Ex Post Facto Research: Design and What It Cannot Prove

Published June 20, 2026 | Last updated October 31, 2026

Quick Answer: Ex post facto research studies a cause after the fact. The exposure and the outcome have both already happened, and the researcher starts from the outcome and looks back at possible causes.¹ It can show that groups differ, but on its own cannot establish why, because nothing was assigned and other differences between the groups remain possible explanations.¹^,² Scope of our worked case: one open-access cross-sectional study of 119 patients with glaucoma, data collected June to August 2018, published 2021 under CC BY 4.0.³ Recomputed in R, the physical health gap between patients with and without visual impairment in the eye with better vision was Hedges g 1.57 (95% CI 1.04 to 2.10). That is a large, unadjusted gap, and a single unadjusted comparison cannot say what caused it.

What Ex Post Facto Research Is

Ex post facto means after the fact. Kerlinger, as quoted in a 1973 ERIC report, defined ex post facto research as “that research in which the independent variable or variables have already occurred and in which the researcher starts with the observation of a dependent variable or variables.” The researcher then studies the independent variables in retrospect for their possible relations to the dependent variable.¹

The same report explains why people use it: an experiment is sometimes impractical or prohibitively costly in time, money and effort, and in other cases it is unethical. It gives examples such as staging fatal airplane crashes to ascertain their causes.¹

Two rows of three steps: an experiment forms groups, gives an exposure, then measures the outcome; ex post facto research starts with an outcome already present, forms groups from what happened, then looks back for candidate causes — The direction of inference. Editverse schematic, following the description of ex post facto research quoted from Kerlinger in Lord (1973).

Ex Post Facto, Causal-Comparative and Quasi-Experimental: How the Labels Are Used

The labels overlap in the sources we read, so it helps to see how each source uses them. This page is about the after-the-fact logic and what a gap between groups can and cannot show. For the mechanics of comparing groups, see our guide to causal-comparative research.

Swipe sideways to see every column.

Label	How the sources we read use it	Manipulation by the researcher
Ex post facto	Kerlinger, quoted by Lord: the independent variables have already occurred and the researcher starts with the dependent variable.	None. Tuckman, quoted by Lord, says the treatment is included by selection rather than manipulation.
Causal-comparative	Lord uses it as a synonym: he writes of the causal comparative method or ex post facto research design. Van Dalen, quoted there, compares a community that has experienced a riot with one that has not.	None. It does not control the variable factors and observes under normal field conditions.
Quasi-experimental	Mouly, quoted by Lord, calls the ex post facto experiment a relatively questionable quasi-experimental design, experimentation in reverse.	Our guide to quasi-experimental designs covers designs in which an intervention is implemented. Check how your target journal classifies yours.

The glaucoma paper we use below shows the overlap in practice: its abstract says 119 patients were included in a causal-comparative, ex post facto research design, and also that the study was observational and cross-sectional.³ We therefore use ex post facto research for the logic and cite each source in its own words.

How an Ex Post Facto Study Is Run

Isaac and Michael, quoted by Lord, give seven steps: define the problem; survey the literature; state the hypotheses; list the assumptions on which hypotheses and procedures rest; design the approach, which includes selecting subjects and sources, choosing data techniques and establishing unambiguous categories; validate the data-gathering techniques; and describe, analyse and interpret the findings.¹ Kerlinger, as quoted there, advised being highly skeptical of any ex post facto study that tests only one hypothesis and said alternative negative hypotheses should be routinely tested, with significant and nonsignificant relations predicted in advance where possible.¹

The weaknesses Isaac and Michael list are the useful part. We matched five of them to the case study below; the right-hand column is our reading of the glaucoma paper.¹^,³

Swipe sideways to see every column.

Weakness listed by Isaac and Michael (as quoted by Lord)	What it looks like in the glaucoma data (Editverse reading)
Lack of control over independent variables	Vision status was measured, not assigned. The groups are the patients who happened to have normal or impaired vision.
Hard to say which of two related variables is cause and which is effect	Visual function and quality of life were collected in one period, June to August 2018, with no follow-up reported. Poor quality of life could follow vision loss, or the reverse could contribute.
Two factors may both relate to an additional factor not observed	Mean age was 59.4 years (SD 18.32), and 37 of 119 patients (31%) were over 70 (Table 1). We found no adjustment for age, disease duration or education in the article.
Dichotomous groups can be vague	The impaired group includes WHO category 1, labelled near-normal vision (6/18 to 6/60). Impaired means categories 1 to 4, so 9 + 4 + 3 + 3 = 19 patients for the better eye, collapsed into one group.
Locating existing groups that differ only in one exposure is difficult	Groups of 19 and 22 impaired patients were compared with 100 and 97 normal-vision patients, from outpatient clinics the authors approached.

A Real Case: Quality of Life in Glaucoma, Recomputed in R

Majerníková and colleagues studied 119 outpatients with primary open-angle glaucoma in Slovakia. They approached selected private ophthalmology outpatient clinics, collected data from June to August 2018 with the WHOQOL-BREF and the NEI VFQ-25 questionnaires, and split patients by visual acuity into normal and impaired, once for the eye with better vision (EBV) and once for the eye with worse vision (EWV).³ The article is CC BY 4.0, which allowed us to redraw its numbers. Its comparisons used Mann-Whitney U and Kruskal-Wallis tests.³

What the published counts say

From the paper’s Table 2, 100 patients had no impairment in the better eye, so 19 did (16.0% of 119), and 97 had none in the worse eye, so 22 did (18.5%).³ Table 5 gives each group’s mean and SD for six WHOQOL-BREF items. We recomputed the differences and Hedges g with 95% confidence intervals from those summary statistics. The paper’s own p values for the better-eye split come from Mann-Whitney U tests, so they are shown beside our figures, not replaced by them.³

Swipe sideways to see every column.

WHOQOL-BREF item	Impaired vision, mean (SD), n = 19	Normal vision, mean (SD), n = 100	Difference, normal minus impaired (95% CI)	Hedges g (95% CI)	Paper p (Mann-Whitney)
Physical health (D1)	12.07 (1.86)	15.12 (1.94)	3.05 (2.09 to 4.01)	1.57 (1.04 to 2.10)	0.00108
Psychological health (D2)	14.33 (1.72)	16.12 (2.02)	1.79 (0.88 to 2.70)	0.90 (0.40 to 1.40)	0.00030
Social relations (D3)	15.65 (1.97)	15.94 (2.55)	0.29 (-0.77 to 1.35)	0.12 (-0.37 to 0.61)	0.1096
Environment (D4)	14.49 (2.35)	16.98 (2.62)	2.49 (1.26 to 3.72)	0.96 (0.45 to 1.46)	0.00081
Overall quality of life (Q1)	4.00 (0.78)	4.52 (0.89)	0.52 (0.11 to 0.93)	0.59 (0.09 to 1.09)	0.0019
Overall health satisfaction (Q2)	2.59 (0.89)	4.00 (0.88)	1.41 (0.95 to 1.87)	1.59 (1.06 to 2.12)	0.0014

Forest plot of Hedges g with 95% confidence intervals for six quality-of-life domains in 119 patients with glaucoma, two splits per domain; physical health 1.57 and 1.18, psychological 0.90 and 0.76, social relations 0.12 and 0.16 (interval includes zero), environment 0.96 and 0.88, overall quality of life 0.59 and 0.60, overall health 1.59 and 1.46 — Recomputed gaps for both splits. Hedges g and 95% CIs computed by Editverse in R from Table 5 of Majerníková et al. 2021 (CC BY 4.0).

Computed by Editverse in R

Table 5 check: for each item the two group sizes sum to 119 (19 + 100 in the better-eye split, 22 + 97 in the worse-eye split).
Five of six items have a better-eye 95% CI that excludes zero. Social relations does not: g = 0.12 (-0.37 to 0.61), consistent with the paper’s nonsignificant p of 0.1096.
Table 1 lists 79 patients with both eyes diagnosed, 26 with the right eye and 11 with the left, which sums to 116, not 119. The paper reports 68.0% for both eyes; 79 of 116 is 68.1% and 79 of 119 is 66.4%.
The article text gives 16.92 as a mean for patients with pathological visus on the environment item in the worse-eye split, while Table 5 gives 14.59 for that group and 16.92 for the normal-vision group. We used Table 5.
The better-eye and worse-eye comparisons use the same 119 people, so the two sets of results are not independent replications.

What this case shows: in this clinic sample, patients with impaired vision reported lower physical health, psychological health, environment, overall quality of life and overall health scores than patients with normal vision, and social relations did not clearly differ. What it does not show: that impaired vision caused the gaps, or how large any causal effect would be. A text search of the full article found none of the words bias, limitation or confounder, and we found no adjustment for age, disease duration or education.³

Why a Group Gap Found After the Fact Does Not Prove Cause

Three separate problems sit behind every ex post facto comparison.

Confounding

The STROBE explanation document says confounding “produces relations that are factually right, but that cannot be interpreted causally because some underlying, unaccounted for factor is associated with both exposure and outcome.” It also says bias and confounding are not synonymous.² In the glaucoma data, age is a candidate: it varies widely in the sample (SD 18.32 years) and could plausibly relate to both vision and physical health, but the article does not report age by group, so we cannot test that here.³

Order in time

Fedak and colleagues describe Bradford Hill’s nine viewpoints: strength of association, consistency, specificity, temporality, biological gradient, plausibility, coherence, experiment and analogy. They report that Hill explained that for an exposure-disease relationship to be causal, exposure must precede the onset of disease.⁴ Our reading is that one cross-sectional comparison supplies strength of association and little else; the other viewpoints need evidence from other studies or data.

A significant gap with no effect at all

To show what confounding does to a test, we ran a simulation. It is not data from the paper. We borrowed three things from the glaucoma article: 119 patients, age mean 59.4 and SD 18.32, and roughly 16% of patients impaired. We then made vision have no effect on the outcome, let older age raise the chance of impairment and lower the outcome, and compared groups in 2,000 simulated studies.³

Two histograms of 2,000 simulated studies with a true vision effect of zero: the naive comparison has median g 0.43 with p below 0.05 in 39.0 percent of studies, while the age-adjusted comparison has median g 0.00 with 4.55 percent — Editverse simulation in R (seed 20261031). The assumed age effect and impairment model are invented for illustration. Age mean, SD and sample size are from Majerníková et al. 2021 (CC BY 4.0).

The naive comparison found p below 0.05 in 39.0% of simulated studies although the true effect was zero. Adjusting for age brought that to 4.55%, close to the 5% expected by chance. Our assumed confounder produced a median g of 0.43, smaller than most gaps in the glaucoma data, so it does not explain them. The point is narrower: a significant gap cannot tell a real effect from a confounded one.

What to Report So Readers Can Judge Cause

STROBE covers cohort, case-control and cross-sectional studies, with 22 items of which 18 are common to all three designs.⁵ The glaucoma article does not mention STROBE; we use five items as a lens on it. Item wording is from the explanation document.² Our STROBE tutorial walks through the full checklist, and our guide to sensitivity analysis for unmeasured confounding covers item 17.

Swipe sideways to see every column.

STROBE item	Wording	Glaucoma article (Editverse reading)
9 Bias	Describe any efforts to address potential sources of bias.	No statement of efforts to address bias found.
12(a) Statistical methods	Describe all statistical methods, including those used to control for confounding.	Mann-Whitney U and Kruskal-Wallis tests; no method to control for confounding found.
16(a) Main results	Give unadjusted estimates and, if applicable, confounder-adjusted estimates and their precision (e.g., 95% confidence intervals). Make clear which confounders were adjusted for and why they were included.	Group means, SDs, Z and p values; no confidence intervals and no adjusted estimates.
17 Other analyses	Report other analyses done—e.g., analyses of subgroups and interactions, and sensitivity analyses.	None found.
19 Limitations	Discuss limitations of the study, taking into account sources of potential bias or imprecision. Discuss both direction and magnitude of any potential bias.	No limitations discussion found in the article text.

The practical rule from these items: put the unadjusted and adjusted estimates side by side, name the confounders you adjusted for and why, and say in which direction any remaining bias would push the gap.²

Need the Analysis Run or Checked? Consult Editverse

Editverse subject-matter experts work to COPE, ICMJE and GPP 2022 guidance, with institutional and country-specific ethics followed.

Editverse

Statistical Analysis

Group gap found and the analysis still to run?

You send: your dataset, your groups and your research questions.

You get: the analysis in SPSS, R or Stata with annotated code, complete output files, an interpretation guide and publication-ready tables, in 4-6 business days.

When: You need the comparison run and reported.

SPSS/R/STATA Statistical Analysis · $400

Analyse my data →

Editverse

Manuscript Writing

Results in hand and the paper still unwritten?

You send: your data, your outline and your references.

You get: a manuscript written by PhD experts, with editing, statistical review, journal selection and submission support.

When: Your study is finished and the paper is not.

Manuscript Writing Services · $1,500

Write my paper →

Editverse

Journal Selection

Unsure where an observational study will fit?

You send: your abstract and your field.

You get: a ranked list of 5-10 journals with the rationale for each, plus a follow-up consultation.

When: You are choosing where to submit.

Scientific Journal Selection Services · $115

Get my journal shortlist →

Free Research Tools from Editverse

Effect Size Calculator: Cohen’s d for two independent groups, computed from the pooled standard deviation.
Odds Ratio & Relative Risk Calculator: an odds ratio or relative risk with its 95% CI from a 2×2 table.
Free STROBE Participant Flow Diagram Generator for Observational Studies: generate the STROBE participant flow diagram.

Frequently Asked Questions

What is ex post facto research?

It is research in which the exposure has already happened and the researcher starts from an observed outcome, then looks back at possible causes. Kerlinger, quoted in a 1973 ERIC report by Lord, defined it as research in which the independent variables have already occurred and the researcher starts with the observation of the dependent variable.

Is ex post facto research the same as causal-comparative research?

The two labels are used together in the sources we read. Lord writes of the causal comparative method or ex post facto research design, and the glaucoma paper used here describes its design as causal-comparative in character with an ex post facto method. Other authors treat the ex post facto experiment as a questionable quasi-experimental design, so state which label you follow and what was and was not manipulated.

Can ex post facto research prove cause and effect?

On its own, a single comparison of existing groups cannot establish cause. Tuckman, quoted by Lord, says a predicted relationship does not necessarily mean the variables are causally related. STROBE explains that confounding produces relations that are factually right but cannot be interpreted causally because an unaccounted for factor is linked to both exposure and outcome.

Is ex post facto research a quasi-experimental design?

Some authors say so. Mouly, quoted by Lord, calls the ex post facto experiment a relatively questionable quasi-experimental design because the researcher has no control over situations that have already occurred. Other sources simply call it non-experimental, so check how your supervisor or target journal classifies it.

What steps does an ex post facto study follow?

Isaac and Michael, quoted by Lord, list seven: define the problem, survey the literature, state the hypotheses, list the assumptions, design the approach with subjects, data techniques and unambiguous categories, validate the data-gathering techniques, and describe, analyse and interpret the findings.

When is ex post facto research the right choice?

Lord gives the reasons: an experiment is sometimes impractical or too costly in time, money and effort, and in other cases it is unethical. Isaac and Michael, quoted there, add that holding every variable but one constant can be unrealistic and can prevent the normal interaction of influential variables.

How should I report an ex post facto study so readers can judge cause?

Use the STROBE items that bear on cause when the design is cohort, case-control or cross-sectional: describe efforts to address bias (item 9), describe the methods used to control for confounding (item 12a), give unadjusted and confounder-adjusted estimates with 95% confidence intervals (item 16a), report other analyses such as sensitivity analyses (item 17), and discuss limitations with the direction and size of any bias (item 19).

What did the glaucoma example show?

In 119 patients with glaucoma, the group with normal vision in the eye with better vision scored higher on physical health, with Hedges g of 1.57 and a 95% confidence interval of 1.04 to 2.10, computed by Editverse from the published summary statistics. The paper compared unadjusted groups, so it shows an association between the groups, not what caused it.

References

Lord HG. Ex post facto studies as a research method. Special Report No. 7320. Syracuse City School District, Syracuse, NY; July 1973. ERIC document ED090962 (the report quotes Kerlinger 1964, Tuckman 1972, Mouly 1970 and Isaac and Michael 1971, which we cite only as quoted there). eric.ed.gov/?id=ED090962
Vandenbroucke JP, von Elm E, Altman DG, et al. Strengthening the Reporting of Observational Studies in Epidemiology (STROBE): explanation and elaboration. PLoS Med. 2007;4(10):e297. doi:10.1371/journal.pmed.0040297. Open access, CC BY 4.0. pubmed.ncbi.nlm.nih.gov/17941715/
Majerníková Ľ, Hudáková A, Obročníková A, Grešš Halász B, Kaščáková M. Quality of life of patients with glaucoma in Slovakia. Int J Environ Res Public Health. 2021;18(2):485. doi:10.3390/ijerph18020485. Open access, CC BY 4.0; tables redrawn and derived numbers computed by Editverse in R. pubmed.ncbi.nlm.nih.gov/33435305/
Fedak KM, Bernal A, Capshaw ZA, Gross S. Applying the Bradford Hill criteria in the 21st century: how data integration has changed causal inference in molecular epidemiology. Emerg Themes Epidemiol. 2015;12:14. doi:10.1186/s12982-015-0037-4. Open access, CC BY 4.0. pubmed.ncbi.nlm.nih.gov/26425136/
von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP; STROBE Initiative. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. PLoS Med. 2007;4(10):e296. doi:10.1371/journal.pmed.0040296. Open access, CC BY 4.0. pubmed.ncbi.nlm.nih.gov/17941714/