Monday, 21 July 2003

Not-so-sweet sixteen

Daniel Drezner has a challenge to those have criticized his take on the whole “sixteen words” theme that the left has been trying to make fly for the last week:

The power of the critique against Bush would be strengthened if it could be shown that a significant fraction of the American public—as well as the legislative branch—supported action against Iraq only because of the claim that Hussein’s regime had an active nuclear weapons program.

Ok, since I’m likely to be terribly bored at some point in the next day or two, and considering I’m sitting not-very-far from the computers the data is housed on, I’ll look at the February 2003 CBS/New York Times Poll, along with several others from the period after the State of the Union, and see what I can find. I can’t give any evidence on the behavior of legislators, but I can at least examine whether the public’s opinion was conditional on WMD, and nukes in particular—assuming the right questions were asked.

Warning for the faint of heart: I may present regression results in addition to the marginals.

Thursday, 24 July 2003

Public opinion is (almost) meaningless

As promised, I went off and played with the data on whether or not voters believed that Iraq's WMD threat justified war. Since it had the most questions on the issue, I used ICPSR Study #3755, better known as the second March 2003 CBS/New York Times telephone poll (1010 total respondents; conducted March 5-7). As I anticipated, no questions directly dealt with whether or not Iraq had been obtaining bits and pieces for nuclear weapons, or even mentioned the word "nuclear" at all; the popular phrase at the time was "Weapons of Mass Destruction," a term left undefined by any of the questions.

I'm extremely reluctant to present results for two main reasons. The first is that there was a huge amount of missing data; many respondents failed to answer a number of the questions, so there is less information available (I worked around this problem by using a Bayesian data augmentation model rather than a typical maximum-likelihood approach). The second reason is that I believe the question wording of most of the items on the survey make a recursive model somewhat inappropriate; many of the questions appear to tap the same underlying dimension, which is basically whether or not the respondent trusts the administration, and the causality is not at all clear. Nevertheless, there are a few interesting findings. So here goes:

Independent Variable Coefficient 95% Credible Interval
UN Handling of Crisis -0.358 -1.063 0.353
US has presented sufficient evidence of WMD 1.335 0.442 2.257
R trusts Bush to handle Iraq issue 1.389 0.558 2.351
R believes Iraq represents a WMD threat to US (3pt) 1.065 0.414 1.771
R believes admin telling all it knows 1.016 -0.097 2.198
R believes Saddam Hussein personally involved in 9/11 0.480 -0.306 1.308
R believes inspectors not making progress (4 pt) 0.511 -0.002 1.033
R party identification (0=Strong Rep; 2=Indep; 4=Strong Dem) 0.026 -0.198 0.262
Male respondent 0.029 -0.361 0.705
Education level of respondent (1-5) -0.066 -0.369 0.226
Age of respondent -0.018 -0.038 0.001
(Intercept - used for model identification) -2.971 -5.322 -0.767

(Coefficients are probit coefficients, as the dependent variable is dichotomous. All variables are yes/no "dummy" variables unless otherwise described.)

A few words for those who aren't accustomed to regression results. The left-hand column is the variable that is believed to have an independent effect on the dependent variable (in this case, support for the U.S. going to war in Iraq). The next column shows the magnitude of the mean effect of that variable. The final columns show the "credible interval" (similar to the "confidence interval" in frequentist interpretations), which basically says that there is a 95% chance that the true coefficient lies within that range of values. If the credible interval doesn't include zero (i.e. both values are positive or both are negative), we can say that at least 19 out of 20 times, the effect in the population at large would be in the direction of the sign (i.e. positive or negative).

What does this model tell us? Generally speaking, people who believed the U.S. had presented sufficient evidence of Iraq's WMD programs, trusted Bush to handle the Iraq issue, and believed Iraq's weapons to the a threat to the U.S. were more likely to support a conflict than those who didn't. (This finding was also robust across all of the various scenarios for war proposed in the survey; the dependent variable here posited no particular configuration of events.) A couple of other effects approach significance: belief that the administration was fully forthcoming, that the UN inspections were ineffective, and the respondent's age (older voters being less likely to support a war).

More interesting is what it doesn't tell us. The effects of education and gender are insignificant; men were no more likely to support the war than women, and more educated people are no less likely than less educated people. Perhaps most interesting, and possibly problematic for anti-war Democrats, especially if the "Bush lied" theme fails to stick, is that there was no independent effect of partisanship; this suggests that "Bush Democrats" (we might call them "WOT Democrats" if we wanted to be cute like Larry Sabato) were just as enthusiastic for war as "Bush Republicans." And, the people who believe Saddam was personally involved in the 9/11 attacks were no more likely to support war than those who didn't.

In my next post, I will present a more interesting model from the same dataset, looking at the question on Saddam and 9/11.

* The data for this analysis were obtained from the Inter-university Consortium for Political and Social Research. All errors in interpretation or analysis are the sole responsibility of the author. The analysis was conducted using GNU R and the MCMCprobit procedure in the MCMCpack package.