Criminogenic Entropy: A Structural Theory of the Crime Decline

Almost every attempt to explain the American crime decline names a cause: lead abatement, better policing, mass incarceration, abortion, smartphones, aging. A new pre-print does something different. It does not propose another cause. It proposes a claim about the shapeof causation — that crime tracks a slow, generational quantity the authors call “criminogenic entropy,” and that once you accept its shape, the search for a single lever starts to look like the wrong question.

We have written before about the decline itself — that violent crime is roughly 60% below its 1991 peak, that this represents on the order of 10,000 Americans not killed each year, and that almost no one is funded to study why. We have catalogued the competing theories. This post is not another entry on that list. It is about a framework that tries to reframe the list itself, laid out by Jeff Asher in his July 20, 2026 Jeff-alytics post “Does Entropy Explain Why Crime Is Dropping?” and drawn from a pre-print by Scott Mourtgos and Ian Adams. It is worth taking seriously — and worth stress-testing hard, which is what the rest of this piece does.

What the theory actually says

Mourtgos and Adams define criminogenic entropy as, in their words, “the number of feasible crime-producing configurations in a society.” Borrow the intuition from physics without leaning on the equations: a high-entropy system has many available disordered states, a low-entropy one has few. Applied to crime, the claim is that societies pass through long arcs in which the number of circumstances that can readily produce a crime rises and falls — and that measured crime rates follow that arc.

They operationalize it with 15 measurable factors. Some push entropy up (the paper cites substance abuse and inequality); some pull it down (homeownership, institutional trust). Aggregated, the authors report that this composite peaks in the 1980s and 1990s and declines thereafter — closely tracking both violent and property crime over the same span. Crucially, the model is fit on data ending in 2023, and its forecasts reportedly matched the 2024 outcomes that followed.

Two properties of the framework do the real work, and both are worth stating plainly.

What it explains that single-cause stories struggle with

The appeal of the framework is not that it fits one curve — plenty of theories fit one curve. It is that it accounts for a set of features of the current decline that have been genuinely awkward for cause-of-the-week explanations.

Breadth.The post-2022 decline is showing up nearly everywhere at once — across cities with aggressive prosecutors and cautious ones, red states and blue, large agencies and small. Asher notes the drop began around mid-2022 and, by his account, possibly even accelerated afterward, spanning administrations. A theory resting on one jurisdiction's policy choice cannot explain a decline that ignores jurisdiction. A slow-moving national structural quantity can.

Slowness and inertia. The decline is a long grind, not a step change. That is what a generational variable predicts and what a policy switch does not.

The 2020–21 spike.This is the framework's cleverest move. Rather than treat the pandemic-era homicide surge as a refutation of any downward trend, Mourtgos and Adams frame it as an “exogenous shock to an otherwise compressing system”— a violent perturbation that the system, because it has memory, absorbed before returning to its prior trajectory. Under that reading, the sharp 2023–2025 fall is not a surprise reversal; it is the system relaxing back toward the path it was already on.

And the data since are at least consistent with that story. The Real-Time Crime Index shows murder down roughly 18% through May 2026 across a sample of 585 agencies covering about 122 million people, with violent crime down 5.6% and property crime down 10.7% year to date. Consistency is not proof — but a theory that forecast 2024 correctly and has not been embarrassed by 2025 or early 2026 has cleared a bar most crime narratives never face.

The skeptic's section

A framework this accommodating deserves suspicion in proportion to its elegance, and Asher — to his credit — supplies most of the doubts himself. Take them in order.

The factor-selection problem.The composite is built from 15 chosen variables. As Asher puts it, a different 15 factors might tell a different story. When a theory's central quantity is an aggregate of hand-selected inputs, the aggregate can be tuned, consciously or not, to track the outcome it is meant to explain. The defense against this is out-of-sample forecasting, which the paper claims — but one matched forecast year (2024) is a start, not a verdict.

The falsifiability problem.A quantity that moves slowly, has memory, and absorbs contrary evidence as “exogenous shocks” is powerful precisely because it is hard to contradict. That same power is a liability: what pattern of future data would the authors accept as showing the theory is wrong? A structural model needs a pre-committed answer to that question, or it risks explaining everything and predicting nothing.

Correlation, still. Fifteen social factors and crime moving together over forty years is a strong correlation, not a demonstrated mechanism. The entropy language supplies a compelling metaphor for why they might co-move; it does not, by itself, establish that the factors cause the crime rather than sharing a common driver with it.

It is not the whole story, and the authors say so. Asher is careful to note that post-COVID community investments and violence-intervention programs likely contributed meaningfully to the magnitudeof the recent drop, even if the structural trend would have produced some decline regardless. His framing is the right one: entropy is, in his words, “an answer,” not necessarily “the answer.” Note too that this cuts against a common political read of the data — Asher adds that the evidence linking proactive policing to the declines is “not super strong.”

Why any of this matters if you build on crime data

The entropy debate can read like an academic parlor game, but it carries a concrete lesson for anyone who consumes crime data operationally — in a real estate model, a safety score, an underwriting memo, an alert. If the dominant force in the numbers is a slow, structural downtrend, several things follow.

• The base rate is drifting under everything.A national structural decline means your historical baseline is not stationary. A neighborhood that looks “improving” year over year may simply be riding the same tide as everywhere else. The interesting signal is the deviation from the national trend, not the raw local direction.
• Resist local over-attribution.If declines are showing up regardless of local policy, then crediting a given city's drop to a specific program — or blaming a specific official for a rise — is exactly the inference the breadth of the decline should make you cautious about. The same caution applies in reverse to your product's narratives.
• A structural downtrend is the right null hypothesis.When a local series ticks up, the question is not “what went wrong here” but “is this above what the national trend and ordinary noise would produce?” Most local spikes, tested that way, are seasonal or statistical — a point we made at length in our piece on crime seasonality.
• Forecasting has a floor and a ceiling.A generational variable changes slowly, which means naive “next year looks like this year, minus a little” forecasts will often be roughly right — and will fail precisely at the exogenous shocks the model warns are unpredictable. Build for both.

None of that depends on the entropy theory being correct. It depends only on taking seriously the possibility that the decline is broad and structural rather than local and causal — which the data, whatever the mechanism, increasingly support.

The useful version of the idea

Strip away the physics vocabulary and the durable insight is modest and probably right: crime rates are the surface expression of deep, slow-moving social conditions, they carry momentum, and they do not turn on a dime for any one intervention. That has been the quiet consensus of careful crime analysts for a while. The Mourtgos and Adams contribution is to give it a formal shape and a testable forecast — and, in doing so, to hand skeptics something concrete to test.

Whether “entropy” survives peer review or turns out to be an evocative label on an ordinary composite index, the framing does real work by pushing back on the reflex to explain a nationwide, decades-long, politically indifferent decline with whatever happened locally last quarter. For anyone building on this data, that reflex is the expensive one to keep. The tide is going out almost everywhere. The job is to measure who is swimming against it.