Decisions Are the Pinch Point

I saw a billboard the other day promoting vegetables “amid cost crisis.” It carried a familiar implication: people eat badly because healthy food is expensive.

Rice, beans, frozen vegetables, eggs and bananas are among the cheapest foods per calorie and per nutrient. The real objection is usually that they do not make a satisfying meal. Once taste, texture, convenience and immediate satisfaction enter the comparison, the unhealthy option wins because people prefer it. The mechanism connecting cost to the meal is a decision.

The shape of the problem

Pick an explanation for why low-SES populations have worse health outcomes: poverty, education, health literacy, food deserts, intergenerational trauma. “Social determinants of health” often bundles these together without specifying how any of them reaches disease.

Some do so directly. Lead in the water, polluted air, a job that wrecks your back and the chronic stress of poverty itself do not need a behavioural mediator. The large modifiable killers driven by diet, activity and substance use have a different shape. Upstream conditions influence the decisions that produce the outcome.

It looks like an hourglass on its side:

The causes sit on the left, the outcomes on the right, and decisions form the narrow middle. The bold lines trace poverty and food environment through food choices to obesity and chronic disease. The faded paths were available but did not activate because the decision went another way.

Poverty matters because it influences decisions. In this model it is an upstream cause, while the decision is the mediator through which most of its effect arrives.

Naming the lever

If healthy food is too expensive, subsidies and price controls are the obvious interventions. They help to the extent that they change what people choose, and the evidence on the size of that change is mixed because cost was never the whole barrier.

Once the question becomes why people make bad food decisions, more useful possibilities appear. They may lack information, work within a bad frame, or face defaults and incentives that make the worse option easy. Each possibility suggests something concrete to change.

Using “social determinants” as the final explanation stops one step early. It points at systemic factors as if they directly produced disease, which removes blame but also hides the lever. If nobody makes a decision, there is nowhere to intervene.

The same test works personally. Reading about nutrition, understanding the biochemistry and following the right accounts amount to theatre unless they change what you do. The relevant question is whether your decisions are different.

Decisions are social too

Where you grew up, what your parents taught you and what the corner shop sells are real inputs outside your control. The pinch point also contains other people’s choices: a child’s nutrition depends on their parents, your diet partly depends on what your partner cooks, and your exercise habits track whether your friends are active or sedentary.

People are unhealthy because of decisions, though not exclusively and never without context. Mechanistically, this is where much of the modifiable harm happens. Better information, framing and defaults can make good choices easier and bad choices harder, which gives the conversation somewhere to go.

Appendix: the causal structure

If you’ve done epidemiology or causal inference, you already see it: the hourglass diagram is a structural claim. Here it is formally.

The wrong DAG

The “social determinants” framing implicitly draws this causal graph:

Poverty causes obesity, and the model stops. It specifies no mechanism or point of intervention and gives individual agency no place. Most public health literature treats the relationship this way, with one arrow from social circumstance to disease.

The DAG with a mediator

The correct graph has a mediator:

The dashed line is the direct effect of poverty on obesity, bypassing decisions. My claim is that this arrow is small beside the mediated path. Poverty influences the decisions that give you diabetes rather than giving you diabetes directly.

In causal-inference terms, decisions are the mediator carrying most of the load. The total effect of poverty on health is mostly indirect, with a smaller direct effect left over.

Why intervention follows the mediator

Judea Pearl’s do-operator makes this concrete. do(good decisions), an intervention on the decision node, blocks the route from poverty through decisions to disease. Poverty and any direct effect remain, but its main route to disease is cut.

Reducing poverty, improving education and fixing food deserts all act on the left side of the graph. They work only when they alter decisions, so the mediator is the direct intervention target.

Mediators, colliders and conditioning

Two distinct problems often get collapsed here. Decisions mediate the poverty-to-disease path, so a study that “controls for behaviour” while trying to isolate poverty blocks the route that carries the effect. This is over-adjustment: subtracting the mechanism makes poverty look harmless by hiding how it acts.

Decisions also have several upstream causes, including poverty, education, genetics and culture, which makes them a collider with respect to those causes. Conditioning on a collider by studying only people who made the same decision can manufacture associations between the causes, making poverty and education look related when they are not. Controlling for behaviour understates poverty’s effect because behaviour is a mediator; collider bias is a separate trap.

The useful approach is to understand how upstream factors flow through decisions and intervene at that point.

The formal claim

The direct causal effect of socioeconomic variables, bypassing behavioural decisions, is the smaller part. Most of the observed association between poverty and disease is mediated by a person’s decisions and those of the people around them.

The exact fraction is debated, but the direction is not: behaviour mediates a large share of the SES-health gradient. The data are not controversial here; the controversy is whether the mechanism may be framed this plainly.