November 2005
Sun Mon Tue Wed Thu Fri Sat
    1 2 3 4 5
6 7 8 9 10 11 12
13 14 15 16 17 18 19
20 21 22 23 24 25 26
27 28 29 30      

Authors' Committee

Chair:

Matt Blackwell (Gov)

Members:

Martin Andersen (HealthPol)
Kevin Bartz (Stats)
Deirdre Bloome (Social Policy)
John Graves (HealthPol)
Rich Nielsen (Gov)
Maya Sen (Gov)
Gary King (Gov)

Weekly Research Workshop Sponsors

Alberto Abadie, Lee Fleming, Adam Glynn, Guido Imbens, Gary King, Arthur Spirling, Jamie Robins, Don Rubin, Chris Winship

Weekly Workshop Schedule

Recent Comments

Recent Entries

Categories

Blogroll

SMR Blog
Brad DeLong
Cognitive Daily
Complexity & Social Networks
Developing Intelligence
EconLog
The Education Wonks
Empirical Legal Studies
Free Exchange
Freakonomics
Health Care Economist
Junk Charts
Language Log
Law & Econ Prof Blog
Machine Learning (Theory)
Marginal Revolution
Mixing Memory
Mystery Pollster
New Economist
Political Arithmetik
Political Science Methods
Pure Pedantry
Science & Law Blog
Simon Jackman
Social Science++
Statistical modeling, causal inference, and social science

Archives

Notification

Powered by
Movable Type 4.24-en


« Social Science and Litigation, Part IV | Main | Applied Statistics - No Meeting »

18 November 2005

British Ideal Points

Mike Kellermann

We have talked a bit on the blog (here and here) about estimating the ideal points of legislators in different political systems. I've been doing some work on this problem in the United Kingdom, adapting an existing Bayesian ideal point model in an attempt to obtain plausible estimates of the preferences of British legislators.

The basic Bayesian ideal point model assumes that politicians have quadratic preferences over policy outcomes; this implies that they will support a proposal if it implements a policy closer to their ideal point than the status quo. Let qi be the ideal point of legislator i, mj be the location of proposal j, and sj be the location of the status quo that proposal j seeks to overturn. The (random) utility for legislator i of voting for proposal j can thus be written as:

sj2 - mj2 + 2qi(mj - sj) + eij

Or re-written as

aj + bjqi + eij

With the appropriate assumptions on the stochastic component, this is just a probit model with missing data in which the legislator votes in favor of the proposal when the random utility is positive and against when the random utility is negative. Fitting a Bayesian model with this sampling density is pretty easy, given some restrictions on the priors.

Unfortunately, applying this model to voting data in the British House of Commons produces results that lack face validity. The estimates for MPs known to be radical left-wingers are located in the middle of the political spectrum. Party discipline is the problem; the influence of the party whips (which is missing from the model) overwhelms the policy utility.

I try to address this problem by moving to a different source of information about legislative preferences. Early Day Motions allow MPs to express their opinions without being subject to the whips. EDMs are not binding, and can be introduced by any legislator. Other legislators can sign the EDM to indicate their support. There are well over 1000 EDMs introduced every year, which greatly exceeds the number of votes in the Commons.

We can't just apply the standard ideal point model to EDM data, however, because there is no way for MPs to indicate opposition to the policy proposed in an EDM. Instead of 'yea' and 'nay', one observes 'yea' or nothing. In particular, it is clear that some Members of Parliament are less likely to sign EDMs, regardless of their policy content. I model this by adding a cost term ci to legislator i's random utility.

sj2 - mj2 + 2qi(mj - sj) + ci + eij

This is a more realistic model of the decision facing legislators in the House of Commons. In this model, the proposal parameters are unidentified; I restrict the posterior distribution for these parameters by assuming a prior distribution that assumes the sponsors of EDMs make proposals that are close to their ideal points.

I'm still finalizing the results using data from the 1997-2001 Parliament, but the results on a subset of the data seem promising; left-wingers are on the left, right-wingers are on the right, and the (supposed) centrists are in the center. These estimates have much greater face validity than those generated from voting data.

If you are interested in this topic, I am going to be presenting my preliminary results at the G1-G2 Political Economy Workshop today (Friday, November 18) at noon in room N401. By convention, it is grad students only, so I hope there are not too many disappointed faculty out there (sure...).

Posted by Mike Kellermann at November 18, 2005 3:09 AM