Oecologia (1994) 100:475-480 9 Springer Verlag 1994 Gary E. Belovsky How good must models and data be in ecology? Received: 3 June 1994/Accepted: 24 August 1994 Abstract Linear programming models of diet selection (LP) have been criticized as being too sensitive to varia- tions in parameter values that have not been or may not be able to be measured with a high degree of precision (small standard error). Therefore, LP's predictions have been questioned, even though the predicted diet choices agree very well with observations in 400 published tests. The philosophical and statistical aspects of this criticism of LP are reviewed in light of the ability to test any non- trivial ecological theory. It is argued that measures of er- ror in field data may not meet simple statistical defini- tions, and thereby, may make sensitivity analyses that use the error measures overly conservative. Furthermore, the important issue in testing ecological theory may not be the statistical confidence in a single test, but whether or not the theory withstands repeated tests. Key words Optimal foraging 9 Linear Programming Philosophy 9 Sensitivity analysis 9 Modeling Introduction Huggard (1994) is critical of my linear programming model of optimal diet selection (hereafter called LP), be- cause it is too sensitive to variation in input values to take concordance between predicted and observed diets seriously; this is an expansion of an earlier criticism (Hobbs 1990). Both criticisms were developed using a portion of the data presented in one attempt to validate LP (Belovsky 1986a). The issue of model sensitivity is important, and Huggard's criticisms provide a good con- text to review statistical and philosophical consider- ations. Assessing an ecological model's validity is a problem fraught with difficulties. This issue is not restricted to G.E. Belovsky Ecology Center and Departmentof Fisheries and Wildlife, Utah State University, Logan, UT 84322-5210, USA model validation, but applies to any hypothesis, because a model is an explicit mathematical statement of a hy- pothesis. Models hopefully reduce the likelihood of al- ternative hypotheses producing similar predictions and increase the likelihood of falsification, as compared with qualitative hypotheses. A problem for model validation is posed by model sensitivity (variability predictions, given variation in input parameters). In more controlled, laboratory sciences, variation in input parameters is equated with measurement error and model sensitivity reflects the potential for error to pro- vide concordance between predictions and observed val- ues (Type I error). However, ecological, especially field, data are notoriously variable, in part due to measurement error, but also due to uncontrollable environmental and individual heterogeneity (e.g., days differ in weather, lo- cations within a habitat differ, individuals differ in size). In addition, an ecological model's input parameters are often computed by combining several different measure- ments; the variances of computed parameters are not measured but estimated, and this can be problematic (Travis 1982). These modeling, statistical, and philo- sophical issues are reviewed in the context of Huggard's criticisms of LR Modeling issues Monte Carlo simulation is an appropriate tool for estimat- ing model sensitivity, but the simulation must be correctly constructed and compared with observed data; Huggard's analysis has several flaws. The first flaw is presented by Huggard as a footnote in his Table 2. Two species (Dissosteira carolina and Bison bison) are excluded from the analysis, because these species' predicted diets did not vary given their standard errors (SEs). Both species are predicted to consume only grass and their observed diets are largely grass, which "anchors" the regression be- tween predicted and observed diets in the region of high grass intake. Therefore, when Huggard discards 2 of 14 comparisons because they are insensitive, he biases his