Cotton species was modeled as a fixed effect, since there are only two possible categories – not enough to meaningfully estimate a random effects distribution. We also included 15 real-valued fixed effect predictor variables that indicate the number of fields, out of the 8 surrounding fields, planted with each of the 15 crops we analyzed. The goal was to control for effects of the surrounding landscape, and thereby avoid spurious correlations between rotational history and yield. Our Bayesian modeling approach required the specification of priors for all parameters whose posteriors were estimated using MCMC. Non-informative priors were used for all fixed effects. The random effects for both field and year were assumed to follow a normal distribution with mean 0 and variance hyperparameters estimated from the data. Since the support of variance parameters is constrained to positive realnumbers, non-informative inverse gamma distributions with shape and scale parameters set to 0.001 were used as the prior for the variance parameter of the top-level stochastic node, and as the priors for the variance hyperparameters of the field and year random effects distributions. Model 2. To help us understand whether any effects of the crop grown in the field the previous year on cotton yield could be due to effects on L. hesperus, we fit the same model as Model 1, but with average June L. hesperus abundance as the response variable. Model 3. Next, to formally assess whether there was an association between the effects of crop rotation on yield and the effects of crop rotation on L. hesperus abundance, we performed a linear regression of the estimated effects on yield against the estimated effects on L. hesperus abundance .
Noninformative N priors were used for the mean and intercept,container raspberries and a noninformative inverse gamma distribution with shape and scale parameters set to 0.001 was used as the prior for the variance. Model 4. A great deal of experimental evidence has demonstrated that crop rotation leads to increased yield compared to successive plantings of a single crop; therefore, we explored whether or not a yield loss was incurred by cotton crops grown in fields where cotton was grown in previous years. For the 782 fields that had complete crop rotational records for the previous 4 years, we calculated the number of consecutive cotton plantings in the 4 years preceding the focal cotton crop. We then fit a model, with yield as the response variable, using the number of consecutive prior cotton plantings as a predictor. Field, year, and cotton type were included as they were in Models 1 and 2. Since the number of prior consecutive cotton plantings could be correlated with the number of cotton fields in the surrounding landscape during the focal year, we avoided a possible spurious correlation between consecutive cotton plantings and yield by also including a fixed effect for the number of cotton fields in the 8 fields adjacent to the focal field. We chose not to explore rotational histories of specific crops for longer than one previous year, since the number of possible rotational histories becomes very large and the number of records for each possible history becomes too small to allow for robust statistical analysis. Model 5. To see if the number of consecutive years of cotton cultivation preceding the focal year was associated with June L. hesperus densities, we fit the same model as Model 4, but with June L. hesperus as the response variable.
Capitalizing on a large existing set of crop records from commercial cotton fields in California, we employed an ecoinformatics approach to explore the effects of crop rotational histories on cotton yield. Our hierarchical Bayesian analyses revealed evidence that several crops, when grown in the same field the year before the focal cotton planting, were associated with either decreased or increased cotton yield , and either increased or decreased early season densities of the pest L. hesperus . Furthermore, crops associated with decreased yield were generally also associated with increased L. hesperus densities, while those associated with increased yield were also associated with decreased L. hesperus densities . These results suggest a possible mechanism for the observed yield effects of these rotational histories. Since L. hesperus preferentially attacks certain crops, a field cultivated with a crop that is heavily attacked by L. hesperus may, if L. hesperus disperse from the focal field, increase the abundance of L. hesperus in nearby fields. These populations may subsequently attack the crop planted in the focal field the following year, explaining the increase in early-season L. hesperus densities that we detected following certain crops. In turn, these increased L. hesperus populations may exert strong herbivorous pressure on focal cotton crops, possibly explaining the corresponding decrease in yield. We believe that the effect of rotational history on early-season L. hesperus likely operates at a landscape scale that is larger than the within-field scale. If cotton was grown in a field the previous year, then farmers in the San Joaquin Valley are required to maintain a 90-day plant-free period prior to 10 March of the following year. This prevents L. hesperus, which overwinter as adults on live host plants, from overwintering in a focal field where cotton was grown the year before.
If a crop other than cotton was grown the previous year, then it could be possible for L. hesperus to overwinter in the focal field on residual plant or weed populations; however,since fields are completely plowed prior to planting cotton in the spring, L. hesperus adults would still need to temporarily leave the focal field. Therefore, we believe that the preferred host crops for L. hesperus increase L. hesperus populations at a landscape scale. Then, when cotton, another target of L. hesperus, is planted in the same field the following year, the cotton field is attacked by this regional population. If regional populations are already large due to lingering effects from crops grown the previous year, L. hesperus populations may move into cotton early in the growing season; this could be particularly damaging given research suggesting that cotton yield is particularly sensitive to L. hesperus densities early in the growing season. Using our data, we were not able to determine at exactly what scale the effects of rotation on L. hesperus likely operate. We do not believe a within-field scale is plausible, but determining a more precise spatial scale for these effects could be an interesting topic for future research. Our findings match expectations of crop yield effects based on previous research on L. hesperus host crop preferences, lending support to our hypothesis that yield effects of crop rotational histories are, at least partially, mediated by effects on L. hesperus. Alfalfa and sugarbeets, both crops for which we found negative effects on yield and positive effects on L. hesperus when grown in a field the previous growing season, are all considered preferred hosts for L. hesperus , and have been shown to also increase L. hesperus populations in nearby cotton fields during an individual growing season. Presumably,draining pots this effect is due to these crops supporting large L. hesperus populations. Large L. hesperus populations are known to build up in alfalfa, and their dispersal following alfalfa harvesting can threaten nearby cotton crops. L. hesperus is also known to emigrate to nearby cotton fields when safflower begins to dry in mid-summer. While the potential for nearby alfalfa and safflower fields to increase L. hesperus populations in cotton fields in a given year has been recognized, our results are the first indication that these landscape effects may extend temporally, affecting L. hesperus populations, and yield, in the next growing season. Tomatoes, associated with increased yield and decreased pest abundance in our data, have likewise been shown to decrease L. hesperus abundances in nearby cotton fields within a given year. While previous experimental work has examined the effects of crop rotations on cotton yield, our work expands on these studies in several ways. First, we explore a much wider diversity of possible crop rotational histories, providing quantitative estimates of the cotton yield effects of cultivating 14 different crops the previous year. Second, since we analyze records from commercial cotton fields, our data have the potential to capture yield effects that could only be detected at this realistic spatial scale. Third, since we have collected data on pest abundances, not only yield, we have also been able to use our data to generate and build evidence for a hypothesized mechanistic explanation of the yield effects we identify. We also found that farmers incurred a decline in cotton yield of about 2.4% for every additional year cotton was grown consecutively in a field preceding the focal season . This is consistent with previous research suggesting that continuous cultivation of cotton in the same location can reduce yield compared to interspersing cotton with other crops.
We also found some evidence that the number of years cotton was grown consecutively in a field was associated with higher June L. hesperus densities: the posterior probability of there being a positive association was about 95%. Identifying the actual mechanism underlying this yield effect is beyond the scope of this study, but would be an interesting avenue for future research. It is possible that the yield decline is not caused by changes in L. hesperus densities, and instead results from the buildup of soil pathogens, especially in light of previous research showing that continuous cotton cultivation increases the densities of fungal pathogens in the soil When interpreting our results, it is important to remain cognizant of the challenges of drawing causal inferences from observational data. The key assumption required to make causal inferences from regression coefficients is that all variables that affect both the treatment assignment and the response variable are included in the model; this ensures that the probability of receiving each treatment becomes, conditional on the predictor variables included in the model, conditionally independent of the response variable. In experimental studies, the treatment assignment is typically controlled by the experimenter, so one can be confident that the only difference between treatment and control groups is in fact the treatment. However, in observational studies, it is impossible to prove definitively that there was no other factor that affected both the treatment assignment and the response variable . As such, we want to be very clear that our hypothesis that the effects of rotation on yield are mediated by effects on L. hesperus densities is exactly that – a hypothesis. While our data do support a negative association between effects on L. hesperus and effects on yield, we cannot prove with observational data that the varying effects on yield are caused by the varying effects on L. hesperus. This could be a fruitful topic for future experimental work. Although causality is impossible to prove using observational data, ecoinformatics paves the way for implementing data-driven agricultural strategies and allows us to mine large datasets to explore important questions that are difficult to address experimentally. While by no means a replacement for experimentation, ecoinformatics can be a cost-effective and realistic complementary approach. In particular, our result identifying the effects of crop rotation on L. hesperus density would have been extremely difficult to reach experimentally. Since L. hesperus readily disperse across spatial scales of more than 1000 meters, an experimental study would have required massive plots comparable to the size of commercial fields in order to adequately capture their spatial dynamics. Growers with knowledge of the crop rotations associated with depressed cotton yield could make more informed decisions, selecting the sequence of crop cultivations that lead to maximized yield. When feasible, cotton plantings could be avoided following crops that decrease cotton yield, and instead limited to fields where crops that increase cotton yield were previously planted. In some cases, market conditions may lead a grower to plant cotton following a yield-depressing crop, even given the knowledge of likely yield loss. In those situations, our results may still be helpful, as an early warning sign of a potential pest problem in a particular field could allow the grower and PCA to focus pest detection efforts on that field and provide time to eliminate the problem before severe yield loss was incurred.