Comparative Assessment Of Statistical Tools For Handling Categorical Response Variables From Farmer Participatory Trials

Abstract

This project looks at responses from participatory on-farm farmer participatory trials that are often measured as ratings farmers score each given treatment on a scale that is ordered but arbitrary or rankings where farmers arrange the treatments in order from most preferred to least preferred. Simple methods such as the preference statistic that uses the proportion of responses where treatment A is preferred to B, Kruskal- Wallis test which is a one way analysis of variance by ranks and the Friedman test that is a two way analysis of variance by ranks are described and the shortcomings with linear model based analyses for these type of data are outlined. The Bradley- Terry model for ranks which is a logit model for paired comparisons is described and used to fit models for plot level covariates. The cumulative logit model has also been given as the appropriate model for analyzing rating data and models have also been fitted to the data. The fitted Bradley-Terry model for treatments was significant but plot level covariatcs did not significantly affect the rankings. The ordinal regression model fitted to treatments was not significant but farmer effects fitted with treatments as main effects were significant.