Recent concern over global amphibian population declines has highlighted a need for more extensive, rigorous monitoring programs. Two sources of variation, spatial variation and variation in detection probability, make the design and implementation of effective monitoring programs difficult. We used Pollock's robust design in a 3-year capture–recapture study to estimate detection probability and temporary emigration for Plethodon salamanders in Great Smoky Mountains National Park (Tennessee/North Carolina), USA. We used 12 competing models to determine the importance of temporary emigration, and we explored temporal and behavioral effects on conditional capture probabilities. The top 4 models all included random temporary emigration, and Akaike model weights indicated that this parameter was the most important. Models that contained behavioral effects in capture probabilities were selected more often than models with equal capture probabilities for marked and previously unmarked individuals. The “best” model contained random emigration and behavioral effects and was selected 4 times as often as any other model. When we included Markovian emigration, the probability of emigrating from the surface usually was less than the probability of remaining an emigrant (73% of site-years). Markovian emigration estimates often were similar and always had overlapping confidence intervals, thus the Markovian model rarely was chosen over the random emigration models (only 9.6% of site-years). Our study is the first to formally estimate temporary emigration in terrestrial salamander populations, and our results verify that significant proportions of terrestrial salamander populations are subterranean. We determined that the probability of capturing salamanders on the surface may also vary temporally within a sampling season. Therefore, we caution against using unadjusted count indices to compare salamander populations over time or space unless detection probabilities are estimated. Temporary emigration models will improve abundance estimates when a large proportion of the population is unavailable for capture during a given sampling period.