The consensus is that home field advantage matters a lot in football. But how much does it matter at the World Cup? How much of an advantage do host countries have? Does being a host cause a country to do better? Let's estimate the causal effect of hosting on tournament performance.
Question
Home field advantage is important in football — playing in front of a home crowd can be a big advantage. Football prediction models always take into account home field advantage, and all the evidence is that it matters a lot. But how much does it matter at the World Cup? How much of an advantage do host countries have?
This is a hard question to answer. We can't just look at the performance records of host countries versus non-host countries. Host countries are systematically different from non-host countries — they're larger and/or richer on average than non-hosts. They have to have the infrastructure (stadiums, transportation, etc.) to host a massive international event like the World Cup.
What makes this question tricky is that the characteristics of hosts also are correlated with better performance. Larger countries have larger player pools and are more likely to produce star players. Football federations in richer countries generally have more resources to spend on training facilities and manager salaries. We need to control for these things so we can isolate the effect of being the host, separate from the effect of being larger and/or richer.
In this article I'm going to use causal inference techniques to estimate the average causal effect of being the host country on a country's performance — how deep a run a country makes. I find that host countries go about 1.5 rounds deeper into the tournament on average that they would have had they not hosted.
Mechanisms
We can't randomly assign host country status, so we'll have to estimate the causal effect of being the host using historical data. If we think of this as an experiment, the treatment is being the host country.
There are several causal mechanisms by which being the host might improve a country's performance at the World Cup. It's not just about getting to play in front of a home crowd; it's a bundle of advantages:
- The host country has a home field advantage at every match — they play every match in front of a home crowd.
- Host countries qualify automatically. They don't come into the tournament having just completed a grueling qualifying campaign. On the other hand, they don't have meaningful competitive matches to prepare for the tournament, so that could cut both ways.
- Hosts are placed in a top seeding pot regardless of their current form, which increases the probability of an easier group.
- Since host countries are announced 6+ years ahead of time, and host countries want to make a good impression, their teams are likely to get higher-than-usual investment. Just look at US Soccer hiring Mauricio Pochettino. If the US hadn't been hosting this tournament, US Soccer probably wouldn't have been willing to pay the kind of salary it takes to hire a manager like Pochettino.
These aren't threats to inference — they're the causal mechanisms by which being a host country should affect a country's performance. When we estimate effect of hosting, we're deliberately estimating the treatment effect of this whole bundle. We're not trying to isolate which pathway matters most — we're just asking whether the whole bundle of benefits that a host gets causes that country to perform better at the tournament than it would have if it hadn't been the host country (the counterfactual).
Data
To estimate the treatment effect of being the host country, I'm going to compare each host's performance in the tournament they hosted to their own performance in other tournaments. I'll need data on the performance of all qualified countries (not just host countries) for every tournament. I'll have one observation per qualified country per tournament. Our sample will include all tournaments (excluding the 2026 tournament, since it's ongoing). I'm going to use data from my World Cup database — available at worldcups.ai.
(Note: I'm going to treat Russia and the USSR as separate countries and Germany and West Germany as separate countries. In my analysis, I'm going to use fixed effects, and fixed effects assume a country's baseline characteristics are time-invariant. Treating these pairs of countries as the same entity, when they have different economic resources, player pools, and political conditions would add measurement error.)
The independent variable of interest in my analysis is whether that country hosted that tournament. The 2002 World Cup had multiple hosts (South Korea and Japan). I'll treat each host nation as a separate treated observation.
The dependent variable is how far they went in the tournament. I'll convert tournament results onto a simple numeric scale — group stage, round of 16, quarterfinal, semifinal, final. This is the right approach here because we need to be able to compare performance consistently over time despite substantial changes in tournament format and size. Note that this measurement strategy assumes that moving from one level to the next is an equal move. It's a reasonable assumption for our purposes.
Methodology
Let's walk through the research design. The quantity I want to estimate is the average treatment effect on the treated (ATT) — the average effect of hosting on the countries that actually hosted (which tend to be larger and/or richer than the average country). This won't tell us the average effect of any country being the host, just the kinds of countries that actually get to host.
(Note: Keep in mind this is a blog article, not a peer-reviewed paper. This isn't going to be a comprehensive analysis.)
There are four threats to inference that I'll need to address.
Threat 1: Confounding from selection into hosting. Hosts aren't randomly assigned. Host countries tend to be larger and/or richer for reasons that are independent from a country's performance in any given tournament. Just comparing the performance of hosts to non-hosts would confound the effect of hosting with these persistent differences.
To estimate the ATT, I'm going to use a differences-in-differences (DiD) design. Specifically, I'm going to estimate a two-way fixed effects (TWFE) regression. The idea is to compare each host country's performance in the year it hosted to that same country's performance in every other tournament it played (and not to other countries) because a country is the best available baseline for itself. This holds constant all time-invariant characteristics of a country, which will allow me to isolate whatever changed specifically in the year it hosted. A TWFE regression makes this comparison across every host, while simultaneously controlling for the things that are constant across all countries within each tournament (e.g., tournament-specific factors). That's why we need data on all qualified countries, not just hosts.
Here's the regression equation:
outcome_it = α_i + γ_t + β·host_it + ε_it.outcome_itis an ordinal variable (1 – 5) that captures how far the country progressed in the tournament (group stage, round of 16, quarter-finals, semi-finals, final).α_i: The country fixed effect. This term absorbs each country's persistent characteristics. This is what neutralizes Threat 1.γ_t: The tournament fixed effect. This absorbs whatever was common to every country in that tournament — tournament format, number of teams, era-specific rules, things like that.host_it: A dummy variable indicating whether the country was hosting the tournament.β: The coefficient on the host indicator. This is the ATT: the average effect of hosting a tournament versus not hosting the tournament.e_it: the error term.
This is a staggered DiD design — 22 hosts, in different years.
β is a weighted average of many within-country before/after comparisons, pooled across every host in the sample. The country and tournament fixed effects are doing what a before/after dummy variable does in a simple two-group DiD.What about controlling for team strength using a country's pre-tournament Elo rating? It's not a simple question. There's a trade-off between absorbing time-varying team quality and introducing a "bad control". On one hand, Elo captures temporary "golden generations" that country fixed effects miss, reducing residual variance and preventing the model from confusing a uniquely talented squad with a hosting advantage. On the other hand, because host countries are announced 6+ years in advance, any pre-tournament Elo rating is already contaminated by the treatment itself — it's affected by all the resources that a host invests in its domestic football programs during those years. Also, because hosts don't play competitive qualifiers, their Elo ratings would be a less responsive measure of their strength because Elo ratings usually place less weight on friendlies. On balance, it's better to omit this control.
Threat 2: Violation of the identifying assumption. The identifying assumption of a DiD design is the parallel trends assumption: if a host country weren't hosting, each host country's performance relative to its own baseline would have moved the same way the rest of the panel did. This doesn't require hosts to be equally good — the country fixed effects already account for that. It just requires that whatever gap that does exist would have stayed constant in the absence of treatment. If host countries were already on a different trajectory before hosting, for reasons unrelated to hosting itself, this assumption would fail and β would pick up that pre-existing trend rather than just the effect of hosting.
I'll test this assumption below with an event-study specification. I'll replace the single
host_it indicator with a set of indicators for tournaments relative to the host year (two before, one before, the one the country hosts, one after, two after). If the pre-treatment coefficients are near zero, then the team's performance wasn't already trending upward before hosting. That would be direct evidence that the parallel trends assumption holds. If these coefficients do show a pre-existing trend, then the DiD estimate is likely contaminated by that violation of the parallel trends assumption.Threat 3: Treatment effect heterogeneity under staggered adoption. A two-way fixed effects DiD with staggered treatment timing can produce biased estimates when treatment effects are heterogeneous across cohorts or change over time. The mechanism is that TWFE implicitly uses already-treated units as part of the comparison group for units treated later. If the effect of hosting in 1970 differs from the effect of hosting in 2010, the 1970 host's post-treatment trajectory contaminates the implicit control group for later hosts, rather than serving as a clean baseline. This is about treated units appearing as controls for other treated units when effects aren't constant across them. Given how much the tournament has changed over nearly a century (format, size, tactics, etc.), assuming a constant host effect across eras is a big assumption. I'll use a synthetic control design as a robustness test to check this.
Threat 4: Unreliable standard errors from too few treated clusters. Since the unit of observation is a tournament-country, the observations for each country aren't independent. I'll need to cluster the standard errors by country. But with only
22 host countries in the sample, I have too few treated groups to rely on the asymptotic properties of cluster-robust standard errors (the number of treated groups matters, not just the total number of groups). Instead, I'll use a wild cluster bootstrap to estimate the standard errors.One other note on independence. One tournament, 2002, had two host countries — South Korea and Japan — whose error terms may be correlated with each other through shared regional or organizational factors that
γ_t doesn't absorb. This risks violating the Stable Unit Treatment Value Assumption (SUTVA). This is unlikely to meaningfully bias the ATT, given it affects just 1 pair of 22 treated observations. Findings
My estimate of the ATT using my TWFE model is
β = 1.453. This means that being the host is associated with going 1.5 rounds deeper into the tournament, on average, across the countries that have hosted. In this model, there are 467 tournament-country observations, 63 team fixed effects, and 22 tournament fixed effects. The model drops one host — Qatar (2022) — because it's a fixed-effect singleton: Qatar has no other World Cup appearance in the sample, so the model can't do a within-country comparison.One of the threats to inference — Threat 4 — is that cluster-robust standard errors can be unreliable when the number of treated clusters is small, which is exactly the situation here:
21 treated countries out of 63 total. To make sure that I'm not estimating overly-precise standard errors, I use three different methods, each resting on progressively weaker assumptions:-
Cluster-robust asymptotic standard errors. This is the standard method, but you're relying on asymptotic theory. You're assuming that there are enough clusters for the sampling distribution to behave predictably. That's not the case here. The standard error is
0.224(p < 0.001). -
Wild cluster bootstraping. In this method, you resample the residuals cluster-by-cluster to build an empirical distribution instead of relying on asymptotic normality. This is important when you only have a few clusters. The standard error is
0.219(p < 0.001). -
Randomization inference. In this method, you don't make any cluster-asymptotic assumptions at all. Instead, you directly simulate the null by randomly reassigning host status among each tournament's actual participants, thousands of times. Then, you build an empirical null distribution from scratch. This is the most conservative of the three methods. The standard error is
0.301(p < 0.001).
The first two agree with each other (
0.224 versus 0.219) but both slightly underestimate the uncertainty. Randomization inference produces a wider null distribution (the standard deviation is 0.301). But the ATT is statistically significant across all three methods.To test the parallel trends assumption, I'll conduct an event study. The idea is to replace the host indicator with indicators for tournaments relative to each host's own hosting year.
Performance is substantively and/or statistically indistinguishable from zero for the tournaments before and after hosting, and jumps sharply only for the tournament where the country is hosting. This is direct evidence that the parallel trends assumption holds. The estimated ATT is
1.497, which is very close to the estimate above. The cluster-robust standard error is 0.279 and the wild cluster bootstrap standard error is 0.269, so the ATT is statistically significant in this model too. Robustness
Threat 3 is that a TWFE model with staggered treatment timing can be biased when the effect of hosting isn't constant across time periods, because already-treated units implicitly serve as part of the comparison group for units treated later. As a robustness test, I use a synthetic control design. It doesn't have this problem because it never uses one host's post-treatment trajectory as a baseline comparison for another host.
The method works case by case, rather than pooling all hosts into a single regression. For each host, I build a weighted "synthetic" version of that country from other teams that played in the same tournament that never hosted a tournament. This way, their own results aren't contaminated by a hosting effect of their own. I'll choose the weights so the synthetic country's pre-tournament trajectory matches the actual host's as closely as possible. Then, I compare the host's actual performance in the tournament it hosted to the performance of the weighted synthetic comparison in that same tournament. The gap between the two is the estimated ATT for that one host. So we get host-specific estimates. These estimates rely on a completely different set of assumptions than DiD, so if they agree, that's a good sign.
Of the
23 host-tournament cases (22 hosts, with co-hosts each counted separately), I can produce a valid estimate for 17. I have to exclude 6, for two different reasons:- No pre-treatment history (3 cases): Uruguay (1930), Italy (1934), Qatar (2022). Uruguay hosted the first World Cup, so no team, host or otherwise, had any prior history; Italy (1934) and Qatar (2022) were debutants as well as hosts.
- Fewer than 5 eligible donors (3 cases): Brazil (1950), Chile (1962), England (1966). I restrict donors to teams that have never hosted to avoid using a comparison country whose own result is itself contaminated by a hosting effect elsewhere. In the small fields of early tournaments, this rule leaves too few eligible donors.
Across the
17 valid cases, the average gap between actual and synthetic performance is 1.44 — almost identical to the ATT from the DiD model. The two designs rely on completely different identifying assumptions, and yet they converge on the same effect size. This is strong evidence that the size of the effect isn't an artifact of the staggered-adoption problem. 
