Is there momentum in the Democratic race? If there is, when did it begin?

This is a question that has been on my mind this week. On Wednesday, I argued that there is evidence that Obama developed momentum in Wisconsin. I also asserted that this did not seem to have been the case in Virginia and Maryland. It was this latter claim that went against the grain. Many thought that Obama did indeed develop momentum for the Potomac primary. I did not.

How could we determine whether momentum has indeed taken root? The most direct way would be to use the exit polls to look at key demographic groups. If Clinton seems to be doing worse with groups she was once strong with - we might conclude that momentum is playing a factor. Many commentators concluded this after Maryland and Virginia. They found that Obama was doing better among Catholics, union workers, women, and "downscale" voters. Accordingly, they declared that he was benefiting from momentum.

But hold on. Demographic groups are not hermetically sealed off from one another. There is a lot of overlap. Any discussion of voting blocs would have to take this into account, especially in a state where one candidate's strongest group is in large supply. In such a situation, the other demographic indicators might be "overwhelmed." This seems to have been the case in both Maryland and Virginia. In fact, Obama had two of his best groups in large supply: African Americans and wealthy voters. This makes a big difference. We need to remember that they too can be union workers, women, etc. Their strength at the ballot box might make it look like Clinton is suddenly losing her groups, but these were never actually her groups.

Compounding this is the fact that there probably was not a great deal of overlap between the two. Obviously, they are not mutually exclusive. Many African Americans make $100,000 or more. However, whites in Maryland make 39% more than African Americans in Maryland; in Virginia, they make 57% more. So, while there is surely some overlap, it stands to reason that African Americans plus wealthy whites made up about half the vote in both states. So, it seems likely that momentum was not a major factor. A careful reading of the exit polls indicates that Obama probably did not "steal" votes from Clinton, but won on the strength of his typical voters.

Of course, we cannot know for sure. As it is, the exit polling data is just not detailed enough to answer our question directly. What to do? One solution is to use more detailed cross-tabulations of exit poll results. Unfortunately, the media does not provide them. We could also wait. We should know in a year or so; when the PhD's who work for outfits with money to do detailed polling publish their write-ups on the Democratic race, we'll get some answers.

A practical, if imperfect, solution for right now is to combine this limited data set with another limited data set. Namely, we can use the exit polls in conjunction with countywide vote returns. This cannot give us a definitive answer (ultimately, momentum has an effect on individual voters - so we would ultimately need survey data to know for sure). However, it can give us something better than what we have now.

Unfortunately, we cannot use the returns by themselves. The biggest problem is with something known as the ecological fallacy. Specifically, our knowledge of how a county voted cannot give us direct knowledge of how individuals in that county voted. This is a real problem. We would have to make an assumption that, for instance, the observation of a poor county voting for Clinton implies that the poor individuals in that county voted for her, too.

This is where the exit polls can be helpful. If the exit polls imply that "downscale" voters are supporting Clinton, this assumption becomes much more reasonable. Conversely, we can use the aggregate voting results, imperfect as they are, to "firm up" the tentative read we get from the exit polls. The idea here is that, while both types of data present inferential problems, we are on fairly solid ground if the exit polls and the aggregate results point in the same direction.

Before we dive into the countywide returns, however, we have to specify matters a bit more. In particular, we face a problem with racially diverse counties. Their returns cannot help us. Consider, for instance, King and Queen County, Virginia. The median income of whites there is about $41,000. Presumably, this favors Clinton. On the other hand, African Americans constitute 35% of the population. This favors Obama, who the county by 44 points. Was his win due to momentum? Or was it due to his voting bloc overwhelming Clinton's bloc? We can never know. So, this type of county is of little use.*

So, we will look at counties where the African American population is small, say less than 5%, to see whether they varied their support according to median white income. Our expectation, following the exit polls, is that because momentum did not seem to exist - Clinton will do well in the "downscale" counties, and Obama will do well in the "upscale" counties.

Let's look at Maryland first. There are three Maryland counties with an African American population of less than 5%. Clinton won Cecil and Garrett counties by more than 20 points. She won Carroll County by just 3 points; however, this is explicable by the fact that the median white income of Carroll County is more than $60,000 per year.

In Virginia, Clinton generally did well in the homogeneously white counties. We find a tight relationship between income and countywide returns, and it is in the direction we expect. The following graph tells the tale. It examines Clinton's margins of victory in homogeneously white Virginia counties according to median white income.

First, note the distribution of the data points. It indicates a strong relationship between income and Clinton's margin of victory.* Note also the small number of homogeneously white counties. Virginia has more than 70 counties (or cities) - and about 40 of them have African American populations of some size. Clearly, Clinton had an uphill battle in this state.

Our real interest is in the four quadrants divided by the black lines. The horizontal line separates Clinton wins (top) from Clinton losses (bottom). The vertical line separates counties where the median income is less than $40,000 (left) from counties where the median income is more than $40,000 (right).*

The array of the data points on the graph is roughly consistent with what we might expect.* We see a large number of observations in the top-left quadrant, and most of the observations on the right-hand side are in the lower-right quadrant. In other words, Clinton is generally winning the counties where voters make less than $40,000; Obama is generally winning the counties where whites make more than $40,000. Both candidates "take" just a few counties that "belong" to the other. Again, by itself this tells us nothing because it provides no indication of how individuals are behaving. However, taking it in conjunction with the exit polls, we have a good indication that momentum was not driving the results in the Potomac primary. Instead both candidates were doing well among their core groups; Obama's groups happened to have been more populous.

Now, let's turn our attention to Wisconsin. As we discussed on Wednesday, the exit polls indicate that Obama may indeed have developed momentum. Specifically, they showed that Clinton seemed to do worse-than-expected among lower income whites. But, again, we cannot be sure because the exit polling cross-tabulations are limited. The aggregate data might be able to supplement it. The following chart reviews returns by median white income for homogeneously white counties.

First, notice that income and county returns still have a reasonably tight relationship, one that moves in the direction we expect (i.e. Clinton does worse as income increases). Second, it appears that Obama benefited from the fact that homogeneously white counties in Wisconsin tend to be wealthier than those in Virginia - we have many more observations to the right of the horizontal line than we had in Virginia. Not surprisingly, they are in the lower-right quadrant. Third, when we look at the four quadrants, we see a pretty significant shift. Obama wins a much larger number of poor counties than he did in Virginia, where Clinton took most of them. We can see this in the large number of observations in the lower-left quadrant. These are counties that, if Wisconsin were following the pattern in Virginia, would be in the upper-right and going for Clinton. Taken in light of the exit polling data, this is strong evidence that there was a shift among "downscale" whites in Wisconsin.

I think this shift is hard to account for without hypothesizing a momentum effect. Consider the following. We know that the Hispanic population, Catholic population, and union population are important factors in how each candidate performs. The difference in the Hispanic population between Virginia and Wisconsin is negligible (4.7% in Virginia to 3.6% in Wisconsin). There are big differences in the Catholic and union populations between the contests, but these should have favored Clinton. The explanatory variable that is not accounted for above that favors Obama is racial homogeneity. It is possible that lower income Wisconsin voters are more amenable to Obama because race is perceived differently in Wisconsin than it is in Virginia. However, I do not think this could explain the whole of this shift. Momentum might have been the critical ingredient.

All in all, the exit polls and the aggregate vote results seem to point in the same direction - a shift in Wisconsin, but not in Virginia and Maryland. Momentum may have been a factor. Again, the best way to approach the question is as we have here - through a combination of the exit polls and the aggregate vote data. In both states, they move together. Neither gives a definitive indication, but both give the same indication. It might be that the indications of the exit polls and the aggregate returns are both wrong, but this is unlikely.

***
Endnotes

^[*] This is not to say that voters in these counties are not following the patterns we have hypothesized. They may be. The trouble is that we just cannot make any determinations because voters in counties like these are probably going strongly in opposite directions.

^[*] However, the relationship is tighter or weaker depending upon Clinton's margin of victory. The clustering is very tight in the top left, but the data points slowly disperse as we move to the bottom right. This property is known as heteroscedasticity, and it would complicate drawing a naïve prediction of a county result based on income. Our purpose here is more modest; so, we need not concern ourselves with this.

^[*] Why $40,000? The Census Bureau only releases income data for the decennial census. Accordingly, the county-level income data is eight years old. To account for this, I have adjusted the line on the graph for inflation. Specifically, this year's exit polls tend to show Clinton winning white voters who make less than $50,000. In 1999 dollars, this works out to about $40,000.

^[*] The key word in this sentence is "roughly." There are three factors that complicate any simple inferences.

(1) As indicated above, we are dealing with inflation-adjusted income statistics. The difficulty here is that we must assume that voters in these counties are making the same real income now as they were eight years ago. This might not be the case.

(2) Party identification and income might be related. For instance, Democrats make less than Republicans in these counties. In this case, the median income for Democrats in a given county would be lower.

(3) Factoring out African Americans does not eliminate our inferential problems. It only reduces it. Specifically, simply because we know the median income of a county does not mean that we know how individuals are voting. It might not be the case that a poor county has gone to Clinton because the poor voters voted for her. We discussed this problem above, and decided that our solution would be to look at the aggregate returns in conjunction with the exit polls.