Studying political parties in different countries

Georgia writes,

I [Georgia] have data on political party attributes for about 55 parties in 15 countries. I will expand this, but at most will have data for about 75 parties in 18 or 19 countries. This is primarily data I collected myself. I also have individual-level data from the CSES survey. For each individual, this includes information about their demographics, which party they identify with, if any, which party they voted for, if any, their assessments (feeling thermometer) of all parties, as well as their placement of their own position and the parties’ positions in their country on a left-right ideological scale.

The question I would like to answer is: what party level and individual level factors make an individual likely to identify with a political party? I am particularly interested in party level factors. For example, are individuals more likely to identify with a party that is ideologically on the left? Or, are individuals more likely to identify with a party if it has a centralized organization (such as a single party manifesto or candidates that are nominated by the national leader)? I think that a lot of these things interact with individual level characteristics. (If you remember, this is the same question John, Eduardo and I wrote about in the PA paper, except we were examining country level, not party level, characteristics. Also, our question then was more “what makes someone a partisan?” and now I’m asking “what makes someone a partisan and to what party are they a partisan of?”)

What I’ve done so far is two things.

First, I’ve used a multilevel model where individuals are within parties within countries. In order to group individuals within parties, I only examine voters and pair them up with the party they voted for in the last election. I drop all non-voters, and I also drop individuals that identify with a party other than the party they voted for. So, I only have voters that identify with the party they voted for and nonvoters. Then, I examine what makes some voters partisans and others nonpartisans. The problem with this model is that I really don’t want to restrict my analysis to those individuals that voted. (Also, I’m pretty sure there is a problem with not dropping nonvoters that would have pid of a party other than the one they voted for if they were to have pid. I do think I could deal with this problem, though, by deleting nonvoters that look like voters that have pid of a party other than the party they voted for using propensity score matching. This isn’t the main problem though since the model isn’t really getting at what I want.)

The second thing I did is use the feeling thermometer variable (which asks respondents to rate a party from dislike=0 to like=10) as my dependent variable. For this analysis, I have multiple observations for each individual since they were asked to rate all parties in their country. So, I have two non-nested levels (individuals and their responses about parties) nested within countries. The problem with this model is that I’m not measuring pid.

The problem that I’m encountering is that I can’t figure out a way to have pid (=1 yes or =0 no) be the dependent variable and include party level factors. If I use the individual as the unit of analysis (with multiple observations), the dependent variables are not only not independent, but they are extremely dependent – a respondent can only have party identification for one party (by survey design).

One thing I thought about doing was for each person in my data, randomly assigning them to one party in their country and using only their responses for that party. That way they’d either be a partisan of that party, or they wouldn’t be, and I wouldn’t have to worry about multiple observations for them. I could even use their attitudes toward other parties (such as the feeling thermometer scores for other parties) as predictors. I think this would mean that I’d have non-nested individuals and parties, nested within countries. Can you let me know if this is a feasible thing to do, or do you have any other suggestions? And, if I do this, how could I go about randomly choosing observations?

My reply: Basically, you have a mulitlvariate outcome (attitudes towards all the parties) and also a multinomial outcome (which party is preferred) along with some other outcomes such as whether the person voted. I certainly would not recommend randomly discarding any data! I think you’d want to model the joint outcome, possibly with some spatial model. Another way to go, as a starting point, is to analyze the responses on each party separately for each country.