As in many insects, gypsy moth larvae that are infected with their baculovirus release infectious particles known as ‘occlusion bodies’ shortly after death, and the occlusion bodies are then available to infect additional larvae (Cory & Myers 2003). A standard method of studying baculovirus transmission is therefore to feed a solution of occlusion bodies to host larvae in the laboratory (Cory & Hoover 2006). In this type of experiment, however, larvae that do not consume the entire dose are discarded; so, there is no allowance for the effects of behaviour. An alternative method is therefore to allow uninfected larvae to consume virus-contaminated foliage in the field (D'Amico et al. 1998; Hails et al. 2002). Field transmission studies, however, have the opposite difficulty of standard methods, in that they do not allow us to disentangle behaviour from susceptibility.
To produce foliage contaminated with virus-infected cadavers we fed a virus solution to hatchling larvae at a dose sufficient to ensure 99% mortality (Dwyer et al. 2005). We then placed these infected individuals on the foliage of red oak trees in the field. To keep the larvae from escaping, we enclosed the branches in mesh bags that allow the passage of air, water and much of the natural spectrum of light. Larvae were left on the leaves for 5 days, to ensure that they were dead, and then were brought into the laboratory. We then used cork-borers to make leaf discs of approximately 1 cm2 in area that contained cadavers, and for controls we made leaf discs from uncontaminated foliage taken from adjacent branches of the same trees.
Pairs of leaf discs with similar vein structures, one contaminated and one clean, were matched and photographed. The trees used in this study were the same as those used in Elderd, Dushoff & Dwyer (2008), which had a level of natural contamination that was effectively zero (less than 0·8% of larvae on control foliage in Elderd et al. 2008 became infected, and anecdotal evidence suggests that these few infections were due to handling in the laboratory). In our first year of trials (2006), we left the clean foliage uncovered by mesh bags. In our second year (2007), however, both control and virus-contaminated leaves were placed inside bags to ensure that differences in foliage quality between the two disc types did not alter larval preferences. In both years, experiments were conducted in July and August when foliage quality is relatively constant both chemically and physically (Hunter & Lechowicz 1992; Salminen et al. 2004), and herbivory at the field site was nearly zero (G. Dwyer, personal observation). It therefore seems unlikely that the observed preference for uncontaminated foliage in the first year was due to the lack of mesh bags on the control foliage compared with the experimental foliage, especially given that the experimental foliage was bagged for only 5 days. More concretely, as we document in the Results section, levels of cadaver avoidance in the second year of trials were indistinguishable from levels in the first year. We therefore include data from both years, and we attribute differences in consumption between the two leaf disc types to the presence or absence of virus particles rather than to some other factor.
To produce uninfected larvae, we hatched larvae from egg masses that had been soaked for 90 min in 10% formalin, which effectively surface sterilizes the eggs (Dwyer & Elkinton 1995). Feral strain insects came from egg masses collected near Gladwin, MI (44·0° N, 84·5°W). Laboratory strain insects were hatched from a strain that has been maintained by the USDA for many generations, and which are consequently of lower heterogeneity than feral insects (Dwyer, Elkinton & Buonaccorsi 1997). All healthy larvae were reared to the fourth instar, and then were used in experiments. To ensure that uninfected larvae were developmentally synchronized, we used only larvae that had moulted to the fourth instar in the previous 48 h (Grove & Hoover 2007).
With full-sibling experiments, differences among families could hypothetically be the result of environmental differences rather than genetic differences. For example, although larvae were reared under identical conditions in the laboratory, it is possible that differences among families were due to the effects of variability in resource quality among female parents in the previous generation. Such variability could affect the susceptibility of offspring, a phenomenon known as a ‘maternal effect’ (Myers 2000). One way to disentangle maternal effects from genetic effects is to mate males to multiple females, and then to test for effects of sire independently of the effects of dam (Lynch & Walsh 1998). In 2007, we therefore mated individual, feral, adult male gypsy moths to two or three feral dams per male, to produce half-sibling groups. We tested 10 half-sibling groups, each with the same sire and two or three dams, with 28–127 individuals in each group.
In 2006, we also tested for effects of spatial structure on detection ability. Full-sibling feral larvae were given a choice between a clean leaf disc and a disc that was tangent to, but did not include, a cadaver-covered leaf disc (see Fig. 1). This allowed us to determine whether virus particles that leak out of a cadaver can be detected and avoided as much as 1 cm away from the cadaver, allowing us to roughly quantify the spatial scale over which avoidance behaviour occurs. We again kept track of full-sibling families, using six families of 25 individuals each.
We intentionally designed our experiments to test for variability in behaviours that affect infection risk, rather than to test for effects of variability in behaviour on infection risk itself, for several reasons. First, we were only able to measure total area consumed, whereas risk of infection is also affected by how close a larva gets to a cadaver while it feeds. Second, we could not control for variability in physiological susceptibility independently of behaviours that affect exposure, yet variability in physiological susceptibility in gypsy moth larvae is known to be quite high (Dwyer et al. 1997). Larvae that ate similar areas of contaminated foliage may therefore have had very different infection risks. We therefore did not expect that our experiments would provide much evidence for effects of behaviour on infection risk. However, in three trials (the two trials using full-sibling feral insects and the trial using a laboratory strain), we nevertheless reared larvae individually on artificial diet for several weeks after exposure to determine which larvae had become infected. The resulting data did indeed show that the amount of leaf area consumed can affect infection risk, but they also showed no effects on infection risk of interactions between family and area eaten, as we expected. It is thus in turn difficult to demonstrate that heritability in cadaver-detection ability alters infection rates. These data are tangential to the main thrust of our work, and so they are presented as Supporting Information.
As we have described, Capinera et al. (1976) provided strong evidence that gypsy moths can detect cadavers. Our main goal was instead to test whether cadaver-detection ability is heritable. Statistically, this meant testing for the effects of either full- or half-sib families on the difference in the amount of foliage eaten between cadaver-contaminated and uncontaminated leaf discs. Some larvae, however, may eat more than other larvae, irrespective of whether a disc is cadaver contaminated or not. The amount of one type of disc that a larvae eats may thus not be independent of the amount that the larva eats of the other type of disc, and it was crucial for our statistical analyses to take this lack of independence into account.
We therefore constructed our statistical models in the following way. In our models, i is the full-sib family, j is the leaf type, 0 for uncontaminated and 1 for cadaver contaminated, and k is the individual larva. For our full-sib experiment, we then write yijk for the average amount of leaf type j eaten by individual k in family i, which depends on the overall average amount eaten μ and the error term εijk. In addition, however, we took into account the lack of independence of contaminated and uncontaminated leaf discs that were fed upon by the same larva, which we represent with the term bk(i). Note that the symbol k(i) signifies that individuals are nested within families (Gomez et al. 2007). Our simplest statistical model is thus,
- (eqn 1)
If this model had fit best, we would have concluded that variability among individuals, including correlations in feeding intensity within individuals, was sufficient to explain our data. The amount eaten, however, may also by affected by the presence of a cadaver, the effect of which we represent with the symbol Dj. We then have that D0=0 for uncontaminated leaves, so that D1 is the change in the average amount eaten due to the presence of a cadaver. Our next most complicated model is therefore,
- (eqn 2)
If larvae avoided cadavers, as we expected them to, then it should be true that Dj<0. As we will describe, larvae do indeed avoid cadavers, and so we refer to this behaviour as avoidance rather than preference.
We further suspected, however, that there would be an effect of family on feeding behaviour. To equation (2), we therefore added the effects of family Fi on the average amount eaten;
- (eqn 3)
In the above equation, we have begun by assuming that the effect of family is the same on both types of disc. If in addition we allow for an effect of family on cadaver-detection ability, the effect of family must instead vary between disc types. In our next model, we therefore added the term FiDj, which represents the interaction between the family effect Fi and the disc-type effect Dj:
- (eqn 4)
Note that for uncontaminated discs, j=0 and D0=0, reducing the full model to yi0k=μ+Fi+bk(i)+εi0k. Our statistical approach was thus to test whether a model that included the interaction effect FiDj, namely equation (4), provided a better fit to the data than did the models that did not include that term, namely equations (1)–(3). We then used the Akaike Information Criterion (AIC) to choose the best model, as we describe in more detail below. We reiterate, however, that we were careful to allow for the lack of independence of discs within a larva, and also that we allowed for direct effects of family and cadaver presence, to test whether direct effects provided a better explanation than the interaction effect which is our main interest. In addition, note that disc is taken to be a fixed effect, while the other variables are taken to be random effects.
In the half-sibling experiment, each individual could also be grouped by sire, and so we added the effects of sire to our models. The response variable is then the amount of foliage of the jth leaf type consumed by the kth individual from the ith family and the lth sire. The effect of sire l on the amount eaten is then Sl, while the interaction effect is SlDj. The term SlDj then represents a difference in cadaver-detection ability between the offspring of different sires, and thus allows cadaver-detection ability to be affected by sire. The full model is then,
- (eqn 5)
Note that here family i is nested within sire l because families arise from multiple dams that are mated to the same sire. As in the full-sib experiments, we compared this model with simpler models in which we deleted all but the average consumption rate μ and the individual effect term bk(il). In the case of the half-sibling experiments, our goal was thus to determine whether there was an effect of sire on cadaver-detection ability, and thus whether cadaver-detection ability is heritable, and again we took into account the lack of independence of leaf discs within a larva. Disc is again considered a fixed effect and all other variables are random effects.
The statistical models that we have described are linear mixed effects models, which we implemented using the package ‘lme4’ in the R programming language (Bates 2007). To choose among the models, we used the AIC. In contrast to tests of statistical significance, the AIC has the advantage that it is based on the assumption that ‘all models are wrong, but some models are useful’ (Box 1979), and it allows us to choose among multiple models at the same time (Burnham & Anderson 2002). AIC is a useful statistical tool in our case because we are not sure which of our many models will best fit our data. The statistical foundations of AIC analyses, however, are quite different from those of significance tests, and so recommended practice is to only include one type of analysis, not both (Burnham & Anderson 2002). As for our purposes AIC is the best choice, we do not present the results of significance tests.
Akaike Information Criterion scores are calculated according to,
- (eqn 6)
where is twice the negative log-likelihood of the parameters given the data y and K is the number of parameters in the model. The best model is the model with the lowest AIC score. Models with more parameters are likely to provide a better fit, and thus a smaller value of the negative log-likelihood, but they will be penalized by the 2K term. The AIC thus operates on the principle of parsimony to find the model that best trades off better fit with less complexity (more precisely, the model with the lowest AIC is the model that minimizes the distance between that model and the true model, Burnham & Anderson 2002). To compare AIC scores between models, we use the ΔAIC for each model (Burnham & Anderson 2002), which is the AIC score of that model minus the AIC of the best model. The model with the best fit thus has a ΔAIC score of zero. To evaluate the relative strength of evidence for different models, we used AIC weights (Burnham & Anderson 2002), such that, of Z total models, model r has weight:
- (eqn 7)
The AIC weight for a particular model is thus a measure of the probability that that model is the best model, and so the relative support for different models can be assessed from the weights.