The special election fervor descends on Arizona’s 8th Congressional District on Tuesday, bringing with it another bout of hypotheses about what these contests mean for the broader 2018 midterm cycle. Democrats don’t have as easy a path to victory here as they had in, say, Pennsylvania’s 18th Congressional District, but anything could happen. Regardless of the winner, what’s most important for November is how much this seat changes from its 2016 baseline. A big swing — 16% (the average swing in Federal special elections so far) or more — gives us another data point that the midterms could be a huge victory for Democrats. A smaller swing here would be a (perhaps more useful) data point saying the opposite.
The candidates are Republican Debbie Lesko and Democrat Hiral Tipirneni. Who’s going to come out on top? Per usual, I’ll walk you through the data.
There is mixed evidence of a competitive race in AZ-08. According to public polling1, the Debbie Lesko (Rep.) has a 5-point lead in the contest. However, error could be large.
|Pollster||Date||Sample||Lesko (Rep.) Margin|
|OH Predictive Insights||4/11/18||500 LV||+10|
Even though special election polling shows Lesko ahead, that same polling has not performed very will in past cycles. In the special election to Georgia’s 6th Congressional District in June 2017, for example, some polling had Democratic Candidate Jon Ossoff up seven points over his Republican opponent Karen Handel. Handel ultimately won with a 4-point margin in the race. More recently, however, polls in the special election to Pennsylvania’s 18th Congressional District nailed the contest, putting Democrat Conor Lamb up 2 points on average (he won by 0.3).
This is to say that there’s a wide distribution of error in special election polling. According to my data set of past polls in the contests, the margin of error on the average in any given cycle ought to be +/- 13%. While polls say Lesko should win by 5 points, she could very well win by 18 or lose by 8. Either one of the extremes would be a big deal, but because of how margins of error work, the odds of a large Lesko win or Tipirneni (Dem.) upset are lower than a modest Lesko win.2
This figure above graphs the daily average of polls in AZ-08 with the shaded margin of error for all polling averages of 21 special election contests.
Polls aren’t the only data in Arizona’s 8th, however. This cycle actually gives us a little (or a big) peak into the result by releasing the partisan breakdown of early and mail-in ballots. What do those say?
2. District Geography and Early Voting
A 5-point win for Republican Debbie Lesko would — and I won’t mince words here — be a very disappointing show for the Grand Old Party. In recent elections, the GOP has had a stranglehold on AZ-08.
In the 2016 U.S. presidential election, Donald Trump won the seat by 21 percentage points. Mitt Romney won by 25. The partisan lean of the seat3 is R+25. Suffice to say that a 16 or 20 point swing to the Democrats would be a huge harbinger of trouble to come in the midterms, but is a Lesko win that small even that likely?
This district geography says no. A “fundamentals” projection — one based on the partisan lean, past vote margin, and current national generic ballot — says this district should give Debbie Lesko a 15-point margin Tuesday night. This would not be all that surprising in terms of a regular open seat election.
Voting so far indicates something along these lines. According to the 150k+ ballots that have been submitted via mail so far, no, the Republicans are not in trouble here. Of those, 21 points more have been cast by registered Republicans than registered Democrats. This compares to a 17-point edge in party registration in the seat, which makes sense as party registration tends to favor Republicans across the country. However, 23% of early ballots cast in AZ-08 have been cast by Independents. It’s possible, though certainly not likely, that those have been cast for Democrats by a margin large enough to offset the Republican lean of these seats. If I were a Democrat, I wouldn’t count on that to save me.
To show this with the data, let’s assume a few things that we’ve learned in recent special elections. First, by my estimates, Conor Lamb’s upset victory in PA-18 was due to about 9 points of persuasion and 13 points differential partisan turnout. If we assume that 4-5 points of early ballots in AZ-08 cast by GOP registrants go to Tipirneni (this would match a 9-point swing in vote margin) she still needs to overcome a 12-point margin in election day votes.
So if (A) Republican Debbie Lesko has a 12-point margin in the early vote (56%) with 150k early/mail-in ballots cast and (B) there are a projected 50k ballots outstanding, then © Democrat Hiral Tipernini needs to win election day voters by 38 points to win AZ-08.4
However, if 33% of the vote is outstanding, with the same early vote breakdown Tipirneni only needs 24% margin in votes cast on election day. This relationship is illustrated below, in one chart for the 25% outstanding votes scenario, and one for the 33%.
Again, I won’t mince words. This doesn’t seem very likely at all. Of course, the split that Independents give Lesko/Tipierneni could change this a lot. And Republicans could be voting for Tipirneni, though no way for us to know how much of this is occurring beforehand. Although my prior is that AZ-08 will have less crossover than PA-18, making this an even harder uphill ballot for Democrats, that is by no means the only possibility. Poll watchers’ estimates of the magnitude of crossover voting, like almost everything else we estimate, has a margin of error around it.
Let me put it this way:
Early voting data are not “real results,” per se, despite what some analysts would have you believe, since partisanship does not equal vote choice. Though they are very correlated in modern America it is not a safe bet to assume all GOP ballots are for GOP candidates, and vice versa for Democratic voters and candidates. Such assumptions would have led us quite astray in the Texas primaries where Democrats cast more early votes than Republicans for the first time since 2010, but cast just 40% of total votes in the D or R primaries.
All in all, this analysis of early voting is useful, I think, for evaluating the impact of (A) higher/lower number of outstanding ballots and (B) varying Republican dominance of early voting. However, I have to rely on a lot of assumptions to make this point and uncertainty certainly remains high. Let’s take a look at some more data and less guessing, this time in the form of special election results from the past 18 months.
3. Special Elections
The results from the past year and a half of special elections are much rosier for Democratic fortunes in Arizona’s eighth than most of the other data. In all of the special elections to federal seats since 2016, the Democrats have swung their margin in the 2016 presidential election by an average of 17 points. These contests are listed below.
|Seat||Partisan Lean (%)||Vote Margin (%)||Dem. Swing (%)|
|South Carolina 5th||R+19||R+3||+16|
|Alabama U.S. Senate||R+29||D+2||+31|
This 17-point overperformance is slightly larger than the Democrats average swing in special legislative elections, which is a more modest 14 points. However, there are caveats here that make me think a projection based on special election data is not the best idea in AZ-08.
This graph shows the swing in special elections (each point is a contest) from the 2016 presidential election in the district (on the y-axis) with a districts partisan lean on the x-axis.
Both the average (and median, for what its worth) special election swings above envelop quite a few contests where the role of candidate quality played a big role in changing the outcome of these contests. In the PA-18 special, for example, Conor Lamb was able to capitalize on ancestral Democratic roots in the area, catering to unions, especially in an otherwise predominantly white, non-college educated, conservative district. In simple terms, he fit the bill there. In the special election to Alabama’s U.S. Senate seat voters were forced to choose between Roy Moore, the Republican candidate who made national headlines for allegations of sexual misconduct with a minor, and Doug Jones, a white, conservative-appearing Democrat who was able to rally a base of black voters in the state. Jones won by 30 points.
In other words, the swing in special elections has been quite extraordinary this year. This is not to say it won’t be predictive in aggregate — indeed, there is a long history of swings in special elections being predictive of broader, nationwide swings in November — but we should be cautious in extrapolating to individual special elections. After all, for an average to be an average, there must be points on both sides (and if there already many pro-Dem points, we may expect to see some land on the other side).
Perhaps what is more important is not what the special elections mean for the outcome of AZ-08, but rather, what the outcome of AZ-08 means for special elections.
This final section wraps up the quantitative indicators and packages them in one graph:
4. Combining Indicators
In forecasting any outcome, it is not wise to rely on any one indicator. It is still unwise to rely on many indicators derived from the same source of data. Formal modeling relies on “Bayesian” techniques to combine probabilistic projections from multiple inputs. However, for the sake of illustrating uncertainty, I’m not going to do that for you all. Instead, I’m going to display all the probabilities separately.
Below, I’ve put together the average projection and expected uncertainty from four different projections into one graph: (1) district-level polls; (2) district geography (or “fundamentals”) plus the nationwide generic ballot and district-level polls; (3) special elections so far; and (4) only the fundamentals plus the generic ballot.
The figure above graphs the estimates and their margins of error for various methods of predicting the outcome in AZ-08.
The average Lesko margin in the projections above is 9 points, and the average Lesko probability of victory is 78%. However, because of the uncertainty here, I am tempted to hedge the probability from the quantitative indicators and place the final odds of a Lesko victory at 3:1, or 75%, or 3-in-4, etc.
Of course, this graphic does not take into account the information we know from early voting. I’ll leave this up to you, the smart reader: should you take the quantitative forecasts alone, or do you want to factor in “real results” via early voting data, also taking in and all the uncertainty and assumptions that go along with it?
In sum, there are a variety of ways to glean insights before polls close in Arizona’s 8th Congressional District. Polling points to a close race, though still gives an upper hand to Republican candidate Debbie Lesko. Other typical indicators in the race (including special election overperformance, seat-level indicators, and national generic congressional ballot polling) are more or less optimistic about Democratic prospects.
And then there’s early voting data in the contest. The bottom line: though it gives us a good idea of how the contest is unfolding thus far, it is by no means a sure reflection of the race. I’d wager to say that the early voting data move my projection of the Lesko margin in the district a few fractions of a percentage point into the red, but the added uncertainty does not make me any more confident about the outcome. I put Lesko’s odds at 75% — a 1-in-4 chance that Democrat Hiral Tipirneni adds AZ-08 to the list of Democratic upsets since 2016 (which is 40 seats long as of the Lamb victory in PA-18).
- I’m omitting internal campaign polls as they’re often incomplete representations of the race. ↩
- Margins of error are typically built around the assumptions of a normal distribution, which report a +/- interval for the 95% confidence interval on an estimate. That normal distribution assigns a higher likelihood that numbers fall near the center of a distribution. So if a poll gives a candidate a 10-point +/-10 lead, they could tie or win by 20, but the most likely expectation is still the 10-point win, and the odds of a big error taper off dramatically as you approach 0 or 20 points. ↩
- Partisan lean is calculated by averaging the relative 2016 and 2012 Democratic/Republican presidential candidate performance in the district, weighted 75%/25% to give more emphasis on the recent cycle. ↩
- (150 * 0.44) + (50 * 0.7) = 101 / 200 = 50.5% ↩