The China Study Criticism: Where the Book Outran Its Data

The China Study, T. Colin Campbell’s 2005 bestseller, drew sweeping conclusions about animal protein and disease from a large ecological survey conducted across rural China in the 1980s. The book has sold millions of copies and become a foundational text for advocates of plant-based diets. But among epidemiologists and nutrition scientists, the book’s core claims have attracted persistent and serious criticism, mostly centered on how far the data can actually take you and how much Campbell left on the cutting-room floor. An independent scientific review scored the book’s accuracy at just 1.9 out of 4, concluding its major claims are “not very well supported.”¹

What the Original Project Actually Measured

The underlying research, formally known as the China–Cornell–Oxford Project, was a collaboration between Cornell University, the University of Oxford, and the Chinese Academy of Preventive Medicine. In 1983, researchers selected two villages at random in each of 65 rural counties across China, then chose 50 families from each village, yielding about 6,500 participants total. One adult from each household provided blood, urine, and detailed dietary information, while food samples from the local markets were also collected and analyzed.² Those individual-level measurements were then compared against county-level death rates from more than 50 diseases, including seven types of cancer, gathered during the period 1973 to 1975.

The project was genuinely impressive for its era. Rural China at the time offered a rare natural experiment: diets varied enormously from county to county but were relatively uniform within each county, and the population was mostly stable, meaning people had eaten much the same way for decades. That made it a valuable snapshot. But a snapshot is what it remained. The study did not follow individuals over time to see who got sick and who did not. It compared county averages to county death rates, which is a fundamentally different kind of evidence from what most readers assume when they hear “the most comprehensive study of nutrition ever conducted.”

The Problem with County-Level Comparisons

The China–Cornell–Oxford Project was an ecological study, meaning it looked at groups (counties) rather than tracking individual people. This distinction matters enormously. When you find that counties with higher average meat consumption also had higher rates of a particular disease, you cannot conclude that the individuals eating meat were the ones getting sick. The high-meat counties might also have had more factory pollution, different cooking methods, more hepatitis B, or a dozen other differences. Epidemiologists call this the ecological fallacy, and it is one of the most well-documented pitfalls in population research.³

In ecological designs, misclassification problems and confounding variables severely limit any attempt to draw conclusions about what causes disease in individual people.⁴ Ecological studies are useful for generating hypotheses, not for proving them. They sit near the bottom of the evidence hierarchy, well below randomized trials and prospective cohort studies that follow individuals over years. Campbell’s book, however, presents the ecological correlations as though they settle the question, which is where much of the scientific frustration originates.

Cherry-Picking the Correlations

The China–Cornell–Oxford dataset is enormous, containing hundreds of variables that can be correlated with one another in thousands of combinations. Critics have pointed out that Campbell highlighted correlations that supported his thesis while ignoring or downplaying ones that did not. The original data, for instance, showed that counties with higher fish consumption tended to have lower cancer rates, and that certain animal-product variables had no meaningful association with the diseases Campbell blamed them for. Meanwhile, wheat flour consumption showed a positive association with several cancers in the dataset, a finding the book does not emphasize.

When you have a dataset with hundreds of variables and dozens of diseases, some correlations will appear by chance alone. Without pre-registered hypotheses and corrections for multiple comparisons, an analyst can construct almost any narrative by selecting the right subset of correlations. This is not necessarily deliberate deception; confirmation bias is a well-known phenomenon in science. But the result is that the book presents a much cleaner story than the data actually tell.

The Rat Experiments and Their Limits

A significant portion of The China Study’s argument rests not on the Chinese survey data at all, but on Campbell’s own laboratory experiments with rats. In those studies, rats were exposed to aflatoxin B1, a potent liver carcinogen produced by mold, and then fed diets containing varying levels of casein (the primary protein in cow’s milk). Campbell found that rats fed 20% casein diets developed far more pre-cancerous liver lesions than rats fed 5% casein diets. The relationship had a threshold: below about 10% casein, lesion development stayed minimal, but between 10% and 12% it jumped sharply upward.⁵

These findings are genuinely interesting in the context of toxicology, but the leap from them to “animal protein causes cancer in humans” involves several large assumptions that critics have dissected at length. First, the rats were dosed with a specific liver carcinogen that is not a significant exposure for most people in developed countries. The protein was not initiating cancer; it was modulating the growth of lesions that a toxin had already started. Second, only casein was tested. The studies did not compare casein to other animal proteins like egg or fish protein, nor did they test whether plant proteins at 20% would behave differently. Third, the 5% casein diet that suppressed lesion growth also stunted the rats’ overall growth. The original paper itself noted that the sharp increase in lesions occurred just above the protein level needed for maximum body weight gain, raising the possibility that caloric restriction or general growth suppression, not something specific to animal protein, was responsible for the protective effect.

Extrapolating from chemically-initiated rat liver lesions to human cancer of all types is a stretch that few cancer researchers would endorse without much more supporting evidence. Campbell’s book, however, uses these experiments as a centerpiece.

Confounders the Book Underplays

One of the sharpest criticisms involves the role of infectious disease in distorting the Chinese data. Hepatitis B virus (HBV) infection was widespread in rural China during the study period and is a dominant risk factor for liver cancer. Counties with high HBV rates would show elevated liver cancer deaths regardless of what people were eating. If those same counties happened to consume more animal protein because they were wealthier or had different agricultural traditions, a spurious correlation between protein and cancer would appear.

Schistosomiasis, a parasitic disease caused by blood flukes, adds another layer. In endemic regions of China, schistosomiasis causes severe liver fibrosis. Research in Hubei province has shown that people who tested positive for schistosoma antibodies had dramatically higher rates of hepatitis B co-infection compared to those testing negative.⁶ The interplay between these two infections creates distinctive patterns of liver disease that can mimic or amplify the kind of damage usually attributed to diet or toxins. Studies of schistosomiasis morbidity in southwest China have acknowledged that hepatitis B virus can produce its own liver pathology and may worsen fibrosis caused by the parasite, though the extent of this interaction remains unclear.⁷

The China–Cornell–Oxford Project did collect data on hepatitis B prevalence, and Campbell has acknowledged the virus as a risk factor. But critics argue that in an ecological analysis, simply “controlling” for a variable like HBV at the county level does not adequately remove its confounding effect, especially when HBV prevalence itself correlates with other dietary and economic variables. The fundamental limitation of the ecological design resurfaces here: you cannot tease apart intertwined causes when your unit of analysis is an entire county.

What the Cardiovascular Data Actually Showed

The China Study makes a strong case that animal foods drive heart disease, but the original research papers paint a more complicated picture. The Cornell China study’s own published findings on coronary artery disease mortality show that death rates were inversely associated with green vegetable intake and certain blood fats, while being positively associated with salt intake and apolipoprotein B. Those apolipoproteins were, in turn, positively correlated with animal protein and meat intake but inversely correlated with plant protein and legume intake.⁸

At first glance, that seems to support the book’s thesis. But the associations are indirect, running through a chain of intermediate biomarkers, and the overall cardiovascular disease picture in the dataset was not as tidy as the book suggests. Some forms of vascular disease showed weak or inconsistent associations with animal food intake in the original publications, a nuance that gets smoothed over in the popular account. The ecological design also means that even associations that do appear cannot be assumed causal. Counties with higher meat intake in 1980s China were also wealthier, more urbanized, and had different levels of physical activity, smoking, and alcohol consumption. Sorting out which factor does the actual damage requires a different kind of study.

The Cholesterol Paradox

One of the more counterintuitive findings in the China data, and one that critics have used to challenge the book’s framing, involves cholesterol. Average blood cholesterol levels in rural China were substantially lower than in Western countries, and within the dataset, lower cholesterol tended to correlate with lower heart disease. Campbell used this to argue for a plant-based diet as a way to drive cholesterol as low as possible. But the relationship between very low cholesterol and overall health is not straightforward.

Research on elderly Chinese populations has found that low LDL cholesterol was associated with higher all-cause mortality, not lower. Studies have linked low LDL to increased risk of depression, cancer, liver disease, and hemorrhagic stroke, though establishing a direct causal link has been difficult.⁹ A large prospective study of Chinese adults found a U-shaped or J-shaped pattern: high cholesterol was primarily linked to coronary heart disease mortality, while low cholesterol was associated with higher all-cause and cancer mortality.¹⁰

This does not mean low cholesterol is bad for everyone, but it does complicate the “lower is always better” narrative. In rural China, very low cholesterol levels could have reflected malnutrition, chronic infection, or liver disease rather than a healthy plant-based diet. The China Study treats low cholesterol in the dataset as evidence for the benefits of avoiding animal foods, but that interpretation ignores the well-documented phenomenon of reverse causation, where illness drives cholesterol down rather than the other way around.

What Larger Studies Found About Meat and Mortality in Asia

If animal protein truly drove cancer and heart disease as powerfully as The China Study claims, you would expect to see strong signals in prospective cohort studies that follow large numbers of individuals over time, the kind of evidence that ecological surveys cannot provide. The results from Asia have been more ambiguous than the book would lead you to expect.

A pooled analysis of Asian prospective cohort studies found no association between total meat intake and all-cause, cardiovascular, or cancer mortality in either men or women. Comparing the highest to the lowest quartile of meat consumption, the hazard ratios were essentially flat: about 1.02 for men and 0.93 for women.¹¹ A study of Chinese adults in Shanghai did find a positive association between red meat and total mortality in men, but not in women, and the authors noted that meat consumption levels in their population were lower than in typical American cohorts. The main red meat consumed in Shanghai was pork rather than beef, and processed meat intake was very low.¹²

These studies do not exonerate red meat or prove it is harmless; they do, however, suggest that the relationship between animal protein and disease is far less dramatic and far more context-dependent than The China Study portrays. The type of meat, how it is processed, how much is eaten, and what else is in the diet all matter. A blanket claim that “animal protein causes cancer and heart disease” does not hold up when tested with stronger study designs.

Where the Book Outran Its Evidence

The most common thread across criticisms of The China Study is not that everything in it is wrong, but that the book consistently presents uncertain, preliminary, or incomplete evidence as though it settles major scientific questions. The ecological correlations become proof of causation. The rat casein experiments become universal statements about all animal protein in all humans. Individual correlations that support the thesis get highlighted while contradictory ones get buried or explained away.

An evaluation of popular diet and health books found that while many American diet books do cite large numbers of references, citing references is not the same as accurately representing what those references show.¹³ The China Study includes an extensive bibliography, and many of the cited studies are real and reputable. The problem lies in the interpretive layer on top: the way correlations become causes, the way animal studies become human prescriptions, and the way a complex, messy dataset gets narrated as a simple morality tale about plants versus animals.

None of this means plant-based diets are unhealthy. There is legitimate evidence, from other study designs, that diets rich in vegetables, legumes, and whole grains are associated with lower risks of several chronic diseases. But that evidence stands on its own merits; it does not need the overreaching framework of The China Study to support it. And plant-based diets carry their own challenges, including the risk of falling short on certain micronutrients if not carefully planned.¹⁴

Why the Book Still Resonates

Despite the substantial scientific criticism, The China Study remains enormously influential, particularly in plant-based and vegan communities. Part of its staying power comes from the fact that it tells a coherent, emotionally satisfying story at a time when nutrition science is notoriously messy and contradictory. “Eat plants, avoid animal products, prevent cancer and heart disease” is a simple, actionable message. The actual state of the evidence, with its endless qualifications about dose, context, confounders, and study design, is much harder to turn into a lifestyle.

Campbell’s personal credibility also plays a role. He is a legitimate scientist with a long career in nutritional biochemistry, and the China–Cornell–Oxford Project was a real and ambitious research endeavor. The criticism is not that Campbell fabricated data or that the project was poorly run. It is that the journey from the data to the book’s conclusions required a series of interpretive leaps that the data themselves do not support, and that the book presents those leaps as though they are settled science. For readers who encounter the book without a background in epidemiology, the distinction between an ecological correlation and a causal finding is not obvious, and the book does little to make it clear.

Nutrition science has moved considerably since the 1980s data collection. Prospective cohort studies, randomized feeding trials, and Mendelian randomization analyses now offer tools that can answer questions ecological surveys never could. The questions Campbell raised about diet and chronic disease remain genuinely important. The answers, though, are turning out to be considerably more complicated than any single book, however comprehensive, could capture.