Outcomes, Exposures, Predictors and Covariates

Overview

Variables acquire methodological roles from the research question. An outcome is the health state, event, experience or quantity the study seeks to describe, explain, compare or predict. An exposure is a naturally occurring, behavioural, environmental, clinical or social factor whose relationship with an outcome is of interest. An intervention is deliberately assigned or implemented. A predictor supplies information used to estimate an unknown or future outcome. A covariate is a broad term for another variable included in description, design or analysis.

These labels are not synonyms. In particular, a predictor need not be a cause, and a covariate is not automatically a confounder. Clear roles prevent researchers from adjusting for every available variable and then attaching a causal interpretation to the resulting model.

Why this matters

Variable roles determine what should be measured, when it should be measured, which variables belong in a model and how an estimate can be interpreted. A post-treatment variable may be useful for prediction but inappropriate for estimating the total effect of treatment. A strong risk marker may improve prediction without being a suitable intervention target. A variable associated with both exposure and outcome may still not be a confounder if it lies on the causal pathway or is caused by both.

Prespecifying the primary outcome and main exposure or intervention also protects against selective reporting. When many outcomes, time points and definitions are available, presenting only the most favourable result creates a misleading account of the study.

Outcome roles

Primary outcome

The primary outcome is the main quantity used to answer the principal objective. It should be clinically or scientifically meaningful, measured reliably and defined with an appropriate time point. A study may have more than one primary outcome, but multiplicity, interpretation and feasibility become more demanding.

Secondary outcome

Secondary outcomes address additional important questions but do not replace the prespecified primary outcome when the latter is unfavourable or imprecise. They may describe safety, mechanism, patient experience, implementation or longer-term consequences.

Exploratory outcome

Exploratory outcomes generate hypotheses or examine less established measures. Their interpretation should acknowledge greater uncertainty and the possibility of chance findings.

Surrogate and intermediate outcomes

A surrogate is used in place of a patient-relevant outcome because it is expected to predict clinical benefit or harm. Association with the clinical outcome is not enough to establish that intervention effects on the surrogate reliably represent intervention effects on the clinical outcome.

An intermediate outcome occurs on a pathway between exposure and final outcome. It may be important for understanding mechanism, but substituting it for the final outcome changes the question.

Composite outcome

A composite counts occurrence of any included component. Its meaning depends on the clinical importance, frequency and intervention effects of all components. The most frequent and least important component may dominate the result.

Outcome hierarchy for the illustrative study

illustrative

Table 1. Prespecified outcome roles

Role	Outcome	Purpose
Primary	Laboratory-confirmed postpartum test completion by day 112	Main service-effectiveness question
Secondary	Time from delivery to completed test	Describes timeliness
Secondary	New diagnosis requiring clinical follow-up	Describes downstream detection
Process	Appointment booked before discharge	Assesses intervention delivery
Exploratory	Participant-reported acceptability	Informs future implementation

Exposure and intervention roles

An exposure is not assigned by the investigator. Examples include air pollution, occupation, existing medication use, socioeconomic conditions and naturally occurring service pathways. Exposure definitions should capture timing, duration, intensity and changes over time.

An intervention is assigned or implemented with the intention of changing an outcome or process. Distinguish:

• allocation or intended strategy;
• actual receipt;
• adherence or fidelity;
• co-interventions; and
• contamination or crossover.

In an observational study, a clinical action can be treated as an exposure, but causal interpretation requires attention to why the action occurred. Patients who receive an intervention may differ systematically from those who do not.

Predictors and prediction

A predictor is a variable used to estimate the probability or value of an outcome for a person or unit. The purpose is predictive accuracy or decision support, not necessarily causal explanation [1,2].

A predictor may be:

• a cause of the outcome;
• a consequence of an earlier cause;
• a proxy for unmeasured information;
• a marker of disease already present; or
• a non-causal correlate that remains useful in a defined setting.

Prediction studies must specify the intended use, prediction time, target population, outcome horizon and information available at the moment of prediction. Variables measured after the prediction time cannot legitimately be used in a model intended for earlier decisions.

Covariates and their distinct purposes

Covariate is an umbrella term. More precise roles include the following.

Confounder

A confounder is a cause of, or proxy for causes of, both the exposure and outcome within the assumed causal structure, and it is not a consequence of the exposure. Uncontrolled confounding can mix the exposure effect with differences in baseline causes of the outcome.

Confounding is question-specific. A variable is not a universal confounder merely because it is associated with the exposure and outcome.

Mediator

A mediator lies on a causal pathway through which the exposure may affect the outcome. Adjustment for a mediator is generally inappropriate when estimating the total effect, but it may be relevant to a carefully defined direct-effect or mechanism question.

Effect modifier

An effect modifier identifies variation in the effect across levels of another variable on a specified scale. It is not a bias to be removed. It can be clinically important because an intervention may work differently across populations or contexts.

Precision variable

A baseline variable strongly associated with the outcome may improve precision when included appropriately, even if it is not a confounder. The reason for inclusion should be stated.

Selection variable or collider

A collider is caused by two other variables. Conditioning on it can create a non-causal association. Selection into a clinic, dataset or complete-case sample may act in this way.

Nuisance or design variable

Site, block, matching factor, calendar period or sampling stratum may need to be included because of design or data structure rather than substantive causal interest.

The same variable can have different roles

illustrative

Table 2. Role depends on the question

Variable	Question	Possible role
Risk perception after counselling	Does booking improve completion overall?	Potential mediator; do not adjust when estimating total effect
Risk perception before counselling	Does booking improve completion overall?	Baseline covariate; possibly a confounder in an observational comparison
Age	What predicts non-completion at discharge?	Candidate predictor
Age	Does age influence completion?	Exposure of interest
Clinic	Is the programme effective across clinics?	Cluster, contextual variable or effect modifier

Baseline and post-baseline variables

Time order is central. A baseline variable is measured before exposure or intervention assignment for the question being asked. A variable measured after exposure may be affected by it even when the variable represents a pre-existing concept imperfectly.

For example, measuring risk perception after an intensive counselling intervention does not produce a clean baseline confounder. The measurement may already contain an intervention effect. Researchers should define the causal time of each variable, not merely the date stored in the dataset.

Explanatory, causal and predictive models

An explanatory model seeks to understand relationships or mechanisms. A causal analysis targets the effect of an intervention or exposure under explicit assumptions. A predictive model seeks accurate predictions for new individuals or settings.

The same regression software can be used for all three, but variable selection and interpretation differ:

• causal models select adjustment variables to address the causal question;
• predictive models select available predictors to optimise and validate performance;
• descriptive models may summarise adjusted associations without supporting causal claims.

Using univariable p-values as the sole gatekeeper is unsuitable for all three purposes. Important confounders may have weak sample associations, and predictive importance requires validation rather than significance testing alone.

DECISION GUIDE

Assign a variable role before modelling

1. What is the primary question: description, causation, explanation or prediction?
2. What is the time origin and ordering of variables?
3. Which variable defines the outcome and at what horizon?
4. Is the factor assigned, naturally occurring or merely informative?
5. Could the covariate cause exposure and outcome, lie on the pathway, modify the effect or result from selection?
6. Is inclusion intended to reduce bias, improve precision, represent design or improve prediction?

Ethical and equity considerations

Variables such as ethnicity, sex, disability, income and geography can represent identity, structural exposure, social position, discrimination or access pathways. Including them without a conceptual rationale may reinforce biological or individualised explanations for structural inequity.

Interface with statistical analysis

The statistical plan should follow the variable-role map. It should identify:

• primary outcome and estimand;
• exposure or intervention contrast;
• timing and analysis population;
• prespecified confounder set for causal questions;
• planned effect-modification variables;
• predictor availability and validation strategy for prediction;
• clustering, repeated measures and design variables; and
• sensitivity analyses for measurement or structural assumptions.

Detailed modelling belongs in the Biostatistics section. The methodological safeguard is to prevent a generic list of “independent variables” from replacing a reasoned analysis plan.

Estimand, estimator and estimateMMed extension

The estimand is the target quantity defined by the question. The estimator is the statistical procedure used to estimate it, and the estimate is the numerical result. ICH E9(R1) formalises this distinction for clinical trials and emphasises alignment among objective, population, treatment conditions, endpoint and intercurrent events [3]. The distinction is useful beyond trials: choosing a regression model does not define the target effect unless the underlying question and variable roles are already clear.

Worked example

WORKED EXAMPLE

Building a variable-role map for appointment booking

Question: Among women with recent gestational diabetes, what is the total effect of arranging a postpartum test appointment before discharge, compared with usual referral, on laboratory-confirmed completion by day 112?

Outcome: completion by day 112.

Exposure/intervention: discharge appointment booking versus usual referral.

Potential baseline confounders in an observational study: clinic capacity, calendar period, maternal risk profile and prior engagement with care, if they influence both booking and completion.

Mediator: reminder receipt or reduced appointment-arrangement burden after booking.

Possible effect modifier: distance to clinic, if the booking effect differs by travel burden.

Design variables: clinic and discharge month.

Variables not adjusted in the total-effect model: post-booking risk perception and reminder receipt, because they may lie on the intervention pathway.

Prediction question: A separate model predicting non-completion at discharge could use baseline risk perception, distance, prior attendance and clinic variables. Its coefficients should not be interpreted as causal effects.

Suggested protocol or reporting wording

Common mistakes

Knowledge check

Is a covariate automatically a confounder because it is associated with both exposure and outcome?

Answer and reasoning

No. Confounding depends on the causal structure and time order. A mediator or collider can also be associated with both, yet adjusting for it may block part of the effect or introduce bias.

Can a strong predictor be a poor intervention target?

Answer and reasoning

Yes. Prediction does not require causation. The predictor may be a marker or consequence of disease, or a proxy for another process. Intervention requires evidence about what happens when the factor is changed.

Why should the primary outcome be specified before analysing the data?

Answer and reasoning

Prespecification limits selective emphasis among many outcomes, definitions and time points. It also aligns sample size, measurement procedures and the principal interpretation with the study objective.

A variable is measured after treatment begins. Can it be treated as a baseline confounder?

Answer and reasoning

Usually not without strong justification. Treatment may already have influenced the variable or its measurement. The team must establish the variable’s causal timing and consider whether it is post-exposure.

Key take-home messages

SUMMARY

Key take-home messages

• Variable roles arise from the question, time order and intended inference.
• Outcomes require a prespecified hierarchy, measure and time point.
• Predictors need not be causes, and covariates are not automatically confounders.
• Mediators, effect modifiers and colliders require different handling.
• Causal, explanatory and predictive models may use similar software but answer different questions.

References

1. Steyerberg EW. Clinical prediction models: a practical approach to development, validation, and updating. 2nd ed. Cham: Springer; 2019.
2. Moons KGM, Altman DG, Reitsma JB, Ioannidis JPA, Macaskill P, Steyerberg EW, et al. Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): explanation and elaboration. Ann Intern Med. 2015;162(1):W1–W73. https://doi.org/10.7326/M14-0698
3. International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use. ICH E9(R1): addendum on estimands and sensitivity analysis in clinical trials to the guideline on statistical principles for clinical trials. Geneva: ICH; 2019. https://database.ich.org/sites/default/files/E9-R1_Step4_Guideline_2019_1203.pdf