear models. We used negative binomial regression to test whether design variables affected species photographic rate, caused omitted variable bias when not included in the model, or affected species relationships with environmental variables. Results Sampled locations in both regions avoided areas high in agriculture and far from roads and preferentially sampled forest (for Europe only at higher elevations). In the USA, sampling further avoided high elevation and precipitation extremes. These same variables also predicted variation in study design, and design variables commonly affected the photographic rates of mammals. Ignoring design variation only rarely caused bias in coefficients of relationships with environmental variables (omitted variable bias), but interactions between design and environmental variables were more common. Main Conclusions For inference beyond the data at hand, researchers must consider spatial sampling bias inherent in multi‐source studies, as is regularly done in the context of citizen science. While ecological inference was often robust to ignoring design variation, researchers should consider its potential to affect, interact with or even mask environmental relationships of interest for their particular data and objectives.