🇧🇷 Leia em português

Lula’s shock re-election on 25 October followed a first-round loss, in which Flávio Bolsonaro won 47% of the vote to Lula’s 45%. Bolsonaro led the opinion polls right up to election day and between-round polling suggested he’d get most of the third-candidate votes. But to many people’s surprise, we woke up to news that Lula had gained 10M votes in the second round, and won with 51.8% of the vote.

It was the first time in Brazilian democratic history that a presidential candidate had won the second round after losing the first. Early write-ups of the election pointed to an energetic between-round campaign by Lula’s supporters, with rallies attended by millions, key “clippable” moments in the TV debates, and late spending promises by the incumbent president.

In particular, there has been a lot of focus on the grassroots online campaigns run by Lula’s supporters. Most notably, André Janones, an allied¹ member of the Câmara dos Deputados, ran a WhatsApp channel called “Exército do Lula” (Lula’s Army) (Dyer et al., 2026). It broadcast mobilisation messages to over 11M supporters, encouraging them to share social media posts to boost pro-Lula content in the algorithms. Alexandre Afonso has written an excellent blog post analysing the content of the channel.

In his book Prototype Politics, Kreiss (2016) describes election post-mortems as a social process of “meaning-making”, in which particular elements of a campaign are picked out and narratives emerge that these elements “won” (or lost) the election. In short, we all tend to pick out particular elements of a campaign and attribute its success to them.

In reality, working out why a campaign won is very challenging. Among political scientists, there is a long-running and unresolved debate about whether campaign periods matter at all for election outcomes, and even if they do, isolating the effects of particular interventions requires careful research design (Foos & Metson, 2026).

Nonetheless, the sheer scale of the mobilisation achieved warrants some exploration. Beyond the “Exército do Lula” WhatsApp mobilisation, other smaller initiatives shed some light on the nature of the between-round campaign. In particular, the webpage “Onde dá pra conversar” (Where you can have a conversation) showed which areas could be swung towards Lula and encouraged voters to reach out to their neighbours to persuade and mobilise them. This tactic echoes the decentralised “relational organising” popularised by Obama’s 2012 campaign (McKenna & Han, 2014).

The site highlighted polling sections (seções de votação) where the number of abstentions and invalid votes was larger than the gap between Lula and Bolsonaro. It also offered detailed suggested scripts for talking to voters who had abstained, spoiled their ballot or voted for another candidate in the first round.

Political scientists may be interested to learn that the advice on persuasion appears to draw very closely on the logic of “deep canvassing” interventions tested in a series of experiments in the US (Broockman & Kalla, 2016; Kalla & Broockman, 2023):

Beyond persuasion, supporters of Lula’s party, the Workers’ Party (Partido dos Trabalhadores, PT) placed a strong emphasis on mobilising people who hadn’t voted in the first round. At first this might seem puzzling: Brazil has compulsory voting for much of the population (literate adults aged 18–70). Even so, turnout does vary. For voters over 70 and illiterate voters, voting is voluntary. Voters can also justify their absence, and the penalty for not voting is small: usually around R$3.51 (US$0.70), though it can be up to R$35 for wealthier voters (Morales, 2026).

On the surface, mobilisation appears to have played a major role in this election: the first round saw the lowest first-round turnout since 2002, and turnout increased substantially in the second round (Figure 1).

Analyses of where Lula gained the most votes between rounds suggest that Lula won by gaining voters in areas that we wouldn’t typically expect to vote for the PT (see posts by Juan Machado and Owen Winter). This doesn’t fit comfortably with a pure mobilisation story, where we’d expect generally supportive areas to have been where gains were strongest.

To dig into this more, I use section-level data to describe how the election played out between the two rounds, and to see what we can learn about how Lula achieved his historic win. Seções de votação (polling sections) are the smallest unit at which results are reported. Each is a single urna eletrônica (voting machine), to which around 300 voters are assigned (μ=318, σ=66).

pr = transpose(proj)
total = src => d3.sum(pr.filter(d => d.source === src), d => d.projected)
share = (src, ch) => d3.sum(pr.filter(d => d.source === src && d.choice === ch), d => d.projected) / total(src)
fmt = x => x ? `${+(x / 1e6).toFixed(1)}M` : "0"
groupForm = src => Inputs.form({
  turnout: Inputs.range([0, 100], {step: 1, value: Math.round(100 * (1 - share(src, "none"))), label: "Cast a valid vote (%)"}),
  lula: Inputs.range([0, 100], {step: 1, value: Math.round(100 * share(src, "lula") / (1 - share(src, "none"))) || 50, label: "Of whom vote Lula (%)"})
})

{
  const settings = {"Third candidates": third, "Blank/null": blank, "Abstained (1st round)": abst}
  const loyal = {"Lula (1st round)": "Lula", "Bolsonaro (1st round)": "Bolsonaro"}
  const rows = Object.keys(fills).flatMap(source => {
    const n = total(source)
    if (loyal[source]) return [{source, choice: loyal[source], projected: n}]
    const v = settings[source].turnout / 100, l = settings[source].lula / 100
    return [
      {source, choice: "Lula", projected: n * v * l},
      {source, choice: "Bolsonaro", projected: n * v * (1 - l)},
      {source, choice: "Blank/null/abstain", projected: n * (1 - v)}
    ]
  })
  const totals = ["Bolsonaro", "Lula", "Blank/null/abstain"].map(choice => ({choice, total: d3.sum(rows.filter(d => d.choice === choice), d => d.projected)}))
  const gap = totals[1].total - totals[0].total
  return html`<div style="color: #8c8c8d; font-family: Inter, sans-serif">
    <p style="font-weight: bold; margin-bottom: 0">Projected second-round vote</p>
    <p>${gap > 0 ? "Lula" : "Bolsonaro"} wins by ${fmt(Math.abs(gap))} votes</p>
    ${Plot.plot({
      width: 760,
      height: 300,
      marginLeft: 130,
      marginRight: 60,
      style: {fontFamily: "Inter, sans-serif", fontSize: "12px", color: "#8c8c8d"},
      x: {label: null, grid: true, nice: true, tickFormat: fmt, domain: [0, Math.max(70e6, d3.max(totals, d => d.total))]},
      y: {label: null, domain: totals.map(d => d.choice)},
      color: {domain: Object.keys(fills), range: Object.values(fills), legend: true},
      marks: [
        Plot.barX(rows, {x: "projected", y: "choice", fill: "source", stroke: "white"}),
        Plot.text(totals, {x: "total", y: "choice", text: d => fmt(d.total), dx: 6, textAnchor: "start", fill: "currentColor"})
      ]
    })}
  </div>`
}

viewof abst = (reset, groupForm("Abstained (1st round)"))

viewof blank = (reset, groupForm("Blank/null"))

viewof third = (reset, groupForm("Third candidates"))

viewof reset = Inputs.button("Reset defaults")

Second-round results

As we now know, Lula did meet the challenge. Figure 5 shows the previous projection with the actual second-round results overlaid. We see that Lula gained 10.4M votes between rounds, 5.3M more votes than the simple projection predicted. Bolsonaro, by contrast, won 2.2M fewer than projected and added just 3.7M to his first-round total. And 3.1M more people than projected cast a valid vote.

The bottom panel of Figure 6 shows that Lula gained votes almost everywhere: he lost votes in only 3% of Brazil’s nearly 500,000 sections. Abstention and spoiled-ballot voting fell on average, but not uniformly, with a sizeable proportion of sections seeing increases.

Runoff elections are, of course, designed so that third-candidate voters can reallocate their votes to the leading candidates. Therefore, we would expect both to have gained votes. The map (interactive version) and top-right density plot in Figure 6 therefore show Lula’s and Bolsonaro’s relative gains by section:

Relative gain = (Lula_R2 − Lula_R1) − (Bolsonaro_R2 − Bolsonaro_R1)

Lula outgained Bolsonaro in 85% of sections and, on average, gained 13 more votes per section than Bolsonaro did between rounds.⁴

⁴ This page shows an interactive version of the map and a lookup table for vote shares and gains for each municipality and city for voters abroad.

Where did these gains come from?

The bulk of each candidate’s runoff votes likely came from their first-round voters. In theory, there are four sources from which candidates could have gained new votes between rounds:

1. Mobilisation: getting people who did not vote in the first round to vote (i.e. get-out-the-vote, or GOTV).
2. Spoiled-ballot persuasion: persuading voters who spoiled their ballot in the first round (by voting blank or null).
3. Third-candidate persuasion: persuading voters who backed a third candidate in the first round.
4. Opponent persuasion: persuading voters who backed their opponent in the first round to change their vote.

We’ve already mentioned the difficulty of attributing the result of an election to a campaign, but interpreting how votes flowed between candidates raises a second problem with the data we have.

To understand how much of Lula’s runoff vote came from each of these sources, we need to know how first-round voters voted in the second round. This type of information is exactly what the poll we used to project the second round (Table 2) contained. However, that poll represents voting intention, and we don’t (yet) have a similar poll for how voters actually changed their vote.

At this point in time, we only have election results at the aggregate level, so we can only look at changes in shares and averages. Methods for ecological inference attempt to use this information to infer how individuals changed their vote, but doing so risks making an ecological fallacy.

To be honest, I’m not an expert in these methods, but very roughly, the idea is to use areas that have very high or low rates of one type of voter in the first round to learn about what those types of voters do in the second (see the appendix for a more detailed explaination of the intuition). Crucially, to transfer information from one area to others, we have to believe that the transition rates are constant across areas (Romero et al., 2020). Unfortunately, this is a very strong assumption and everything that follows should be read with that caveat in mind.

Descriptive predictors of Lula’s gain

Figure 7 plots correlations between Lula’s relative gain and various predictors. It shows that Lula tended to gain more in polling sections where there were larger increases in turnout and larger decreases in blank/null votes. In sections where turnout rose by 10 more voters, Lula’s gain over Bolsonaro was about 3 votes larger; where blank/null votes fell by 10 more, about 9 votes larger.

Interestingly, Lula’s gains also appear to be stronger in sections with more first-round votes for Cury, who, recall, endorsed Bolsonaro for the runoff. For every 10 more first-round Cury voters in a section, Lula’s relative gain was 5 votes larger.

Ecological inference model

To directly estimate vote transfers, I turn to a more formal ecological inference model. To be clear, these methods allow us to produce estimates, and can improve on simple associations, but none of them solve the fundamental problem of identifying vote shifts from aggregate data.

I use an LPHOM (“Linear Programming under Homogeneity”) estimator (see Romero et al., 2020), mostly because it is relatively fast to run and has a neat R package to implement it.⁵ To further ease fitting the model, I aggregate the data up to much larger electoral zones (zonas eleitorais), which contain a median of 46,651 voters.

⁵ Admittedly, this is not my finest estimator-selection moment… sorry! But similar analyses (e.g. Daniel Schteingart) by people with higher standards and bigger computers, using different estimators, appear to concur with what I find.

Figure 8 shows the estimated shares of how each first-round group voted in the second round (see the appendix for a comparison to the polling estimates we used earlier).

The model suggests that both candidates retained almost all of their first-round voters. Voters who spoiled their ballot are estimated to have overwhelmingly split to Lula in the second round, alongside voters for Cury, Renan Santos and the left-wing third candidates. Amongst those who abstained, we predict 8% voted for Lula and 3% for Bolsonaro. Whilst these are small percentages, remember that this group is much bigger than any third candidate’s vote.

Building on that intuition, multiplying these shares by each group’s first-round size gives us an estimate of the composition of Lula’s second-round vote, shown in Figure 9.

Figure 9 suggests that, aside from retained first-round voters, Lula made most of his gains from third-candidate and spoiled-ballot persuasion. The estimated votes gained from these two groups of voters alone would have been enough to swing it for Lula. These models match the patterns suggested in Figure 7, where Lula’s gains correlated most strongly with falls in blank/null votes, Cury vote share and changes in turnout. Mobilisation is estimated to have added 2.7M votes which, whilst sizeable, would not have been enough alone to win the runoff.

Why did third-candidate and blank/null voters switch to Lula?

What might explain this pattern? A chunk of these voters may have been using the first round as an opportunity to send a clear, and potentially costly, signal to Lula that they were dissatisfied with him, but disliked Bolsonaro more, so voted tactically for Lula in the runoff.

These voters are both groups who showed up to vote, but deliberately chose against the top two in some way. This explanation makes most sense for voters who voted for a third candidate in the first round. The runoff is only called if no candidate reaches 50% of valid votes. Voting for a third candidate counts as a valid vote and increases the chance of a runoff round. Spoiling the ballot risks one of the top-two candidates winning in the first round. However, since the runoff system was introduced, all Brazilian elections but two presidential elections (1994 and 1998) have gone to a runoff, so voters probably don’t perceive this as a big risk.

Political science research does show that Brazilian voters use the runoff system strategically. In mayoral elections, there is only a runoff in municipalities with more than 200K voters. A paper by Fujiwara (2011) uses this cut-off to compare how many votes third candidates get in municipalities just above and below this threshold. The reproduced figure from the paper (Figure 10) shows evidence that Brazilian voters do indeed use the runoff system in this way.

Sources for third-candidate endorsements

Candidate	Party	Second-round endorsement	Source
Augusto Cury	Avante	Flávio Bolsonaro	Correio 24 Horas
Ronaldo Caiado	PSD	Flávio Bolsonaro	Correio 24 Horas
Romeu Zema	Novo	Flávio Bolsonaro	Correio 24 Horas
Clariana Barão	DC	Flávio Bolsonaro	Correio 24 Horas
Wilson Grassi	Democrata	Flávio Bolsonaro	Gazeta do Povo
Renan Santos	Missão	Neutral	Correio 24 Horas
Samara Martins	UP	Lula	Correio 24 Horas
Edmilson Costa	PCB	Lula	Correio 24 Horas
Hertz Dias	PSTU	Lula	Correio 24 Horas
Rui Costa Pimenta	PCO	Null vote	Correio 24 Horas
Leonardo Avalanche	PRTB	None (withdrew before first round)	Mais Goiás

Polling of third-candidate voters

In a survey by Quaest, run between 10 and 13 October (source), respondents were asked whether and how they voted on 4 October, and their second-round voting intention (Table 2).

First-round group (share of sample)	Bolsonaro	Lula	Undecided	Blank/null/won’t vote
Caiado (2%)	42%	23%	2%	33%
Renan Santos (2%)	40%	14%	3%	43%
Augusto Cury (2%)	30%	34%	6%	30%
Others (<1%)	38%	33%	7%	22%
Abstained or blank/null	38%	43%	2%	17%

The poll also reported runoff vote intention among voters who had abstained or spoiled their ballot in the first round. 43% said they would vote for Lula and 38% for Bolsonaro. However, we might have reason to doubt these estimates. Only 17% of them said they would abstain or spoil their ballot again.

It is unclear what population this group really represents, which makes using it for extrapolation difficult. It is very unlikely to be a representative sample of all first-round abstainers and spoiled-ballot voters.

If it were, it would imply a runoff valid-vote share of 93.6%, whereas the actual figure was 78.1%. Given how hard it is to reach non-voters in general, and especially in a compulsory-voting context, I assume this subgroup mostly represents voters who spoiled their ballot.

It may also include 16- and 17-year-olds, for whom voting is voluntary. For the projection, I therefore apply these figures only to first-round blank and null votes. Even applying them to abstainers as well, the projection still has Bolsonaro winning (see Figure 11).

How do the ecological inference and between-round polling estimates compare?

Figure 12 compares transition rate estimates from our ecological inference model and the between-round poll. The most striking difference is that the ecological inference estimates predict that almost no blank or null votes went to Bolsonaro. The ecological inference estimates also suggest far more Renan Santos voters split to Lula than predicted.

Neither of these is a “ground truth” of how voters voted in the second round. The poll may be based on an unrepresentative sample, or have missed changes that happened closer to the second round. The ecological inferences are subject to strong assumptions, as discussed.

We will have to see which, if either, was closer once high-quality individual-level post-election polling is released.

My understanding of the intuition behind ecological inference

We are trying to discover what proportion of each first-round group voted for each option in the second round (i.e. the empty cells in Table 3). This would let us see where Lula got his additional votes from: multiplying the size of each first-round group by the proportion of that group who voted for Lula in the second round. However, with only aggregate vote counts, we can’t observe these proportions directly.

First round ↓ / Second round →	Lula	Bolsonaro	Blank/null	Abstained	First-round total
Lula	?	?	?	?	53.9M
Bolsonaro	?	?	?	?	56.1M
Third candidates	?	?	?	?	9.3M
Blank/null	?	?	?	?	6.0M
Abstained	?	?	?	?	33.5M
Second-round total	64.3M	59.8M	4.2M	30.5M	158.8M

From aggregate data, we can see the totals from each round (the margins), but not the combinations of first- and second-round choices (the cells).

The idea of ecological inference is to use variation across areas to estimate these cells. Areas with a very high or very low share of a group tell us most about how that group behaves. If we assume that people transfer their votes at similar rates everywhere (a strong assumption), we can in principle recover the national transition rates. The toy example below shows an extreme version of this, and how it can go wrong.

Example

Imagine there are only two candidates, A and B, and everyone votes in both rounds. There are three areas with 1,000 voters each. We want to know what share of each first-round group voted A in the second round. The true rates, which we never get to see, are:

• 90% of first-round A voters vote A again, and 10% switch to B.
• 30% of first-round B voters switch to A, and 70% stay with B.

Across the three areas, we observe the totals for each round, but not how many voters switched between rounds (Table 4).

Round 1 ↓ / Round 2 →	A	B	Round 1 total
A	?	?	1,500
B	?	?	1,500
Round 2 total	1,800	1,200	3,000

These totals are consistent with many possible transition rates, so we can’t recover the true rates from this information alone. For example, all 1,500 first-round A voters could have voted A again, with 300 B voters switching to A. Or 1,200 A voters could have stayed, with 600 B voters switching. Both give exactly the same totals.

Now let’s break the results down into the three areas. Imagine that in the first round one area voted 100% for A, another voted 0% for A (i.e. 100% for B), and the third was split 50/50 (Table 5).

Area	First-round A share	Second-round A votes	Second-round A share
1	100%	900	90%
2	0%	300	30%
3	50%	600	60%

Area 1 contains only first-round A voters, so we can tell that 90% of them voted A again. Area 2 contains only first-round B voters, so A’s 30% share there in the second round means 30% of B voters must have switched to A.

If the rates are the same everywhere, then in Area 3, 90% of first-round A voters and 30% of first-round B voters vote A. That gives A a second-round share of 0.5 × 90% + 0.5 × 30% = 60%. All three areas lie on a straight line (Figure 13). Its value where first-round A share is 0% is the B → A rate, and its value at 100% is the A → A rate.

In real data we have far more areas, but we almost never see areas made up entirely of one group. Instead, we fit a line through many mixed areas and extend it to estimate the rates at each end, where the group makes up 0% or 100% of voters. This logic depends on transition rates being the same across areas.

Figure 14 shows simulated data from 200 areas in two worlds. In green, the circles and solid line show what we would observe if transition rates were the same everywhere. Extending the line to 0% and 100% first-round A share correctly recovers the true rates (the red crosses). In orange, the triangles and dashed line show a world with the same average rates, but where behaviour varies with local support: where A did well, A voters were more loyal and B voters were more likely to switch to A.

Extending this line underestimates the B → A rate (18% rather than 30%) and overestimates the A → A rate (103% rather than 90%).

This example also shows how linear extrapolation can lead to implausible estimates (we estimate a transition rate for A → A of 103%).

Crucially, the orange points still sit on a tidy straight line, so nothing in the data warns us that something is wrong. The ends of the line tell us how voters behave in the most A-heavy and B-heavy areas, and in this world those voters aren’t typical.

The methods used for ecological inference in practice are far more sophisticated than this. They can control for background features of places, partly relax the assumption that transition rates are the same everywhere, and rule out impossible estimates. But fundamentally, none of them can get around the need to make some strong assumptions.

References

Broockman, D., & Kalla, J. (2016). Durably reducing transphobia: A field experiment on door-to-door canvassing. Science, 352(6282), 220–224. https://doi.org/10.1126/science.aad9713

Dyer, G., Pooler, M., & Beatriz, L. (2026). How lula turned his election fortunes around in three weeks. Financial Times. https://www.ft.com/content/629894c7-e753-4e77-9794-c363c64dce9a

Foos, F., & Metson, L. (2026). Election campaigns. Preprint. https://doi.org/https://osf.io/preprints/socarxiv/sf7cv_v1

Fujiwara, T. (2011). A regression discontinuity test of strategic voting and duverger’s law. Quarterly Journal of Political Science, 6(3), 197–233. https://doi.org/10.1561/100.00010037

Kalla, & Broockman. (2023). Which narrative strategies durably reduce prejudice? Evidence from field and survey experiments supporting the efficacy of perspective‐getting. American Journal of Political Science, ajps.12657. https://doi.org/10.1111/ajps.12657

Kreiss, D. (2016). Prototype politics: Technology-intensive campaigning and the data of democracy. Oxford University Press.

McKenna, E., & Han, H. (2014). Groundbreakers: How obama’s 2.2 million volunteers transformed campaigning in america. Oxford University Press.

Morales, A. (2026). Punishment for not voting in brazil and how to fix it. The Rio Times. https://www.riotimesonline.com/punishment-for-not-voting-in-brazil-2026/

Romero, R., Pavía, J. M., Martín, J., & Romero, G. (2020). Assessing uncertainty of voter transitions estimated from aggregated data. Application to the 2017 french presidential election. Journal of Applied Statistics, 47(13), 2711–2736. https://doi.org/10.1080/02664763.2020.1804842

Wikipedia. (n.d.-a). 2026 brazilian general election. Wikipedia. Retrieved October 26, 2026, from https://en.wikipedia.org/wiki/2026_Brazilian_general_election#Campaign

Wikipedia. (n.d.-b). André janones. Wikipedia. Retrieved October 26, 2026, from https://en.wikipedia.org/wiki/Andr%C3%A9_Janones

The average study testing a campaign message goes something like this: a sample of (mostly) human¹ online respondents is randomly selected to be shown a message or video and then asked questions about whatever the campaign is interested in finding out about, often how they will vote. The analyst then compares the answers of those who saw the message with those who didn’t to work out the effect of the message.

Hewitt et al. (2024) study an archive of such message tests conducted by the company Swayable during the 2018 and 2020 US elections. Their conclusion is simple: pre-testing online messages creates huge electoral advantages by allowing campaigns to identify the most persuasive ads and focus their budget on these.

But there’s a slight problem. From academic research at least, we don’t know that any online ads really persuade voters. Our best guess is that, when deployed in the real world, online ads increase vote share by about 0.7 of a percentage point (Coppock et al., 2022). Moreover, we can’t take it for granted that other, more intensive forms of voter contact - such as door-to-door canvassing - persuade voters to change their vote either (Kalla & Broockman, 2017).

A more direct comparison comes from a brilliant paper by two researchers interested in how to persuade the public to support policies addressing the climate crisis. Carnes & Henderson (2025) essentially replicate a pre-testing-to-deployment pipeline, but carry on testing the interventions as they are sent out into the world. They found that messages which looked promising in pre-testing didn’t work when mailed out to people.

So what goes wrong? In a note, Ben Tappin has outlined a framework for understanding, from the perspective of practitioners, what surveys are useful for. Surveys can be misleading when the population who take them differs from the population who receive the message in the real world, or when the message simply doesn’t reach its audience in the real world. For an extensive description of all these features, do make sure to read the note.

Carnes & Henderson (2025) focus on attention. In a follow-up survey experiment they vary the extent to which respondents are forced to pay attention to the treatment, and find that messages only work when respondents are made to pay attention to them.

Attention requires people to actually receive a message. In experimental design, we call this “compliance”. Broadly, analysts have two options when dealing with non-compliance: ignore it or model it. Ignoring it, we can look at the effect of trying to deliver the intervention - the intent-to-treat (ITT) effect. Modelling it, we can first look at the effect of trying to give people the intervention on whether they actually get it, and use this to scale the ITT. This gives us the complier-average-causal-effect (CACE): the effect of the intervention among the type of people who would receive it if you attempted to contact them.

Often, however, there are multiple ways to define compliance in each case, so the CACE can mean very different things between studies. For example, when can we say someone has actually ‘received’ a political leaflet - when it’s gone through their letterbox, or only when they’ve seen it? What if they’ve seen it but not read it, or read it but not processed it? The point is that we can only measure one of these, so the CACE in this case would tell us something about the effect of the leaflet among people who can be successfully reached by mail.

It’s not just about attention, though. Sometimes treatments are simply easier to execute in a survey setting. In a project I’m working on, we randomly assigned campaign supporters to messages tailored to the issue they had previously stated they cared most about. The problem here is that the treatment (tailored messaging) requires the sender to know something about the person receiving it - in this case, the issue they cared about most. In a survey, this is trivial: ask them. But in the real world this type of information is hard to obtain, and quickly gets outdated. So whilst this type of issue tailoring has been shown to work in surveys, we found it didn’t work in the field (unless the message explicitly mentioned that the recipient had previously raised the issue).

What can practitioners do? Firstly, they should pay close attention to what a message test can tell them. In Tappin’s (2026) framework, they should reflect on whether they want to know whether to deploy a particular message, or which message to deploy. Message tests from surveys are rarely useful for understanding whether a message will be effective overall, because so many factors differ between a survey and the real world that the effect size estimated in a survey is often much larger than the effect in the field.² In political campaigns, this is generally not the question practitioners are asking (although perhaps it should be, given how little we know about the effects of persuasion interventions). Instead, they are often interested in finding out which message to use. Surveys can be more helpful here, as it’s less likely that something about the survey setting changes the relative effectiveness of messages. Secondly, practitioners are generally well placed to test things in the field, as they are delivered. This can be costly, and is certainly more costly than message testing in surveys. However, doing a few well-executed field tests can buy valuable knowledge about how future surveys may, or may not, tell us what might work.