On the robustness of N-mixture models

Area of Science:

• Ecology
• Wildlife population monitoring
• Statistical modeling

Background:

• N-mixture models offer an alternative to mark-recapture for estimating population size and detection probability using count data.
• These models do not require individual animal identification, making them appealing for certain wildlife studies.
• However, N-mixture models are known to be sensitive to their underlying assumptions.

Purpose of the Study:

• To investigate the impact of common assumption violations on N-mixture model inference.
• To quantify the bias introduced by specific, realistic violations of N-mixture model assumptions.
• To provide guidance on data collection for robust population size estimation.

Main Methods:

• Simulated count data under scenarios with violations of N-mixture model assumptions.
• Evaluated effects of double counting, unmodeled temporal variation in population size, and unmodeled temporal variation in detection probability.
• Assessed the detectability of these violations using goodness-of-fit tests.

Main Results:

• Small, realistic violations of N-mixture model assumptions can lead to substantial biases in population size and detection probability estimates.
• The considered violations were often qualitatively small and difficult to detect with standard goodness-of-fit tests.
• Bias in estimates was pronounced even with minor deviations from model assumptions.

Conclusions:

• N-mixture models require careful consideration of their assumptions due to high sensitivity to violations.
• Standard goodness-of-fit tests may fail to detect assumption violations that cause significant estimation bias.
• For reliable population size estimates, especially in critical applications, collecting additional data like marked recaptures for direct detection probability estimation is recommended.