This study shows the utility of using matrix population models to combine multiple life history traits into a single estimate of the population growth rate (λ) when testing for genetic trade-offs between parasite resistance and lifetime fitness. The most important result of this study is that our estimates of λ suggest that a population of refractory mosquitoes will be half the size of a population of highly susceptible mosquitoes in just 23 days. Hence, complete resistance to Plasmodium can be costly for A. gambiae and may explain why natural populations of mosquitoes maintain genetic variation for malaria resistance. The strategy to combat malaria by replacing natural mosquito populations with transgenic, malaria-resistant mosquitoes will fail if the transgenics carry similar fitness costs.
The population growth rate of the refractory genotype was always lower than that of the highly susceptible genotype (Figure 1). This suggests that the immune mechanisms required for complete refractoriness to P. yoelii nigeriensis are costly for A. gambiae and would not evolve under the laboratory conditions used in this study. This conclusion is supported by Hurd et al.'s  laboratory evolution experiment where mosquitoes fed exclusively on infected mice over 22 generations did not evolve refractoriness. Likewise, two recent population genetic studies estimated the strength of selection on 8 different anti- Plasmodium defence genes in the A. gambiae species complex and concluded that there was no evidence for strong directional or balancing selection on these genes [26, 27]. Such non-significant patterns of selection are expected if the costs of evolving Plasmodium -resistance genes are similar to the costs of Plasmodium infection. More generally, our results are consistent with numerous studies on other host-parasite systems that have found that the evolution of anti-parasite defence mechanisms in the host comes at the expense of other life history traits [4, 6, 7].
The conclusions of this study differ from Hurd et al.  because we combined all the life cycle parameters into a single measure of fitness, λ. Numerous authors have pointed out the importance of combining the components of fitness (survival, reproduction, development rates) into a measure of lifetime fitness such as λ [11, 12, 28]. Univariate analyses of multiple fitness components do not provide insight into lifetime fitness because they do not account for the conditional dependence of later expressed components of fitness (e.g. fertility) on those expressed earlier (e.g. survival to reproduce). Furthermore, such analyses are likely to give conflicting results because negative correlations between fitness components are common [10, 29]. The ambiguity of this approach is illustrated by the analysis of Hurd et al. , which found no clear pattern of differences in fitness components among the three genotypes. Similarly, none of the pair wise t-tests of the 7 life cycle parameters were statistically significant after correcting for multiple comparisons (Table 3). However, after combining the life cycle parameters into a single measure of fitness (λ) we found significant differences among the three genotypes. This study shows that lifetime fitness is the product of many parts and that small, statistically insignificant but consistent differences in these parts can add up to large differences over the course of a life cycle.
Our results are consistent with the only other study to test for genetic trade-offs between malaria resistance and other life history traits in a mosquito . Yan et al.  found that their refractory strain of Aedes aegypti was smaller, had lower survivorship and laid fewer eggs than the highly susceptible strain in both the presence and absence of the avian malaria parasite, P. gallinaceum. In contrast to Yan et al. , the major strengths of this study were that we (1) selected the refractory and highly susceptible strains from the same population, (2) included unselected control genotypes allowing us to rule out inbreeding effects, and (3) repeated the experiment three times (i.e. the black, red and green groups). One limitation of this study is that there was no replication of mice within the 9 combinations of group and environment. It is therefore possible that the significant group:environment interaction on λ (model 5 in Table 1) was caused by random variation among mice. It is well known, for example, that the gametocyte density in the vertebrate host influences the infectivity of the blood meal and the subsequent oocyst load in the mosquito . Fortunately, because the three genotypes were blocked by the factor 'mouse', the limitations of the experimental design do not affect the conclusion that the population growth rate of the refractory genotype is lower than that of the other two genotypes.
The life cycle parameters that reduced λ the most for the refractory genotype were post blood-feeding survival and hatching success (Figure 2). The lower post blood-feeding survival suggests that the refractory mosquitoes evolved immune responses that harm both "self" and "non-self" (i.e. autoimmunity costs; ). For example, following an infected blood meal, Anopheles females upregulate expression of nitric oxide synthase producing levels of nitric oxide  that limit ookinete development  but may also be toxic for the mosquito . Similarly, the phenoloxidase cascade responsible for the melanization of oocysts in the midgut produces phenol by-products that may be cytotoxic for the mosquito [35, 36]. The induction of the melanization response in gravid A. gambiae females also reduces the deposition of protein (e.g. vitellin) in the eggs . Studies with Plasmodium -infected Anopheles females have shown reduced vitellin provisioning of eggs, which may result in lower hatch rates [38, 39]. Hence, a trade-off between vitellin egg provisioning and an upregulated immune system post-blood feeding in refractory females is one explanation for their lower egg hatch rates.
We found no evidence that Plasmodium infection reduced the population growth rate of A. gambiae. The main effect of environment (presence versus absence of Plasmodium) was not statistically significant and explained only 8.2% of the variation in λ (Table 2). In contrast, the main effect of genotype accounted for 21.7% of the variation in λ (Table 2). Although it has been repeatedly shown that Plasmodium reduces egg production in Anopheles mosquitoes [15–17] including the Keele population from which the refractory and highly susceptible genotypes were selected by Hurd et al. , our sensitivity and elasticity analyses (Figure 3) show that λ is minimally affected by changes in fertility. The effect of Plasmodium on mosquito survival is more controversial . Environmental factors can also influence the virulence of the mosquito- Plasmodium interaction. For example, Lambrechts et al.  showed that A. stephensi infected with P. yoelii yoelii suffer more than uninfected individuals when fed on low glucose levels. Other studies have shown that Plasmodium -induced mortality is influenced by humidity, temperature, diet, larval density, and bacterial infection (reviewed in ). How this environmental variation structures the virulence of the mosquito- Plasmodium interaction in the field is an open question. In this study, the stressed environment was supposed to mimic field conditions and included cold temperatures, reduced sugar, and cage shaking to induce flight. However, because the experiment did not include an uninfected and stressed treatment it was not possible to determine whether environmental stress increased or decreased the cost of infection.