Ask Mallard
Study design

Synthetic control

4 min read

Synthetic control compares a treated place with a weighted blend of untreated places that resembled it before an intervention.

NHMRC EVIDENCE LEVEL · 2009Not yet classifiedProvisional design classification. Study quality and certainty of findings require separate appraisal.
Why this level?Specify the research question: intervention, diagnostic accuracy, prognosis, aetiology, or screening intervention. Other question types are outside this hierarchy. NHMRC 2009 source

Synthetic control is a family of nonrandomized comparisons. Its evidence depends on donor choice, time history and identification assumptions, not the method name.

Fit

When to use it, and when not to

Use it when

  • One or a few places adopt a policy, and comparable untreated places have long outcome histories.
  • You can identify an intervention date and plausible donors before viewing outcomes after treatment.

Not when

  • Many treated and untreated units support a shared trend comparison. Consider difference-in-differences, with its own assumptions. Read: Difference-in-differences
  • No credible untreated donor exists, but a long time series does. Consider interrupted time series and its limits. Read: Interrupted time series

Before you start

What you need

Decide these first
  1. The treated unit, date and outcome scaleFix the follow-up window and contrast before fitting. Define whether the effect concerns a place, a rate or a population.
  2. Eligible untreated places and their historiesSelect donors using policy knowledge, comparable measurement and freedom from spillovers. Record exclusions before fitting.
  3. The fitting method and pre-treatment periodClassic synthetic control uses nonnegative weights summing to one. Other variants use different constraints and need their own justification.
  4. A plan for missing periods and changing coverageRecord data gaps before choosing donors. A complete series can still reflect a change in who was measured.
It rests on these
  1. The donor blend approximates the treated unit without treatment.How to check: Check pre-treatment fit over a long period and held-out pre-treatment times. Good fit supports plausibility but does not prove validity.
  2. Donors remain untreated and unaffected by the intervention.How to check: Check anticipation, movement across borders and other policies. Exclude contaminated donors and repeat the analysis.
  3. No unmodeled later shock affects only the treated unit.How to check: List concurrent changes and test alternative dates and donors. A later gap may have another cause.
  4. Missing records are handled under a named assumption.How to check: Missing completely at random (MCAR) means gaps are unrelated to recorded and unrecorded values; this is rarely plausible. Missing at random (MAR) means gaps depend only on recorded factors. Under MAR, consider multiple imputation; it is not automatically better than a justified complete-case analysis. Missing not at random (MNAR) is plausible if unobserved values themselves influence whether records are absent. Under MNAR, test departures with a delta or tipping-point analysis that shifts the unobserved values.

Plain language

How it works

Choose the treated unit and donors first. Fit weights using outcomes and predictors measured before treatment. Apply the fixed weights to donor outcomes after treatment, then subtract the weighted outcome from the observed outcome. Plot both paths and their gap. Classic weights cannot be negative and sum to one; extensions can differ.

Fit before treatment, compare afterwards
Untreated donors and pre-treatment histories determine weights. Fixed weights then form a post-treatment comparison; the outcome gap requires counterfactual assumptions.Fit before treatment, compare afterwardsPre-treatment historiesand eligible donorsFit weights beforetreatmentCompare observed andweighted later outcomesA good pre-treatment fit does not establish the unobserved outcome without treatment. Check spillovers andconcurrent shocks.
Conceptual diagram. It shows the structure of the idea, not data from any study.
Assess how fragile the gap is

Show fit before treatment, exclude donors one at a time, and compare placebo places or dates where credible. Similar gaps in untreated places weaken the interpretation. Uncertainty depends on the method and assumptions; donor counts are not patient sample sizes. Do not treat a close fit as evidence that later shocks are controlled.

Worked example

A worked example

Invented for illustration
Setting
One region changes its referral policy. Two untreated regions form a simple donor blend.
Question
How far is the treated region from its comparison in the follow-up year?
Approach
Weights fixed from pre-policy data are 0.6 and 0.4. Donor admission rates in follow-up are 20 and 30 per thousand.
Result
The weighted comparison is 24 per thousand; the observed rate is 18. The observed-minus-comparison gap is -6 per thousand.
What it means
The gap of -6 is an estimate under the design assumptions. It is not a confidence interval or proof of a policy effect.

In practice

Recent studies PubMed search · not vetted

Searching PubMed…

Watch out

Common mistakes

The one to remember

A close fit before treatment does not prove the weighted comparison represents what would have happened without treatment. Donor contamination, spillovers and later shocks can break the comparison.

  • Choosing donors because they give the desired post-treatment result.Instead: Fix the donor rules and fitting period before examining the later gap.
  • Treating placebo gaps as automatically valid randomization p values.Instead: Assignment was not necessarily random. State the exchangeability argument and limitations of the comparison.
  • Ignoring a poor fit before treatment.Instead: Show the pre-treatment gaps and sensitivity to donors. Reconsider the design if credible fit is not possible.

Inside Mallard

How Mallard uses this

Plans do not link to this design yet. Mallard has no dedicated synthetic-control sizing form. An analysis naming synthetic control triggers its advanced causal methods complexity cue.

No Mallard code example points to this page.
Atlas node design.synthetic_control, the id a plan cue records.

How to report it

No reporting guideline dedicated to synthetic control was found in a PubMed search on 2026-11-06. The search used the design name and reporting-guideline or checklist terms in titles. Use STROBE to structure reporting of the observational comparison, while adding donor eligibility, the fitting period, weight constraints, pre-treatment fit, later gaps and sensitivity analyses. State the identification assumptions and uncertainty method. A reporting checklist cannot establish a valid counterfactual. A reporting guideline is not a design manual: use it to make the study transparent, not as a substitute for careful design. Check the target journal's own requirements as well, which this page cannot know.

Related Learn pages

Related designs and methods

Evidence

Sources & further reading

  1. The Effect of Nationwide Organized Cancer Screening Programs on Gastric Cancer Mortality: A Synthetic Control Study (opens in a new tab) · Gastroenterology · doi 10.1053/j.gastro.2023.11.286
  2. The Long-Term Impacts of an Integrated Care Programme on Hospital Utilisation among Older Adults in the South of England: A Synthetic Control Study (opens in a new tab) · Int J Integr Care · doi 10.5334/ijic.6475
  3. Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies (opens in a new tab) · BMJ · doi 10.1136/bmj.39335.541782.AD
  4. Multiple imputation for missing data in epidemiological and clinical research: potential and pitfalls (opens in a new tab) · BMJ · doi 10.1136/bmj.b2393
Prepared with AI assistance. Each source was verified by PMID against PubMed on 2026-11-06, matching title, authors, year and DOI. Record verification does not establish that a paper supports every use. No independent professional or methodological approval is claimed.
Last source review: 2026-11-06
AI-generated educational content
Mallard uses reputable methodological and reporting sources, but this material may contain errors or omit important context. Verify important methodological decisions with the cited sources and, when appropriate, a qualified research methodologist or statistician.