Observational Studies

OPBSResearch

Must Know

• Observational studies describe outcomes and predictors (risk factors), then study their associations to infer prognosis.
• Prospective and retrospective cohorts are the commonest observational designs.
• Cross-sectional studies give prevalence only, never incidence.
• Prevalence is the proportion with a condition at one time point. Incidence is the proportion newly developing it over a set period.
• Prevalence equals the prior probability of disease before workup.
• Age and sex are never dependent variables, so they always act as predictors.
• Cohort studies are the observational design that can measure incidence.
• Cohorts record predictors before the outcome, so causal inference is stronger than in other observational designs.
• Not knowing the outcome when measuring predictors reduces bias in prospective cohorts.
• Recall bias is much lower in prospective than retrospective studies.
• Confounders relate to both predictor and outcome and can produce false associations.
• Known confounders can be modelled if the sample is large enough. Unknown confounders cannot.
• Controlling unknown confounders is the main statistical advantage of an RCT.
• Case-control studies are mainly retrospective and suit rare outcomes.
• The primary result of a case-control study is an odds ratio.

Comparing Observational Designs

Design	Strengths	Weaknesses
Cross-sectional	Quick, cheap. No loss to follow-up. Useful baseline data inside RCTs and cohorts	Poor for causation. Impractical for rare diseases
Case series	Describes disease features in patients who all have the condition	Cannot answer comparative questions
Prospective cohort	Prognosis and predictors. Incidence. Better measurement	Expensive, slow. Inefficient for rare outcomes. Confounding
Retrospective cohort	Faster and cheaper than prospective	Less control of sampling and follow-up. Poorer baseline data. Confounding harder to adjust
Case-control	Rare outcomes. Cheap. Generates hypotheses	One outcome only. No incidence or prevalence. More selection and recall bias

Design Choices

• Inclusion criteria that are too strict slow enrolment. Criteria too loose make the sample too heterogeneous.
• The STROBE diagram reports patients screened, excluded with reasons, and enrolled.
• Multicentre studies raise enrolment and generalisability but are harder to run.
• Single-centre studies are easier to run but slower to enrol.
• To refine an already successful treatment, use a single-centre prospective cohort.
• For a rare condition or outcome, use a case-control design.
• A nested case-control study draws cases and controls from an existing cohort at risk.
• Most cohort analyses need multivariate modelling, ideally with a statistician.

Numbers to Remember

• Sher 1995 MRI study of 96 asymptomatic shoulders found cuff tears in 54 percent of those over 60.
• MOON enrolled over 3,500 ACL reconstructions with follow-up at 2, 6 and 10 years.
• MOON follow-up above 80 percent allowed a level I prognostic cohort.
• Prospective cohort costs rise when follow-up above 70 percent is sought.
• McCullough MOON study found 63 percent of high school and 69 percent of college players returned to football.

More on the example studies

• MOON predictors of worse 10-year outcome were low baseline PROMs, higher BMI, smoking, prior meniscal surgery, lateral meniscectomy, revision ACLR, grade 3 or 4 cartilage defects and later ipsilateral surgery.
• MOON PROMs improved by 2 years and then held steady at 6 and 10 years.
• MARS found allograft revision ACLR had worse 2-year PROMs and more graft rerupture than autograft.
• In McCullough, half of those not returning to football cited fear of reinjury.
• Beynnon 2015 matched 88 noncontact ACL tears with 88 controls on MRI. Small ACL volume in males and narrow notch with steep tibial cartilage slope in females raised risk.
• Westermann found combined ACL and MCL injuries managed surgically had lower 2-year PROMs and more revision for arthrofibrosis.
• Framingham is the best known prospective cohort outside orthopaedics.

Exam Traps

• Case-control studies cannot give incidence or prevalence because the investigator sets the case to control ratio.
• Cohort studies can examine many outcomes. Case-control studies examine one.
• Bias is a bigger problem in case-control than in cohort studies.
• Loss to follow-up threatens cohorts and RCTs but not cross-sectional studies.
• Retrospective cohorts carry a higher risk of confounding and missing data than prospective cohorts.

Related pages. Level of Evidence

Recent Review. Multivariable Versus Multivariate Models

• Multivariable describes a model with several independent variables (predictors or covariables) and one outcome.
• Multivariate describes a model with several dependent variables (outcomes) analysed together.
• Logistic regression of one outcome adjusted for age, sex and BMI is multivariable.
• Modelling mortality and infection together as joint outcomes is multivariate. True multivariate models are rare in orthopaedics.
• A confounder is linked to the exposure, linked to the outcome and not on the causal pathway between them.
• Observational studies adjust for confounding by entering several risk factors into one multivariable model.
• In 600 articles from JBJS, AJSM, BJJ and CORR, 75 used multivariate. Of these, 92 percent meant multivariable.
• Only 2 of 75 studies (2.7 percent) used multivariate correctly. Multivariable was always used correctly.

More on the terminology study

• Articles were the first 50 per journal in 2010, 2015 and 2020.
• Of 120 articles using either term, 61 percent were retrospective cohorts and level III evidence was the most common.
• In 2010, 90 percent of articles using either term chose multivariate.
• Hidalgo and Goodman set out the distinction for the wider statistical community in 2013.
• For complex models, the clearest approach is to list the outcome and every covariable in the methods.
• The authors suggest editors and reviewers screen submissions for misused statistical terms.

Kolin DA, Landy DC, Watkins A, Fabricant PD, Chalmers BP. The terms “multivariate” and “multivariable” are used incorrectly and interchangeably in orthopaedic publications. Should we care about the distinction? J Bone Joint Surg Am. 2023;105(11):896-9. PubMed

Recent Review. Causal Language in Observational Research

• Causation can in general only be shown with randomisation. Observational studies cannot remove selection bias or unmeasured confounding.
• Association means a statistically significant relationship between two variables where one need not cause the other.
• AMA journals bar causal wording such as effect of, impacts, increases or leads to, except in well powered RCTs.
• Unadjusted observational analyses support only descriptive statements.
• Adjusted analyses (matching, multivariable regression, propensity scores) control measured confounders and can at most suggest causation.
• Well executed RCTs are the level at which causation can be credibly claimed.
• Instrumental variables, regression discontinuity and Mendelian randomisation can estimate causal effects from observational data under set assumptions.
• Observational data linked early hip fracture surgery with lower mortality. HIP ATTACK found no such benefit, implying residual confounding.
• Of 400 observational studies in top orthopaedic journals, 60 percent used causal language, most often in abstract conclusions.

More on causal language

• Causal wording appeared in 48 percent of abstract conclusions, 37 percent of titles and 10 percent of abstract results.
• Inter rater concordance on titles was 99.3 percent, Cohen kappa 0.984.
• Raising HDL cholesterol was long thought protective from observational data until RCTs showed no benefit.
• Nearly half of orthopaedic trainees could not assess study design and 83 percent could not judge strength of association from an odds ratio.
• Studies using large registries and administrative datasets are growing by more than 35 percent a year.
• Social factors with no biological component cannot by themselves cause biological complications, so a mechanism should be discussed.
• The Bradford Hill criteria frame causal reasoning across a body of evidence.

Varady NH, Feroe AG, Fontana MA, Chen AF. Causal language in observational orthopaedic research. J Bone Joint Surg Am. 2021;103(19):e76. PubMed

Source

Cantrell WA, Spindler KP. Observational Studies. In. Aaron RK, editor. Orthopaedic Basic Science. Foundations of Clinical Practice. 5th ed. Philadelphia, Wolters Kluwer, 2021.

Practise Type X questions on this topic

Reviewed by Professor Phong Tran, Head of Orthopaedic Surgery, Western Health. Last updated 17 October 2026.