Using Logistic Regression to Analyze the Sensitivity of PVA Models: a Comparison of Methods Based on African Wild Dog Models

Abstract

We used logistic regression as a method of sensitivity analysis for a stochastic population viability analysis model of African wild dogs ( Lycaon pictus ) and compared these results with conventional sensitivity analyses of stochastic and deterministic models. Standardized coefficients from the logistic regression analyses indicated that pup survival explained the most variability in the probability of extinction, regardless of whether or not the model incorporated density dependence. Adult survival and the standard deviation of pup survival were the next most important parameters in density-dependent simulations, whereas the severity and probability of catastrophe were more important during density-independent simulations. The inclusion of density dependence decreased the probability of extinction, but neither the abruptness nor the inclusion of density dependence were important model parameters. Results of both relative sensitivity analyses that altered each parameter by 10% of its range and life-stage-simulation analyses of deterministic matrix models supported the logistic regression results, indicating that pup survival and its variation were more important than other parameters. But both conventional sensitivity analysis of the stochastic model which changed each parameter by 10% of its mean value and elasticity analyses indicated that adult survival was more important than pup survival. We evaluated the advantages and disadvantages of using logistic regression to analyze the sensitivity of stochastic population viability models and conclude that it is a powerful method because it can address interactions among input parameters and can incorporate the range of parameter variability, although the standardized regression coefficients are not comparable between studies. Model structure, method of analysis, and parameter uncertainty affect the conclusions of sensitivity analyses. Therefore, rigorous model exploration and analysis should be conducted to understand model behavior and management implications.Utilización de la Regresión Logística para Analizar la Sensibilidad de Modelos de AVP: una Comparación de Métodos Basados en Modelos de Perros Silvestres Africanos Resumen: Utilizamos la regresión logística como un método de análisis de sensibilidad par a un modelo de análisis de viabilidad poblacional de perros silvestres Africanos ( Lycaon pictus ) y comparamos estos resultados con análisis de sensibilidad convencionales de modelos estocásticos y determinísticos. Coeficientes estandarizados de los análisis de regresión logística indicaron que la supervivencia de cachorros explicaba la mayor variabilidad en la probabilidad de extinción, independientemente de que el modelo incorporara la denso-dependencia. La supervivencia de adultos y la desviación estándar de la supervivencia de cachorros fueron los parámetros que siguieron en importancia en simulaciones de denso-dependencia, mientras que la severidad y la probabilidad de catástrofes fueron más importantes durante simulaciones denso-independientes. La inclu...

Using Latent Selection Difference to Model Persistence in a Declining Population

“…In a hypothetical system, Cross and Beissinger [5] used logistic regression to uncover mechanisms behind extinction or persistence among simulated PVA models for African wild dogs ( Lycaon pictus ). In this case, some model populations persisted and others did not, providing a binary classification appropriately analyzed using logistic regression.…”

Metamodels for Transdisciplinary Analysis of Wildlife Population Dynamics

“…One approach for sensitivity testing of many potentially interacting variables is to generate a large number of scenarios, each with parameters sampled from the distributions describing the ranges of uncertainty. Regression analyses can then be used to test the impact of each factor (including interactions) on model predictions [62–64]. …”

Accounting for Management Costs in Sensitivity Analyses of Matrix Population Models

“…Although our analytic method requires relatively simplistic matrices to produce amenable analytical solutions, the procedure can also be followed numerically if the vital rates and the relative costs of managing them can be estimated, as seen above. Indeed, any of the range of methods of calculating sensitivity coefficients (e.g., McCarthy et al 1995; Cross & Beissinger 2001) could be modified to incorporate management costs by dividing the coefficients by the marginal costs of management. In a similar vein, the method can be adapted to pursue objectives other than the maximization of a deterministic population's growth rate.…”

“…Biological (Caswell 2000) and mathematical (Benton & Grant 1999) constraints on parameter values that describe vital rates can lead to a trade‐off between elasticity values and parameter variation, with parameters yielding high elasticity values often having the least range of natural or management‐induced variation (Cross & Beissinger 2001 and references therein). For most long‐lived species the dominant eigenvalue is most sensitive to changes in the adult survival rate (e.g., Heppell et al 1996; Fisher et al 2000; Hunt 2001; Hebblewhite et al 2003).…”