### Comparing Ideology and Partisanship Across State Legislatures

Bumped--Chris

Andrew Gelman at Statistical Modeling, Causal Inference, and Social Science points to a very interesting article by Christopher Berry and Nolan McCarty Mapping State and Congressional Ideology looking into ideology and partisanship as measured by voting patterns within state legislatures. What is really clever and useful is that the authors have scaled the State ideological dimension to the US Congress, allowing us to compare the partisanship between state and federal levels as well as between different states.

How do they do the scaling? Using "bridge" politicians who have graduated from their state legislature to the US Congress. As Andrew comments:

Cool. This is sort of like those things where people compare Babe Ruth to Mike Schmidt, or whatever: Ruth played with Gehrig, who played with etc etc., going up to the present time. I guess the next thing to do is check that the "bridge actors" identified by Shor et al. are not systematically different than other legislatures, for example, in changing their attitudes when moving up to the big House.

Not all states provide good databases of voting records, so they only have a few states mapped: PA, FL, MI and CA. The short of it? Measured by median voting records, MI Republicans are very conservative, CA Democrats are fairly Liberal, and PA Dems are relatively more to the middle, and both PA and FL Reps share the same, moderately conservative voting record.

For deep statistical discussions, I really enjoy the blog Statistical Modeling, Causal Inference, and Social Science which writes a lot on statistics applied to political analysis. These guys are <hush> Bayesians </hush>, a branch of statistics with alchemical reputation when I was back in school.

Here's the abstract:

Two major problems exist in applying ideal point estimation techniques to state legislatures. First, there has been a scarcity of available longitudinal roll call data. Second, even where such data exists, scaling ideal points within a single state suffers from a basic defect. No comparisons can be made across institutions, whether to other state legislatures or to the US Congress. Our project is a solution to both of these dilemmas. We use a new comparative data set of state legislative roll calls beginning in the mid-1990s to generate ideal points for legislators. We then take advantage of the fact that state legislators sometimes go on to serve in Congress to create a common ideological scale between Congress and the various legislatures. These "bridge actors" are similar in concept to members of the House who go on to serve in the Senate, thereby providing the "glue" necessary to scale the House and Senate together. We have successfully prototyped this approach for California, Pennsylvania, Michigan, and Florida. Using these bridge actors, we create a new state-federal congressional common space ideological scores. We conclude by using these common space scores to address important topics in the literature.

The money quotes start at page 16 in the PDF of the article:

Now, for the first time, we can directly compare the results from different states with each other, as well as with the US House. We do so first by comparing the range of ideological preferences in each institutional setting. The US House is constrained by the NOMINATE procedure to lie in the (-1,1) range. Unlike scaled scores, predicted scores can range beyond the (- 1,1) range of NOMINATE. Therefore, California's and Michigan's most conservative Republicans are quite conservative indeed by congressional standards, reaching out as far as 1.5 on the first dimension. In contrast, Florida has an ideological range that look more like that of Congress.

Second, we compare medians of each state's bichamber "legislature." The congressional median is at 0.5, understandable as the House was dominated by Republicans over the course of 1996-2006. Florida, Michigan, and Pennsylvania also have moderately conservative institutional medians over their respective time periods, while California has a moderately liberal median. The results largely echo Berry et al. (1998)'s elite ideology scores for the states (see below).

Third, we can compare the party medians. Michigan's Republicans stick out by being extremely conservative, while the other states' Republicans mirror the House. Democrats are far more diverse, ranging from less liberal in Pennsylvania and Michigan to most liberal in California. Florida sticks out again, and looks most like a microcosm of the US House, with identical Democratic and Republican medians.

Fourth, we can say something about partisan polarization which has become such a hot topic in American politics (McCarty, Poole and Rosenthal 2006). The baseline remark, of course, is that preferences are distributed bimodally. Beyond this simple fact, what more can we say? One way to compare polarization across institutional settings is to check the distance between party medians. Pennsylvania's party medians are closest together, California's are furthest apart, and Florida's look like the US House.

##### The Comparison Method Comes From DW-Nominate

I haven't checked it out yet, so I don't know if the origin is direct.  But the DW-Nominate scoring method was developed to compare congress members across congresses from the first congress down to today.

Check out the website:

www.voteview.com/dwnomin.htm

and the text file

http://www.voteview.com/page2a.htm

which goes into its origins and limitations.

by Paul Rosenberg 2007-04-21 11:20AM | 0 recs
##### Or...

To normalize House to the Senate ideology.

by MetaData 2007-04-21 01:43PM | 0 recs
##### Please note that Blue-Red is reverse of normal.

Here in Colorado, I would guess that the Republicans would be very Conservative and the Democrats moderately Liberal, at least within the legislative branch.

Another interesting thing in the article, is the observation that Legislator ideology doesn't evolve, rather it changes as candidates die or are replaced. That is why the bridge politicians provide such a highly correlated estimator for mapping from the Federal level back to the states.

by MetaData 2007-04-21 01:51PM | 0 recs

I think for the NOMINATE model, the stability of a legislator is assumed rather than deduced.  I believe that DW-NOMINATE allows for linear trends in legislator ideal points over time, but when you do that, you tend to find that they don't move much; NOMINATE just assumes that they stay put, as I vaguely recall.  The cool thing to me is that, looking at Figure 5 of the paper, when you regress the NOM scores for bridge politicians in the state and House legislatures, boy, they match pretty closely -- even the slopes and intercepts are pretty close to 1 & 0.  I wonder if and when politicians will learn to use this data in choosing who to help boost up to a House seat.

by brackdurf 2007-04-21 03:08PM | 0 recs
##### Even more than that,

(Sorry, in the above plots the Blue-Red coding follows the standard. OTHER plots in their paper have the colors flipped.)

They did look at the possibility of other dimensions, which are vaguely suggested in their statistical analysis, but abandoned them. The correlation of the Liberal-Conservative dimension is so strong, that bringing in additional dimensionality has virtually no effect.

The whole paper is a must-read for MyDD wonks.

Watch the time-series of partisan divide for the four states. Even before the 2006 elections, California steadily goes more partisan. It is interesting to note that in Pennsylvania, the least polarized state of the study, the divide remains virtually stable over time.

by MetaData 2007-04-21 04:35PM | 0 recs
##### Re: Even more than that,

What's odd to me is that, from figure 9 panel 1, California's partisan divide obviously goes way up over time.  But looking at Figure 1, panel 1 -- the skree plot for California alone -- it seems to have the most multi-dimensionality going on of all the states.  I don't quite know how to reconcile these two things, though you can see a little of what's going on in the horizontal gap in the middle of figure 2, panel 1 (ie, there are clearly two democratic and republican factions, as well as the d/r divide).

Also, on an even more technical note, an APRE of 53% for PA suggests that something is wonky with that state.  Don't know what that means either though...

by brackdurf 2007-04-21 04:52PM | 0 recs
##### Damn, I'm just fevered (literally)

When I can't read colors and interpret them, its time to hang up the keyboard.

by MetaData 2007-04-21 04:37PM | 0 recs