Case Study 2 — The Soda Tax: What a Policy Experiment Can and Can't Tell You

A policy case. ⚠️ The taxes described here are real — Mexico (2014), the UK Soft Drinks Industry Levy (2018), and several US cities. I describe their design and the general direction of findings rather than quoting precise effect sizes, because the estimates vary substantially by study, method and follow-up period, and inventing a number would break this book's rule (Chapter 2 §2.10).

Setup

This is the best population-level natural experiment nutrition has.

Chapter 2 spent a chapter explaining why nutrition evidence is weak: you can't randomize people to diets for thirty years, you can't blind them, and self-reported intake is unreliable.

⚠️ A tax sidesteps nearly all of that. It changes one variable, at a known date, for an entire population, with an unexposed population next door. Purchases are measured by scanner data rather than by asking people what they ate.

So this case study is partly about soda and mostly about what a natural experiment can carry.

The three designs

They are not the same intervention, and the difference between them is the most interesting thing here.

💡 The UK design is the clever one, and it's worth understanding why.

By tiering the levy on sugar content and announcing it two years early, ⚠️ the UK made it cheaper for manufacturers to reformulate than to pay. The target wasn't the shopper. It was the recipe.

And a large fraction of the UK soft drinks market was reformulated below the levy threshold before the levy was even collected — which means a substantial part of the effect happened to people who never changed their behaviour, never saw a price rise, and in many cases never knew a policy existed.

⚠️ That is a completely different mechanism from a consumption tax, and it produces a different distribution of who benefits.

What the evidence supports

Purchasing: strong

Across jurisdictions, taxed beverage purchases fell. This is measured with scanner data rather than recall, has been replicated in multiple settings, and is about as solid as this kind of finding gets.

Two important qualifiers, both of which cut against enthusiasm:

1. Some of the reduction is substitution rather than elimination. People buy untaxed drinks — including diet versions, juice, and in some settings bottled water. ⚠️ §18.13's comparator question, arriving as a policy outcome: a switch to diet soda is probably an improvement, a switch to juice is probably not much of one, and a switch to water is a clear one. Which of these dominates varies by setting.

2. Cross-border shopping is real in small jurisdictions. ⚠️ A city tax measured only within the city will overstate the effect — some of the "reduction" walked across a boundary. Studies that included surrounding areas find smaller effects than those that didn't. This is a measurement artefact with a direction, which is the most useful kind to know about.

Reformulation: strong, in the UK

⚠️ The sugar content of the UK soft drinks market fell substantially. This is a direct measurement of products rather than an inference about behaviour.

And it is the finding that generalizes least, because it depends entirely on the tiered design. A flat tax gives a manufacturer no reason to reformulate at all.

Health outcomes: 🟡 not yet established

⚠️ This is where honesty is required and where most reporting on soda taxes fails.

Purchasing changes are measurable within months. Weight, dental caries, and type 2 diabetes incidence change over years to decades, and by the time they do, ten thousand other things have also changed.

Some studies have reported associated changes in outcomes like dental extractions in children. ⚠️ These are encouraging, they are observational, and the follow-up periods are short relative to the diseases in question.

🔬 Claim → Evidence → Verdict

The claim: "Soda taxes work."

The verdict depends entirely on what "work" means, and this is the whole case study:

"Work" =	Verdict
Reduce purchases of taxed drinks	✅ Well supported
Cause reformulation (tiered designs)	✅ Well supported
Reduce total sugar intake	🟢 Probably true — ⚠️ net of substitution, smaller than the purchasing effect
Improve health outcomes	🟡 Unclear — not yet established. Plausible, mechanistically supported, and not yet demonstrated at the outcome level
Reduce health inequality	🟡 Contested — see below

📉 Evidence quality: Strong quasi-experimental evidence on the proximal outcomes; the distal outcomes are not yet measurable with confidence.

⚠️ A claim that soda taxes "reduce obesity" is running ahead of the evidence in the same direction the enthusiasts want it to run — which is exactly the failure mode Chapter 2 §2.6 describes, and it does not become acceptable because you like the policy.

The regressivity argument

The strongest objection, and it deserves a fair hearing because it is usually dismissed by people who agree with the tax.

The objection: ⚠️ SSB consumption is higher in lower-income groups, so a consumption tax takes a larger share of income from people who have less. That's a regressive tax, and it is a real cost paid by real households — the Ortiz-Lindqvists, on $180 a week.

The counter-arguments, and they are not weak:

1. ⚠️ The health burden is also concentrated in the same groups — obesity, T2D and dental disease. If the tax works, the benefit is progressive even though the payment is regressive.

2. Price responsiveness is higher in lower-income groups, so they reduce consumption more — meaning they pay less tax than a static calculation implies and get more of the health benefit.

3. Earmarking changes the arithmetic. Several jurisdictions direct revenue to school meals, water fountains, or health programmes in affected communities.

4. ⚠️ And the UK design largely sidesteps the objection, because reformulation delivers the benefit without anyone paying more for anything.

The honest position: ⚠️ the payment is regressive, the benefit may be progressive, and whether the net effect is progressive depends on an empirical question — how much health improvement actually occurs — that §18's evidence cannot yet answer.

Anyone who tells you this is settled, in either direction, is telling you about their politics.

What this case demonstrates about evidence

Four things, and they matter beyond soda.

1. ⚠️ Natural experiments are strong where randomization is impossible — and they measure what they measure. Scanner data on purchasing is excellent. Scanner data on diabetes incidence does not exist.

2. The proximal-distal gap is where over-claiming happens. ⚠️ Purchases → intake → weight → disease is four links, and the evidence gets thinner at every one. Most reporting quotes link one and concludes at link four.

3. Design determines mechanism. ⚠️ A tiered pre-announced levy and a flat consumption tax are different interventions with the same name, and lumping their results together — which meta-analyses sometimes do — obscures the most useful finding available.

4. And it's a rare case where you can see the counterfactual. ⚠️ Untaxed jurisdictions next door, untaxed product categories in the same shop. That's the thing nutrition epidemiology almost never has, and it's why this case is worth more than its subject matter.

Analysis

1. "Soda taxes work" has five different answers depending on the outcome, ranging from ✅ to 🟡. ⚠️ §18.1's four-claims problem, at policy scale.

2. The UK levy targeted manufacturers, not shoppers — and most of its effect arrived through reformulation, before collection began. A completely different mechanism from a consumption tax.

3. Substitution and cross-border shopping both shrink the true effect, and studies that don't account for them overstate it. ⚠️ Both artefacts have a known direction, which is unusually helpful.

4. Health outcomes are not established, and the gap between "purchases fell" and "people got healthier" is four inferential links wide.

5. The regressivity objection is real and its resolution depends on an empirical question that hasn't been answered.

6. ⚠️ And this is the best evidence nutrition has for a population-level dietary intervention. Sit with that. It's a quasi-experiment with strong proximal outcomes, unresolved distal outcomes, and a contested distributional analysis. If this is the top of the evidence hierarchy for food policy, calibrate accordingly — in both directions.

Discussion Questions

1. ⚠️ The UK levy achieved much of its effect through reformulation, before it was collected. Is that a better or worse outcome than reducing consumption? Who benefits under each, and who is left out?

2. "Soda taxes work" splits into five claims with different verdicts. Write the sentence a health minister should say that is both accurate and useful in a press conference. ⚠️ Is such a sentence possible?

3. The regressivity objection: payment regressive, benefit possibly progressive. What evidence would settle it? How long would you have to wait, and what would confound it in the meantime?

4. ⚠️ A tax reduced purchases of taxed drinks and increased purchases of fruit juice. Per §18.3, has anything improved? How should a policy be designed to prevent this, and what does your design do to the politics?

5. Cross-border shopping means city-level studies overstate effects. Does that make city taxes bad policy, or bad experiments, or both?

6. ⚠️ This is described as the best population-level evidence nutrition has. Does that raise your confidence in soda taxes, or lower your confidence in nutrition policy generally? Defend your answer against the opposite one.

Your Turn

1. Find out whether you live under one. Search for a sugar tax or levy in your country, state or city. ⚠️ If you do, find out its structure — flat or tiered? On the manufacturer or the consumer? Is the revenue earmarked? Most people who live under one cannot answer these.

2. Do the reformulation check. If you're in a jurisdiction with a tiered levy, ⚠️ look at the sugar content on a mainstream soft drink and compare it to what you remember. In several markets the number changed and nobody announced it.

3. Find one news story about a soda tax and grade it against the five-outcome table. Which outcome does the headline claim? Which one does the study actually measure? ⚠️ How many of the four inferential links does the article skip?

4. And the harder one — the one that tests whether §18's method is doing any work on you:

⚠️ Write down the strongest argument against the policy position you currently hold on soda taxes.

If you favour them, write the best version of the regressivity and paternalism objections. If you oppose them, write the best version of the concentrated-health-burden and reformulation arguments.

Then answer §18.15's row: what evidence would change your mind?

If you can't fill that row, you're holding a position the same way the claims in Chapter 17 were held — and being right about the conclusion wouldn't change that.