Stefano Garzon

Research · Oct 23, 2026

Historical controls and non-randomized studies

Concept

A trial needs a comparison group. Randomization is the preferred way to form it. It removes investigator and participant influence on who receives the intervention, tends to balance measured and unmeasured prognostic factors, and supports statistical tests without further assumptions about group comparability.

There are two main alternatives. In a non-randomized concurrent control, groups are treated in the same period, but allocation depends on institution, physician preference or patient choice. In a historical control, a new series of patients receives the intervention and is compared with an earlier series treated differently. Such a control is, by definition, non-randomized and non-concurrent.

Non-randomized concurrent control

The central weakness is the lack of assurance of comparability. The investigator must assume that all important prognostic factors are known and measured. Matching on more than a few factors is impractical. In small studies, real baseline differences go undetected. In large studies, uncertainty about unmeasured factors remains. Subtle physician selection, including unconscious selection, can steer patients with different prognoses into each group.

Historical control

The case for it is ethical and practical: all new participants receive the intervention, recruitment takes roughly half the time, and the treated group can be larger. Control data come from the literature or from files and charts of earlier studies.

The problems fall into three groups.

Pocock identified 19 instances in which the same intervention was used in two consecutive trials at the same institution. Mortality differences between these identically treated groups ranged from −46% to +24%, and 4 of the 19 were significant at 5%.

Empirical evidence

Sacks and colleagues gathered six therapies assessed in 50 randomized and 56 historically controlled trials. The therapy was judged better in 79% of historically controlled trials and in 20% of randomized ones. The difference came mainly from the control groups: historical controls had worse outcomes. Adjustment for prognostic factors did not change the result.

In the review of anticoagulation in acute myocardial infarction, 15 of 18 historically controlled studies and 5 of 8 non-randomized concurrent studies favored treatment. Pooled non-randomized studies suggested a mortality reduction of about 50%; the six pooled randomized trials, about 20%.

Hybrid designs and databases

Pocock proposed combining historical and randomized controls, allocating a smaller fraction to control. This requires the same entry and evaluation criteria, the same center and recent data. If historical-control bias is large, more participants may need randomizing than in a fully randomized trial.

Registries and electronic health records broaden the population studied and help detect adverse events. They depend on the analysis model and suffer from confounding by indication, missing data and multiple comparisons. Bailar and colleagues list what strengthens a study without internal controls: a hypothesis and analysis plan set before the data.

There are exceptions in which the effect is large enough to make concurrent comparison unnecessary, such as penicillin in pneumococcal pneumonia and treatment of malignant hypertension.

Common errors when reading a paper

What to check when reading a trial

  1. How the control group was formed and in what period.
  2. Whether entry criteria, outcome definitions and assessment methods were the same in both groups.
  3. Whether background therapy, diagnostic criteria or coding changed between eras.
  4. Whether hypothesis and analysis plan were set before data collection.
  5. Whether the control event rate is consistent with contemporary randomized trials.

Related topics: allocation concealment, subgroups, confounders and DAGs.

References