Recent river resoration projects have removed dams that have isolated some populations of alewife, a native Connecticut fish. Restored rivers allow migration of alewife to and from the ocean, making secondary contact between the landlocked and ocean-living alewife possible. This could lead to hybridization of the two separate life history forms of fish, but many barriers remain including the possibility that they spawn at different times of the year. We constructed spawning time distributions from field observations collected over 3 years in 5 lakes in Connecticut and fit time-to-spawning (survival analysis) models to the data. We found that the fish in landlocked lakes, with no connections to the ocean, spawn later in the summer. However the large variation in spawning date creates a short overlap between the two populations, allowing for the possibility of hybridization as more rivers are restored and more lakes are once again connected to the ocean.

Complex population processes may require equally complex models, which can lead to analytically intractable estimation problems. Approximate Bayesian computation (ABC) is a computational tool for parameter estimation in situations where likelihoods cannot be computed. Instead of using likelihoods, ABC methods quantify the similarities between an observed data set and repeated simulations from a model. A practical obstacle to implementing an ABC algorithm is selecting summary statistics and distance metrics that accurately capture the main features of the data. We demonstrate the application of a sequential Monte Carlo ABC sampler (ABC SMC) to parameter estimation of a general stochastic stage‐structured population model with ongoing reproduction and heterogeneity in development and mortality. Individual variation in demographic traits has considerable consequences for population dynamics in many systems, but including it in a population model by explicitly allowing stage durations to follow a realistic distribution creates a complex model. We applied the ABC SMC to fit the model to a simulated representative data set with known underlying parameters to evaluate the performance of the algorithm. We also introduced a systematic method for selecting summary statistics and distance metrics, using simulated data and receiver operating characteristic (ROC) curves from classification theory. Evaluations suggest that the approach is promising for model inference in our example of realistic stage‐structured population models.