A meta-analysis of preference–performance relationships in phytophagous insects

Authors


  • Present address: Tomas Roslin, Department of Applied Biology, PO Box 27 (Latokartanonkaari 5), FI-00014 University of Helsinki, Finland

* E-mail: sofia.gripenberg@zoo.ox.ac.uk

Abstract

Ecology Letters (2010) 13: 383–393

Abstract

The extent to which behavioural choices reflect fine-tuned evolutionary adaptation remains an open debate. For herbivorous insects, the preference–performance hypothesis (PPH) states that female insects will evolve to oviposit on hosts on which their offspring fare best. In this study, we use meta-analysis to assess the balance of evidence for and against the PPH, and to evaluate the role of individual factors proposed to influence host selection by female insects. We do so in an explicitly bitrophic context (herbivores versus plants). Overall, our analyses offer clear support for the PPH: Offspring survive better on preferred plant types, and females lay more eggs on plant types conducive to offspring performance. We also found evidence for an effect of diet breadth on host choice: female preference for ‘good quality plants’ was stronger in oligophagous insects than in polyphagous insects. Nonetheless, despite the large numbers of preference–performance studies conducted to date, sample sizes in our meta-analysis are low due to the inconsistent format used by authors to present their results. To improve the situation, we invite authors to contribute to the data base emerging from this work, with the aim of reaching a strengthened synthesis of the subject field.

Introduction

Although behavioural and evolutionary ecologists have long been studying the principles dictating resource use by heterotrophs, the extent to which behavioural choices reflect fine-tuned evolutionary adaptation remains an open debate (e.g. Jaenike 1978; Pyke 1984; Charnov & Stephens 1988; Mangel 1989; West & Sheldon 2002; Shuker & West 2004; Burton-Chellew et al. 2008). In the context of plant–insect interactions, host selection by ovipositing females offers a central theme (Bernays & Chapman 1994; Schoonhoven et al. 2005; Tilmon 2008). Here, the range of patterns detected has stimulated a wide range of theories on forces moulding female choice with respect to offspring performance (reviewed by e.g. Thompson 1988; Courtney & Kibota 1990; Jaenike 1990; Mayhew 1997; Craig & Itami 2008).

The basic premise: female preference reflects offspring performance

From an evolutionary perspective, the basic setting for host plant choice is a simple one: given that plants differ in their suitability as food for insects, and that the immature stages of developing insects are often rather immobile, natural selection should favour females with an ability to discriminate between hosts of different suitability for larval development (Levins & MacArthur 1969; Jaenike 1978; Thompson 1988; Mayhew 1997). According to the preference–performance hypothesis (also known as the ‘naïve adaptionist hypothesis’ or the ‘mother-knows-best hypothesis’), females will then maximize their fitness by laying their eggs on plant types on which their offspring perform the best. This is our first, general hypothesis (H1; cf. Table 1) – that female preference reflects offspring performance, as dictated by the food quality of the host plant. From this, we also derive our first testable prediction: that – setting any effects of natural enemies and competitors on performance aside – offspring will perform better on plants chosen by females (prediction P1.1, Table 1).

Table 1.   Hypotheses tested in this study along with specific predictions
HypothesisPrediction
H1. Female preference reflects offspring performanceP1.1 Offspring performance higher on preferred plants
H2. Preference–performance coupling is modified by ecological and/or life history factorsP2.1 Coupling tighter with narrower diet
P2.2 Coupling tighter on woody than on herbaceous plants
P2.3 Coupling tighter for sessile than for mobile offspring
P2.4 Coupling tighter for gregarious than for solitary offspring
P2.5 Coupling tighter for non-feeding than for feeding adults
P2.6 Coupling tighter across than within plant species

Numerous studies have empirically assessed the relationship between female preference and offspring performance. In some cases, female choice seems clearly adaptive, with offspring performance being high on plant types preferred by the females. Nevertheless, tens of studies report cases where female preference and performance appear uncoupled, or where the relationship is surprisingly weak (e.g. Rausher 1979; Valladares & Lawton 1991; Underwood 1994; Fritz et al. 2000; Faria & Fernandes 2001). What factors may explain such apparent contradictions?

The plot thickens: the strength of the preference–performance relationship is modified by ecological and/or life-history factors

Several evolutionary and ecological considerations have been proposed to explain apparent mismatches between choice and performance (Thompson 1988; Courtney & Kibota 1990; Thompson & Pellmyr 1991; Mayhew 1997; Craig & Itami 2008). Our second general hypothesis (H2; Table 1) is then that the strength of the general preference–performance relationship is modified by ecological and/or life-history factors. To specify this general notion into testable predictions, we will next pool factors proposed to cause deviations from the preference–performance association into two groups, depending on whether (1) there is a true lack of preference–performance link and female choice is not adaptive, or (2) there is no link between female preference and host plant quality from the offspring’s perspective, but female choice is still adaptive in other ways. We will then identify factors likely to accentuate a positive coupling between preference and performance. For each of these considerations, we derive a specific prediction (P2.1–P2.6; cf. Table 1) which we test by meta-analysis. In this context, we note that the literature offers many further predictions (see the above-mentioned review papers) – but for the sake of conciseness, we have restricted our survey to ideas for which we have been able to obtain quantitative data for preliminary assessment.

Female choice is non-adaptive and there is a true lack of preference–performance coupling

As evolutionary change will always start with the raw material provided by past evolution, and as organisms evolve as wholes rather than as independent parts, several factors may constrain evolution from reaching an optimal solution (Gould & Lewontin 1979). In the context of female choice, several constraints may emerge from limitations on the information processing capacity of the insect nervous system (e.g. Levins & MacArthur 1969; Fox & Lalonde 1993; Bernays 1998; Janz 2003; Egan & Funk 2006). If selecting between plants becomes more complex the more characteristics they differ in, then one may predict that the link between female preference and offspring performance varies with the degree of diet specialization (prediction P2.1, Table 1). Monophagous insects would then be predicted to be better decision-makers than insects with a broader diet, because appropriate decisions would seem harder to make for females faced with multiple stimuli. A similar relationship between diet breadth and choosiness might arise from several additional physiological mechanisms and morphological characters, allowing taxa with fewer hosts to be more ‘finely tuned’ to them, as opposed to generalized adaptations or plasticity expected in polyphages (e.g Roslin & Salminen 2008 with references therein).

Even in cases where insects are not trapped by evolutionary or physiological constraints, preference and performance may not necessarily be linked. For example, as it may take many generations for evolution to filter out females that choose poor quality hosts, the preference–performance link can be weak (or even negative) where insects interact with novel plant species (Chew 1977; Thompson 1988). As another example, if the relative quality of host plant types varies unpredictably in time, females may fail to evolve an ability to choose the plant type that would be most suitable for their developing offspring (Cronin et al. 2001; Gripenberg et al. 2007a). If female choice is constrained by temporal variation in host plant quality, we may predict that females may find it particularly difficult to assess the quality of herbaceous plants (Craig & Itami 2008), because they are thought to change more during a growing season than are woody plants (but see Salminen et al. 2004; Gripenberg et al. 2007b). We test this prediction as P2.2 (Table 1).

Female behaviour may be adaptive although it is not manifested in selecting for food quality

Most studies assessing the relationship between female preference and offspring performance do so by asking whether females rank their host plants based on their food quality for offspring. Nevertheless, neither female choice nor offspring performance takes place in isolation, but in ecological contexts rendered more complex by an unequal distribution of host plants, microclimatic conditions, mutualists, competitors and/or natural enemies. While many studies explicitly assume that females base their choice on the quality of the plants, the distribution of other taxa may also have a profound impact on both female preference and offspring performance (e.g. Lima & Dill 1990; Nomikou et al. 2003; Heisswolf et al. 2005; Van Mele et al. 2009). Thus, a plant may be only marginally suitable as food, but provide a safe haven from natural enemies (Björkman et al. 1997) or competitors (Wise & Weinberg 2002), or a source of valuable mutualists (Atsatt 1981) – and thus be highly preferred for adaptive reasons. In this context, we note that this study is explicitly focused on bitrophic interactions: on whether variation in plant quality from an offspring’s perspective is reflected in female choice. While this solution is dictated by practical considerations – i.e. by the amount of empirical studies accumulated to date – the extent to which natural enemies and competitors will mould preference–performance relations is an interesting question. From a perspective of quantitative analysis, it will still have to wait until further evidence has been accumulated.

Mechanisms promoting the formation of a positive preference–performance relationship

While the factors discussed above are likely to blur the link between female preference and offspring performance, other ecological traits may favour female choosiness. In particular, limited offspring mobility has been highlighted as a potentially important factor promoting female preference for good quality hosts (Thompson 1988; Craig & Itami 2008). The extreme case is insects with sessile larvae (like gallers and leaf miners), where larvae are confined to the very resource unit selected by their mother, and where the preference–performance coupling is thus expected to be particularly strong. This generates prediction P2.3 (Table 1): If limits to offspring mobility accentuate the importance of female choice, we expect the strength of preference–performance coupling to be strongest in taxa with sessile larvae.

Another factor hypothesized to favour a strong link between preference and performance is the aggregation of offspring. Selecting high-quality hosts may be more important for females laying their eggs in clutches than for females distributing their eggs singly. In the former case, much might be lost with a single bad decision, whereas the laying of single eggs will be more akin to a risk-spreading strategy (Mangel 1987; Hopper 1999). In that case, a few bad decisions will result in the loss of few offspring only. This equals prediction P2.4 (Table 1): If the aggregation of offspring increases the variance in offspring performance, the association between female preference and offspring performance should be more accentuated in taxa laying their eggs in batches than in taxa laying their eggs singly.

Finally, if insect females have the potential to feed as adults, their fecundity may be less dependent on food resources acquired at previous stages (e.g. Wheeler 1996; Jervis et al. 2008). The potential for such ‘refuelling’ is expected to lessen the importance of larval food resources on adult reproduction and longevity, and hence to weaken the coupling between female preference and offspring performance (P2.5, Table 1). A weaker association between female preference and offspring performance in species where adult females consume no food could also support the notion of a conflict between parent and offspring: if adults and offspring feed on different resources, females may actually increase their fitness by being ‘selfish’ at the expense of the performance of their offspring (e.g. Jaenike 1986; Mayhew 2001; Scheirs & De Bruyn 2002; Janz et al. 2005).

Certain traits of the host plants may also promote a link between female preference and offspring performance. Naturally, the larger the difference in offspring performance on different host types, the more likely are females to evolve an ability to select hosts of high quality (Craig & Itami 2008). As one may intuitively assume that differences in offspring performance are larger between than within plant species, one might expect stronger evidence for the preference–performance hypothesis to emerge from studies assessing preference–performance relationships in an interspecific than an intraspecific setting (P2.6, Table 1).

Aims

Although preference–performance relationships have been assessed in a large number of plant–herbivore interactions, empirical data to support or reject general hypotheses are scattered. In this paper, we aim to assess (1) to what extent the literature provides support for the overall preference–performance hypothesis as played out in a bitrophic setting (H1; Table 1) and (2) the importance of selected biological factors proposed to influence the observed strength of preference–performance relationships (H2; Table 1). Given the rich body of theory and the vast literature on the topic, the ‘broader, systematic approach to testing general hypotheses about the evolution of preference–performance relationships’ recently advocated by Craig & Itami (2008) is admittedly challenging. Consequently, our intention is not to provide an all-encompassing synthesis, but to complement previous narrative reviews on the subject. We focus on the two general and six specific predictions derived from the ideas discussed above (cf. Table 1), and use meta-analysis to establish extant, quantitative support for each of them. For practical reasons, our selection of hypotheses is rather based on the availability of data to test them than on the current status of each hypothesis.

Methods

The relationship between female preference and offspring performance can be seen from two perspectives: one can either assess offspring performance in relation to the choices made by females, or female responses in relation to plants differing in their suitability as hosts for the offspring. The general assumption in the PPH literature is that differences in offspring performance generates a selection pressure resulting in female preference, but still, many studies conducted to date have used an inverse approach: the authors have first established which plant is preferred versus not preferred by the female, then compared offspring performance between these plant types. Importantly, this is not the same as first establishing which plant type the offspring performs best on, then testing whether the female has generated a preference for it – and in fact, the two approaches could yield different conclusions. We therefore performed two types of analyses, using either a measure of offspring performance or of female preference as the response. In the ‘performance analyses’, our aim was to assess whether offspring perform better on host types preferred by ovipositing females than on less preferred host types. In the ‘preference analysis’, we tested whether females are more prone to lay their eggs on host types on which their offspring perform well than on host types of poorer quality. As most studies assess the preference–performance relationship on two – or at most a few – plant types, our effect sizes are based on pairwise comparisons.

Data generation

To survey the literature on preference and performance, we started out from studies collected for a previous narrative review on the subject (Mayhew 1997), along with papers collected for a preliminary meta-analysis on the topic (Parnell 2004). These papers were obtained by searching Entomology Abstracts in Cambridge Scientific Abstracts, Blackwell Synergy, and Science Direct, using the search terms ‘oviposition preference’, ‘oviposition’, ‘host plant preference’, ‘host selection’ and ‘preference–performance’. Based on abstract and title information, 350 papers were reviewed in depth to assess suitability. To include more recent studies, we examined all papers having cited Mayhew (1997) in ISI Web of Science (last accessed 2 September, 2009; n = 140 studies).

For a study to be included in our general data set, it had to contain sufficient information to allow calculation of the relevant effect size (see Performance analyses and Preference analysis), and to allow the ranking of plants in terms of preference and/or performance (see Appendix S1). To avoid problems associated with pseudoreplication, we took a conservative approach and included only one data point per insect species and published study. In other words – in cases where preference and performance were assessed on more than two plant types, we randomly chose one plant pair to be included in our analyses. Information was extracted from both numerical and graphical representations. When the information was graphical (e.g. variation around a mean shown as error bars), the image was digitized and data extracted using the software ImageJ.

Specific criteria for inclusion of studies in individual analyses are given below (Meta-analysis). In addition, we excluded five types of studies which matched our general criteria, but which we found too deviant to describe any general associations between herbivorous insects and their natural host plants. First, studies involving insects feeding on dead plant material (such as logs and stored peas) were excluded. Second, we excluded a number of studies in which female preference for certain plant types had been inferred from adult feeding preferences rather than from information on oviposition behaviour. Third, we excluded cases where the host plant species upon which preference and performance were tested was derived from an area well beyond the current distributional range of the herbivore. (for such cases, observed preference will hardly reflect recent selection pressures, because the insect will never encounter the plant in nature). Fourth, to avoid the results being influenced by the inclusion of ‘negative controls’ (i.e. plant species upon which the herbivore does not typically feed), we excluded studies that include plant types that are available, but not used as hosts in the field. Finally, studies examining preference and performance on anthropogenic cultivars were excluded, because the average length of such plant–herbivore associations is likely of younger origin than the average length of extant insect–plant associations.

As many studies reported multiple, partly overlapping measures of preference and/or performance, we used a series of clear-cut criteria to reduce the number of non-independent data points. These criteria are identified in Appendix S2.

Meta-analysis

Performance analyses

To compare performance on preferred versus non-preferred plants, plant types were ranked according to how acceptable they were among ovipositing females (Appendix S1). Performance analyses were focused on three metrics derived from the literature: Offspring survival, i.e. the fraction of offspring surviving from the onset of a study until the completion of the study (the exact period over which survival was assessed varied among studies); size, often assumed to reflect fitness (Honĕk 1993; Tammaru et al. 1996), and development time, reflecting the time the insect is exposed to natural enemies and, for species with multiple generations, the potential for population growth (e.g. Feeny 1976; Clancy & Price 1987; but see Benrey & Denno 1997; Williams 1999).

For each performance metric, we derived a separate metric of effect size. For survival data, we used the (ln-transformed) odds ratio (Rosenberg et al. 2000). Thus, for a study to be included in this analysis, it had to provide sufficient information for us to infer the number of living and dead individuals on the preferred and non-preferred plant type. For offspring size, we used the standardized difference (Hedges’d; Hedges & Olkin 1985) between the mean weights of offspring on plants preferred by females and offspring on less preferred plant types. To be included in this analysis, a study had to report the mean, a measure of variation around this mean (SD or SE) and sample sizes for offspring weight on two plant types differing in their attractiveness to ovipositing females. For development time, we used Hedges’d of mean development time between preferred and less preferred plant types as our effect size.

In each of the above analyses, we used the less preferred plant type as the ‘control group’, and the more preferred plant as the ‘experimental group’ (Rosenberg et al. 2000). Thus, for survival and weight, a positive effect size implies that offspring fare better on the host type preferred by females. As fast development is assumed to reflect higher fitness, a negative effect will support the preference–performance hypothesis in the analysis of development times.

Preference analysis

To assess whether female insects prefer host types on which their offspring perform best, we compared the number (or density) of eggs per plant unit between ‘good quality’ and ‘bad quality’ plants (for a definition of ‘good’ and ‘bad’ quality plants, see Appendix S1). Again, we used Hedges’d as our effect size (cf above), and defined the ‘bad quality plants’ as the control group, and the ‘good quality plants’ as our experimental group. Hence, here a positive effect size will support the preference–performance hypothesis.

Statistical models

To examine the support for individual predictions (Table 1; P2.1–P2.6), we tested whether effect sizes differ among studies assigned to different groups (Rosenberg et al. 2000). To test P2.1, we classified target taxa as being monophagous, oligophagous or polyphagous. In this context, insects specialized on a certain plant genus were classified as monophagous, insect species feeding on plants from several genera within a family as oligophagous, and insect species known to feed on plant species from several families as polyphagous. To test P2.2, we noted whether the studied plants were woody or herbaceous. We tested P2.3 by assigning the studied insect species to either of two broad groups: free-feeding or ‘sessile’ (the latter group consising of leaf miners, gallers, stem borers and sap-suckers with immobile immatures). To test P2.4, we noted whether females lay their eggs in batches or as single eggs. To test P2.5, we grouped studies according to whether the females take up nutrients as adults. In cases where this information was not explicitly mentioned, we assumed that species with fully developed mouth parts do indeed feed as adults. Finally, to test P2.6, we classified studies according to whether the host plants compared were of the same or of different species (i.e. whether the comparison was intra- or interspecific). In cases where the information needed for assigning studies to different groups was not available in the primary publication, we contacted the authors for additional information, or consulted with experts of the respective taxonomic group.

As we were more interested in variation among effect sizes than in group-specific means, we used random effects models. For each response variable, we first conducted a summary analysis with no data structure (i.e. no categorical variables). If the total heterogeneity of the sample of effect sizes (QT) was significant (indicating that there is structure in the data not captured by such a simple model), we conducted analyses with our set of pre-defined explanatory variables (separate analyses for each explanatory factor). If the between-group heterogeneity (QB) was significant, we inferred that the mean effect size differed between groups. Nenety-five per cent confidence limits around mean effect sizes were obtained by bootstrapping (bias corrected confidence limits based on 4999 iterations; cf. Adams et al. 1997). An effect is statistically significant if its confidence intervals do not overlap with zero. Following conventional practice, the size of a mean effect size was considered small when less than 0.2, moderate when 0.5 and large when greater than 0.8 (Gurevitch & Hedges 1993).

All analyses were performed in MetaWin version 2 (Rosenberg et al. 2000). Studies included in the analyses are listed in Appendix S3.

Results

The data compiled allowed us to compare the relationship between female preference and offspring performance on 21–29 pairs of plant types, depending on the response (Appendix S3). Visual exploration of the data revealed no signs of publication bias (Fig. 1): setting aside some extreme outliers (notably in Fig. 1c), the variation around the overall effect size decreased with sample size, effect sizes were independent of sample size, and at any given sample size, individual studies were normally distributed around the overall effect size (cf. Rosenberg et al. 2000).

Figure 1.

 Funnel plots showing relationships between effect size (Hedges’d or ln OR) and sample size in the analyses of offspring (a) weight, (b) development time, (c) survival and (d) female preference. (a–c) Sample sizes (n) refer to the number of individuals used to measure performance on the plant type preferred by females (in most studies, sample sizes on the less preferred plant are similar). (d) n shows the number of units (typically individuals) of ‘good quality’ plants on which oviposition was assessed.

The majority of studies in our data base concern Lepidoptera, with scattered entries of studies involving other insect orders (Appendix S3). Effect sizes did not differ significantly between insect orders (Table 2), although in the preference data QB was almost significant (P = 0.06). This pattern was mainly due to the small effect sizes observed in two studies on Hemiptera (Fig. 2). Based on these results, we conclude that pooling data from studies conducted on different insect groups is unlikely to have biased our results.

Table 2.   Total heterogeneity (QT) and between-group heterogeneity (QB) of effect sizes in studies comparing offspring weight, development time, and survival on preferred and less preferred plant types, and female preference (i.e. no. eggs laid) for good versus poor quality hosts. QT values are given for models without data structure (i.e. no explanatory variables). As QT was non-significant in the summary analysis on survival, no further explanatory factors were examined for this response
ResponseGrouping variableQTQBd.f.P
Weight 51.83 270.003
Insect order1.5520.46
Diet specialization0.9820.61
Woody or non-woody plant0.1810.67
Offspring mobility0.6210.43
Egg distribution0.0110.91
Feeding by adults2.2610.13
Inter- or intraspecific study0.0510.83
Development time 77.27 21< 0.00001
Insect order3.0920.21
Diet specialization0.1420.93
Woody or non-woody plant
Offspring mobility0.00310.95
Egg distribution0.2210.64
Feeding by adults0.2010.65
Inter- or intraspecific study0.0210.90
Survival 18.90 200.53
Female choice 53.68 280.002
Insect order8.8540.06
Diet specialization14.8320.0006
Woody or non-woody plant0.5110.55
Offspring mobility1.1310.29
Egg distribution0.2410.62
Feeding by adults0.0410.83
Inter- or intraspecific study0.5910.43
Figure 2.

 Mean effect sizes (with 95% bias corrected confidence intervals) for studies conducted on different insect orders in the preference analysis.

Overall support for the bitrophic PPH (Hypothesis H1)

The summary analyses provide clear support for a link between female preference and offspring performance (i.e. for Hypothesis H1 in Table 1). Overall, offspring were more likely to survive on plant types prefered by ovipositing females. Likewise, the number of eggs laid by females was higher on plants on which offspring perform well (Fig. 3). There was also a trend for offspring to be larger and to develop faster on plant types preferred by ovipositing females, although for these performance traits the overall effect sizes did not differ significantly from zero (Fig. 3).

Figure 3.

 Overall effect sizes for the four response variables addressed in this study. For offspring weight and development time, and for female choice, the effect size used is Hedges’d; for offspring survival, it is ln(OR). Bars are 95% bias-corrected confidence limits. Where confidence limits do not overlap with zero (grey line), effect sizes are considered statistically significant.

Where significant, the mean effect sizes in the models without any data structure were 0.73 (preference) and 0.95 (survival), and were hence moderate to large (Gurevitch & Hedges 1993). For all response variables, except for offspring survival, the total heterogeneity of effect sizes (QT) was statistically significant when tested against a χ2 distribution with appropriate degrees of freedom (Table 2), indicating that the variance in effect sizes was greater than could be explained by sampling error alone.

Ecological and life history modifiers of the preference–performance relationship (Hypothesis H2)

As the distribution of effect sizes revealed no sign of unexplained variation around joint mean survival (non-significant QT¸Table 2), we dropped this response variable from our analyses of ecological and life-history factors influencing the preference–performance relationship. For the other response variables, most of the explanatory factors examined (Table 1; P2.1–P2.6) had no detectable influence on the link between preference and performance (Table 2, Fig. 4). However, when insects were grouped according to their degree of diet specialization, effect sizes differed significantly between groups in the preference analysis (Table 2, Fig. 4a). Here, the mean effect size was larger for studies conducted on oligophages than for studies conducted on polyphages or monophages (Fig. 4a). In other words, the link between female preference and offspring performance was stronger for insects specialized on plants within a certain family than for insects with either broader or narrower diets.

Figure 4.

 Mean effect sizes (with 95% bias corrected confidence intervals) for studies grouped according to explanatory variables. In each panel, the results from analyses of each response variable (weight, development time, female preference) are displayed. The different panels show the effect of (a) diet specialization (P2.1 in Table 1), (b) type of host (i.e. woody or herbaceous; P2.2), (c) offspring mobility (P2.3), (d) pattern of egg distribution (P2.4), (e) feeding by adults (P2.5) and (f) whether the study assessed the preference–performance relationship inter- or intraspecifically (P2.6). All but one study in our data set on development time was conducted on non-woody plant species, preventing us from examining the effect of this explanatory variable on this particular response.

Discussion

This study provides a first quantitative synthesis of evidence pertinent to the preference–performance relationship in herbivorous insects. For both performance (survival) and preference, overall effect sizes were moderate to strong (Gurevitch & Hedges 1993) and significantly differed from zero (Fig. 3). As all data points in our meta-analysis emanated from studies conducted in isolation from natural enemies, competitors and mutualists, this study can be seen as offering unequivocal support for the preference–performance hypothesis in a bitrophic setting. Given the effort taken to explain situations where female choice does not match patterns of offspring performance (e.g. Thompson 1988; Courtney & Kibota 1990; Mayhew 1997; Craig & Itami 2008), this is in itself a remarkable result. Compared with previous narrative reviews of the topic, this study then illustrates the advantage of quantitative research synthesis as a way of seeing the wood for the trees. Although there are convincing examples of cases where female choice does not match offspring performance (e.g. Rausher 1979; Valladares & Lawton 1991; Underwood 1994; Fritz et al. 2000; Faria & Fernandes 2001 and examples in Appendix S3), the general rule is clearly a positive association – be it with modifications.

Based on the information available in our data base, we were able to test six specific factors proposed to influence the strength of the preference–performance relationships. Of these, the current data provide support only for a role of diet breadth: In the preference analyses, effect sizes were larger for oligophagous than for polyphagous insects. This pattern supports the ‘neural constraints hypothesis’ (e.g. Bernays 1998; Bernays & Funk 1999; Janz 2003), which posits that the behaviour of insects may be constrained by the limitations of their neural system to process complex information. It is equally compatible with the idea that specialization may allow ‘fine-tuning’ to the characteristics of individual host species although multiple physiological and morphological means (e.g. Roslin & Salminen 2008). In such a case, the relationship between diet breadth and the accuracy of female choice might not necessarily be due to non-adaptive constraints, but could also suggest that oligophages are under stronger selection to develop a preference for high quality plants than are insects with broader diets. Nevertheless, the observed pattern came with a curious twist: effect sizes for monophages were significantly smaller than for oligophages (and approximately the same size as for polyphages), suggesting that the preference–performance link may be strongest for insects with an intermediate level of diet specialization. The comparably weak preference–performance relationship observed in monophages may be a methodological artefact. Clearly, plants belonging to different genera and families are likely to differ in more ways than are plants of the same species. As monophages were defined as insects specialized on plants within a genus, the plant types in plant pairs on which the preference–performance relationship is assessed will likely be more similar than is the case in studies on oligophagous and polyphagous insects. In studies conducted on monophagous insects, 100% of the data points involve comparisons on plants belonging to the same species or the same genus, whereas the corresponding figure for oligo- and polyphagous insects was 45–54%, depending on the response. At this point, we consider this the most plausible explanation for the observed pattern.

Of the five other predicitions tested, none gained support. Whether these negative results reflect true patterns or whether they are partly due to the lack of statistical power remains open to debate – but at this point, this study does not back them up as major modifiers of the general relationship between preference and performance. Instead, among the wealth of specific hypotheses put forward to account for variation in the strength of preference–performance coupling, this study pinpoints diet breadth as a promising avenue for future research.

While the patterns discussed above serve to bring out the advantages of meta-analysis, there are also limits to this approach – the most important of which are due to the motley methodology and style of reporting which is common in PPH research. From a practical perspective, the variable approaches taken to measure the link between preference and performance as well as in presenting the results makes it challenging to combine data from different studies. Consequently, only a fraction of the studies conducted to date could be included in our meta-analyses. Compared with the general volume of studies conducted in this field (see Data generation), the set of n = 21–29 useful studies identified by us per response will seem vanishingly small (Appendix S3). Moreover, data to test additional factors proposed to influence the strength of the preference–performance relationship such as the effect of time limitation (e.g. Ward 1987), phylogenetic constraints (e.g. Price 1994, 2003) and the co-evolutionary history of the plant–insect associations tested (e.g. Agosta 2006) are not readily available. As a result, there are currently limits to the kind of questions that can be answered by means of meta-analysis. Hence, we urge our colleagues to continue addressing the many intriguing questions flourishing within the subject field – and when reporting the fruits of their work, to present their data in a coherent format. Most importantly, they should remember to include at least group-specific means, sample sizes and standard deviations – a recommendation which may seem trivial, but is rarely met. To stimulate this development, we will continue building on the data base we have now established (Appendix S3), with the aim of eventually addressing new questions with additional data, and exposing old ones to the critical test of additional data. We believe that such a compilation of data will be helpful in advancing the field, and invite our colleagues to join in on the effort by communicating their quantitative findings to us.

Acknowledgements

We thank all authors that have responded to our queries on specific details on individual study systems, and the following people for sharing their expertise on particular insect taxa: Anders Albrecht, Heikki Roininen, Marko Mutanen, Sören Nylin and Stig Larsson. Bob O’Hara provided statistical advice and three anonymous referees offered valuable suggestions for improvement. Financial support from the Academy of Finland (grant number 129636 to the Centre of Excellence in Metapopulation Research 2009-2011, and grant numbers 126296 and 129142 to SG and TR, respectively), and from the Ella and Georg Ehrnrooth foundation (to SG) is gratefully acknowledged.

Ancillary