Chapter 22: Down-Ballot and Global Forecasting

The Senate race is in some ways the easiest electoral forecasting problem in American politics. There are only a few dozen competitive statewide races per cycle. Each state-level contest receives substantial polling attention. The candidates are well-known quantities with years of public record. The universe of potential outcomes — two candidates, one winner — is binary and well-defined.

The moment you move down the ballot, almost everything gets harder.

The U.S. House of Representatives has 435 seats. In any given election cycle, roughly 60–80 of those seats are genuinely competitive; the rest are safe for one party or the other by margins large enough to make individual-race polling impractical or unnecessary. For those 60–80 competitive races, the polling coverage is dramatically thinner than for Senate contests. Many individual House races receive no public polling at all. The candidates are often unknown outside their districts. The district boundaries change every ten years through redistricting and may change more frequently through court-ordered remaps.

State legislative and gubernatorial forecasting multiply these challenges further: there are approximately 7,000 state legislative seats in the United States, distributed across 50 states with different election cycles, different redistricting histories, and wildly different levels of data availability.

And then there is the international dimension. The methods developed in the preceding three chapters are deeply rooted in the American electoral context — binary choice between two major parties, a presidential system with direct popular vote, detailed polling infrastructure developed over eighty years of industry practice. What happens when you try to apply the same analytical framework to a parliamentary election in Germany, a presidential runoff in Brazil, or a multi-party coalition environment in Israel?

This chapter extends the forecasting framework in both directions — toward the small-district complexity of down-ballot American elections and toward the structural diversity of international electoral systems. Along the way, it surfaces a theme that runs through both domains: the data environment shapes the forecasting methodology as much as the political system does.

22.1 House Forecasting: The Challenge of 435 Individual Races

A House forecasting model is not simply 435 Senate models running simultaneously. The structural differences between House and Senate forecasting are significant enough that a Senate approach applied wholesale to House races will produce badly miscalibrated results.

22.1.1 The Scale Problem

The most obvious challenge is scale. The political analytics industry simply cannot conduct high-quality likely-voter polling in 435 districts while maintaining methodological consistency across races. The 60–80 competitive districts receive the bulk of polling attention, but even those are underpolled relative to Senate races. A competitive Senate race in a large state might receive 25–30 polls in the final two months of the campaign. A competitive House district — represented by a smaller geographic area but often just as nationally consequential — might receive 3–5 polls, some of them of questionable quality.

The practical consequence is that House forecasting relies much more heavily on indirect signals than Senate forecasting. Rather than building individual-race models from a rich polling environment, forecasters use:

• Historical district-level partisan lean (previous presidential and House results in the district)
• Generic ballot polling (the national question "which party do you prefer for Congress?" rather than any specific race)
• Incumbency effects and candidate quality metrics
• Fundraising data (which the Federal Election Commission makes public on a quarterly basis)
• Redistricting adjustments (how new district lines changed the partisan composition)
• Forecaster ratings from organizations like the Cook Political Report, the Rothenberg-Stewart Political Report (now Inside Elections), and Sabato's Crystal Ball

These inputs are combined into a district-level estimate that incorporates both race-specific signals (where they exist) and the broader national political environment.

💡 Generic Ballot as the Anchor. The generic ballot — "If the election were held today, which party's candidate would you vote for in your district for the House of Representatives?" — serves as the primary anchor for national-level House forecasts. It captures the overall partisan wind that blows across individual races: when Democrats lead the generic ballot by 6 points nationally, their candidates tend to outperform their district-level historical baselines, and vice versa. The generic ballot is widely and consistently polled, making it a high-quality input even when individual-race data is scarce.

22.1.2 The Structural Environment of House Races

Before building a House model, it is worth understanding the structural features that make House races systematically different from Senate races:

Geographic sorting and incumbency. The American population has sorted itself geographically in ways that make most House districts non-competitive by design. Urban districts are overwhelmingly Democratic; rural districts are overwhelmingly Republican. Only suburban and exurban swing districts generate genuine competition. This reduces the forecasting problem from 435 races to roughly 60–80 — but it means that errors in the competitive district universe propagate into the overall seat projection.

Name recognition asymmetry. In a Senate race, both candidates are typically known quantities with significant public profiles. In competitive House races, challengers frequently have very low name recognition, which introduces uncertainty about candidate quality that does not appear in a name-recognition-adjusted baseline.

Redistricting risk. House district boundaries can change dramatically every ten years. A district that was safely Republican in the previous cycle may become competitive or even Democratic-leaning after a court-ordered remap. Any House model built using historical district-level baselines must carefully account for boundary changes.

Campaign finance effects. House race outcomes are highly sensitive to the relative spending levels of the two campaigns and affiliated outside groups. FEC quarterly data provides a useful leading indicator of race competitiveness: challengers who raise substantial money are competitive; challengers who cannot raise money are not.

22.1.3 Detailed Methodology for District-Level Forecasting

Building an individual House district forecast requires assembling multiple data layers and combining them through a principled model. The following walkthrough describes the methodology used by leading forecasting organizations, using a hypothetical competitive district as an example.

Establishing the baseline partisan lean. The starting point for any district forecast is its baseline partisan lean — an estimate of how the district votes relative to the national average in a "neutral" environment where neither party has a structural advantage. The standard approach uses a weighted average of recent presidential results in the district's current boundaries.

For a district where Biden won 52% in 2020 and Clinton won 49% in 2016 (translating to +2 Democratic and -1 Democratic relative to the national popular vote, respectively), a simple average gives a baseline lean of approximately +0.5 Democratic, or a district that performs roughly at the national average. More sophisticated approaches weight more recent results more heavily and adjust for boundary changes from redistricting.

Applying candidate quality adjustments. Incumbency is the most important candidate quality variable. Incumbents typically outperform their district's partisan baseline by 3–6 percentage points — the "incumbency advantage" that has been extensively studied in political science. For open seats, candidate quality is approximated by prior electoral experience, fundraising success in the early reporting period, and ratings from independent forecasting organizations.

In the Garza-Whitfield campaign's network of down-ballot interest (a competitive House district in the state's northern suburbs), Nadia Osei's analytics team applies a standard 4-point incumbency adjustment for the Republican incumbent, then adjusts upward for an unusually well-funded Democratic challenger who has raised more in the pre-primary period than the previous three Democratic challengers combined.

Incorporating individual-race polling. When district-level polling is available, it is incorporated as a direct estimate of candidate preference — but with calibration for pollster quality and recency. A single poll from an unrated pollster conducted three months before the election receives much lower weight than a recent poll from an A-rated organization using a registered voter sample with transparent methodology.

The standard Bayesian updating approach treats the baseline model prediction as a prior and individual polls as likelihood updates. Concretely: if the baseline model predicts a Republican advantage of 3 points, and a recent poll shows a 1-point Democratic advantage, the blended estimate — weighted by the reliability of each source — might settle around a 1-point Republican advantage, reflecting greater weight on the poll as the more recent direct measurement.

Simulating uncertainty. District-level forecasts are presented as probability distributions, not point estimates. The sources of uncertainty are multiple: - Uncertainty in the generic ballot forecast (which propagates to all districts) - Polling error (historical district-level polling has proven substantially less accurate than state-level Senate polling, partly because of smaller sample sizes and lower-quality polling organizations) - Candidate-specific uncertainty (what happens if a scandal emerges, a debate goes badly, or turnout patterns deviate from historical norms)

A Monte Carlo simulation of 10,000 election scenarios, where each scenario draws random errors for the generic ballot, each district's polling, and candidate-specific factors, produces a probability distribution over the district outcome — typically presented as something like "the Democrat wins in 38% of simulations," equivalent to saying the Republican is a modest favorite but the race is genuinely competitive.

📊 The Forecasting Organization Ecosystem

The Cook Political Report, founded by Charlie Cook in 1984, is the grandfather of American district-level electoral forecasting. It rates each House, Senate, and gubernatorial race on a seven-category scale from "Solid Democrat" to "Solid Republican," with intermediate categories of "Likely," "Lean," and "Toss-up." The ratings are based on a combination of polling, fundraising data, and the expert analysis of Cook's staff — an explicitly hybrid quantitative-qualitative methodology. Inside Elections (formerly Rothenberg-Stuart) and Sabato's Crystal Ball at the University of Virginia provide similar expert-judgment ratings that serve as important inputs to quantitative seat projection models.

22.2 From Generic Ballot to Seat Projections

The generic ballot gives you a national signal. Converting that signal into a seat projection requires understanding the historical relationship between national popular vote margin and seat outcomes — a relationship that is real but highly variable.

22.2.1 The Efficiency Gap Problem

The most important reason why popular vote margin does not translate cleanly into seats is geographic efficiency: the way votes are distributed across districts. A party that wins large majorities in its core districts "wastes" votes — margins of 80-20 in urban districts do not produce more seats than margins of 60-40 would. Meanwhile, narrow wins in many competitive districts are more seat-efficient than a few massive wins.

This distributional feature of vote geography has been formalized in the concept of the efficiency gap, originally developed by Stephanopoulos and McGhee (2015) as a mathematical measure of partisan advantage in district maps. But the practical implication for forecasters is more straightforward: you need to know not just what the national swing is but how that swing is distributed across districts.

A national Democratic swing of 4 points will produce very different seat outcomes depending on: - How many districts are currently at 50-55% Republican (in range to flip with a 4-point swing) - How the swing is distributed geographically (uniform national swing is different from a concentrated swing in suburban districts) - Whether the swing occurs in already-safe districts (wasted votes) or in competitive districts (seat-changing)

📊 The Seats-Votes Curve. Political scientists have long studied the relationship between a party's national vote share and its seat share in legislative elections. The "seats-votes curve" is typically nonlinear: small shifts in vote share around 50 percent produce disproportionate changes in seats (a "swing ratio" greater than 1), while shifts in the tails (from 20% to 25%, say) produce smaller seat changes. The current polarized geography of American House elections has substantially reduced the swing ratio compared to the 1970s and 1980s, meaning that a given national vote shift produces fewer seat changes than it would have fifty years ago.

22.2.2 Building a Seat Projection Model

A simple seat projection model for the House follows these steps:

Step 1: Establish a district-level baseline. For each of the 435 districts, calculate the expected partisan lean based on recent presidential and House results. Many forecasters use a combination of the last two presidential results in the district (appropriately averaged) and the most recent House result if an incumbent is running.

Step 2: Apply a national swing. The generic ballot forecast provides a national swing expectation — the degree to which Democratic or Republican candidates are expected to outperform their historical baseline. Apply this swing uniformly (or with district-type adjustments) to all 435 districts.

Step 3: Apply race-specific adjustments. For the 60–80 truly competitive districts, layer in individual-race polling, fundraising data, and forecaster ratings to adjust the uniform swing prediction up or down.

Step 4: Run probabilistic simulation. Just as with Senate races, run Monte Carlo simulations that incorporate uncertainty in the generic ballot, the swing-to-seat translation, and individual-race measurements. The output is a probability distribution over seat totals for each party.

✅ Best Practice: Transparent Uncertainty in Seat Projections. A responsible seat projection presents a range: "Democrats are expected to hold 215–235 seats, with a median projection of 225 and a 90 percent probability of winning between 200 and 248 seats." Point-estimate projections ("Democrats will win 221 seats") communicate false precision and mislead consumers of forecasts.

22.3 Senate Forecasting: The Small-Sample Problem

Senate forecasting falls between House and presidential forecasting in terms of data availability. Statewide Senate races receive substantially more polling than House races, but the total number of competitive Senate races per cycle is small — typically 10–15 genuinely competitive contests.

22.3.1 Challenges Specific to Senate Races

Small cycle-to-cycle sample. Unlike the House, where 60-80 competitive races provide substantial data about the national political environment, Senate cycles involve only a handful of competitive contests. A forecaster trying to assess whether a 3-point lead in Senate polls is sufficient to win cannot easily calculate a historical accuracy rate for Senate polling in the current environment — the sample of cycles is simply too small.

Staggered elections. Senate elections occur in three classes: Class 1 (up every 6 years, last in 2018), Class 2 (last in 2020), and Class 3 (last in 2022). In any given cycle, only one-third of senators face re-election. This means that the map — which states are defending seats — is determined six years in advance and varies dramatically in its competitive potential from cycle to cycle.

State-level idiosyncrasy. Each Senate state has its own political culture, regional media landscape, and electoral history that limits the applicability of national patterns. A 5-point generic ballot advantage for Democrats means something very different in a cycle where competitive Senate seats are concentrated in Republican-leaning states vs. a cycle where they are concentrated in swing states.

Candidate quality effects. Senate candidates are high-profile enough that individual quality matters substantially — far more than in an individual House race. A weak Senate candidate can underperform the partisan baseline by 5 or more points; an exceptional one can outperform by similar margins. Modeling candidate quality systematically requires extensive historical data on candidate characteristics (incumbency, previous electoral experience, fundraising trajectory) that is not always available or consistent.

🔴 Critical Thinking: The 2022 Senate Case. The 2022 Senate cycle illustrated candidate quality effects dramatically. In Pennsylvania, Georgia, and Arizona, Republican candidates nominated through primary processes were sufficiently weak (by general election standards) that they underperformed the Republican baseline substantially in states that the party's strong candidates could have won. No generic ballot model adequately captured this, because generic ballot models measure partisan environment rather than candidate quality. The miss was not a polling failure in the same sense as 2016; it was a model structure failure.

22.4 State Legislative and Gubernatorial Forecasting

Below the Senate level, data scarcity becomes the dominant challenge.

22.4.1 The Data Desert of State Legislatures

Approximately 7,000 state legislative seats are contested in a given election cycle. Professional polling in state legislative races is essentially nonexistent except in the handful of states with well-funded partisan interests and competitive chambers. Most forecasting of state legislative outcomes relies on:

• Presidential performance in the district (how did the district vote in the last one or two presidential elections)
• Registration data (the relative proportion of registered Democrats, Republicans, and independents)
• Generic state partisan environment (analogous to the national generic ballot but for state-level races)
• Fundraising data (available through state-level disclosure systems of highly variable quality)

The result is that state legislative forecasting is almost entirely a statistical exercise based on structure and environment, with minimal individual-race signal. This makes it accurate in aggregate — you can predict that Democrats will win approximately 48 percent of state legislative seats nationally in a given environment — but unreliable for individual races.

22.4.2 Why State Legislative Forecasting Is the Hardest Problem

State legislative races represent the hardest individual forecasting problem in American electoral politics. The reasons compound:

Redistricting every decade creates baseline discontinuity. A district that existed in 2020 with certain partisan characteristics may be entirely different in 2022 after redistricting. Historical comparison is compromised wherever district lines change substantially.

Candidate filing uncertainty. State legislative races frequently feature uncontested races (only one party files a candidate), late candidate withdrawals, and primary outcomes that profoundly affect general election competitiveness. Forecasting tools calibrated on competitive races cannot easily handle this dynamic.

No consistent data infrastructure. Unlike federal elections, where the FEC provides standardized campaign finance reporting and vote data, state elections are administered by 50 different state election authorities with wildly different data release standards. Some states publish precinct-level results within hours of polls closing; others take weeks and provide only county-level data. Some state legislatures have campaign finance disclosure; others do not. The absence of a standardized data layer makes systematic state legislative modeling extremely difficult.

Local candidate effects dominate. In small districts — many state legislative districts have 20,000–50,000 residents — candidate personal qualities, local reputation, and community ties matter substantially more than in larger-electorate races. A well-liked local business owner running as a Republican might win a district that would otherwise go Democratic; a controversial incumbent of either party might underperform by 10+ points. These idiosyncratic effects cannot be systematically modeled from available data.

Turnout is small and variable. State legislative primary elections in off-years can see turnout below 10% of registered voters, making outcomes highly sensitive to which specific voters happen to participate. Small-scale campaign tactics — door-knocking, direct mail, phone banks — can shift these small electorates in ways that polling cannot reliably capture even when polling is attempted.

The practical implication for analysts: state legislative forecasting should be presented as rough probabilistic assessment of chamber-level outcomes (will Democrats gain enough seats to take a majority?) rather than reliable individual-race predictions. When Nadia Osei's team builds a state legislative targeting model for down-ballot races in Garza's state, they use it to identify the 15–20 districts most likely to be competitive and concentrate field resources there — not to predict winners and losers in each individual district.

22.4.3 Gubernatorial Forecasting

Gubernatorial races fall in an intermediate position: they receive more polling than state legislative races but less than U.S. Senate races. The structural environment for governors includes several features that differ from Senate forecasting:

No staggered elections. Governors in most states face election on the same cycle, providing a larger cross-sectional dataset in any given cycle (38 states elect governors in midterm years, 11 in presidential years, and 1 — Kentucky — on an odd-year cycle).

State economic conditions matter more directly. While Senate race outcomes are substantially driven by national presidential approval, gubernatorial outcomes show a stronger connection to state-level economic conditions and the incumbent governor's approval rating. A forecasting model for governors should therefore apply state-level economic data more aggressively than a Senate model would.

No national partisan trend override. A strong national Democratic wind lifts Senate candidates significantly but has a more muted effect on gubernatorial outcomes, where voters are more willing to split tickets between the presidential and gubernatorial ballot. The correlation between generic ballot and gubernatorial outcomes is real but weaker than the correlation with Senate outcomes.

22.5 Multilevel Regression and Poststratification (MRP) for Subnational Estimates

One of the most important methodological innovations in electoral forecasting is Multilevel Regression and Poststratification (MRP), which allows high-quality subnational estimates from national-level surveys. MRP is the statistical engine underlying YouGov's constituency-level estimates in the UK, and it has become increasingly standard in American academic and commercial forecasting for state and district-level applications.

22.5.1 The MRP Framework

The core challenge MRP addresses is this: a national survey with 3,000 respondents contains perhaps 12–20 respondents in any given congressional district. A sample of 15 people cannot reliably estimate district-level opinion.

MRP solves this problem in two stages:

Stage 1: Multilevel regression. Estimate a statistical model that predicts individual vote preference as a function of individual characteristics (age, education, race, gender) and geographic unit characteristics (state partisan lean, regional demographics, urban/rural classification). The multilevel structure borrows statistical strength from across the full dataset: the model's estimate of how 35–44-year-old college-educated white women in the Midwest vote is informed not just by the handful of such respondents in any particular district but by all respondents with those characteristics across the full sample.

Stage 2: Poststratification. The Census Bureau's American Community Survey provides demographic profiles of every congressional district in the country. Apply the regression model's predictions to each possible combination of demographic characteristics in each district, weighted by how many people in the district have those characteristics. The result is a synthetic district-level estimate assembled from demographic building blocks.

💡 MRP Intuition. Think of MRP as constructing each district's opinion estimate from a national survey the way you might construct a demographic portrait from Lego blocks. Each demographic cell (young Black women with college degrees in Southern cities; older white men without college degrees in rural Midwest; etc.) is a Lego block with an estimated political preference. You assemble the district's opinion by combining the blocks in the proportions that match the district's actual demographic composition. The power of the approach is that you don't need many survey respondents in any particular district — you need to accurately estimate the preference of each demographic group nationally, then apply those estimates to the district's known demographic profile.

22.5.2 Limitations of MRP

MRP is a powerful tool but not a universal solution. Its limitations are important:

Validity of the regression model. MRP produces accurate estimates only if the relationship between demographics and political preferences is stable across geographic units. In practice, there are substantial "place effects" — political behavior in rural Wisconsin is not fully predicted by the demographics of rural Wisconsin alone, because local political culture, history, and candidate-specific factors matter independently. Models that do not include geographic random effects will underestimate geographic variation.

Demographic ceiling. MRP works best when vote preference varies substantially across demographic groups. In a highly polarized environment where a 65-year-old white man in Montana and a 65-year-old white man in Georgia vote similarly regardless of local context, demographic prediction does a good job. But in an election where local economic conditions, candidate-specific factors, or mobilization campaigns produce substantial geographic variation independent of demographics, MRP will miss those patterns.

Sparse cells. If a district has very few residents in certain demographic cells (say, a rural district with very few Black college graduates), the Census-based poststratification may assign substantial weight to demographic groups whose political behavior is estimated from respondents elsewhere in the country with very different local contexts.

📊 YouGov's MRP in UK Elections. YouGov's constituency-level MRP model produced accurate predictions of the 2017 UK hung parliament — a result that simple polling averages had missed — and continued to perform well in 2019 and beyond. The 2017 success came from capturing the substantial geographic variation in the Brexit-era vote shift, which was strongly demographically patterned: Leave-voting areas with high proportions of older, less-educated white voters shifted to the Conservatives, while Remain-voting areas with high proportions of young, educated, diverse populations shifted to Labour. MRP, with its demographic disaggregation, was well-positioned to capture this realignment. Standard constituency-level polling, with its small samples, was not.

22.6 International Election Forecasting: Similarities and Differences

The models developed in Chapters 17–21 are products of a specific electoral context: the American two-party system with its well-developed polling infrastructure, federal election data reporting, and decades of empirical political science research. Applying these models internationally requires confronting fundamental differences in electoral systems, data availability, and political culture.

22.6.1 What Translates Across Systems

Some features of election forecasting are genuinely universal:

• The value of poll aggregation. Regardless of political system, individual polls are noisy, and averaging multiple polls from diverse sources reduces noise. The exponential decay weighting, sample-size weighting, and quality adjustments described in Chapter 21 apply with minor modifications to most polling environments.

• Fundamentals-based priors. Economic voting — the tendency of voters to reward or punish the incumbent government for economic performance — is documented across dozens of countries and political systems. A fundamentals-based prior incorporating economic conditions and government approval is applicable to most electoral environments.

• The need for uncertainty quantification. The gap between point estimates and probability distributions is equally important in international forecasting. Overconfident point estimates are as misleading in Brazilian presidential forecasting as in American Senate forecasting.

🌍 Cross-National Evidence for Economic Voting. Lewis-Beck and Stegmaier (2000) reviewed evidence for economic voting across 23 countries and found consistent relationships between economic performance and incumbent vote share in presidential and parliamentary systems alike. The magnitude varies — economic voting is stronger in systems with clear incumbency attribution (single-party governments) than in those with diffuse responsibility (multi-party coalitions) — but the basic relationship is robust.

22.6.2 What Does Not Translate

Several features of American election forecasting do not transfer cleanly to other contexts:

Binary vs. multi-party competition. The American two-party system allows a forecast to express election outcomes as a single margin between two candidates. Parliamentary systems with three or more significant parties require modeling multiple party vote shares simultaneously, with the constraint that all shares must sum to 100 percent. More critically, the electoral outcome in a parliamentary system is government formation — which coalition will form, who will be prime minister — which depends on post-election negotiations that are not directly predictable from vote shares alone.

Presidential vs. parliamentary accountability. Fundamentals-based models assume that voters know who to hold accountable for economic conditions. In a presidential system, accountability is relatively clear: the president is the chief executive. In a parliamentary system with a coalition government, responsibility for economic outcomes is distributed across coalition partners, and voters may punish or reward individual parties differently based on their role in the coalition. German coalition governments, for example, often see the junior partner lose votes even when the coalition's economic record is strong, because voters differentiate between "who ran things" and "who was nominally present."

Data availability. American election data is extraordinarily rich: precinct-level vote totals, individual-level voter registration data, quarterly FEC filings, and decades of consistent polling infrastructure. Many countries have nothing approaching this. In Brazil, precinct-level results are available publicly but in formats that require substantial processing. In India, the world's largest democracy, state-level polling coverage is inconsistent, district-level polling is rare, and polling infrastructure is comparatively underdeveloped relative to the scale of the electoral system. Forecasting in data-scarce environments requires different methodological strategies than forecasting in data-rich ones.

22.7 Comparative Case Studies: Four International Models

22.7.1 United Kingdom: YouGov's MRP Revolution

The evolution of British election forecasting is among the most instructive international stories in the field. Following the 2015 disaster (described in Chapter 20), the British polling industry undertook substantial methodological reform. The most significant innovation was YouGov's development and deployment of a constituency-level MRP model, first applied at scale in the 2017 snap election.

YouGov's model uses a large national poll (typically 50,000–100,000 respondents collected over multiple weeks) as the raw input. The multilevel regression estimates individual voting intention as a function of age, education, social class, 2016 Brexit referendum vote, and geographic region. The poststratification step applies these estimates to each of the 650 constituencies using Census data on the demographic composition of each constituency.

The 2017 result: YouGov's model predicted a hung parliament; virtually all other polling averages predicted a comfortable Conservative majority. The actual result was a hung parliament. YouGov's model was directionally correct where the industry consensus was wrong.

The mechanism: The 2017 election featured an unusually strong swing among young, educated voters toward Labour — a shift that was demographically concentrated in ways that constituency polls (which were either absent or had very small samples) missed. MRP, by pooling demographic data across constituencies and applying a model that explicitly included age and education interactions, captured this pattern.

22.7.2 The 2017 YouGov Model in Depth

The YouGov 2017 constituency model offers a rich case study in how MRP captures what standard polling misses, and what methodological choices drove its success.

The sample size advantage. YouGov fielded approximately 50,000 interviews across the 2017 campaign — far larger than any individual constituency poll and providing enough national-level data to estimate demographic group preferences with high reliability. This sample size is the foundation: without it, the multilevel model lacks the statistical power to estimate the key interaction terms (the relationship between age, education, Brexit vote, and party preference) with sufficient precision.

The Brexit vote variable. The inclusion of respondents' 2016 Brexit referendum vote as a predictor was analytically crucial. Brexit vote was both strongly predictive of 2017 voting intention (Remain voters swung sharply toward Labour; Leave voters held more strongly for Conservative) and strongly geographically concentrated in ways that constituency demographic composition only partially captured. By including Brexit vote directly in the model, YouGov captured a political dimension of variation that was invisible to standard demographic-only MRP.

The constituency-level geographic random effect. Beyond demographics, the model included constituency-level random effects that allowed local geographic variation not explained by demographics — local candidate quality, regional economic conditions, the specific pattern of 2015 UKIP vote collapse — to influence predictions. Districts with similar demographics but different local contexts received different predictions, reflecting the model's ability to learn from patterns in the available constituency-level data.

Where the 2017 model succeeded and why. The model captured the hung parliament outcome because it correctly estimated the geographic distribution of the Labour surge. Young, educated, urban constituencies in large cities and university towns swung dramatically to Labour; these constituencies were correctly identified by the model's demographic profile as Remain-voting, high-education areas likely to follow the national trend for that demographic. Standard constituency polls, fielded in relatively safe Conservative constituencies in expectation of a comfortable majority, missed the Labour surge in these demographically predictable but polling-sparse areas.

Limitations of the UK model. The YouGov MRP has not been uniformly successful. In the 2019 election, its final estimate was directionally correct (Conservative majority) but underestimated the Conservative margin in a way that echoed the 2015 failure — some of the demographic relationships had shifted between 2017 and 2019, and the model did not fully capture the further collapse of the Brexit Party vote in Conservative-held seats. More fundamentally, the model's success in 2017 depended on demographics being the primary driver of the realignment; if political change is driven by non-demographic factors (candidate quality, local scandal, targeted mobilization), MRP's demographic machinery offers fewer advantages.

🌍 The UK System's Unique Features. The first-past-the-post electoral system in the UK produces constituency-level outcomes from plurality vote. Unlike proportional representation systems, a 3-point national swing can flip dozens of seats simultaneously — or relatively few, depending on geographic distribution. MRP is particularly well-suited to FPTP systems because it directly estimates constituency-level outcomes rather than national vote shares.

22.7.3 Germany: Coalition Complexity and Multi-Party Vote-Share Forecasting

German federal elections present a fundamentally different forecasting challenge. The German electoral system combines single-member constituency seats with proportional representation (a "mixed-member proportional" system). The Bundestag's composition is determined primarily by party list vote shares, with the constituency seats redistributed to ensure overall proportionality. In practice, this means that forecasting the German election requires:

1. Estimating vote shares for six or seven significant parties (CDU/CSU, SPD, Greens, FDP, AfD, Die Linke, and potentially others)
2. Translating those vote shares into seat allocations (which is relatively straightforward given the proportional system)
3. Forecasting which coalition is most likely to form — which requires modeling post-election negotiation dynamics that are not reducible to vote shares

The 5% threshold problem. German election law requires parties to win at least 5% of the national list vote to receive proportional representation seats (or to win at least 3 direct constituencies). This threshold creates a distinctive forecasting complication: a party polling at 4.5% is not safely below the threshold — its actual result could plausibly land at 3.8% (far below, no seats) or 6.2% (clear entry, significant seat share). The uncertainty around threshold-crossing parties has an outsized effect on coalition formation probabilities, because a party that misses the threshold eliminates one potential coalition partner and redistributes its votes to the remaining parties.

Modeling coalition formation. The third step — forecasting coalition formation — is where German election forecasting has no clean methodological solution. Coalition formation depends on ideological compatibility, party leadership preferences, the specific results (which parties cleared the 5% threshold), and post-election bargaining that cannot be predicted from polling alone. Forecasters typically present scenario analyses: "If CDU wins 30% and SPD 25%, the most likely coalitions are X, Y, or Z, with probabilities estimated from historical coalition negotiation patterns."

Historical coalition patterns in Germany provide some predictive information: CDU/CSU and FDP have a long established relationship; SPD, Greens, and FDP formed the "traffic light" coalition in 2021; CDU and SPD have entered grand coalitions (GroKo) when the arithmetic forced it. But historical patterns are only moderately predictive of future coalitions, particularly in fluid political environments with new parties (AfD's presence has constrained coalition options by being widely considered beyond the pale of coalition eligibility).

📊 The 2021 German Election. The 2021 Bundestag election is illustrative of both the opportunities and limitations of German election forecasting. The CDU/CSU, which had led polls for much of the cycle, fell dramatically in the final months of the campaign to finish at 24.1% — a historic low — while the SPD surged from below 20% to win 25.7%. The final polling averages from the major German institutes captured this trend reasonably accurately, predicting SPD ahead of CDU by 1–3 points. But the coalition outcome — the traffic light coalition of SPD, Greens, and FDP — was one of three plausible configurations and was only settled after weeks of negotiation. The "forecast" of the election outcome in any meaningful sense required predicting not just vote shares but coalition formation.

22.7.4 France: Runoff Systems and Two-Stage Forecasting

France's presidential election uses a two-round system: if no candidate wins an outright majority in the first round, the top two finishers advance to a runoff two weeks later. This creates a distinctive two-stage forecasting challenge.

First-round forecasting: With five or more significant candidates, first-round French presidential polls must estimate vote shares for each candidate simultaneously. The challenge is twofold: pre-first-round polls are trying to predict outcomes in a multi-candidate race where strategic voting (and its absence) is crucial, and third-party candidates frequently collapse or surge in the final weeks as voters make strategic decisions.

Second-round forecasting: Once the two finalists are known, second-round forecasting resembles a standard two-candidate race — but with a key complication. The second-round electorate is not identical to the first-round electorate. Voters who supported eliminated first-round candidates must now choose between the two finalists, often reluctantly. Their transfer patterns — how Le Pen supporters vote when their candidate is eliminated, or how Mélenchon supporters vote when his is — are systematically measured through "voting intention in the second round" polling but are subject to significant uncertainty, particularly regarding turnout among the eliminated candidates' supporters.

2022 example: The 2022 French presidential election illustrates first-round forecasting challenges. Marine Le Pen and Emmanuel Macron advanced as expected, but the first-round votes for Éric Zemmour (7.1%) and Valérie Pécresse (4.8%) significantly underperformed their polling peaks earlier in the campaign. Pre-campaign polling had suggested Zemmour might reach 15%; the final result was less than half that. This kind of late-campaign collapse among new or polarizing candidates is common in multi-candidate systems and is notoriously difficult to forecast.

🌍 Transfer Vote Uncertainty. In two-round systems, the largest source of forecasting uncertainty is often not the first-round result but the transfer rates. A small shift in how first-round voters of eliminated candidates transfer to the two finalists can substantially change the outcome. This uncertainty is irreducible from polling alone — it requires either direct measurement of second-round intention among first-round supporters of each candidate, or historical analysis of transfer patterns in comparable elections.

22.7.5 Brazil: Data Quality, Polarization, and the 2022 Forecasting Challenge

Brazil's 2022 presidential election — between incumbent Jair Bolsonaro and former president Lula da Silva — was one of the most intensely watched and technically challenging international election forecasting problems of recent years.

What made Brazil 2022 difficult:

Massive polling divergence. In the months before the first round, individual polls from reputable Brazilian firms showed Lula's first-round lead varying from 6 to 18 points. This is not a marginal methodological difference; it is a fundamental disagreement about the structure of the electorate. Some of this variance reflected genuine methodological differences — telephone vs. in-person polls, different likely voter screens, different approaches to Bolsonaro's historically lower willingness to participate in surveys.

Systematic Bolsonaro underestimation. Brazilian forecasters faced an analog of the American 2016 and 2020 problem: Bolsonaro supporters appeared less likely to participate in surveys, or less likely to disclose their true preferences, than Lula supporters. Multiple cycles of Brazilian polling had shown Bolsonaro outperforming pre-election polls.

Regional concentration of support. Brazil's enormous geographic and demographic diversity — a country of 215 million people spanning a continent — creates severe geographic concentration in voting patterns. Polling that accurately represented São Paulo State might dramatically misrepresent Mato Grosso. National polling averages conceal substantial regional variation.

The actual first-round result: Lula 48.4%, Bolsonaro 43.2%. Most polls had shown Lula above 50% in the first round, which would have ended the election immediately; the actual result forced a second round. This miss — similar in structure to American polling errors — reflected the same underlying mechanisms: Bolsonaro nonresponse and regional sampling challenges.

⚖️ Ethical Analysis: Polling and Democratic Integrity. In Brazil, as in other polarized environments, polling has become itself a political battleground. Bolsonaro and his allies repeatedly attacked polling firms as biased and coordinated campaigns to discourage his supporters from participating in polls. This is a direct analog to Trump's 2020 attacks on polling. When political leaders actively undermine polling participation among their supporters, the partisan nonresponse bias discussed in Chapter 20 is not merely a methodological artifact — it is an intentional political strategy. Forecasters and polling firms operating in this environment face an ethical obligation to be explicit about this mechanism and its implications for their estimates.

22.8 Specific Data Challenges in International Forecasting

A recurring challenge in international forecasting is the data-scarce environment: elections where polling coverage is thin, historical data is limited, or the underlying data infrastructure has fundamental quality problems.

22.8.1 The Spectrum of Data Availability

International electoral data availability ranges enormously:

• Data-rich environments: United States, UK, Germany, France, Canada, Australia. High-frequency polling from multiple independent firms, extensive historical records, detailed precinct-level vote data, voter registration or administrative data.

• Moderately data-available: Brazil, Mexico, India, South Korea, Japan. National polling available but with quality concerns; regional polling inconsistent; historical records present but requiring substantial cleaning.

• Data-scarce: Much of sub-Saharan Africa, Central Asia, many island nations. Polling infrastructure weak or nonexistent; election administration data available only at the national or regional level; historical records limited or unreliable.

22.8.2 Cross-National Differences in Polling Infrastructure

Even within the "data-rich" category, substantial differences in polling infrastructure affect forecast quality in ways that cross-national comparisons often obscure.

Polling frequency and the late-breaking information problem. American forecasters benefit from high-frequency polling — in a competitive cycle, new polls arrive daily or near-daily in the final weeks. This provides forecasters with rapid detection of late-breaking shifts in opinion (scandal, debate performance, economic news). Many European systems have lower polling frequency, with "polling blackout" periods in the final days before election day mandated by law (France prohibits publication of polls in the 24 hours before voting; Spain bans them in the five days before). These blackout periods create information gaps that increase uncertainty in final forecasts.

House effects and the absence of a reference standard. American forecasters can estimate "house effects" — systematic tendencies for individual pollsters to show results that skew Democratic or Republican — because there are enough polls from different organizations to compare. When a country has only two or three polling firms, estimating house effects is impossible: any systematic difference between firms could reflect house effects, true methodological differences, or random sampling variation. Without a reference standard, aggregate estimates are less reliable than they would be with a larger polling ecosystem.

Language and linguistic diversity as a systematic data gap. Countries with significant linguistic minorities face a structural gap in polling coverage that is difficult to close. In Canada, English and French polls are often reported separately (and their results can differ substantially), but polling in indigenous languages is essentially nonexistent. In Belgium, Flemish and Walloon polls are conducted separately, but the Brussels capital region — linguistically mixed and politically distinctive — is chronically underpolled relative to its electoral importance. These linguistic gaps are not random: they systematically exclude populations with distinctive political profiles from the polling universe.

22.8.3 Strategies for Data-Scarce Forecasting

When polling data is unavailable or unreliable, forecasters must rely more heavily on structural approaches:

Expert elicitation. Political scientists, country specialists, and diplomatic observers can provide probabilistic assessments of electoral outcomes based on qualitative knowledge of local conditions. Structured expert elicitation, following protocols developed by Tetlock and others, can produce calibrated probability estimates even when quantitative data is absent.

Historical analog matching. If current election conditions (incumbent approval, economic conditions, political context) resemble those preceding a past election in the same country or a similar country, the historical outcome provides a baseline probability estimate.

Social media and alternative data. Search trends, social media sentiment, and mobile communications data have been explored as partial substitutes for polling in data-scarce environments. Results are mixed: these data sources capture engagement and salience but not necessarily voting intention, and they are subject to severe selection biases in countries with uneven internet access.

Cross-national calibration. Estimates of economic voting effects or incumbency effects derived from data-rich countries can be imported to data-scarce environments, adjusted for structural features of the specific electoral system. This approach works best when the data-scarce country shares structural features with the countries from which the parameters are derived.

✅ Best Practice: Be Explicit About Data Limitations. A responsible international forecast in a data-scarce environment explicitly quantifies the uncertainty attributable to data quality rather than hiding it in a general confidence interval. Saying "we estimate the incumbent wins with 60% probability, but given limited polling coverage, our uncertainty is substantially higher than in comparable data-rich forecasting environments" is more honest and more informative than either refusing to forecast or presenting a confident number without caveat.

22.9 Forecasting in Parliamentary vs. Presidential Systems

The structural difference between parliamentary and presidential systems shapes forecasting methodology in ways that go beyond data availability.

Presidential systems (United States, Brazil, Mexico, France, South Korea) have relatively clear accountability: a single chief executive is elected by popular vote or electoral college. Forecasting the presidential election is a well-defined problem with a binary or near-binary outcome. Down-ballot races are independent (senators, representatives, governors have separate electoral mandates).

Parliamentary systems (UK, Germany, Canada, Australia, Scandinavia, most of the rest of the world) elect a legislature, and the government emerges from the legislature's composition through coalition formation. Forecasting a parliamentary election requires:

1. Estimating party vote shares (a multi-dimensional estimation problem)
2. Translating vote shares to seat allocations (which depends on the specific electoral formula — FPTP, proportional, mixed-member)
3. Estimating government formation probabilities (which depends on post-election negotiation)

None of these steps is independent. A 0.5-point shift in one party's vote share might cross a threshold (the 5% minimum in Germany, the regional seat threshold in Spain) that dramatically changes seat totals and therefore coalition possibilities.

The "electoral forecast" in a parliamentary system is therefore genuinely more complex than in a presidential system, because the object of interest (which government will form) is further removed from the directly observable input (vote shares). A forecaster who publishes "we give the Conservative Party a 62% chance of forming a government" is making a much more complex prediction than one who publishes "we give the Republican candidate a 62% chance of winning the presidency."

🔵 Debate: Is Parliamentary Forecasting Harder? Some analysts argue that parliamentary systems are easier to forecast because vote shares are more predictable (given proportional representation, votes translate more linearly to seats, reducing the geographic concentration effects that make American House forecasting complex). Others argue that coalition formation is genuinely harder to forecast than a two-candidate race. The evidence is mixed: in systems with stable coalition patterns (Scandinavia), government formation is relatively predictable; in systems with novel or fragile coalitions (Israel, Italy), it is essentially random conditional on vote shares.

22.10 A Practical Guide to Following Down-Ballot Races Analytically

For students and practitioners who want to track down-ballot races systematically, the following framework identifies the most informative data sources and the key analytical questions to ask at each stage of the cycle.

22.10.1 The Early Cycle (6-18 Months Before Election Day)

At this stage, the most valuable activities are structural rather than horse-race oriented:

Identify the universe of competitive races. For House races, the Cook Political Report's initial ratings provide a starting list. For state legislative races, redistricting analysis from the Princeton Gerrymandering Project or the Brennan Center identifies which chambers are structurally competitive based on map geometry.

Establish baseline partisan lean. For each race you're tracking, calculate the district's recent presidential performance (from Dave Wasserman's Cook Report data or Voting and Election Science Team precinct-level results). This provides the structural context for everything that follows.

Monitor candidate filing. In many cycles, the most important early news is who files to run. A well-funded challenger filing in a nominally safe seat is an early signal of potential competitiveness. An incumbent's failure to file for re-election creates an open-seat race with fundamentally different dynamics.

22.10.2 The Primary Period (3-12 Months Before Election Day)

Follow fundraising filings. FEC quarterly and monthly filings for federal races are the most reliable early indicators of competitive potential. Look for: challengers who reach $500,000+ in the first filing period (indicates a viable campaign); incumbents with unusually low fundraising (financial vulnerability signals); outside groups filing Independent Expenditure reports for specific districts (sophisticated organizations' money is a leading indicator of expected competitiveness).

Track primary outcomes. The outcome of primaries — particularly which type of candidate wins — substantially affects general election competitiveness. A Republican primary that nominates a candidate positioned at the median of the district electorate is different from one that nominates a candidate positioned at the median of the Republican primary electorate. These positions are not always the same.

Watch for early polling. Academic groups (YouGov, Cooperative Election Study) and some advocacy organizations conduct early-cycle surveys. These are imprecise but directionally useful for identifying whether any districts outside the expected competitive universe are beginning to shift.

22.10.3 The General Election Period (0-90 Days Before Election Day)

Integrate all available signals. With competitive races identified and candidates known, the analytical task is integrating: any available polling (weighted by quality and recency), fundraising data (financial position through the most recent FEC filing), field organization signals (observed canvassing activity, volunteer density), and the evolving generic ballot environment.

Watch for late-cycle signals. In the final 30 days, pay attention to: sudden increases in outside group spending (suggests professional forecasters have shifted their assessment of competitiveness), media organization endorsement decisions (local newspaper endorsements have diminished but not zero effect on competitive margins), and any candidate-specific events that might move a district out of or into the competitive range.

Maintain calibrated uncertainty. Even at the very end of the cycle, down-ballot races carry substantially more uncertainty than statewide Senate races. A House district model that assigns 70% probability to one candidate is offering a substantially less confident prediction than a Senate model with the same probability — because individual-race polling is less reliable and local effects are harder to model.

📊 Key Data Resources for Down-Ballot Tracking

22.11 Who Gets Counted: A Global Perspective

The analytical theme of "Who Gets Counted" takes on particular resonance in international election forecasting. The question of whose preferences are captured by polls — and whose are systematically excluded — is not merely a methodological problem. It is a question of democratic representation.

In Brazil, rural and economically marginal voters have historically been underrepresented in polling samples because telephone and online polls underrepresent populations without reliable telephone or internet access. The voters most likely to be missed by standard sampling are precisely the voters who have benefited least from the formal economy and who are often the most consequential swing voters in elections where economic populism is a central theme.

In India, polling in Hindi-speaking northern states with better infrastructure is substantially more accurate than polling in linguistically diverse southern states or in rural areas with low cellular coverage. The "Indian polling average" that international observers cite is systematically more representative of urban, educated, connected voters than of the agricultural populations that constitute a majority of the electorate.

In sub-Saharan Africa, election-day prediction based on vote tallying from randomly selected polling stations (parallel vote tabulation) has in some cases produced more accurate early estimates of outcomes than pre-election polling — because the actual votes, once counted, represent the population better than any pre-election survey could.

The methodological implication is the same in all of these contexts: the measurement tool shapes the map of public opinion that forecasters work from, and that map is always a selective representation of reality. Understanding who is systematically absent from the map is a prerequisite to understanding the limits of the forecast.

22.12 Conclusion: The Forecasting Frontier

The progression from Senate forecasting (Chapter 21) to House forecasting to international election forecasting illustrates a consistent principle: the forecasting methodology must be matched to the data environment, not the other way around. A Monte Carlo simulation grounded in 17 high-quality polls from a state-level Senate race produces meaningful output. The same simulation applied to a Brazilian state with 2 polls of uncertain quality is false precision wearing statistical clothing.

The appropriate response to data scarcity is not to refuse to forecast — the need for probabilistic assessment does not disappear because the data is thin. It is to calibrate uncertainty honestly, lean more heavily on structural and historical information, and communicate explicitly about what the forecast does and does not capture.

The expanding global reach of electoral analytics also confronts the "Who Gets Counted" theme in its most literal sense. The populations whose preferences are most difficult to measure — the rural poor, the linguistically marginal, the institutionally mistrustful — are precisely the populations whose electoral participation may be most consequential in highly competitive environments. Building forecasting methods that can reach them is not just a technical problem; it is a democratic one.

The map is always incomplete. The territory is always larger than the map. What distinguishes a rigorous forecaster from a careless one is not the ability to eliminate this gap but the discipline to know where it is, how large it is, and what it implies for the use of the forecast.

Summary

• House forecasting relies primarily on generic ballot signals, historical district baselines, incumbency effects, and fundraising data, rather than individual-race polling — which is thinly distributed across 60–80 competitive seats.
• The generic ballot translates imperfectly to seat projections because of geographic vote distribution (efficiency), the specific location of competitive districts, and the nonlinear seats-votes relationship.
• Senate forecasting faces a small-sample problem per cycle and is particularly sensitive to candidate quality effects that generic environmental models cannot capture.
• State legislative forecasting is almost entirely structural — demographic, registration, and historical presidential results — because individual-race polling is nearly absent. It is the hardest forecasting problem in American politics at the individual-race level, due to redistricting discontinuity, inconsistent data infrastructure, and large local candidate effects.
• MRP (Multilevel Regression and Poststratification) allows subnational estimates from national surveys by modeling individual demographics and poststratifying to the Census distribution; YouGov's UK constituency model is the leading applied example, with the 2017 hung parliament prediction its signature success.
• International election forecasting faces three distinctive challenges: multi-party competition requiring simultaneous vote-share estimation, electoral system translation from votes to seats, and government-formation prediction in parliamentary systems.
• Data quality and coverage vary enormously across countries; forecasting methodology must adapt to data availability rather than assuming a standardized polling environment.
• International data challenges include not just data quantity (polling frequency) but data quality (house effects, linguistic coverage gaps, regulatory constraints like polling blackout periods).
• The "Who Gets Counted" theme is universal: the populations hardest to survey are often the populations whose votes are most consequential in close elections.

Key Terms

Generic ballot — The national survey question asking which party's candidate a respondent would support for Congress; used as the primary national-level input to House seat projections.

Seats-votes curve — The historical relationship between a party's national vote share and its seat share in a legislature; typically nonlinear, with higher swing ratios near the 50-50 threshold.

Efficiency gap — A measure of partisan advantage in legislative maps that captures the differential "wasting" of votes between parties across districts.

Multilevel Regression and Poststratification (MRP) — A statistical technique that combines individual-level survey data with population-level demographic data to produce subnational opinion estimates without requiring direct sampling in each subnational unit.

Poststratification — The step in MRP where regression-based predictions for each demographic cell are combined with census data on the demographic composition of each geographic unit to produce weighted estimates.

Parliamentary system — A system of government in which the executive emerges from the legislature and is accountable to it; contrasted with presidential systems where the executive is independently elected.

Parallel vote tabulation — A method of election verification that samples a random selection of polling stations and tallies results in real time to project overall outcomes independent of official counts.

Transfer rates — In two-round or preferential voting systems, the proportion of first-choice voters for an eliminated candidate who transfer their support to each remaining candidate.

5% threshold — The minimum national vote share required for a party to receive proportional representation seats in the German Bundestag; creates distinctive forecasting uncertainty for parties polling near this boundary.

Polling blackout — Legally mandated periods before an election during which publication of new polls is prohibited; common in European democracies and increases late-stage forecasting uncertainty.