Over a 3-year period, we assessed 2 sampling designs for estimating grizzly bear (Ursus arctos) population size using DNA capture–mark–recapture methods on a population of bears that included radiomarked individuals. We compared a large-scale design (with 8 × 8-km grid cells and sites moved for 4 sessions) and a small-scale design (5 × 5-km grid cells with sites not moved for 5 sessions) for closure violation, capture-probability variation, and estimate precision. We used joint telemetry/capture–mark–recapture (JTMR) analysis and traditional closure tests to analyze the capture–mark–recapture data with each design. A simulation study compared the performance of each design for robustness to heterogeneity bias caused by reduced capture probabilities of cubs. Our results suggested that the 5 × 5-km grid cell design was more precise and more robust to potential sample biases, but the risk of closure violation due to smaller overall grid size was greater. No design exhibited complete closure as estimated by JTMR. The results of simulation studies suggested that CAPTURE heterogeneity models are relatively robust to probable forms of capture-probability variation when capture probabilities are >0.2. Only the 5 × 5-km designs exhibited this capture-probability level, suggesting that this design is preferred to ensure estimator robustness when population size is <100. The power of the CAPTURE model selection routine to detect capture probability variation was low regardless of sampling design used. Our study illustrated the trade-off between intensive sampling to ensure robustness and adequate precision of estimators while being extensive enough to avoid closure violation bias.