Associate Professor Sama Low-Choy
This workshop comes fourth in a series of workshops on the statistical analysis of Likert-scaled survey data. Although all of these workshops use R, we focus on interpretation of outputs, which are similar across packages. So participants can ignore the "R" code and sit next to someone else producing the outputs.
This session builds on three previous workshops:
1. "Excellent Graphics" - Tufte's principles of how to create effective visualisations of your data: before, during and after statistical analysis.
2. "Qualitative validation of surveys" - Sometimes a set of items in a survey all measure different aspects of the same concept or "construct". This session starts with simplistic use of Cronbach's alpha and progresses to correlations (numbers or pictures) and a taste of factor analysis.
3. "Analysis of Survey Data: beyond Chi-squared tests" - Whilst Likert scale data is commonly obtained in surveys, its analysis requires finesse. This session starts with simplistic use of Chi-squared tests, and shows their limitations. We examine how to "construct" a score from multiple items, and consider how this score depends on various factors, via regression (of constructs) and decision trees.
This workshop takes the next step but revisits these three previous workshops where relevant, to put things into context.
Instead of considering a score, we focus on an individual Likert-scale item, and analyse how it may relate to other factors (including demographivs, other Likert-scaled items or numeric variables like Age or other constructs). These models are called multinomial regressions because they model how an outcome, having multiple possible categories (e.g. Strongly disagree to Strongly Agree), depends on other factors.
A multinomial regression for ordinal data is related to binomial regression for binary data, which is covered in detail in two other workshops (Bayesian Logistic Regression, Part 1: Meaning & Elicitation; Part 2: Inference & Computation.)