Published  |  Last updated

Quick Answer: A causal-comparative research design compares groups that already differ on a characteristic the researcher did not assign, such as having a movement pattern or not, and asks how an outcome differs between them.1 Because nobody randomised the grouping, a difference between groups can reflect other baseline differences, so plan in this order: define the groups before you look at outcomes, check baseline balance with standardised differences, choose an adjustment, and report an effect size with its interval.2,3 The worked case is one open-access paper (CC BY 4.0, BMC Musculoskeletal Disorders 2026): 60 older men, 30 with and 30 without dynamic knee valgus, redrawn and recomputed in R. Group differences ran from g = 1.43 to 2.83, but baseline height differed between groups and one printed baseline test statistic does not match its own means and SDs.4

Causal-Comparative Research: What the Design Is

The University of Phoenix research group deck quotes an encyclopedia definition: a causal-comparative design seeks relationships between independent and dependent variables after an event has already occurred, by comparing two or more groups of individuals, and it is also referred to as ex post facto research.1 The deck contrasts it with two neighbours, as the table shows. The word “causal” in the name promises more than the design delivers: the deck says neither causal-comparative nor correlational research can definitively state that a true cause and effect relationship occurred,1 and the authors of our worked case wrote that causality cannot be inferred because of their cross-sectional design.4 For the ex post facto label and the reporting items that help readers judge cause, see our guide, Ex Post Facto Research: How the Design Works and What It Cannot Prove.

Swipe sideways to see every column.

DesignWho forms the groupsWhat is compared
ExperimentalThe researcher manipulates the independent variable in the experimental group.Outcomes in the manipulated and comparison groups.
Causal-comparativeThe subjects are already in groups because the event has already occurred.Two or more groups on a dependent variable.
CorrelationalOne group of individuals, with no group formed by the event.Relationships among variables within that one group.

The three rows paraphrase the deck, which in turn quotes the Encyclopedia of Research Design; we have not read the encyclopedia itself.1

How to Plan a Causal-Comparative Study: Five Decisions

The table lists five decisions to write down before data collection. The five-step frame is ours; the right-hand column shows what the worked case below did, and the source for each rule.4,2,3

Swipe sideways to see every column.

DecisionWrite downHow the worked case handled it
1. Define the groupsThe grouping rule, who applies it and when.Knee valgus was medial displacement of the knee relative to the second toe during a squat, classified as present if seen in at least two of three trials, by one trained examiner. The authors list lack of blinding and no formal reliability analysis as limitations.
2. Choose the outcomesOne measure per question, with its method.Sharpened Romberg time (eyes open and closed), Timed Up and Go, photogrammetry-based knee proprioception error and the 16-item Falls Efficacy Scale-International.
3. Check baseline balanceStandardised differences for background variables, not p-values.Table 1 compared age, height, weight and BMI with t-tests; we recompute standardised differences below.
4. Choose an adjustmentMatching, weighting, covariate adjustment or none, and why.The authors used independent t-tests with no adjustment; they restricted by inclusion criteria (men aged 60 to 70, no falls history, no regular exercise).
5. Report size and intervalEffect size with a 95% interval, plus what the design cannot show.The paper reports t, df and p. Hedges g and intervals are our computation.

Matching, ANCOVA or Propensity Scores: Choosing an Adjustment

Austin defines the propensity score as the probability of treatment assignment conditional on observed baseline covariates, and calls it a balancing score: among subjects with the same score, the measured baseline covariates have a similar distribution in the two groups.2 He describes four methods built on it: matching, stratification, inverse probability of treatment weighting and covariate adjustment.2 The same text says the key assumption is that all variables affecting both assignment and outcome have been measured, and that this assumption also underlies regression-based approaches.2 ANCOVA has its own warning. Miller and Chapman call it a widely misused approach to substantive group differences on potential covariates and say that in many cases there is no means of “correcting” or “controlling for” real group differences on a covariate.5

Swipe sideways to see every column.

ApproachWhat it doesCaution from the sources
Matching or weighting on a propensity scoreBuilds comparison sets in which measured baseline covariates are similarly distributed.Balances measured covariates only; the no-unmeasured-confounders assumption remains.
ANCOVA or regression adjustmentEstimates the group difference at a common covariate value.Miller and Chapman: often misused when groups differ on the covariate itself; read their paper before using it for intact groups.
Restriction by inclusion criteriaRemoves variation in a variable by design, as in a fixed age band.Our description of what the worked case did; it cannot adjust for variables that still differ, such as height.

Whichever route you take, Austin describes the standardised difference as a way to compare covariate balance, and notes that significance testing for balance is criticised because p-values depend on sample size.2 Our suggestion: show balance before and after adjustment in one figure, and name the confounders you could not measure.

A Real Case: Dynamic Knee Valgus and Balance in Older Men, Recomputed in R

Al-Saadi and Roshani compared older men with and without dynamic knee valgus in Urmia, Iran, in a study they describe as causal-comparative.4 The paper is CC BY 4.0, so we could redraw its numbers. Everything below comes from the printed article, and every derived number was computed by Editverse in R. Scope: 60 men aged 60 to 70, 30 per group, published 22 April 2026.

How the study was built

Swipe sideways to see every column.

ElementAs reported in the article
Sample60 men aged 60 to 70 living in Urmia, Iran, described as purposively selected in the methods and as a convenience sample in the limitations; 30 with and 30 without dynamic knee valgus.
ExclusionsNo history of falling, tricyclic medication, lower-limb fracture or dislocation, or physiotherapy or regular exercise in the previous six months.
AnalysisShapiro-Wilk for normality, then independent samples t-tests in SPSS version 26 at the 0.05 level; no covariate adjustment.
Authors’ conclusionDynamic knee valgus may be associated with poorer balance, proprioception and fear of falling; causality cannot be inferred because of the cross-sectional design.
DataAvailable from the corresponding author on reasonable request; we did not request them.

How large were the differences?

Every outcome was poorer in the valgus group and every printed p-value was below 0.001.4 We recomputed each t statistic from the printed means and SDs (n = 30 per group): all five agreed with the printed values to within 0.05, and the three non-integer degrees of freedom printed (46.50, 43.89 and 35.16) matched Welch values we recomputed of 46.47, 43.89 and 35.14. The other two rows print 58 degrees of freedom; we did not see the variance tests, so the choice of t-test variant per row is inferred from the degrees of freedom.

Forest-style chart of Hedges g for five outcomes in 30 older men with and without dynamic knee valgus: eyes-open balance 2.67, eyes-closed balance 2.80, Timed Up and Go 2.83, knee proprioception error 1.77, fear of falling 1.43, with 95% intervals all above zero
Hedges g with 95% intervals, computed in R from Table 2 of Al-Saadi and Roshani (CC BY 4.0). The paper reports t-tests only; effect sizes and intervals are Editverse computations, with the normal-approximation interval.

The five values of g run from 1.43 to 2.83 pooled SDs, with the widest interval (0.87 to 2.00) for fear of falling. Hedges g is Cohen d with a small-sample correction.3 With 30 per group, intervals are wide even when the point estimates are large.

Were the groups alike at baseline?

The text under Table 1 says none of the variables differed significantly (p > 0.05), but the table itself prints p = 0.042 for height (we recompute 0.041). Expressed as a standardised difference, the groups differ by −0.54 pooled SDs in height, well outside the 0.1 band that Austin says has been taken to indicate a negligible difference.2 Age (0.07) is inside that band and weight (−0.14) is just outside it.

Dot chart of baseline standardised differences between the valgus and non-valgus groups: age 0.07, height −0.54, weight −0.14 and BMI 0.85 from printed means and SDs but 0.34 from the printed t value, with a shaded band at plus or minus 0.1
Baseline standardised differences computed in R from Table 1 of Al-Saadi and Roshani (CC BY 4.0). The ring shows the difference implied by the printed t value.

A table that does not add up

The BMI row prints means of 25.86 (SD 0.61) and 25.39 (SD 0.49) with t = 1.31 and p = 0.196. From those means and SDs we recompute t = 3.29 and p = 0.002. The printed t of 1.31 would correspond to a pooled SD of about 1.39, more than twice the printed SDs. The printed SDs also look narrow beside the SDs of weight and height: by a first-order approximation they would require weight and height to correlate at roughly 0.89 and 0.93 within the two groups. We cannot tell whether the SDs, the t value or the p-value is misprinted, we did not have the raw data, and we did not contact the authors. The other three rows reproduce: age, height and weight recompute to t = 0.28, −2.09 and −0.52 against printed 0.29, −2.08 and −0.53. If the BMI imbalance is real, it matters, because BMI is a plausible confounder of balance performance; that reading is ours.

How much could an adjustment change the answer?

A scenario calculation shows the limit. With one covariate, an ANCOVA-adjusted group difference equals the unadjusted difference minus the within-group correlation between covariate and outcome times the covariate’s standardised difference, in outcome SD units. The table applies that rule to the height imbalance (−0.54) at assumed correlations. These are assumptions, not estimates from the data: it assumes one covariate, a linear relation and a common slope, and it is not an analysis of the raw data.

Swipe sideways to see every column.

OutcomePrinted-data gLowest g if |r| = 0.3Lowest g if |r| = 0.5Lowest g if |r| = 1.0
Static balance, eyes open2.672.512.402.13
Static balance, eyes closed2.802.632.532.26
Dynamic balance (Timed Up and Go)2.832.662.562.29
Knee proprioception error1.771.601.501.23
Fear of falling (FES-I)1.431.271.160.90

Even a correlation of 1.0 would move g by no more than 0.54. That is reassuring only for this one measured imbalance. It says nothing about confounders nobody measured, about other covariates acting together, or about whether the classification of valgus and the outcome testing were free of observer bias. The authors flag the lack of blinding as a limitation.4

What this case shows: in this convenience sample of older men, those classified as having dynamic knee valgus performed worse on all five measures, by large standardised margins. What it does not show: that valgus causes poorer balance, that the groups were balanced on every background variable, or that the results apply to women or to other regions. The authors say the same about causality and generalisability.4

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 logo

Editverse

Methodology Consultation

Group comparison, or a stronger design?

You send: your research question, the groups you plan to compare and the background variables you can measure.

You get: three 45-minute sessions with a PhD methodologist covering research design review and recommendations, sampling strategy and a written methodology framework document.

When: You need to defend the design to a committee or reviewer.

Methodology Consultation Package · $400

Book my consultation →

Editverse logo

Editverse

Statistical Analysis

Need the group comparison run and reported?

You send: your dataset, your groups and your preferred software.

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

When: You want results you can place straight into a results section.

SPSS/R/STATA Statistical Analysis · $400

Analyse my data →

Editverse logo

Editverse

Manuscript Writing

Analysis finished and the paper still unwritten?

You send: your data, your design notes and your target journal.

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

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

Manuscript Writing Services · $1,500

Write my paper →

Free Research Tools from Editverse

Frequently Asked Questions

What is causal-comparative research?

The University of Phoenix slide deck quotes an encyclopedia definition: a causal-comparative design looks for relationships between independent and dependent variables after an event has already occurred, by comparing two or more groups of individuals. The researcher does not assign the groups. In our worked case, 60 older men were grouped by whether they showed dynamic knee valgus during a squat, and their balance, proprioception and fear of falling were compared.

How does causal-comparative research differ from experimental and correlational research?

The same deck, quoting the encyclopedia, says that in experimental research the researcher manipulates the independent variable, while in causal-comparative research the subjects are already in groups because the event has already occurred. It adds that correlational research works with one group and examines variables within it, while causal-comparative research compares two or more groups.

Can causal-comparative research prove cause and effect?

The deck says that neither causal-comparative nor correlational research can definitively state that a true cause and effect relationship occurred. The authors of our worked case wrote that causality cannot be inferred because of the cross-sectional design. Austin notes that in observational studies the baseline characteristics of treated subjects often differ systematically from those of untreated subjects, so a group difference can reflect those characteristics instead of the grouping.

How do I check that my groups are comparable at baseline?

Compare background variables such as age, height and weight with standardised differences. Austin explains that a standardised difference below 0.1 has been taken to indicate a negligible difference, that there is no universally agreed threshold, and that significance tests are confounded with sample size. In our case the height difference was 0.54 pooled SDs, well outside that band.

Should I use ANCOVA, matching or propensity scores to adjust for group differences?

Austin describes four propensity score methods: matching, stratification, inverse probability weighting and covariate adjustment. All rest on the assumption of no unmeasured confounders, which he says also underlies regression-based approaches. Miller and Chapman warn that ANCOVA is widely misused for real group differences on a covariate and that in many cases there is no way to correct for them. Our suggestion is to report balance before and after any adjustment and to say which confounders could not be measured.

How do I report the size of a two-group difference?

Report an effect size with its interval, not only a p-value. Lakens explains that Cohen d is biased in small samples and that Hedges g multiplies d by a correction factor equal to 1 minus 3 divided by the quantity 4 times (n1 plus n2) minus 9. For the five outcomes in our case, Editverse computed Hedges g from the printed means and SDs and obtained values from 1.43 to 2.83, with 95% intervals between 0.87 and 3.54.

What did the knee valgus case show, and what did it get wrong?

With 30 older men per group, those with dynamic knee valgus had poorer static balance, dynamic balance and proprioception and greater fear of falling, and every printed p-value was below 0.001. Editverse recomputed every t statistic from the printed means and SDs and found agreement for all five outcomes. In the baseline table, however, the BMI t value printed as 1.31 does not match the printed means and SDs, which give 3.29. We could not tell which printed number is wrong.

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

The University of Phoenix deck quotes the encyclopedia as saying causal-comparative research is also referred to as ex post facto research. Our separate guide on the ex post facto label explains how that design works and what it cannot prove.

References

  1. Lawrence FP. Causal-Comparative Research. Slide deck, University of Phoenix Research Methodology Group; 20 April 2023. The deck quotes a QuestionPro blog post and the Encyclopedia of Research Design (Sage, 2010); we cite the deck and have not read those two sources. phoenix.edu/content/dam/edu/research/doc/2023/causal-comparative-research.pdf
  2. Austin PC. An introduction to propensity score methods for reducing the effects of confounding in observational studies. Multivariate Behav Res. 2011;46(3):399-424. doi:10.1080/00273171.2011.568786. Full text read on PubMed Central (PMC3144483). pubmed.ncbi.nlm.nih.gov/21818162/
  3. Lakens D. Calculating and reporting effect sizes to facilitate cumulative science: a practical primer for t-tests and ANOVAs. Front Psychol. 2013;4:863. doi:10.3389/fpsyg.2013.00863. Full text read on PubMed Central (PMC3840331). pubmed.ncbi.nlm.nih.gov/24324449/
  4. Al-Saadi YMR, Roshani S. Comparison of balance, proprioception, and fear of falling in older adults with and without a compensatory dynamic knee valgus. BMC Musculoskelet Disord. 2026;27:479. doi:10.1186/s12891-026-09883-x. Open access, CC BY 4.0; figures redrawn and derived numbers computed by Editverse in R. Licence read in the PMC XML (PMC13238032). pubmed.ncbi.nlm.nih.gov/42021197/
  5. Miller GA, Chapman JP. Misunderstanding analysis of covariance. J Abnorm Psychol. 2001;110(1):40-48. doi:10.1037//0021-843x.110.1.40. We read the abstract on PubMed; the full text was not available to us. pubmed.ncbi.nlm.nih.gov/11261398/

Editverse Research Team | Updated October 2026