Skip to main content

Welcome to a Grasp lesson

Create a free account to keep studying and create your own study path.

Lesson 4

Evaluating Causal Claims in Chronic Disease Research

Recommended resources

Welcome to the module on testing chronic-disease claims. MAHA connects several proposed causes—diet, environmental exposures, inactivity, and aspects of medical care—to chronic illness. This lesson provides a way to evaluate any specific causal claim without assuming that the whole account stands or falls together. The focus is on study design, confounding, and the possibility that a change in diagnosed disease reflects a change in detection rather than a change in underlying illness.


From a claim to a comparison

Consider the hypothetical claim: “A particular dietary pattern increases the risk of type 2 diabetes.” To assess it, a study needs to specify the dietary pattern, the people being compared, the period of exposure, and the outcome. “Diagnosed diabetes within five years” is measurable, but it is not quite the same thing as the biological onset of diabetes: diagnosis also depends on whether and when someone is tested.

The central difficulty is that researchers cannot observe the same people twice over the same five years, once with the dietary pattern and once without it. Instead, they compare groups and try to make the groups comparable in other respects. An association means that exposure and outcome occur together more or less often than expected. A causal effect means the outcome would have differed if the exposure had been changed while relevant other conditions remained comparable.

Study design determines how close a comparison can come to that ideal.

Module 3: Epidemiologic Studies: A General Overview

Watch “Module 3: Epidemiologic Studies: A General Overview” from the Centers for Disease Control and Prevention for its comparison of study designs and its account of common errors. Its examples concern radiation; the design principles also apply to chronic-disease claims.

Start with study designs: follow how cohort studies begin with exposure, case-control studies begin with disease, and ecological studies compare groups rather than individuals. Then watch sources of error for the distinctions among chance, bias, and confounding. Focus on what information each design actually collects, rather than treating the video’s ranking of designs as a rule for every research question.

The diagram below adds a useful time perspective. Its “retrospective study” column depicts looking back through records to reconstruct exposure groups and later disease occurrence—a retrospective cohort approach. That differs from a case-control study, which selects people according to whether they already have the disease.

Five study designs arranged against past, present, and future time: cross-sectional studies measure disease prevalence at a snapshot; case-control studies compare earlier exposures among people with and without disease; retrospective and prospective cohorts compare later disease incidence across exposure groups; randomized trials assign groups before comparing outcomes.

In a cross-sectional study, exposure and disease are measured at roughly the same time. It can show, for example, that people reporting a dietary pattern are more likely to currently have diabetes. It usually cannot establish whether the diet preceded the disease. People may also have changed what they eat because they were diagnosed.

A case-control study starts with people who have the disease and a suitable comparison group who do not, then examines earlier exposures. It is useful when an outcome is uncommon or takes a long time to develop. Its credibility depends heavily on selecting appropriate controls and obtaining reliable exposure histories. People with a diagnosis, for instance, may recall or report past habits differently from people without one.

A cohort study starts with exposure groups and compares their subsequent disease occurrence. A prospective cohort follows people into the future; a retrospective cohort uses records to reconstruct a sequence that has already happened. Either can establish that a recorded exposure came before a recorded outcome, provided the dates and measurements are dependable. Neither automatically removes differences between the groups.

In a randomized trial, researchers assign an intervention by chance. This makes pre-existing characteristics less likely, on average, to explain differences between groups. Trials are especially informative when the proposed exposure can be changed ethically and maintained long enough to measure the relevant outcome. They still depend on good follow-up and comparable outcome measurement. Many long-term chronic-disease questions cannot be answered by a single feasible trial.

Finally, an ecological comparison might find that places with higher consumption of a product also have higher diabetes rates. It does not establish that the individuals who consumed more were the individuals who developed diabetes. Such comparisons can generate hypotheses, but differences in age, income, health care, or other factors between places can produce misleading patterns.



Confounding: when a third factor explains the pattern

A study may measure exposure before disease and still mistake association for causation. Confounding occurs when another factor is related both to the exposure and, independently, to the outcome. Age is a frequent candidate in chronic-disease research: age distributions can differ between comparison groups, and many chronic diseases become more common with age.

The following numbers are invented solely to illustrate the problem, not findings about any real diet:

Five-year comparison People New diabetes diagnoses Diagnosed risk
Higher exposure 1,000 120 12%
Lower exposure 1,000 80 8%

Taken alone, the table suggests that diagnosis was times as common in the higher-exposure group. Now suppose the groups have different age distributions:

Age group Higher exposure Lower exposure
Older 112 diagnoses among 800 people: 14% 56 among 400: 14%
Younger 8 among 200: 4% 24 among 600: 4%

The overall 12% versus 8% difference remains mathematically correct. Yet within each age group, the diagnosed risk is identical across exposures. The higher-exposure group simply contains more older people. These data do not support an effect of the exposure on diagnosis; age accounts for the observed overall difference in this example.

Researchers address known confounders by comparing people within relevant groups, matching participants, or adjusting statistically for measured differences. These methods require accurate information about the factors being controlled. They cannot guarantee that an unmeasured factor—or a poorly measured one—has been removed. Nor should every variable associated with an outcome automatically be adjusted away: a factor that is itself a consequence of the exposure may be part of the effect under investigation.

Analyzing and Interpreting Data | Field Epi Manual

Read the CDC Field Epidemiology Manual’s account of why an observed association may not be causal, then its guidance for weighing causal evidence. Although written for field investigations, the distinctions apply to chronic-disease studies.

In “Interpreting Data from a Field Investigation,” begin with the paragraph opening “A computer can crunch numbers” and continue through “Epidemiologic Interpretation Checklist 1.” Read the error checklist, noting how chance, selection bias, information bias, and confounding differ. Then read the discussion preceding and within “Epidemiologic Interpretation Checklist 2,” from the paragraph beginning “Now suppose that an elevated RR or OR” through the dose-response paragraph: causal considerations. Focus particularly on exposure preceding disease, consistency across studies, and the fact that dose-response is supportive rather than mandatory.

The manual separates several threats that are easy to collapse into a single word, “uncertainty.” Chance is random variation; a confidence interval helps convey how precisely a study has estimated an association. A narrow interval cannot, by itself, rule out systematic error. Selection bias concerns who enters or remains in a study: if participants lost during follow-up differ in both exposure and health risk, the remaining comparison may be distorted. Information bias concerns how exposure or disease is measured: one group might remember past diet differently or receive more diagnostic testing. Confounding concerns an alternative explanation rooted in differences between the groups, as in the age example.

After investigating these alternatives, causal confidence grows when exposure clearly precedes disease, results recur in independently conducted studies, and the proposed mechanism fits with other evidence. A stronger association or a credible dose-response pattern can add support. None is an automatic verdict. A plausible mechanism cannot rescue badly measured data, while the absence of a simple dose-response pattern does not prove that an effect is impossible.



A rise in diagnoses is not necessarily a rise in disease

Claims about a chronic-disease “epidemic” often begin with a trend. Before attributing that trend to a diet, chemical, or policy, separate three quantities:

  • Number of cases: a larger population can have more cases even if each person’s risk is unchanged.
  • Incidence: new cases over a stated period among people at risk.
  • Prevalence: people living with a condition at a given time. This can rise because more people develop the condition, but also because people live longer with it.

Reported incidence and prevalence additionally depend on how conditions are defined, detected, and recorded. More screening can find previously undiagnosed disease. A changed diagnostic threshold can classify additional people as cases. Improved reporting can make a longstanding problem more visible. None of these possibilities implies that every observed increase is artificial; real and detection-related increases can occur together.

Principles of Epidemiology: Lesson 5, Section 5|Self-Study Course SS1978|CDC

Read the CDC self-study course’s explanation of apparent changes in disease occurrence. It supplies a practical check before treating a reported trend as evidence that an exposure has caused harm.

Under “Interpreting results of analyses,” locate the paragraph beginning “Observed increases or decreases in incidence or prevalence.” Read artifactual changes, including the intervening list and the comparison of Figures 5.7 and 5.8. Focus on case definitions, diagnostic procedures, health-seeking behavior, and reporting. The figures illustrate why similar-looking increases need not have the same explanation.

For a claim that diabetes diagnoses have risen, a stronger trend analysis would use rates with consistent population denominators, examine rates within comparable age groups, and check whether screening practices and diagnostic definitions changed. It would also examine evidence of underlying disease burden, while recognizing that any additional measure has its own limitations. An age-adjusted rise under stable definitions and testing would be more persuasive evidence of a real increase than a rise in raw case counts alone. It still would not identify which exposure caused the rise.



Bringing the evidence together

A measured judgment is more useful than either “correlation proves it” or “observational evidence proves nothing.” For a particular chronic-disease claim, the reasoning can be stated in this order:

  1. Define the claim precisely. Name the exposure, the disease outcome, the population, and the time period. Keep “more diagnoses” distinct from “more underlying disease.”
  2. Identify the comparison and its timing. Determine whether the evidence compares places, takes a snapshot, reconstructs past exposures, follows exposure groups, or assigns an intervention. Establish whether exposure was measured before disease.
  3. Test alternative explanations. Examine chance, selection and measurement bias, plausible confounders, and any change in diagnostic or reporting practices. Note which concerns the study actually addressed.
  4. Weigh the wider evidence. Give weight to replicated findings from methods with different weaknesses, consistent timing, and relevant experimental evidence when available. State what remains uncertain.

Suppose a report shows that both consumption of a product and diagnosed diabetes increased over two decades. That observation can justify investigation, but it cannot by itself show that the product caused diabetes: the comparison is at the population level, exposure histories are unknown, and disease detection may have changed. If well-conducted individual-level studies then repeatedly find that exposure precedes new disease, with comparable testing and careful attention to confounding, the causal case becomes stronger. If the association disappears when groups are made comparable, or tracks a change in screening rather than disease occurrence, the claim becomes weaker.

This standard is not a demand for perfect evidence before taking any action. It is a way to state accurately what the current evidence supports, what it leaves open, and what new finding would change the assessment.



Conclusion

Study design tells you what a comparison can establish; confounding and bias test whether its apparent effect has another explanation; and stable diagnosis and reporting are essential for interpreting disease trends. An increase in reported illness and an association with an exposure are important observations, but neither alone establishes a cause.

The next lesson applies this framework to ultra-processed foods, distinguishing what controlled feeding trials can show about eating behavior from what observational studies can establish about long-term chronic disease.