A few days ago, I wrote a post that included a link to some data on the health care vote assembled by the Washington Post. Wonk and researcher that I am, I decided to crunch some numbers to find out if the data could be used to predict how House members voted on the bill and to predict which members changed their vote between the original House bill and the House's vote on the Senate bill.
The variables available to me in this limited dataset were: the two votes on health care reform, the state the member represents, the member's political party, financial contributions the member received from health care lobbyists, and the percent of persons uninsured in the member's c congressional district.
A quick look at the descriptive statistics finds that 41.3% of members are Republicans, while 58.7% are Democrats. The average House member comes from a district in which 16.9% of persons are uninsured (the median was 15.8%). The average amount of total contributions from health care lobbyists received by a House member was $487,800, although this figure was pulled upwards by a handful of members--the median for total contributions was quite a bit lower at $334,733. Only 14 members changed their position on the issue between the first and second House votes: 8 switched from "No" to "Yes" and 6 switched from "Yes" to "No."
Now for the more interesting stuff: predicting votes. In most cases with a dichotomous (i.e., 0/1) outcome, it is appropriate to use what is known as a logistic regression model (which is non-linear). However, because all Republicans voted against the bill, being Republican is a "perfect predictor" of the outcome and the model drops all the Republicans from the analysis. Because I didn't want that to happen, I decided to run instead what is called a linear probability model. There are some drawbacks to using an LPM (e.g., such models can predict impossible outcomes that fall outside of the 0/1 range) but for this type of back-of-the-envelope analysis, it's an acceptable way to go.
First, I tried to predict the outcome of the vote using party affiliation, monetary contributions, and proportion of uninsured constituents as explanatory variables. I found that we have explained nearly 73% of the variation in the voting outcomes using just these few variables. The catch is that it looks like party affiliation (as indicated by "rep") is doing all the explaining. Neither the percent uninsured or the financial contributions are significant. So, in this first model, how a House member voted was not influenced by the money they received from lobbyists or the needs of their constituents, it was purely political.
But what if party affiliation is overshadowing these other important factors? For example, what if the percent uninsured in the community matters, but it only matters for Democrats? I assessed this by running a model interacting both percent uninsured and contribution amount with party affiliation, but neither interaction was significant (results not shown).
It's possible, however, that there may be a number of unobserved factors that affect the outcome of the vote that are not captured by these very few variables. Many of these things may exist at the level of the state. For example, members from Nebraska might have been more likely to vote for the bill if it still contained the "Cornhusker Kickback" but less likely given the removal of that provision promised by reconciliation. For this reason, I ran a model with state-level fixed effects that controls for anything and everything at the state level that doesn't change over time but may effect the outcome of the vote. I found that such state level factors do seem to exist, but controlling for them hasn't changed the original conclusion that party affiliation is the only thing that matters.
Next, I turned my attention to trying to predict who changed their votes. I ran models to predict any change as well as specific changes from "No" to "Yes" and from "Yes" to "No." Again, I found that only party affiliation was significant, with Republicans less likely to have changed their vote. That's pretty understandable when you consider that none of them changed their vote, having consistently voted against the bill.
So what can one take away from this rather simplistic analysis? I think there are two conclusions: First, either money doesn't matter at all, or it matters equally and the enormous sums spent on both sides of the aisle offset each other at the end of the day. Second, our elected officials did not vote for health care reform because a large number of their constituents are uninsured, nor did they vote against health care reform because most of their constituents are already insured. While one might expect that the greater the need for reform as observed on the ground in a member's district the greater the likelihood that he or she would have voted for reform, this just isn't borne out by the data. Actually, I don't think that's too surprising. After all, how likely is it that the uninsured will be turning out in droves this November?
So the optimists would like to think that our representatives represent the needs of their constituents--but that doesn't seem to include the uninsured here. And the pessimists would like to think that our representatives represent the wishes of the wealthy organized interests, but that didn't seem to matter here, either. At the end of the day, it came down to politics. Of course, that captures a lot of things--beliefs, philosophies, world views, values--but that appears to have made all the difference... this time.