More Conflicts, More Civilized?

Introduction

Political violence is commonly tracked through two headline figures: the count of recorded events and the count of fatalities those events produce. Intuitively, the two should share similar trends of dynamics — more conflict, more deaths. However, based on the Armed Conflict Location & Event Data (ACLED) project's weekly regional files from 2014 to 2025, we found that trends between the number of conflicts and fatalities are not statistically identical. The number of recorded events globally rose roughly sixteenfold (from 25k to 409k per year), while fatalities grew only fivefold (from 47k to 243k). Average per-event lethality collapsed from 1.9 to 0.6. Why is that the case? Does that mean, sarcastically, while we are observing a resurgence of global conflicts, conflicts are turning to a more civilized version, resulting in less lethality?

Is that true? To discover the real attribution to this discrepancy, we ask three questions in turn:

1. Is the divergence real and statistically meaningful?
2. Is it explained by a shift toward low-lethality event categories, or by within-category changes?
3. Is the pattern global or driven by particular regions?

A supplementary section extends the analysis to the full Africa series (1997–2025) where coverage is deepest.

Data & Methods

We use six ACLED regional weekly aggregation files (Africa, Asia-Pacific, Europe & Central Asia, Latin America & the Caribbean, Middle East, US & Canada), dated March 2026. Regional coverage is uneven with regard to the time frame. Taking the most parsimonious converging manner, we restricted the cross-region analyses to 2014–2025. A separate Africa-only frame, loaded before the 2014 filter, is used for the long-run supplement.

For the first question, we aggregate to annual global totals and compare the two series via Pearson and Spearman correlation, year-over-year growth, and a pooled log-linear regression with a series-by-year interaction that formally tests whether the two trend slopes differ. For the second question, we decompose the change in aggregate lethality into a composition effect, a within-type effect, and an interaction. This is the standard Oaxaca-style shift-share decomposition; because aggregate lethality is exactly a share-weighted average of type lethalities, the decomposition is an identity, not an approximation. For the third question, we repeat both analyses inside each region. All analysis is in Python, with tokens kindly offered by the Anthropic (the antonym of generosity).

Finding 1 — The global discrepancy is real

Across 2014–2025, annual events and fatalities remain strongly positively correlated (Pearson r = 0.856, p < 0.001; Spearman ρ = 0.811), but grow at sharply different rates. A log-linear fit implies events growing at a rate of 27.5 % per year, while the fatalities grow at 12.3 %. In a pooled regression with a series-year interaction, the slope difference is −0.127 log-units (t = −3.06, p = 0.006, thereby standing as a rejection of the null that events and fatalities follow a common trend). The mechanical consequence is that per-event lethality has fallen from 1.91 fatalities per event in 2014 to 0.59 in 2025, a 69 % decline. The sharpest inflection is between 2017 and 2018, when lethality roughly halves from 1.62 to 0.84.

Two readings are on the table：
Substantively speaking, the world records more conflict events, but those events are on average smaller.
Methodologically speaking, ACLED's coverage has expanded, and the added events are disproportionately low-lethality (protests, small clashes) that would previously have gone unrecorded. Both would produce the pattern observed. Finding 2 examines whether this “collapse” is concentrated in the composition of events or within event types.

Finding 2 — Within-type lethality change does most of the work

The shift-share decomposition of lethality from 2014 to 2025 attributes 82% to within-type lethality changes, 27% to composition, and −10% to interaction (shares sum to 100). Both channels operate, but the within-type channel dominates. Composition did shift as expected — the share of events that are Protests rose from 29% to 38%, Strategic developments from 5% to 10%, while Violence-against-civilians fell from 20% to 10% and Battles from 24% to 16%.

Yet the bigger change is that per-event lethality fell inside every single category:

1. Battles from 4.23 to 2.16 (−49%);
2. Explosions/Remote violence from 2.58 to 0.58 (−78%);
3. Violence-against-civilians from 2.89 to 1.14 (−61%).

Notably, Explosions/Remote violence actually grew as a share of events (+11%) — compositionally, that would raise lethality — but its own per-event deadliness collapsed so sharply that the net effect is still downward.

Robustness with the coarser disorder type classification gives a stronger within-type share (92%). In other words, it implies that the character of recorded combat has changed — more frequent but smaller engagements, better civilian-protection norms, or more successful remote-violence interception. A measurement reading would say that ACLED is increasingly capturing the long tail of low-casualty incidents within each type. Without an independent ground-truth series, both stories remain on the table and must be treated as joint explanations. We can certainly find other databases for a cross-check or complement the types of casualty, but that will be beyond this study.

Finding 3 — The pattern is real but heterogeneous across regions

Repeating both analyses by region shows the global story is not a single mechanism. In Africa (2014–2025), events grew 13.6% per year against fatalities at 8.4% per year; the decomposition mirrors the global split (25% composition / 81% within). The Middle East shows the same shape, more sharply: events +50% per year, fatalities +26% per year, lethality rate is −1.34, with 78% composition and 62% within (they add to more than 100 because the interaction term is strongly negative).

Latin America tells a different story: trend slopes barely differ (+38% per year vs +35% per year), and what little lethality decline there is comes almost entirely from composition shift (~102%) — per-type lethality barely moved. This is consistent with a region where the type of unrest has shifted toward protests, not where combat has become less lethal.

Europe-Central-Asia is the exception that is developing in an “uncivilized” manner. Fatalities grew 183% per year, against events at 66% per year — per-event lethality rose, driven by the war in Ukraine from 2022. Any global story about “declining violence intensity” breaks here.

US-and-Canada enters only in 2019 and its ΔL is effectively zero — the series is dominated by protests at near-zero lethality and is not interpretable in the same frame. Asia-Pacific looks like a hybrid (66 % composition, 57 % within). The overall conclusion is that the global discrepancy is not a single mechanism. Africa and the Middle East drive the within-type lethality collapse narrative; Latin America is pure composition; Europe-CA is an active-war outlier.

Africa, 1997–2025

Africa's coverage extends back to the late-1990s, allowing a longer view. Over 1997–2025 events grew 12.0 %/yr while fatalities grew only 2.8 %/yr — a divergence much larger than the post-2014 window shows. Per-event lethality fell from 8.63 to 1.42 (ΔL = −7.21), driven 95 % by within-type change.

Splitting the series reveals two regimes. In 1997–2013, events grew slowly (+5.5 %/yr) while fatalities actually fell (−7.5 %/yr) — this is the period that contains the 1998–1999 Second Congo War spike, with lethality peaking at 33.9 fatalities per event in 1999. In 2014–2025, events grew rapidly (+13.6 %/yr) and fatalities grew modestly (+8.4 %/yr) — more recording, moderate lethality. The 2014 cut therefore averages across a structural break that is obvious in the long series.

The long-run view reinforces the cross-region finding: even net of the DRC/Rwanda mass-violence episode and pre-2011 thin coverage, most of the long-run lethality decline is a within-type change, not a mix shift. That said, coverage-expansion bias is mechanically stronger over a 29-year window than a 12-year one, so the real-world vs measurement ambiguity is also larger.

Discussion & Limitations

Three caveats bound these findings. First, coverage is not uniform across regions or time: Europe-CA, LatAm, and US-Canada enter the panel late, and pre-2014 coverage (used only for Africa) is thinner than post-2014. Analyses on the common 2014–2025 window mitigate but do not remove this. Second, ACLED's sourcing has expanded over the period — an endogenous increase in the event denominator mechanically reduces average per-event lethality without any change in underlying violence. Finding 2's within-type effect is consistent with either a real change in combat character or the within-type equivalent of coverage expansion, and the data cannot distinguish them. Third, the shift-share decomposition is descriptive, not causal: it partitions ΔL into algebraic components, not into structural mechanisms. Any claim that "violence has become less deadly" rather than "the recording of violence has broadened" would need an external benchmark this report does not construct.

References

• Raleigh, C., Linke, A., Hegre, H., & Karlsen, J. (2010). Introducing ACLED: An Armed Conflict Location and Event Dataset. Journal of Peace Research, 47(5).
• ACLED Codebook (2024 edition, acled-codebook.html).
• ACLED regional aggregation files, downloaded week-of 21/28 March 2026.

Mathematical formulation

A.1 Annual aggregation

Let $w$ index weeks and $t$ index calendar years. For each region $r$ (or for the global panel) we aggregate weekly cells to annual totals of events $E_t$ and fatalities $F_t$:

$$E_t \;=\; \sum_{w \in t} \text{EVENTS}_w, \qquad F_t \;=\; \sum_{w \in t} \text{FATALITIES}_w.$$

The aggregate per-event lethality and the year-over-year growth rate are

$$L_t \;=\; \frac{F_t}{E_t}, \qquad g^Y_t \;=\; \frac{Y_t}{Y_{t-1}} - 1 \quad\text{for } Y \in \{E, F\}.$$

A.2 Correlation and log-linear trend

Between annual series $\{E_t\}$ and $\{F_t\}$ we report Pearson's $r$ and Spearman's $\rho$. The log-linear trend for each series is

$$\log Y_t \;=\; \alpha_Y + \beta_Y \cdot t + \varepsilon_t, \qquad \hat{g}_Y \;=\; e^{\hat\beta_Y} - 1,$$

where $\hat{g}_Y$ is the annualised percentage growth implied by the fitted slope.

A.3 Pooled interaction test for slope equality

To formally test whether the events and fatalities trends share a common slope, we stack the two log-series and fit

$$\log Y_{t,s} \;=\; \beta_0 + \beta_1 \cdot t + \beta_2 \cdot \mathbb{1}[s = F] + \beta_3 \cdot \bigl(t \cdot \mathbb{1}[s = F]\bigr) + \varepsilon_{t,s},$$

where $s \in \{E, F\}$ identifies the series. The coefficient $\beta_3$ is exactly the difference between the fatalities and events log-slopes; the null hypothesis of interest is

$$H_0: \beta_3 = 0 \quad\text{vs.}\quad H_1: \beta_3 \neq 0,$$

tested by the usual OLS t-statistic $t = \hat\beta_3 / \widehat{\text{SE}}(\hat\beta_3)$.

A.4 Shift-share decomposition

Let $k$ index event categories (EVENT_TYPE or DISORDER_TYPE). Define the category share of events and the category per-event lethality as

$$s_{k,t} \;=\; \frac{E_{k,t}}{E_t}, \qquad l_{k,t} \;=\; \frac{F_{k,t}}{E_{k,t}}.$$

Because aggregate lethality is exactly a share-weighted average of category lethalities,

$$L_t \;=\; \frac{F_t}{E_t} \;=\; \sum_k \frac{E_{k,t}}{E_t} \cdot \frac{F_{k,t}}{E_{k,t}} \;=\; \sum_k s_{k,t} \, l_{k,t},$$

the change $\Delta L = L_1 - L_0$ between a base period $t=0$ and an end period $t=1$ admits the Oaxaca-style identity

$$\Delta L \;=\; \underbrace{\sum_k (s_{k,1} - s_{k,0})\, l_{k,0}}_{\text{composition}} \;+\; \underbrace{\sum_k s_{k,0}\, (l_{k,1} - l_{k,0})}_{\text{within-type}} \;+\; \underbrace{\sum_k (s_{k,1} - s_{k,0})(l_{k,1} - l_{k,0})}_{\text{interaction}}.$$

This is an algebraic identity, not a regression-based approximation: the three terms sum exactly to $\Delta L$. Reported percentages ("27 % composition / 82 % within-type") are each term divided by $\Delta L$.

A.5 Regional and long-run application

For RQ3 all of A.1–A.4 are applied to each region independently, using that region's earliest and latest years in the 2014–2025 window as $(t_0, t_1)$ for the decomposition. For the Africa supplement, $t_0$ and $t_1$ are the first and last years of the full 1997–2025 series, and sub-window decompositions use $t_0 = 1997, t_1 = 2013$ and $t_0 = 2014, t_1 = 2025$.