Submitted to: Precision Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 9/16/2008
Publication Date: 2/1/2009
Citation: Willers, J.L., Jenkins, J.N., McKinion, J.M., Gerard, P., Hood, K.B., Bassie, J.R., Cauthen, M.D. 2009. Methods of analysis for georeferenced sample counts of tarnished plant bugs in cotton. Precision Agriculture. 10:189-212.
Interpretive Summary: This study suggests statistics derived from complete enumeration distributions or count models may be better for making pest management decisions than the mean (or confidence interval) of the original samples from the cotton field. Both methods modeled the sample counts as functions of categorical explanatory variables, which were the labels for two samplers and the cotton habitat categories simply called Marginal, Good or Best. Both methods demonstrated differences in TPB numbers among categorically defined cotton habitats derived from remote sensing imagery. These methods also showed that two observers preferentially allocated their sampling efforts among these habitats to answer different questions. Traditional correlation analyses of these counts failed to find an association between TPB counts and quantitative classes of cotton vigor. It is recommended that researchers and industrial investigators would benefit from applications of count model methods to analyze insect sample counts, while farm applications would benefit from the complete enumeration analysis. Both methods apply whenever large numbers of zeros occur among samples, exhibit a skewed distribution, or are not otherwise normally distributed.
Technical Abstract: The problem of how to analyze cotton pest insect samples when a large percentage of the collected samples are zero valued counts is examined. Geo-referenced samples (n=63) collected by two observers for tarnished plant bug (TPB; Lygus lineolaris [Palisot de Beauvois] (Heteroptera: Miridae)) were analyzed by several methods and compared. First, a traditional approach using correlation analysis was employed. This was followed by a complete enumeration analysis where three scenarios were established depending upon whether or not imagery or observer information was utilized. The first scenario assumed the insect samples were non-stratified. Here, a distribution of sample averages was built by complete enumeration of all combinations of samples taken 4 at a time. The second scenario used geo-referenced imagery of the cotton fields to apportion them into three cotton growth categories labeled ‘Marginal’, ‘Good’, or ‘Best’. Based upon their coordinate locations, these insect samples were completely enumerated using allocations of 4, 6, 8, or 10 samples at a time from various habitat sample sizes to complete a sensitivity analysis of how different allocations affected results. It was learned that results changed little for allocations greater than 4. Therefore, using allocations of 4 samples at a time, a third scenario apportioned the samples by the 2 observer and 3 habitats to build 6 other complete enumeration distributions. These enumeration distributions are non-parametric estimators of the sampling distribution of (1) the pest insect population average for the field, (2) the average of each cotton habitat or (3) the averages for each observer by habitat. To substantiate the enumeration analyses, these insect samples were further analyzed by Poisson regression models. The Poisson regression analyses indicated differences between the mean counts of TPB by the two observers over the habitats, while the observer by habitat interaction was non-significant. For every combination of observer and cotton growth category, the Poisson regression model estimated the mean rate of TPB numbers. These means were similar to the corresponding modes of the complete enumeration distributions. Both the enumeration and the Poisson regression results indicated the two observers, while using the same map for variability in cotton growth, independently selected different sample locations. This finding was confirmed by GIS overlays of the sample locations on the habitat map. Both methods demonstrated that TPB numbers differed by habitat categories, despite the occurrence of zero valued samples, while the correlation analysis approach failed to find differences.