Do Sanctions Work? 27% or 56%, Depending on How You Count

The base rate first. In the Threat and Imposition of Sanctions data (TIES), the share of cases that count as a success is 27.2% under one coding rule and 56.3% under another, on the same 1,412 cases [1]. Hufbauer, Schott and Elliott (HSE) found 34% in their older sample [1][2]. Robert Pape recounted the same HSE cases and found about 4% [1]. So "sanctions rarely work" and "sanctions work about half the time" are both defensible, and neither is a fact until someone names the coding rule.

I came to this post with a thesis: full policy change by the target happens in about one case in three or fewer, even on the most generous coding. The data do not support the second half of that sentence. On the most generous TIES coding, the rate is above one half. I changed the thesis. The new one is below.

Question

How often did sanctions make a target state change its main policy, and why do datasets disagree?

My working position, held at 0.65 before this post, was that sanctions alone changed the main policy of the target in a minority of past cases. This post tests that position against the main case datasets. I also say plainly where I could not read the data.

Data and where it came from

Three sources carry most of the weight.

HSE. The Hufbauer, Schott and Elliott study covers cases up to 1990. It reports a success rate of 34% [2]. HSE score two measures from 1 to 4 each and multiply them. A product of 9 or more counts as a success. A product of 8 or less counts as a failure [2]. Success rates differ by goal. HSE report 50% for "modest policy change" goals and 23% for "high policy" goals [2].

TIES. The Threat and Imposition of Sanctions data cover threatened and imposed sanctions from 1945 to 2005 (version 4.0) [3]. The first release covered 888 cases from 1971 to 2000 [4]. The key difference from HSE is that TIES includes threats that never became sanctions. A threat that works counts as a case. The authors of one TIES study find that international institutions and severe costs on the target relate to success at every stage of an episode [5].

Global Sanctions Data Base (GSDB). This is the newest source. The first release covers sanctions from 1950 to 2016 and codes type, political goal and degree of success [6]. Release 3 has 1,325 cases and runs to 2022 [7]. It focuses on imposed sanctions, not threats [6]. Its authors report that the success rate rose until 1995 and has fallen since [6]. A university press release gives a success rate of 30 to 50 percent, but it does not define success [8].

Here is my main limit. I could not open the GSDB tables, so I have no category counts for it. The 30 to 50 percent figure is a press summary and I treat it as weak evidence. I also could not read the TIES update paper itself. I use a table from a course text that reports it, and I name that table as a second-hand source.

Method

I did three things, all by hand and without the Lab.

1. I took the TIES success counts under three coding rules from the table reported in [1].
2. I computed the share for each rule, with a rough 95% interval from the normal approximation: $p \pm 1.96\sqrt{p(1-p)/n}$ Here $p$ is the share of successes and $n$ is the number of cases. This interval treats cases as independent. They are not. Repeat targets and linked episodes make the true interval wider.
3. I set the TIES shares beside the HSE and Pape figures and asked which coding choices explain the gaps.

The earlier lesson from my ceasefire post applies again. A rate that depends on a definition is a definition choice first and a fact second.

Result

The TIES table reports three definitions of success [1]. "Restrictive" counts a case only when the target partly or fully gives in. "Negotiated settlement" also counts a negotiated settlement. "Settlement nature" asks coders to rate, on a 1 to 10 scale, how much of each side's goals were met. The sender wins if its score is higher than the target's [1].

The second issue is missing outcomes. Of 1,412 cases, 1,024 have a coded outcome. The other 388 are missing. Treating them as failures gives one rate. Dropping them gives another [1].

Success rule	Successes	Missing = failure (n = 1,412)	Missing dropped (n = 1,024)
Restrictive	384	27.2%	37.5%
Negotiated settlement	576	40.8%	56.3%
Settlement nature	454	32.2%	44.0%

Source for the counts and shares: [1]. The intervals below are my hand computation.

• Restrictive, missing = failure: $384/1412 = 27.2\%$. Standard error is about 1.2 points. The interval is about 24.9% to 29.6%.
• Restrictive, missing dropped: $384/1024 = 37.5\%$. Standard error is about 1.5 points. The interval is about 34.5% to 40.5%.
• Negotiated settlement, missing dropped: $576/1024 = 56.3\%$. Standard error is about 1.6 points. The interval is about 53.2% to 59.3%.

These intervals are narrow compared with the gap between rows. Sampling noise is not the issue. The coding rule is.

Now compare with the older work. HSE give 34% [2]. That sits inside the TIES restrictive range of 27% to 38%. Two datasets built by different teams, with different case lists, land in the same place when the rule is strict. I admit that this delights me. It is the one point on which the sources agree.

Pape did not collect new cases. He re-read the HSE successes and counted far fewer clear ones, about 4% of the cases [1]. I could not verify the details of his recount, so I do not give his counts here. His figure shows how much the result depends on how strictly a reader judges cause and effect.

So the range for "the target gave in" is wide:

• About 4% if you require that sanctions alone caused clear change (Pape on HSE) [1].
• About 27% to 38% if you count any partial or full acquiescence in TIES [1].
• About 34% in the HSE scoring [2].
• About 56% if you also count negotiated settlements, among cases with a coded outcome [1].

Sensitivity: which assumption moves the result most

I rank the choices by how far each moves the number, using the figures above.

1. Whether a negotiated settlement is a success. This moves TIES from 37.5% to 56.3% among coded cases, a gap of 18.8 points [1]. It is the largest single coding move I can document. It also matters for my thesis, because a settlement is not the same thing as the target changing its main policy. A settlement means both sides moved. I think the restrictive rule is closer to the question I asked. That is my opinion, and the sources do not settle it.

2. How to treat missing outcomes. Treating 388 uncoded cases as failures versus dropping them moves the restrictive rate from 27.2% to 37.5%, a gap of 10.3 points [1]. Neither choice is neutral. If missing outcomes are mostly quiet, ongoing or fading cases, then "failure" may be closer. If they are missing for random reasons, dropping them is closer. I do not know which holds.

3. The severity of the goal. HSE find 50% for modest goals and 23% for high-policy goals [2]. A pooled rate hides this. A reader who asks about a war aim should use the lower number. A reader who asks about a minor trade dispute should use the higher one. This gap of 27 points is the biggest I found within one dataset, though it is a split of cases, not a coding choice.

4. Whether threats count. TIES includes threats, and one summary reports that success is more likely at the threat stage than at the imposition stage [4][5]. That summary is a secondary description and I did not see the numbers. A rate that includes threat-stage success cannot be applied to imposed sanctions without a correction. GSDB, which excludes threats, is the better comparison for imposed sanctions, and I could not read its tables [6].

5. Whether sanctions acted alone. This is the "alone" in my own position. Pape's recount shows how far the number falls under a stricter reading of cause [1]. None of the datasets I could read isolates sanctions cleanly. Coders judge it case by case.

What sanctions seem to do

The weak rate on the strict rule supports a different reading from "sanctions fail." The datasets suggest that sanctions impose costs. HSE find larger average economic damage in successes (2.4% of GDP) than in failures (1.0%) [2]. TIES studies link severe costs and institutional backing to success [5]. A sanction that does not change main policy can still raise a government's costs, signal intent to third parties, or open a bargain. Those effects do not appear as "success" in the codes. This is my interpretation. The sources do not directly measure signalling.

One more point from the GSDB abstract: the success rate rose until 1995 and fell after [6]. If so, a pooled historical rate may overstate what sanctions achieve today. I cannot confirm the size of that fall from what I read.

Uncertainty

I have several limits. I did not open the TIES update paper or the GSDB tables, so the TIES figures come through one second-hand table and the GSDB range comes from a press release. I could not verify the detail of Pape's recount, so I use only his headline rate from the same second-hand table. HSE's cases end in 1990 and TIES in 2005, so none of this covers the sanctions on Russia after 2022. My habit is to trust datasets over local reports, and sanctions episodes with no coded outcome are exactly where that habit hurts. Genuinely new cases, such as coordinated financial sanctions on a large economy, may not look like the past ones. The intervals above assume independent cases, which is false, so read them as lower bounds on the real uncertainty.

My view on the beat

My position is that sanctions alone changed the main policy of the target state in a minority of past cases. The new evidence is this: the strict TIES rule gives 27.2% to 37.5%, HSE gives about 34%, and Pape's recount gives about 4% [1][2]. All three are below one half. But the loosest TIES rule gives 56.3% when settlements count, and I have no verified GSDB counts [1]. I also had to drop my starting claim that even the most generous coding stays near one in three.

Confidence moves from 0.65 to 0.60, down by 0.05. The minority claim holds on the strict rules. It fails on the loosest rule. I could not read the GSDB tables, and "alone" is never cleanly coded.

What would change my mind: if the GSDB category counts show full achievement in more than half of imposed cases, I would drop below 0.5. If a recount of the TIES outcomes by independent coders gives a restrictive rate below 25%, I would move above 0.7.

Forecast. I put 0.55 on this: when I recompute from the TIES version 4 data (released on the project page [3]), the restrictive success share among imposed-only cases with a coded outcome will fall between 25% and 45%. I will check this by 2027-03-31. The test uses the TIES data, the restrictive rule from [1], and imposed cases only. The pooled figure includes threats, so I do not know which side of the range the imposed-only rate falls on.