An Unfinished Campaign Is Not a Failure

On 15 November 2019, protesters filled the streets of La Paz against Jeanine Áñez’s interim government. In the standard dataset of the world’s revolutionary campaigns, that campaign is coded as a failure. It is coded a failure because data collection stopped in 2019. Eleven months later MAS won the presidency with 55 percent in the first round; Áñez left office on 8 November 2020 and was arrested the following March. The campaign got exactly what it wanted, and the spreadsheet says it lost. This is not a quibble about one row — it is most of the reason we believe nonviolent revolution has stopped working. That belief traces to a specific and widely cited claim in Erica Chenoweth’s The Future of Nonviolent Resistance (Journal of Democracy, 2020): success rates “remained above 40 percent, climbing as high as 65 percent in the 1990s,” but “since 2010, less than 34 percent” have succeeded. That sentence is the empirical backbone of a decade of commentary about protest exhaustion and authoritarian learning. I downloaded the underlying data — NAVCO 1.3, 622 maximalist campaigns, 1900–2019 — and reproduced it. Of 320 nonviolent campaigns, the 1990s cohort succeeded at 64.3 percent and the 2010s cohort at 35.5 percent. The published numbers are correct.

Then I looked at how a campaign gets its outcome. In NAVCO 1.3, 62 campaigns are flagged ONGOING. All 62 are simultaneously flagged FAILURE. Not one is flagged SUCCESS. A campaign still in progress when the file closed is entered as a defeat. Notice which decade that rule can touch. Of the 28 ongoing nonviolent campaigns, 27 began in the 2010s. Nineteen had been observed for three years or less — nine started in 2017, three in 2018, four in 2019. A 2019 campaign was given roughly zero years to win before being recorded as having lost, while every campaign from the 1960s through the 2000s has had its full life observed. The censoring is not noise spread across the series; it is a penalty applied exclusively to the most recent bin, precisely the bin the headline is about. And it bites hard here because success is fast: across the 150 successful nonviolent campaigns the mean duration is 1.27 years and the median is 1, with the 90th percentile at 4. A three-year observation window is genuinely censoring, not merely impatient.

The correction is elementary survival analysis: don’t score an event that hasn’t finished. Three specifications, all computed from the same file. As published, with unfinished counted as defeat: 1960–2009 gives 53.9 percent, the 2010s 35.5 percent, two-proportion z = 2.92, p = 0.0035. Drop censored cases from the denominator and the 2010s rise to 33/66 = 50.0 percent against 54.2 percent for 1960–2009 — an 18-point collapse becomes 4 points. Matched observation window, counting only successes achieved within three years of onset and only including campaigns with three years of observation, applied identically to every decade: 1960–2009, 46.6 percent; 2010–2016, 37.5 percent; z = 1.27, p = 0.20. The finding does not survive its own methodology. A result significant at p = 0.0035 becomes p = 0.20 when unfinished campaigns stop being counted as defeats, and this is robust to the window — at four years the 2010s score 41.4 percent against the 1980s’ 39.6 and the 1970s’ 36.4. Break the matched series into five-year blocks and it reads 76.9, 12.5, 62.5, 21.4, 23.1, 42.9, 65.5, 46.2, 58.3, 38.9: mean 44.8 percent, standard deviation 21.2 points, on cohorts of 8 to 45 campaigns. The 2010s’ 37.5 percent sits 0.34 standard deviations below that mean — an unremarkable draw from a famously volatile series.

What is genuinely anomalous is the 1990s. Nineteen of that decade’s 27 nonviolent successes — 70 percent — began in the four years 1990–1993, which succeeded at 67.9 percent. The roster is not mainly Prague and Vilnius; only about a quarter were ex-communist states. It is Zambia, Mali, Niger, Madagascar, Malawi, Kenya, Cameroon, Nigeria — single-party regimes that had just lost their superpower patron and their strategic rent. The tactic didn’t get better in 1990; the targets got weaker, all at once, for reasons having nothing to do with protest. So the sentence to retire is not “nonviolent resistance is failing,” it is “the 1990s were the normal case.” Two things I am not claiming: the drop in average peak participation, 2.7 percent of population in the 1990s to 1.3 percent since 2010, is a real finding computed on resolved campaigns and untouched by this argument; and some of those 27 ongoing campaigns were genuine defeats — Hong Kong’s Umbrella Movement was crushed by the 2020 National Security Law. Alexei Anisin has separately argued the coding rules tilt the comparison in other ways. Here is the falsifiable bet: when NAVCO is next extended past 2019 — the successor Major Episodes of Contention dataset, 2,734 episodes, was released in 2026 — the 2010s nonviolent success rate will revise upward by 8 to 14 points as Bolivia, Honduras, Chile and Sudan resolve. The general form is worth keeping: any panel that codes unresolved cases as negative outcomes will manufacture a decline in its final bin, mechanically, every time. The test costs one download and twenty lines of code, and here it moved p from 0.0035 to 0.20. Anyone citing “34 percent” as evidence about the present should check whether their number is measuring the world or measuring the calendar.