Observational Studies and Confounding

What observational studies are

In an observational study, researchers watch what happens to people without assigning them a treatment. Common types are cohort studies (follow a group over time), case-control studies (compare people with and without a condition) and cross-sectional studies (a snapshot). [1] Much biomedical research is observational. [1] These studies are essential for rare harms, long-term patterns and questions where a trial is not possible. Their weakness is cause and effect.

Association is not cause

An association means two things occur together. A confounder is a third factor linked to both the exposure and the outcome that can create or hide an apparent effect. For example, people who choose a supplement may also exercise more, eat differently, see doctors more often or have more money, and any of those could explain better health on their own. Sorting this out requires adjusting for all confounding factors needed to ensure comparability, a standard that is hard to meet in practice. [2]

Other traps include people who stay on a treatment being different from those who stop. Researchers describing a hormone-therapy analysis noted that the group of current users can be depleted of women susceptible to harm. [2]

A worked example

Hormone therapy for postmenopausal women had decades of observational evidence behind it, yet the Women's Health Initiative authors said the balance of risks and benefits was still uncertain. [3] In their randomized trial, 16,608 women aged 50 to 79 received an estrogen-plus-progestin tablet or placebo. The trial was stopped after a mean of 5.2 years because breast cancer risk crossed the stopping boundary and the overall index suggested risks exceeded benefits. The hazard ratio for coronary heart disease was 1.29 (1.02 to 1.63) and for invasive breast cancer 1.26 (1.00 to 1.59). [3]

This is one trial of one specific product in one population. It shows why a randomized comparison was needed, not that every observational finding is wrong. Across highly cited research, 5 of 6 nonrandomized studies were later contradicted or found stronger effects than later work, versus 9 of 39 randomized trials. [4]

How researchers try to close the gap

• Adjust statistically for known confounders, which cannot handle those not measured.
• Design the analysis to mimic a randomized trial. Hernan and Robins describe causal analysis of observational data as an attempt to emulate a target randomized experiment, and they urge judging studies by how well they do so. [2]
• Look for repeat findings across settings and for randomized trials.

Reading checklist

1. Was anyone assigned the treatment at random? If not, treat results as a lead.
2. Which factors were adjusted for, and which obvious ones were not?
3. Does the study report according to a standard such as STROBE, a 22-item checklist? [1]
4. Do headlines use "causes" or "prevents" for a study that found only a link? See research hype and press releases.

Related guides

Randomization, blinding and placebo explains how trials control for confounding. Systematic reviews and meta-analyses explained shows how studies are combined. Reading longevity studies applies these ideas to aging research.