03 · What You Need to Know
How Explanatory and Predictive Objectives Change Prognostic Research
What Does Explanation Mean in Prognostic Research?
Explanation concerns understanding relationships among characteristics, processes, and subsequent outcomes.
For example, researchers may investigate why some patients recover more slowly than others or why certain students are more likely to withdraw from an academic program.
Such research may examine whether baseline severity, socioeconomic circumstances, treatment access, or other factors are associated with future outcomes.
However, the term explanation requires care. An adjusted association may provide evidence about a relationship without establishing the mechanism responsible for it.
When researchers seek causal explanation, they must articulate the causal question and address assumptions concerning confounding, selection, measurement, and temporal ordering.
Thus, explanatory prognostic research may be descriptive, associational, or explicitly causal, but these levels of inference should not be conflated.
What Does Prediction Mean in Prognostic Research?
Prediction concerns estimating an outcome that has not yet occurred using information available at a specified prediction time.
For example, a model might estimate the probability of hospitalization within one year or academic withdrawal during the first semester.
The primary concern is whether the predicted outcomes correspond sufficiently well to what subsequently occurs in relevant individuals.
Prediction research may use regression, machine learning, or other statistical approaches.
It does not require every predictor to have a causal relationship with the outcome.
A characteristic may be useful because it summarizes underlying risk, even if changing that characteristic would not change the outcome.
How Do Explanatory and Predictive Questions Differ?
| Feature |
Explanatory Objective |
Predictive Objective |
| Central question |
How or why is a characteristic related to the outcome? |
How accurately can the future outcome be estimated? |
| Primary focus |
Relationships, mechanisms, and interpretation |
Predicted outcomes and performance |
| Variable selection |
Guided by theory, research questions, and relevant causal or associational structure |
Guided by predictive information, availability, stability, and intended use |
| Typical output |
Association or causal effect estimate |
Risk estimate, score, or prediction function |
| Key evaluation |
Identification assumptions, bias, precision, and substantive interpretation |
Calibration, discrimination, validation, and usefulness |
| Principal limitation |
Association may not be causal |
Accurate prediction may not explain underlying mechanisms |
The distinction should be established before selecting a statistical model. The same method can be used for different purposes, but the evidence required to support the conclusions differs.
Why Does a Strong Association Not Guarantee Accurate Prediction?
A variable may have a substantial association with an outcome without distinguishing individuals who will experience that outcome from those who will not.
For example, a prognostic factor may double the relative risk of an outcome, but the absolute risks in the relevant groups may remain low and their outcome distributions may overlap substantially.
The factor can therefore be scientifically informative while providing limited individual predictive discrimination.
Likewise, an association may be estimated precisely in a large sample without materially improving a model's predictions.
Researchers should distinguish the magnitude of an association from the predictive contribution of a characteristic.
This is central to the difference between prognostic factor studies and prediction model studies.
Why Can a Useful Predictor Be Noncausal?
A predictor can contain information about an outcome without causing it.
For example, a laboratory result may reflect disease severity. It can predict deterioration even if directly changing the laboratory measurement would not alter the underlying disease process.
Similarly, prior academic performance may predict future achievement without being the sole or direct cause of that achievement.
Predictive usefulness depends on the relationship between available information and subsequent outcomes, not necessarily on whether the predictor is an intervention target.
However, noncausal predictors may be sensitive to changes in measurement practices, population composition, or institutional processes.
Researchers should consider whether their predictive relationships are likely to remain stable in the intended setting.
Why Is Variable Selection Different for Explanation and Prediction?
In explanatory research, variables should be selected according to the scientific relationship being investigated.
When causal interpretation is intended, researchers should consider confounders, mediators, colliders, and the causal structure linking exposures to outcomes.
Including a variable simply because it is statistically significant may introduce inappropriate adjustment or obscure the estimand.
In prediction research, the primary concern is whether variables improve predictive performance and are available when predictions must be made.
A variable may be included even without a causal interpretation, provided its measurement and use are appropriate.
Nevertheless, predictive variable selection should also avoid excessive complexity and outcome-driven overfitting.
The two objectives therefore require different selection rationales, even when they use the same dataset.
Should Explanatory Research Prioritize Statistical Significance?
Statistical significance may contribute to assessing an association, but it should not be the sole criterion for explanatory importance.
Researchers should examine effect magnitude, confidence intervals, theoretical coherence, measurement validity, and potential bias.
A small association may be statistically significant in a large dataset while having limited substantive importance.
Conversely, an imprecise estimate may remain compatible with several meaningful effects.
Explanatory interpretation should therefore emphasize the strength and limitations of the evidence rather than treating p-values as definitive statements about mechanisms.
Should Prediction Research Prioritize Statistical Significance?
Not as its primary performance criterion.
A predictor may improve model performance despite an imprecisely estimated individual coefficient.
Conversely, a statistically significant coefficient may add little useful information beyond other predictors.
Prediction models should be evaluated using measures appropriate to their intended outcomes and decisions.
These include calibration, discrimination, and, where relevant, decision-analytic usefulness.
Selection procedures based solely on univariable p-values can produce unstable models and should not be treated as a general prediction-model development strategy.
What Is Calibration in Predictive Prognostic Research?
Calibration concerns the agreement between predicted outcome probabilities and observed outcome frequencies.
For example, if a model assigns approximately 20% risk to a group of individuals, a well-calibrated model should produce an observed event frequency consistent with that prediction, subject to uncertainty.
A model may rank individuals reasonably well while systematically overestimating or underestimating their absolute risks.
Such a model may be problematic when decisions depend on specific risk thresholds.
Calibration should therefore be evaluated explicitly rather than inferred from regression significance or discrimination alone.
What Is Discrimination?
Discrimination concerns a model's ability to distinguish individuals with different outcome experiences.
For binary outcomes, the area under the receiver operating characteristic curve is commonly used to summarize ranking performance.
However, a high AUC does not establish that predicted probabilities are numerically accurate.
For time-to-event outcomes, appropriate discrimination measures must account for follow-up and censoring.
The importance of discrimination, calibration, and clinical usefulness depends on the prediction model's intended application.
Why Does Prediction Require Validation?
A prediction model may perform well in its development dataset because it captures sample-specific patterns that do not generalize.
Internal validation, using appropriate resampling methods, can help estimate optimism and assess model stability.
External validation evaluates the fixed model in independent data relevant to its intended use.
These procedures are central to predictive research because apparent development performance is often optimistic.
Explanatory research also benefits from replication and independent evidence, but replication of an association is not identical to validating a prediction model's calibration and discrimination.
The evaluation strategy should follow the intended scientific claim.
Can a Model Explain Outcomes Well but Predict Poorly?
Yes. A model may contain scientifically meaningful relationships yet provide limited predictive accuracy in new individuals.
For example, several factors may have credible associations with recovery while leaving substantial unexplained variation at the individual level.
In addition, a model may fit its development data well but perform poorly elsewhere because of overfitting, measurement differences, or changes in population characteristics.
Measures of explanatory fit should not automatically be interpreted as evidence of out-of-sample predictive performance.
The distinction is especially important when researchers move from publishing associations to proposing operational decision tools.
Can a Model Predict Well Without Explaining the Outcome?
Yes. A model may use complex relationships or proxy variables that provide accurate predictions without identifying the mechanisms responsible for the outcome.
For example, a machine-learning model might predict deterioration from patterns in repeated measurements while offering limited insight into the biological processes producing those patterns.
Such predictions may still be useful for certain decisions.
However, researchers should consider interpretability, measurement stability, fairness, and the consequences of acting on the predictions.
Predictive success does not justify causal explanations that the model was not designed to establish.
Does Explanation Require Causal Inference?
Not every explanatory analysis establishes a causal effect.
Researchers may explain statistical variation by describing associations or investigating theoretical relationships without identifying a causal estimand.
However, claims about what would happen if a factor were changed require causal reasoning.
For example, finding that low engagement predicts academic withdrawal does not establish that increasing an engagement score would prevent withdrawal.
The observed association may reflect prior achievement, institutional circumstances, financial constraints, or other determinants.
Researchers should distinguish statistical explanation from causal explanation and use designs appropriate to the strength of inference sought.
Why Can Prediction Fail After an Intervention Changes the System?
Prediction models are often developed using data generated under existing practices.
When institutions begin acting on predictions, those actions may change the relationship between predictors and outcomes.
For example, a university may provide additional support to students identified as having high withdrawal risk. If the support is effective, some students who would previously have withdrawn may remain enrolled.
The model's original risk estimates may then become less well calibrated under the new intervention policy.
This does not necessarily mean the intervention failed. It may mean that the process generating outcomes has changed.
Researchers should therefore monitor predictive performance after implementation and distinguish model evaluation from intervention-effect evaluation.
Can One Prognostic Study Pursue Both Explanation and Prediction?
Yes, but the objectives should be specified separately.
A study may investigate the association between a theoretically important factor and an outcome, then develop or validate a prediction model using several characteristics.
The explanatory component requires appropriate interpretation of associations and, where relevant, causal assumptions.
The predictive component requires performance assessment, overfitting control, and validation.
The same regression coefficients should not automatically be interpreted as both causal effects and validated predictive contributions.
When data are limited, pursuing both objectives may also increase analytical demands and opportunities for selective interpretation.
How Does the Research Objective Affect Sample Size?
Explanatory and predictive analyses may require different sample size calculations.
A prognostic factor study may be planned around the precision or power required to estimate a specified association.
Prediction-model development requires consideration of model complexity, anticipated predictive performance, overfitting, and precision of absolute risk estimates.
Researchers should not assume that a dataset adequate for estimating one association is also adequate for developing a multivariable prediction model.
The principles governing outcome-event requirements in prognostic studies should therefore be applied according to the intended objective.
How Should Explanation and Prediction Be Reported?
Researchers should state whether their primary objective concerns associations, causal effects, prediction, or a combination of these.
For explanatory analyses, reporting should clarify the theoretical rationale, estimand, adjustment strategy, assumptions, effect estimates, and uncertainty.
For prediction models, reporting should describe predictor availability, model development, validation, calibration, discrimination, and intended use.
STROBE provides general guidance for observational studies, while TRIPOD+AI provides reporting guidance for clinical prediction models developed or evaluated using regression or machine-learning methods.
Transparent reporting helps prevent conclusions from exceeding what the study design and analyses can support.
