A country-year record of national elections since 1945, and a first look at whether elections change political division, indifference, mobilization, and the quality of democracy. The story comes in three parts: how often the world votes, what voting more often does, and what holding an election does. Explore each below.
Before any consequences, the raw record. Elections have grown far more common since 1945, and a handful of countries go to the polls strikingly often.
National polling occasions worldwide each year, 1945 to 2026. The rise tracks decolonisation, the third wave of democratisation, and the post-communist transitions. 2026 is still running, so its count is partial and the series is drawn dashed there.
Total national elections in the chosen period, collapsed to one polling occasion per day. Choose the period and how many countries to show; hover for the count.
Here the treatment is how often a country votes. Two outcomes anchor the question: turnout, which we read as a proxy for indifference, and polarization, a proxy for division. The turnout pattern holds up across the models; the division one doesn't.
Each dot is a country, averaged over the chosen period. Move between the two outcomes, switch the period, toggle regime types, or hover a country. The fitted line and correlation recompute from whatever is showing, so you can see how much any group drives the pattern. The turnout panel also lets you swap the turnout source, so you can see whether the wider IDEA-based series and V-Dem's own series tell the same story.
The frequency coefficient (per one standard deviation) with country and year fixed effects, standard errors clustered by country. Turnout stays negative across every recent-frequency measure; division sits near zero and changes sign. Left of the line means more elections predict lower turnout or less division.
Rather than trust one model, each coefficient below is re-estimated across every reasonable combination of outcome, treatment, controls, fixed effects, and sample. Each dot is one specification, ranked by size; the dashed line is the median. This is the honest test of whether a result survives analytical choices.
Multiple imputation (Amelia, m set to the maximum missingness) pooled with Rubin's rules, so imputation uncertainty is carried into the intervals, against the complete-case estimate. Standard errors clustered by country, the treatment level.
Now the treatment is a single election year. What changes when a country goes to the polls? Mostly mobilization and the quality of democracy, less so division, and how much depends on the regime.
A country-year model with country and year fixed effects, tracing political division in the years around a national election. If elections drove division, the jump would land at the election year. Instead division drifts up before the vote too, so this looks like a build-up, not an effect of the election itself. Hover for the numbers.
The same country-year model, run for five annual outcomes, each standardised. Positive means more of the outcome in an election year. Elections mobilize, and sometimes inflame, but they do not clearly divide. Hover for the robustness across specifications.
V-Dem's Electoral Democracy Index and its five sub-indices, in an election year versus a normal one. Because the outcome is democracy, these models control for income and education but not for democracy itself. Elections coincide with gains almost everywhere, most of all in filling elected offices.
The same election effect, re-estimated within each Regimes-of-the-World category, from closed autocracy to liberal democracy. Each small panel is one outcome; each dot an effect for one regime type, with its 95% interval. The pattern the project cares about is the hump in mass mobilization: elections mobilize most in the contested middle, electoral autocracies and electoral democracies, and least at the two extremes.
Pick an outcome, then pin any analytical choice or leave it open. The curve and the summary recompute from the specifications that match, so you can check for yourself whether a result depends on the choices, rather than taking one model's word for it.
Everything here is real, drawn straight from the analysis outputs. Sources are attributed; nothing is invented.