Multi level modelling, conceptual questions I am studying national election campaigns of individual candidates by means of an elite survey (N around 500). Due to the nature of my country case, there are several contextual factors i would like to place special emphasis upon within my IVs. These factors also give the data an hierarchial structure. Candidates in constituencies (over 200) and constituencies in regions (16), that is. Further, candidates are, of course, members of different parties. What i am especially interested in is, whether my DV (some latent variable measuring personalization of the campaign) varies over constituencies and regions and whether the effect of different party memberships on the DV varies over the context factors.
My question now is, if a multilevel model would be adequat for investigating these relationships. Naturally there are very low sample sizes on the constituency level, but from what i understood, this should not be a problem for multilevel modelling. Also, how do i link the party with the district/ regional level. By means of interactions?
Kind regards
 A: Multilevel models work well with observational survey data as you've described. 
Wrt small sample sizes and depending on the functional form of the hierarchical model more (or less) information is required for each cross section. E.g., frequentists such as Singer and Willett in their book Applied Longitudinal Data Analysis argue that at least 3 data points are required to estimate any effect or trend, the classic position. On the other hand Gelman and Hill in chap 13 of their book Data Analysis Using Regression and Multilevel/Hierarchical Models contend that by using a Bayesian approach effects for cross sections even with an n of 1 are still estimable -- in the posterior. 
Linking party at the district/regional level is a structual data issue that requires party information within that greater level of granularity. It sounds like your data would have that structure. Given that, ANOVA-type interactions of these categorical fields would be the way to go, subject to the limitations of your approach (frequentist vs Bayesian) and the progressively smaller sample sizes as more and more cross-classifications are tested. 
You don't indicate how many parties are present but, for the sake of argument, let's assume a three party system for your country. With an N of 500, 200 constituencies, 16 regions and 3 parties, you will have to be very careful which interactions are tested as even the simplest interaction of 16 regions by 3 parties results in 48 possible cells to populate with an expected 10 respondents per cell (500/(16*3)). Of course your data will not be so smooth and evenly distributed so there will be many cells with zero respondents and a few cells with a large number of respondents.
