A systematic review and meta-analysis of unimodal and multimodal predation risk assessment in birds

Abstract

Despite a wealth of studies documenting prey responses to perceived predation risk, researchers have only recently begun to consider how prey integrate information from multiple cues in their assessment of risk. We conduct a systematic review and meta-analysis of studies that experimentally manipulated perceived predation risk in birds and evaluate support for three alternative models of cue integration: redundancy/equivalence, enhancement, and antagonism. One key insight from our analysis is that the current theory, generally applied to study cue integration in animals, is incomplete. These theories specify the effects of increasing information level on mean, but not variance, in responses. In contrast, we show that providing multiple complementary cues of predation risk simultaneously does not affect mean response. Instead, as information richness increases, populations appear to assess risk more accurately, resulting in lower among-population variance in response to manipulations of perceived predation risk. We show that this may arise via a statistical process called maximum-likelihood estimation (MLE) integration. Our meta-analysis illustrates how explicit consideration of variance in responses can yield important biological insights.

Fig. 1. Illustration of three types of multimodal cue integration.

We assume that the unimodal cues differ in information quality (i.e., certainty), such that stimulus II has higher certainty and elicits a stronger response on its own compared to stimulus I. A This illustrates signal redundancy (or equivalence), whereby the multimodal stimulus does not increase certainty relative to the higher certainty stimulus (II) on its own. B This illustrates enhancement, where the multimodal stimulus increases certainty relative to either stimulus on their own, thereby eliciting a stronger response. C This illustrates antagonism, whereby the multimodal cue results in a lower estimation of risk than the more certain unimodal cue on its own. Note that any reduction in the response to the multimodal cue relative to the more certain stimulus (II) would be considered antagonism even if it is higher than the response to the lower certainty cue (I).

Fig. 2. Illustration of phylogenetic and geographic breadth of estimates included in meta-analysis.

A Shows the phylogenetic relationships used in the meta-regression, grouped by order and the associated mean effect size for response to manipulations of perceived predation risk for k estimates from K studies. B Shows the geographic distribution of studies, where the colour of the country on a gradient from yellow to red represents the total number of studies (n). Grey is used for countries from which no estimates were obtained. Silhouettes representing different bird orders were obtained from PhyloPics, with artist credits and copyright: T. Michale Keesey (PDM 1.0), Ferran Sayol (CC0 1.0), Gabriela Palomo-Munoz (CC BY-NC 3.0) Andy Wilson (CC0 1.0) and Alexandre Vong (CC0 1.0). Detailed copyright information for all images can be accessed at: https://www.phylopic.org/permalinks/4d2aebec1e2f2da818396c344eb377c61d6ce0d70ddb15d09d7671defdf00ed2.

Fig. 3. Orchard plot of meta-analytic mean effect sizes, standardised mean difference (SMD or Hedge’s g) for each of six treatment levels for experimental manipulations of perceived predation risk: A = acoustic, AV = acoustic + visual, AVO = acoustic + visual + olfactory, O = olfactory, OV = olfactory + visual and V = visual.

The circles denote the meta-analytic means. Note that the black rectangles represent the 95% confidence intervals, while the whiskers denote the 95% prediction intervals. A Shows results from meta-analysis including all treatment levels. B, C Illustrate results from analyses restricted to the three most common cue types (A, V and AV). B Shows estimated effects from homoscedastic model, and (C) shows estimated effects from heteroscedastic model. Total number of estimates (k) is given on to the right of each plot with the number of studies contributing estimates in parentheses.

Fig. 4. Illustration of the effect of significant moderators of the effect of manipulations of perceived predation risk in birds.

Magnitude of response varies as a function of (A) response type (behaviour, life history, or physiology), and (B) declines with increasing treatment duration. However, different treatment durations tend to be associated with different response types as shown in (B), making it difficult to tease apart their effects. In (A), the circles denote the meta-analytic means. Note that the black rectangles represent the 95% confidence intervals, and whiskers denote the 95% prediction intervals. In (B), the regression is plotted with 95% confidence intervals (inner dotted line) and 95% prediction intervals (outer dotted line). Total number of estimates (k) is given on to the right of each plot, with the number of studies contributing estimates in parentheses.

Fig. 5. Assessing publication bias.

A Funnel plot. B Egger regression to assess funnel asymmetry. 95% confidence intervals are depicted by the two outer dotted lines. C Regression to test time lag effect of published effect sizes, with 95% confidence intervals depicted by the two inner dotted lines and 95% prediction intervals depicted by the two outer dotted lines (these are non-linear as the predictions are derived from multi-moderator models). Total number of estimates (k) is given at the top of each plot, with the number of studies contributing estimates in parentheses.

Fig. 6. Illustration of multimodal cue integration under two scenarios.

For example, (A) an acoustic cue provides a lower mean estimate of risk and higher uncertainty/variance (blue dotted line) compared with a visual cue (red dashed line). The estimated risk that integrates both these sources of information using maximum likelihood estimation (MLE) integration will have lower variance than either alone, and the mean will be closer to the mean of the higher certainty unimodal cue (solid black line). B An acoustic cue (blue dotted line) and a visual cue (red dashed line) provide similar means and variances in estimated risk. Under multimodal cue integration using MLE integration (solid black line), the mean estimated risk remains unchanged but has lower variance relative to both unimodal cues.

Fig. 7. Illustration of how maximum likelihood estimation integration (MLE) could result in lower among-study variance in response to manipulations of perceived predation risk when two redundant cues are integrated relative to the among-study variance when either cue type is presented alone.

Each panel illustrates the same five hypothetical populations (shown in five distinct colours). If there is across-study heterogeneity in the probability function associated with study-specific unimodal cues, as shown in (A) (Acoustic) and (B) (Visual), then even if the mean and among-study variance in response to each of the two unimodal cues are identical, maximum likelihood integration will result in lower among-study variance, as shown in (C).