Observational vs Experimental Studies

In an experimental study the investigator controls and allocates the exposure — classically by randomisation (the RCT). In an observational study (cohort, case-control, cross-sectional) the investigator only measures exposures and outcomes that occur naturally, without intervening in who is exposed.

The same evidence pyramid, split by who controls the exposure. The top tiers (RCTs and meta-analyses of them) are experimental — the investigator randomises. The lower tiers (cohort, case-control, case series) are observational — exposure is whatever happened in real life.

Experiments rank higher because randomisation balances both measured and unmeasured confounders, isolating cause. Observational studies sit lower because exposure is self-selected: even with statistical adjustment, residual and unmeasured confounding can manufacture — or hide — an effect.

Randomisation is the only tool that controls confounders nobody has measured, so experiments give the cleanest causal answer. But they are not always possible: when randomising would be unethical or impractical, observational designs are the only option — and they better reflect real-world, unselected practice. The trade is feasibility and realism against confounding.

• Experimental = investigator assigns exposure (randomises)
• Observational = watch only → vulnerable to confounding
• Randomisation balances unmeasured confounders; adjustment cannot

HRT and coronary heart disease — observational cohorts vs the WHI RCT. Large observational studies (e.g. the Nurses’ Health Study) reported that postmenopausal women taking hormone replacement therapy had roughly half the rate of coronary events, and HRT was widely prescribed for cardioprotection. Then the randomised Women’s Health Initiative (estrogen + progestin, JAMA 2002, 16,608 women) was stopped early: CHD was increased (hazard ratio ≈ 1.24), along with stroke, VTE and breast cancer. The observational “benefit” was largely a healthy-user effect: women who chose HRT were healthier, slimmer and better-screened. Randomisation in the WHI removed that selection — and the apparent protection reversed into harm.

• Confounding by indication — reading a treatment–outcome association from observational data as a causal effect.
• Immortal time bias — the period before a patient could be classed as “treated” is wrongly credited to the treated group, fabricating benefit.
• Over-reading observational causal claims — statistical adjustment cannot remove unmeasured confounding.