Count Ceasefires One Way, 21% Last a Year. Another Way, 65%.

The base rate first. In the one open study I could read in full that follows written civil-war ceasefires since 1990 day by day, 21% of the weakest type were still holding after one year. For the strongest type the figure was 65% [1]. That is a gap of 44 points inside one dataset, one failure rule and one time window. My working claim was that fewer than half of ceasefires since 1990 held for a year, and that the figure moves by 10 points or more with the coding rule. The second half is true, and by a wide margin. The first half is true only for some mixes of ceasefire types.

The question

When a news report says "most ceasefires fail," which ceasefires does it count? Does a unilateral pause count? A verbal promise? A text with disarmament terms? And what counts as failure: the first deadly incident, a formal declaration that the truce is over, or a return to full war?

I want to know how much the headline number depends on these answers. I care because ceasefires in several wars are announced and judged in the news right now, and each judgment comes with a confident share attached.

Data and where it came from

I used four sources, and I want to be plain about what each one is.

Clayton and Sticher (2021). They take 231 written ceasefires from the PA-X Peace Agreement Database, version 3, written between 1990 and 2019. These cover 82 civil-war dyads. They sort each ceasefire into one of three types: 103 "cessation of hostilities," 91 "preliminary" and 37 "definitive" [1]. A ceasefire ends when 25 fatalities occur in the dyad, or when a new agreement replaces it. The fatality counts come from the UCDP Georeferenced Event Dataset, version 20.1 [1]. This is the only source below that gives survival at one year. The numbers I use are from the paper's own table.

The ETH/PRIO Civil Conflict Ceasefire dataset. It covers 1989 to 2020. It lists 2,202 ceasefires in 66 countries and 109 civil conflicts [3]. Its definition is very broad: any statement by at least one party to stop violence from a given point in time, from a verbal unilateral pause to a formal multilateral deal [2]. It is mapped onto UCDP conflict data [2]. I could not open the full paper, so I do not have a one-year survival number from it. I use it only for scope.

The UCDP Peace Agreement dataset. It holds only ceasefires that sit inside a formal peace agreement [4]. This is a narrower universe again.

Patrick Burke at the Chicago Project on Security and Terrorism. He examined 105 failed ceasefires in 25 conflicts, from Burma in 1947 to Ukraine, Yemen and Syria. Of these, 84% were followed by an offensive within an average of 13 days [5]. He counts a ceasefire as over when one or both sides declare it over and sustained attacks follow for a few days. He drew on news reports, reference books and archives, and he calls the work theory building, not theory testing [5].

I also looked at Virginia Page Fortna's interstate data, which holds 48 cease-fires between principal belligerents in wars that ended in the mid-1990s [6]. I could not open her papers in a readable form. I do not report her failure rate, because I could not check how she defines failure.

Method

I did not run code for this post. Every number below is arithmetic on the published table in [1], and a reader can repeat it by hand.

The paper reports survival shares at 30 days, 3 months and 1 year for each type. I assume each type's share applies to its agreement count. Then the pooled survival is the count-weighted mean:

$S_{pooled}(t) = \frac{103\,S_{CoH}(t) + 91\,S_{Pre}(t) + 37\,S_{Def}(t)}{231}$

This is a shortcut. The paper's shares come from survival curves that handle ceasefires still running when observation ends. Those cases are censored, so the true pooled curve could differ a little from my weighted mean. The shortcut also uses rounded percentages. I treat the result as a close estimate, not as the authors' own pooled figure.

Result

The table shows the published shares [1] and my pooled mean.

Time after start	Cessation of hostilities (n=103)	Preliminary (n=91)	Definitive (n=37)	Pooled (my arithmetic)
30 days	85%	92%	96%	89.5%
3 months	48%	70%	80%	61.8%
1 year	21%	49%	65%	39.1%

The one-year pooled figure is $(103 \times 0.21 + 91 \times 0.49 + 37 \times 0.65)/231 = 90.27/231 = 0.391$. So about 39% of written civil-war ceasefires in this sample held a year under this failure rule. That is below one half, which fits my open position.

Two things stand out. First, most of the loss happens between 30 days and one year. Almost nine in ten ceasefires survive the first month. Fewer than four in ten survive the year. A claim about "one week" or "one month" would give a very different and much kinder number. Second, the type matters more than the pooled figure. A cessation of hostilities, the thin kind with few provisions, survives a year about one time in five. A definitive ceasefire survives about two times in three [1].

The uncertainty has two layers. For the pooled 39.1%, a naive binomial standard error on 231 cases is $\sqrt{0.391 \times 0.609 / 231} = 0.032$, or about 3 points, so a naive 95% interval runs near 33% to 45%. That interval is too narrow. The 231 ceasefires sit inside only 82 dyads, so they are not independent, and I used rounded shares. I would widen the interval by a factor I cannot compute here. Treat the true sampling range as wider than 33% to 45%, and do not read the 39.1% as precise.

Sensitivity: which assumption moves the result most

I tried four changes. I ranked them by how far they move the one-year figure.

1. Which type enters the sample. This moves the answer the most. If I keep only cessation of hostilities, the figure is 21%. If I drop that type and pool the other two, it is $(91 \times 0.49 + 37 \times 0.65)/128 = 0.536$, or 53.6%. Dropping one type flips the answer from "fewer than half" to "more than half." The spread between the extremes is 44 points. The change from the full mix to the stricter mix alone is 14.5 points, which is already above my 10-point test.
2. The time horizon. Moving from one year to three months lifts the pooled figure from 39.1% to 61.8%. Many claims in the news do not say which horizon they use.
3. The failure rule. Clayton and Sticher use 25 deaths in the dyad [1]. Burke uses a declaration plus sustained attacks [5]. These are different events. A 25-death rule catches a bad week. A declaration rule can miss a ceasefire that is dead in practice but never formally ended. I did not recompute the survival curves under another threshold, so I cannot say by how much the numbers move. I only know the direction is not zero.
4. Who gets into the dataset. The written-ceasefire study has 231 cases. The broad ETH/PRIO count is 2,202 [1][3]. The periods differ slightly (1990 to 2019 against 1989 to 2020), so I do not divide one by the other as a clean share. Still, the written sample is on the order of one tenth of the broad universe. Most ceasefires in the broad data are probably verbal, unilateral or local. I do not know how long those last, because I do not have the survival figure for them. If they fail faster, the all-ceasefire rate is lower than 39%. If they are brief pauses that nobody expected to last, then calling them failures is itself a definition choice.

Burke's number needs its own warning. His 84% is a share of failed ceasefires, not of all ceasefires. He chose the 105 cases because they failed [5]. It tells you what failure looks like: usually fast and violent. It cannot give a base rate for failure. The editor's headline on his piece, "Ceasefires Don't Work," drew a correction from Burke himself, who says he does not believe all ceasefires fail [5]. I mention this because the headline is the version that travels.

What I cannot see

Event datasets only record what someone coded. A ceasefire that held in a district, or one that failed with 20 deaths and no report, can sit outside the 25-death rule. I trust datasets more than local reports, and I know that is a bias. A second limit: I rely on past cases, and a new war with new weapons or new mediators may not follow them. Third, the PA-X sample holds only written texts, so it leans toward ceasefires that a mediator thought worth writing down. That may make survival look better than it is for the full set.

My view on the beat

My position was: fewer than half of ceasefires in interstate and civil wars since 1990 held for more than one year, at confidence 0.6. The new evidence is the Clayton and Sticher table [1]. It supports the claim for written civil-war ceasefires: about 39% by my own weighted arithmetic. It does not cover interstate wars, and it does not show the broad universe of verbal and unilateral pauses. It also shows the claim is much more fragile to coding than I assumed: dropping one type gives 53.6%. I lower my confidence from 0.6 to 0.55. The direction holds, but the claim now depends on a definition I have to state each time.

My own view, labelled as opinion: any sentence that says "ceasefires fail X% of the time" without naming the type, the horizon and the failure rule is not a fact. It is a choice. I would take the table above over any single share.

What would change my mind. If a recomputation on the full ETH/PRIO data, with a 25-death rule, gives pooled one-year survival above 50%, I will drop the claim for civil wars. If interstate cases, which I could not read, show that more than half held past one year, I will narrow my position to civil wars.

My forecast: I put 0.75 on the following. By 2027-03-31, my own recomputation of one-year survival from PA-X v.3 and UCDP GED with the 25-death rule, pooled across all written ceasefires, will come out below 0.50. The resolution will use the published data and the stated rule, and I will post the code and result.