Statistics for Categorical Surveys—A New Strategy for Multivariate Classification and Determining Variable Importance

Abstract

Surveys can be a rich source of information. However, the extraction of underlying variables from the analysis of mixed categoric and numeric survey data is fraught with complications when using grouping techniques such as clustering or ordination. Here I present a new strategy to deal with classification of households into clusters, and identification of cluster membership for new households. The strategy relies on probabilistic methods for identifying variables underlying the clusters. It incorporates existing methods that (i) help determine the optimal cluster number, (ii) directly identify variables underlying clusters, and (iii) identify the variables important for classifying new cases into existing clusters. The strategy uses the R statistical software, which is freely accessible to anyone

Adaptive Rangeland Decision-Making and Coping with Drought

“…This method uses tree structured regression analysis to identify variables likely to predict membership probabilities for each previously identified latent class. This analysis provides a non-parametric class of regression trees that accommodates multiple data types (including categorical and numeric data) and large numbers of candidate predictor variables, as well as allows examination of potential interactions [41][42][43][44].…”

When households stop logging — Evidence for household adaptation from East Kalimantan

“…This enabled the development of a decision tree (with a limited set of variables) through which a typology could be assigned to each household in the area. Details of the cluster analysis are provided in Herr (2010) and Smajgl and Bohensky (2011).…”

On-Ranch Grazing Strategies: Context for the Rotational Grazing Dilemma

“…This analysis uses treestructured regression to identify variables likely to predict respondents' membership probabilities (continuous response variable) for each class of grazing strategy determined from the LCA analyses. This analysis accommodates nonparametric data and large numbers of candidate predictor variables, is appropriate in cases that include data collected in categorical and nominal forms, allows examination of data that potentially interact in a complicated and nonlinear fashion, and recursively partitions the overall variance to form groups of similar responses (Cutler et al, 2007;De'ath and Fabricius, 2000;Herr, 2010;Hothorn et al, 2006;Strobl et al, 2009).…”