Tuesday 16th July
Wednesday 17th July
Thursday 18th July
Friday 19th July
Download the conference book
Surveys and compositional data |
|
| Convenor | Dr Germa Coenders (University of Girona) |
| Coordinator 1 | Dr Valentina Hlebec (University of Ljubljana) |
Statistical compositions consist of positive data arrays with a fixed sum and which only convey information on the relative importance of each component. The commonest examples are proportions of the set of components of a total, whose sum can only be 1. Composition indicators are frequent in surveys. Among others we find surveys measuring compositions of household budgets, time-use surveys, and compositional indicators used in social network surveys, usually expressed in percentages of family members, friends, neighbours, co-workers and others in a social network.
Statistical analysis of compositional data is a challenging task because compositional data lie in a restricted space and components cannot vary independently from one another ("all other things constant"): the relative importance of one component can only increase if the relative importance of at least one other component decreases. A popular solution is to transform compositional data by means of logarithms of ratios of components prior to applying standard analysis methods, while taking great care in the interpretation of the results.
Simpler statistical methods such as ANOVA, linear regression and cluster analysis have a well documented tradition in compositional data analysis in the fields of geology and biology, among others. Less has been done in the survey research field, regarding, for instance, measurement models or structural equation models on compositional data. In these fields, the naive analysis of raw proportions is of common practice even if it is plagued with statistical problems (inconsistent inferences, heteroskedasticity, non-normality, censoring, perfect collinearity, and unclear interpretation, among others). The session aims to bridge methodological knowledge between the natural and social sciences in order to narrow this gap.
Proportions of a total, including social network compositions (proportions of partner, family, friends, etc.,) lie in a restricted positive unit-sum space, which challenges statistical analysis. Even if specialised techniques for compositional data analysis are starting to appear, compositional data can also be transformed so that they can be subject to standard and well understood statistical techniques. This is the approach we take here.
Network compositions can be both dependent and explanatory variables and are usually measured with error by survey instruments. Structural equation models make it possible to correct measurement error bias. With this purpose, Coenders et al. (2011) fitted a confirmatory factor analysis model to transformed network compositions using additive log-ratio transformation.
In this paper we use a more interpretable transformation called isometric log-ratio. We also extend the factor model to a full structural equation model including predictors and outcomes of network compositions. The findings and hypotheses in the literature can be reformulated with isometric log-ratios in a more interpretable manner, as no component can increase without some other decreasing. For instance, we find relationships of gender with partner support as opposed to the remaining family members, relationships of education and extraversion with friend support as opposed to colleagues, neighbours etc..., and relationships of family support as opposed to non-family support with tie multiplexity and tie closeness.
The increasing consolidation of Low Cost Airlines (LCA) means a significant reduction in travel costs, thereby modifying the composition of the traveller's trip budget. It is vital for destination marketers to understand an LCA traveller's expenditure composition. The trip budget shares considered are transportation and basic and discretionary expenses. In this article we use data from an official statistics survey of air travellers to Spain in 2010, the compositional data analysis methodology and latent class models to classify LCA expenditure shares into market segments. Then, we relate the classes to external variables by using a multinomial Logit for the respondent background predictors using the pseudo-draw method. We have found six classes of LCA users, ranging from those who spend almost nothing in discretionary expenses to those who mostly make their savings in transportation or to those who are quite similar to legacy airlines users. All the considered background variables affect the probability of being a member of the different classes.
The use of survey questions with children may sacrifice quality for quantity, if responses to closed questions are not valid. Three surveys (with adults, 16 year-olds, and 10/11 year-olds) in Northern Ireland asked respondents about caring responsibilities. The question wording varied slightly across surveys, reflecting differing age-groups and mode of administration. Counter-intuitively, a higher percentage of children than older respondents identified caring responsibilities, suggesting that younger children interpreted the question differently. This is borne out in the follow-up open ended question. This paper will examine the possible implications of adapting closed questions for children
There is substantial interest in the relationship of cognitive ability and ethnicity. The purpose of this study was to identify problems associated with cognitive ability measures in respondents of multiple ethnic groups, with or without English as their primary language, before implementing a module of cognitive measures in the UK Household Longitudinal Study (UKHLS).
Cognitive interviewing was conducted with 43 participants who varied in age, gender, primary language, and ethnic group. The cognitive ability measures included tests of prospective, working and episodic memory; fluid intelligence; phonemic fluency; and numeric ability. Problems identified in questions, instructions, or response format included lack of comprehension of specific words, and problems with English spelling. There were also reports of variation in familiarity with the format and tasks of cognitive ability tests as shown in statements and requests for repetition and clarification. Problems with the cognitive ability measures were more clearly associated with language than with self-identified ethnic group. Collection of several auxiliary variables to characterize the interview situation was recommended. Solutions recommended for the identified problems were implemented in the UKHLS. These modifications may reduce some sources of differential measurement error confounding ethnic comparisons on cognitive ability without unduly modifying measures that have been implemented in other surveys. The validity of cognitive ability measures for different ethnic groups, particularly those not interviewed in their primary language has not been assessed.