Containment testing

Host range testing methods

Statistical analysis

Most of the traits to be analysed in biocontrol do not follow a Gaussian ('Normal') distribution, and thus standard t-tests, analysis of variance (ANOVA) or regression analyses cannot be used with validity to statistically test the effect of a treatment. All these common 'classical' methods assume that the distribution of residuals around the fitted model (i.e., the error distribution) is normal. Thus data need to be appropriately transformed (Fernandez 1992) to first achieve a Gaussian distribution or preferably different statistical approaches have to be used (Hoffmeister 2005).

Non-parametric tests like Mann-Whitney U tests or Kruskal-Wallis tests are often the most appropriate test for host specificity testing data, and are more than adequate if the results are clear-cut. However non-parametric tests lack statistical power, which can mean that real, but more minor, differences existing between, for example, test species and target (control) species will tend to be interpreted as not significantly different by a non-parametric test, when in fact this might not be the case. Non-parametric tests will almost certainly have insufficient power to analyse data if sample sizes were forced to be small due to other factors out of the researchers' control. Generalized Linear Models can be used to predict responses both for dependent variables that are not normally distributed and for dependent variables which are nonlinearly related to the predictors (Hoffmeister 2005). Expert statistical advice will probably be required to both appropriately utilise and interpret such models.

Choice tests are often invaluable for undertaking tests on the preference hierarchies shown by ovipositing or feeding insects. However choice tests are problematic for statistical analyses (Mansfield and Mills 2004, Withers and Mansfield 2005). The same individual insect is confronted with both target and non-target species, and we usually wish to obtain comparative acceptance rates for both target and non-target species as a result of such a test. Thus, a repeated measure design must be used. This renders the data dependent (not independent as required for Gaussian statistics, see above). The acceptance of non-target hosts or prey within such a test may well depend upon the frequency of target and non-target hosts within the cage. Therefore every target that is eaten or accepted and that is not immediately replaced alters the experimental conditions of the experiment, and the acceptance of any given host or prey may depend on the current ratio of availability of alternative hosts or prey. Thus if exploited hosts or prey cannot be replaced immediately, simultaneous choice tests may become almost impossible to interpret (Hoffmeister 2005).

An alternative to simultaneous choice tests are sequential no-choice tests (Singer 1986). This design does have appropriate statistical analysis methods available to treat and analyse the data (Hoffmeister 2005). However sequential no-choice tests generate a multitude of potential behavioural complications which may be almost impossible to control for or adequately measure (Barton-Browne and Withers 2002, Withers and Barton-Browne 2004, Withers and Mansfield 2005).

References

Barton-Browne L. and Withers T.M. (2002). Time-dependent changes in the host-acceptance threshold of insects: implications for host specificity testing of candidate biological control agents. Biocontrol Science and Technology 12: 677-693.

Fernandez G.C.J. (1992). Residual analysis and data transformation: Important tools in statistical analysis. Horticultural Science 27: 297-300.

Hoffmeister T.S. (2005). From design to analysis: effective statistical approaches for host range testing Pp. 672-682 In: Second International Symposium on Biological Control of Arthropods, Davos, Switzerland, 12-16 September, 2005, M.S. Hoddle (Ed.) United States Department of Agriculture, Forest Service, Washington.

Mansfield S. and Mills N.J. (2004). A comparison of methodologies for the assessment of host preference of the gregarious egg parasitoid Trichogramma platneri. Biological Control 29: 332-340.

Singer M.C. (1986). The definition and measurement of oviposition preference in plant-feeding insects. Pp. 65-94 In: Insect-Plant Interactions, J.R. Miller and T.A. Miller (Ed.) Springer-Verlag, New York.

Withers T. and Mansfield S. (2005). Choice or no-choice tests? Effects of experimental design on the expression of host range. Pp. 620-633 In: Second International Symposium on Biological Control of Arthropods, Davos, Switzerland, 12-16 September, 2005, M.S. Hoddle (Ed.) United States Department of Agriculture, Forest Service, Washington.

Withers T.M. and Barton-Browne L. (2004). Behavioral and physiological processes affecting the outcome of host range testing. Pp. 40-55 In: Assessing host ranges for parasitoids and predators used for classical biological control: a guide to best practice, R.G. Van Driesche and R. Reardon (Ed.) USDA Forest Service, Morgantown, West Virginia.