Temporal or spatial variation in selection has the potential to explain long standing evolutionary problems such as evolutionary stasis and the maintenance of genetic variation. Long-term field studies of plants and wild vertebrates have provided some insights, but multigenerational measures of selection in wild invertebrates remain scarce. Short-lived ectothermic animals are likely to experience more pronounced environmental variation across generations than longer-lived and endothermic species. As a result, variation in selection may be particularly significant in these groups. Over ten years, we have monitored an individually tagged population of wild crickets (Gryllus campestris) using a network of up to 133 day-night video cameras. The over a million hours of video that we watched allowed us to capture detailed information about naturally and sexually selected traits and life-history parameters. Over ten discrete generations, population size ranged from 51 to 546 adults. There were also substantial differences among years in the average values of traits including adult emergence date, body size, lifespan, and several behavioral traits. We combined measurements of these traits with individual fitness, measured as the number of adult offspring inferred from genetic-marker-based parentage assignments. This revealed substantial variation in selection gradients across years in several traits, with evidence that in one trait, adult emergence date, selection switched from positive to negative among years. Our findings suggest that fluctuations in selection gradients are common but complete reversals in the direction of selection may not be very frequent.

Study system

Our data are the product of WildCrickets, a long-term project monitoring of a wild population of field crickets G. campestris in a meadow in northern Spain (Rodríguez-Muñoz et al. 2019d). This species has a single generation each year, with the first adults emerging in mid to late April and the last adults dying in mid-July. Individuals of both sexes build burrows as a refuge from predation and bad weather. Most interesting events occur at burrow mouths (Rost and Honegger 1987) with individuals spending only short periods moving between them. This lifestyle allows us to record the adult lives of the entire population in great detail, by attaching unique tags to individuals as they become adult and monitoring the population through daily surveys and a network of up to 133 day/night video cameras. During the adult season, males call from their burrows to attract females and both sexes move around the meadow, displacing members of the same sex from burrows and sharing burrows with a single member of the opposite sex (Fisher, Rodríguez-Muñoz, & Tregenza, 2016). We take a DNA sample at adult emergence, allowing us to assign parentage of each generation from amongst the adults sampled in the previous breeding season. Details on how the meadow is managed every year, our monitoring protocol and parentage assignment to estimate fitness, are available in Rodriguez-Muñoz et al. (2010) and Rodriguez-Muñoz et al. (2019d). The data included in this study cover eight generations within the period 2006-2014, and comprise 364,902 hours of video where a cricket was present under the camera. We did not include the years 2009, 2015 and 2016 as we do not have parentage assignments for offspring from those years.

Parentage assignment

Genetic profiling with microsatellite loci was performed to conduct parentage analysis. Details of this procedure are provided elsewhere (Bretman et al. 2011; Rodríguez-Muñoz et al. 2019d). Briefly, we used between 14 and 21 autosomal loci. Genotyping was performed on an ABI3730 capillary sequencer, using standard protocols (Ball et al. 2010) and scoring was performed using GeneMapper v3.7 software. Parentage analysis was performed using genotype data combined with spatial and mating information in a Bayesian framework using the MasterBayes package (Hadfield et al. 2006; Koch et al. 2008). We estimated the pedigree on a year-by-year basis rather than as a single run, as field crickets are annual and thus generations are not overlapping. Using the modal parentage assignment for each individual, maternity was assigned to a sampled individual in the population for 1,326 out of 1,568 individuals (0.85 of the population) and paternity to a sampled individual for 1,441 individuals (0.92 of the population). These figures, and our observation that during the breeding season, new adults (that we have not observed having overwintered there) occasionally appear in the meadow, indicates that there is limited immigration into our meadow (see also Bretman et al. 2011). The median confidence of maternity assignments, to known individuals, was 0.985 and in paternity assignments was 0.987.

Description of traits

We explored variation in selection by analyzing the relationship between fitness (response variable) and eight morphological, behavioral and life history traits (predictors) that could affect reproductive success. We quantified fitness as the number of offspring produced per adult cricket in year t that survived to adulthood in year t+1. This fitness measure has the weakness that it is affected by both parent and offspring traits (as offspring traits will affect their survival to adulthood). This means there is the potential for our selection metrics to be affected by direct and indirect effects of traits on juvenile survival. However, our metric does provide a clear measure of each parent’s count of offspring that have the potential to pass their genes on to the next generation. In common with most invertebrates, counting the eggs laid by female crickets living in the wild is an impossible task, as each female injects hundreds of eggs into the soil in multiple locations. Our predictors included adult size, timing of adult emergence and longevity, and traits that quantified the intensity of polygamy, mobility (in relation to both time and space), and effort in attracting mates (for males only).

In our study population, sexual activity does not start until about five days after adult emergence (Rodríguez-Muñoz et al. 2019a). We therefore only included data from events that happened at least five days post adult emergence. Adult crickets move frequently between burrows, leaving one burrow and arriving at another a few minutes later (median duration of each visit to a burrow is 1.2 h). Because we often have more burrows than cameras to monitor them, there are some crickets where we only have a small amount of observational data, perhaps from a single period during the individual’s life. Cricket behavior is dependent on the weather (for instance crickets avoid leaving their burrows when it is raining). It is also affected by the seasonal changes that occur as spring progresses and there are systematic changes in behavior with age (Makai et al. 2020). These effects mean that parameter estimates from crickets where we only have a few hours of observational data are likely to be very unreliable. To avoid the noise from these poorly sampled, but otherwise random individuals masking biologically relevant patterns, we excluded crickets observed for less than 96 h over their reproductive lives. For those living more than 35 d, at least 24 h out of the total observation time had to happen after that age, this meant that we had between 787 and 1374 unique individuals for each trait that we studied. This data exclusion is not based on any characteristics of the crickets, but on chance in relation to their visits to our cameras, so we do not expect it to cause any bias. The number of individuals per year and sex before removal of poorly sampled crickets is shown in Table S2. We quantified our predictors as follows:

Thorax width: Measured from a digital picture using imageJ analysis software (Schneider et al. 2012).

Emergence date: Day of the year when the cricket reached adulthood (counted from day 1 on 1 Jan).

Lifespan: Number of days alive as an adult. For individuals where death date was not observed, we assume they died on the day after the last available observation.

Mates per day: This is the mean number of unique partners a cricket mated with per day, over its whole life. We calculated it by dividing the total number of different mates an individual was observed mating with by the total number of days of video observation we had of that cricket.

Burrows per day: Mean number of different burrows visited per day. We calculated it by dividing the total number of unique burrows visited when under observation by the total number of days we observed the cricket with a camera.

Distance among burrows: Maximum distance between any two pairs of burrows from among all the burrows where we observed the cricket during its adult life. Moving exposes crickets to risk of predation and reflects investment in reproduction, as food resources are hyper-abundant through the meadow. To control for the effect of time under observation, we divided the maximum distance by the total number of days the target individual was observed.

Time at burrow: Median duration of time spent on each visit to any burrow, calculated from all the visits recorded for this individual over its life. Usually, sexual activity starts a few days after adult emergence. Hence, for crickets with known emergence date, we excluded all the movements happening before 5 days of age.

Calling effort (males only): We recorded point samples every two minutes during the first 10 minutes of every hour where we observed the target male with a camera. If we saw the male calling at any of those point samples, we recorded him as calling. We then calculated calling effort as the proportion of hourly observations when the male was calling. As for the previous trait and for the same reason, we excluded the first five days after adult emergence. We also excluded all individuals with less than 10 samples in total to avoid very unreliable estimates due to small sample size.

Statistical analyses

To compare selection among years we need a common model to describe the relationship between traits and fitness. Inspection of graphs for individual years indicated that the dominant patterns we observed were either no effect or directional selection within years, we therefore estimated linear selection gradients. To test for variation in selection gradients, we followed the common approach of running generalized linear mixed models independently for each trait. To simplify the analyses, we ran separate models for each sex. This approach is also appropriate because it is likely that the same traits are under different selection regimes in males and females; for instance, movements around the meadow have a completely different function in mate-searching males than they do in females (who are surrounded by the singing of potential mates). We did not include age in our analyses because our species has discrete generations; adults emerge in early to mid-spring and die between late spring and early summer within the same year. We ran analyses in R (R Development Core Team 2020, v. 4.0.3) and R Studio (R Studio Team 2020, v. 1.3.1093 ) using the lme4 package (Bates et al. 2015) with a Negative binomial family distribution to cope with overdispersion. This approach uses a log-link function by default, which puts fitness on a relative scale equivalent to the traditional approach of dividing by mean fitness and using linear regression (e.g. Lande and Arnold 1983; Bonnet and Postma 2018). This means that our fitness gradients can be directly compared with those estimated using earlier methods. We checked whether the output of the models showed overdispersion by using the method proposed by Harrison (2014). Before the analyses, we centered all traits within years (by subtracting the mean of the year) and standardized them across years by dividing by standard deviation of the centered values. The phenotypic traits of small ectotherms like these crickets are extremely dependent on weather conditions which vary substantially from year to year. A cold winter will affect the size and emergence date of the entire cohort and failing to control for this means that variation is dominated by environmental effects. In view of this, we centered traits within years. For those traits where means are similar across years, centering within years will not make a relevant difference. However, in the supplementary information, we have also included the same analyses based on data centered among years, so that it is possible to compare results between both approaches.

We estimated selection gradients separately for each trait using mixed models with Fitness as the response variable, focal Trait (as described above) as a fixed effect, and Year as a random effect. In each case, our main ‘random slopes’ model allowed both the average Fitness and the effect of the focal Trait on Fitness to vary across each Year in the random effects, i.e., (Fitness ~ 1 + Trait + (1 + Trait|Year)). We then ran a second model whereby the effect of the Trait on Fitness was constrained to be the same across all Years, with the ‘random intercepts’ allowing Years to vary only in their average Fitness, i.e., (Fitness ~ 1 + Trait + (1|Year)). We then compared these models using a likelihood ratio test  to assess whether the contribution of the random slopes term for Year was significant. We followed  Visscher (2006) and assumed the difference in loglikelihood between the models was distributed as a 50:50 mix of  and . This contribution indicates whether the selection gradient varies among years. When a gradient varies, it can do so without changing the direction of selection, so that the differences are just in the intensity of selection. Alternatively, variation can involve changes in the direction of selection (reversals); i.e. a trait can have a positive or negative effect on fitness depending on the year. To provide insights into how selection gradients vary among years, we extracted the coefficients per year from the random effects, and estimated their confidence intervals by bootstrapping using the bootMer function in lme4.