Massachusetts Senate Primary Analysis

First Takeaways

• Turnout was even higher than in the March Presidential primary, and higher than expected by many observers. This was true across the state. It will be interesting to see the precinct by precinct analysis of in-person voting on election day versus early voting, in person and by mail.
• Markey had strong results, as expected, in CD 5, his former district, as well as CD 7, the most progressive district in the state.
• More surprisingly, Markey also won the more moderate districts, CDs 3, 6, and 8.
• In those precincts that turned out more votes in September than in March, Kennedy won overall. Many of these precincts are in the gateway cities.
• Kennedy also won overall in the precincts of color.
• Simple regression analysis suggests a distinction between Sanders voters and Warren voters, where Sanders voters were much more inclined to vote for Kennedy than Warren voters.

Results by Congressional District

Markey won 7 of the 9 congressional districts, including Kennedy’s home CD, with big margins in 5 and 7, as well as a strong showing in the more moderate CD 6th.

Maps: Markey (red), Kennedy (blue)

color intensity indicates margin

Kennedy Wins Precincts of Color

Across the 263 precincts I have identified as precincts of color, Kennedy outpolled Markey:

POCs gave Kennedy 9.4% of his overall vote total, versus 5.4% for Markey.

Turnout Relative to March Presidential Primary

Leading up to election day, those of us trying to estimate the likely turnout looked at past primary results:

I figured turnout was going to be around 1,000,000 — more than in 2018 but not approaching either the 2016 or 2020 Presidential primaries. I didn’t see this election as generating the same level of interest and engagement as the Presidential races.

September 2020 turnout: 1,427,868

Wow, the turnout last week narrowly exceeded the March 2020 turnout by 10,000 votes. Clearly, sending out vote-by-mail ballots to every voter boosted turnout way beyond what would have happened in a pre-COVID environment.

Turnout By Congressional District

Overall, last week’s turnout was comparable to last March’s turnout statewide. How did this play out at the level of the nine congressional districts?

This table shows that at the level of the nine congressional districts, turnout last week was not that different from turnout in the March Presidential primary --- with the key exception of the western CD 1, where House Chair of Ways and Means faced an AOC-endorsed primary challenger, and turnout there was 21% higher than in March.

Of note is that turnout in CD 4 and CD 5 — Kennedy’s current district and Markey’s old district — were essentially the same as in March, even with a hotly contested primary in CD 4. To me this underscores the intense turnout we saw last March, and that carried through to last week.

Kennedy Wins in Precincts with Higher Turnout

On a precinct by precinct basis, I compare turnout in March with turnout in September, and split the results into two groups: those precincts where the September turnout is higher than in March, and those where it is lower. (I leave out precincts with totals of less than 25 votes.)

Kennedy narrowly won overall in those precincts with higher turnout, but trailed by 25 points in those with lower turnout.

Looking at the map where the turnout was 105% or more in September versus March, it shows many of the gateway cities:

Statistical Models

This is the start of some exploratory analysis to look at vote totals at the precinct level and compare them to vote totals from the Presidential primary in March.

There were 14 candidates on the March ballot, and I group seven of them as “moderates” and the seven other as “progressives”. Of course, by election day many of the candidates had suspended their campaigns and most of the votes went to either Biden, Sanders, Warren or Bloomberg. I grouped Biden and Bloomberg with the moderates, and Sanders and Warren with the progressives. Note that out of the 1.4 million votes cast in March, they were evenly split between the moderates and progressives.

By comparing the vote totals received by Kennedy, and Markey, at the precinct level with precinct results from the Presidential primary, we can get some sense of who was supporting the Senate candidates.

Model 1: using Moderate and Progressive vote totals

In this first model, I estimate the coefficients for predicting the Senate candidate vote totals by the March moderate and progressive totals in each precinct:

JK  =  mjk * MODERATE   +  pjk * PROGRESSIVE   + model error
EM  =  mem * MODERATE   +  pem * PROGRESSIVE   + model error

The results of a standard regression fit are:

If you limit yourself to the moderate-progressive split of the vote in March, statewide the Kennedy vote totals are much mor highly correlated with the moderate totals than the progressives, while Markey is highly correlated with the progressive totals while getting some weight from the moderate totals as well.

This model can also be run separately for the nine congressional districts, as well for the precincts of color, and the variations in the fits help to describe the differences in the voting groups.

So, the results of this simple moderate-progressive model is that overall Kennedy won among those who voted for moderate candidates in March while Markey won among the progressives while also cutting into the moderate votes. But there are regions where this simple story isn’t quite as strong, say in CDs 3 and 6, plus Markey clearly enjoyed a “home court” edge in CD 5.

Model 2: using Biden, Sanders, Warren and Bloomberg votes

This model uses only the vote totals from the four top candidates. The story is a bit different:

In this model, Kennedy’s precinct vote totals are best estimated by the Biden and Sanders vote totals, while the largest weights for Markey’s totals are on the Warren and Bloomberg totals.

TO DO: beyond simple regression models

The two models show basic correlations for the two candidates separately. The next step is to look at models that estimate both candidates vote totals as part of a single model, and then use that to look at over- and under- performance at a precinct level.