My thesis-partner and I examine context level and individual level predictors effect on populist right-wing voting (i.e., Danish People's Party in Denmark) applying a multilevel logistic regression model with three levels (individuals nested in municipality-elections nested in municipalities). We use a pooled sample encompassing five elections (2001, 2005, 2007, 2011, 2015).
We follow the approach of Fairbrother (2014), separating the context effects into a cross-sectional and a longitudinal component, so change is measured at the upper-level. In this model, we were not able to estimate cross-level interactions - possibly due to low between-cluster variance (our model did not converge).
However, applying a multilevel linear regression model by substituting the binary dependent variable with a scale variable (from 0 - 10, with 0 indicating no sympathy for the party and 10 indicating sympathy for the party), we were able to include two cross-level interactions and an accompanying random slope of educational attainment (this is a robustness check).
Our question is how to interpret the results?
We interact the individual level variable educational attainment (six point scale), with the longitudinal component of the context variable, unemployment rate. The included variables are grand-mean centered.
We find that the conditional slope of the individual level predictor, educational attainment, to be significant at p
Three comments that are meant to be helpful!
1 I find it much easier when there are interactions to interpret results by using a predicted value plot of the variables involved in the interaction, keeping all the other variables in the model at say their modal or mean value. It is particularly hard to interpret confidence intervals around the predicted line without some visualization. The SEs for the individuals dummies are of course at some value of the continuous interacting value and will be different at different values, that is the point of the interaction. And you cannot read that directly froma table of results. To that end we built a tool - the 'customised prediction' facility in MLwiN to do exactly that
Article Manual supplement for MLwiN Version 2.14
see Article Investigating the Macro Determinants of Self-Rated Health an...
in which all the graphs were produced with MLwiN2 On non convergence with logistic,; I usually use MCMC estimation in this situation ; Book Developing multilevel models for analysing contextuality, he...
3 Given the importance of time when it is included, I would have interacted education with time, but you may well have done that and found it not to be important.
- Badges
- Science topic
- Similar topics
- Social Science
- Sociology