Prediction Versus Explanation in Clinical Models

Prediction and explanation are different jobs. Prediction asks what is likely to happen for a patient, sample, image, or encounter. Explanation asks what produces an outcome or what would change if an intervention changed; the same dataset and equation can be used in both conversations, but that does not make the answers interchangeable.

This distinction matters because a reliable predictor need not identify a cause. Conversely, a well-designed causal study can estimate the effect of an intervention while leaving substantial uncertainty about which individual patient will experience an event, and clinical models should be judged against the question they were designed to answer.

Start with the verb#

Words often tell you the actual task. Predict, classify, detect, forecast, and estimate risk usually signal a prediction question. Prevent, cause, treat, explain, mediate, and change usually signal a causal question.

Consider two questions about hospital readmission. “Which patients are likely to return within 30 days?” is predictive. “Would a home-visit program reduce 30-day readmission compared with usual discharge planning?” is causal. A risk model may help allocate the program, but it does not estimate the program's effect; a trial of the program may estimate an average effect, but it does not automatically provide accurate individual risk.

The questions can work together; prediction may identify when an action is needed, causal evidence may identify which action is effective, and implementation research may show whether the action can be delivered reliably. Combining these stages is useful; collapsing them is not.

What prediction is trying to optimize#

A prediction model maps information available at a defined time to an outcome that is not yet known at that time, and examples include estimating the probability of sepsis in the next six hours, classifying an image, or forecasting one-year cardiovascular risk.

Predictive performance must be assessed in patients who did not determine the fitted rule. The model's intended use specifies the population, setting, time zero, outcome, horizon, and available inputs. Without that frame, a high score tells you nothing stable.

Good prediction is more than ranking. Discrimination asks whether higher-risk patients tend to have more events. Calibration asks whether predicted probabilities agree with observed frequencies. A model can distinguish higher from lower risk while consistently overstating absolute risk. Threshold measures and decision analysis ask what happens when predictions trigger an action.

Transport matters because predictor meanings and outcome rates change. A medication order can signal severity at one institution and routine prevention at another. A laboratory value can be measured with a different assay. A code can reflect a local billing practice. External validation tests whether performance holds across a boundary relevant to intended use.

What explanation is trying to estimate#

Causal explanation needs a contrast. “What would the outcome have been under intervention A compared with intervention B for a defined population and period?” is more precise than asking whether A affects health.

The target can be an intention-to-treat effect, a per-protocol effect, an average treatment effect, an effect among treated people, or another estimand. Each requires assumptions and data aligned to that target. Randomization can make treatment groups comparable at baseline on average, but protocol deviations, missing outcomes, treatment switching, and measurement still require attention.

Observational causal analyses need a defensible design. Investigators identify common causes of treatment and outcome, define time zero consistently, avoid conditioning on consequences of treatment, and address selection and measurement. Statistical adjustment cannot create information on an unmeasured confounder.

A causal estimate can be modestly predictive. If many factors determine an outcome, learning that one treatment changes average risk may explain only a fraction of individual variation. That does not make the causal result weak. It answers a different question.

A predictor does not become a cause#

An umbrella predicts rain because umbrellas are carried when rain is expected. Removing umbrellas would not stop the weather. Clinical data contain many analogous signals.

A clinician's order for an urgent test may strongly predict deterioration because the order encodes concern. Preventing the order would not protect the patient. A billing code can predict outcome through care intensity. A neighborhood measure may proxy access, resources, structural conditions, or measurement practices without specifying which pathway is causal.

Predictive models can use proxies if they improve performance and their use is lawful, ethical, and safe. Yet using a proxy can reproduce inequity or fail when practice changes. Documentation should make clear what a variable measures operationally, not just its column label.

The reverse is also possible. A causal factor may add little predictive value. Smoking has causal effects on several diseases, but adding smoking to a short-horizon model dominated by current physiology may barely change discrimination. Predictive contribution does not measure causal importance.

Feature importance is not intervention evidence#

Model explanations such as coefficients, permutation importance, saliency maps, and Shapley-style attributions can reveal how a fitted model's output varies with inputs; they can help detect unexpected reliance, debug systems, and communicate behavior.

They do not normally identify what would happen if you changed the input. Features are correlated. The model may respond outside the data distribution. A measurement may be a consequence of disease rather than a cause. Altering a field in software is not the same intervention as changing the underlying biological or social process.

A positive coefficient is not automatically a causal effect either. Its interpretation depends on model form, coding, included variables, selection, and assumptions. Adjusting for a mediator can remove part of an effect, adjusting for a collider can create association, and omitting a confounder can distort a causal estimate.

The term “explainable AI” should therefore be read narrowly. It can mean an account of model behavior or an interface that helps users understand output. It should not be treated as proof of biological mechanism, fairness, or benefit.

Variable selection follows the question#

For prediction, a variable is useful if it improves performance in the intended setting, is available at prediction time, is measured reliably, and can be used appropriately. A noncausal marker can qualify. A variable should not be included if it leaks future information or makes deployment infeasible.

For causal estimation, variable choice is guided by the assumed causal structure. Common causes of treatment and outcome often need control. Instruments, mediators, and colliders require different treatment. Adding every available variable can increase rather than reduce bias.

This difference explains why generic advice to “control for everything” is unsafe. In a predictive model, broad features may be useful but can overfit. In a causal model, automatic selection based on association can omit needed confounders or include harmful controls.

Domain knowledge is needed in both tasks. For prediction, it defines timing, feasible inputs, clinically relevant errors, and transport risks. For causality, it informs the graph of plausible relations and whether identification assumptions are credible.

Time zero prevents impossible models#

Every clinical question needs a moment when eligibility is established and information becomes available. For prediction, the model can only use information known by that moment. For causal estimation, treatment strategies and follow-up should begin from a consistent time.

Violating time order creates misleading accuracy. A discharge diagnosis cannot predict a decision made on admission. A culture result finalized two days later cannot support a bedside score claimed to run at presentation, and even a timestamp before the outcome can leak information if it was generated because the outcome process had already begun. In causal analyses, inconsistent time zero can instead create immortal-time bias or compare people at different disease stages, which is why defining eligibility, assignment, and follow-up together helps align an observational study with the trial it is trying to emulate.

Different validation questions#

Predictive validation asks whether a fixed model performs in new data. Internal validation estimates optimism. External validation tests a new time, place, or population. Evaluation should include calibration, discrimination, uncertainty, subgroup results, and consequences at proposed thresholds.

Causal validation cannot rely on prediction scores alone. Investigators examine exchangeability, positivity, consistency, measurement, missingness, and whether the analysis estimates the planned contrast. Negative controls, sensitivity analyses, triangulation across designs, and replication can probe assumptions, but none proves that all bias is absent.

Randomized trials provide strong design-based protection against baseline confounding when conducted and analyzed appropriately. They do not guarantee broad transport. Participants, comparators, adherence, and care settings can differ from routine use.

A model can include both components. A heterogeneous treatment-effect model predicts how outcomes differ under alternative treatments. This is a causal prediction task, not ordinary risk prediction. It needs data that identify treatment effects, adequate variation, careful validation, and uncertainty that reflects the harder target.

Accuracy is not clinical usefulness#

Even a well-calibrated risk estimate may not improve care. An output needs a user, a decision, and an action that can change the pathway. If every patient receives the same appropriate care, an additional risk ranking may add workload without benefit.

The human-model team also changes performance. Users may defer to a score, selectively accept recommendations, enter data differently, or develop alert fatigue. DECIDE-AI focuses reporting on early live clinical evaluation, including workflow and human factors. International Good Machine Learning Practice principles likewise emphasize testing under clinically relevant conditions and monitoring after deployment.

Benefit should be demonstrated at the level claimed. A lower prediction error is a technical outcome. Faster review is a process outcome. Fewer complications is a patient outcome. Each requires an appropriate comparator and design.

Four common category errors#

First, “the model predicted it, so the feature caused it.” Prediction establishes neither direction nor mechanism.

Second, “the coefficient was not statistically different from zero, so the variable cannot predict.” A causal parameter test and out-of-sample predictive contribution are not the same evaluation.

Third, “the model is transparent, so the recommendation is safe.” A short score can be biased, poorly calibrated, or linked to a harmful action.

Fourth, “the trial showed benefit on average, so the model can identify who benefits.” Average effects do not establish reliable individualized treatment effects. These errors are avoidable when the protocol names the question before analysis and reports changes to it.

A review checklist for either model#

Read the objective and rewrite it as one sentence with a population, time zero, input or intervention, outcome, horizon, and purpose. If you cannot, the model's target is unclear.

For prediction, ask whether every input is available at use, whether evaluation data were isolated from development, and whether calibration and decision consequences are reported. The overfitting guide explains why a protected evaluation matters.

For causal claims, ask what alternatives are compared, how confounding and selection are handled, and which assumptions connect the observed data to the counterfactual contrast. The confounding-by-indication guide develops one common clinical problem.

For both, inspect missing data, subgroup uncertainty, transport, human factors, and monitoring. The site's research overview places these checks within a wider evidence-appraisal framework.

The clearest model report does not promise to answer every question: it says which question was asked, why the design can answer it, what uncertainty remains, and which decision the result can responsibly support.

References#

1. Shmueli: To Explain or to Predict?
2. TRIPOD+AI statement
3. PROBAST+AI risk-of-bias and applicability tool
4. FDA summary of Good Machine Learning Practice guiding principles
5. DECIDE-AI reporting guideline
6. Hernan and Robins: What If

This article is educational and does not establish that any model is suitable for clinical use. Model evaluation should be matched to its intended population, decision, jurisdiction, and workflow.