Descriptive, Analytical, Explanatory and Predictive Research

Overview

Health research can serve different purposes. A study may describe the distribution of a health problem, examine relationships between variables, seek an explanation for why an outcome occurs, or predict an outcome for a new individual. These purposes overlap in practice but are not interchangeable.

Confusion commonly arises because the same dataset and statistical model can be used for different aims. A regression model may describe adjusted associations, estimate a causal contrast under strong assumptions, or generate predictions. The mathematical form alone does not determine the research purpose. The purpose should be declared first because it affects the question, design, variable selection, validation, analysis and interpretation [1,2].

Why this matters

A descriptive study can be excellent when decision-makers need to know the magnitude and distribution of a problem. It becomes weak only when it is interpreted as proving why the problem occurs. An analytical association study may identify important patterns but still be unable to support a causal explanation. A model that explains past data well may predict new patients poorly, while a highly predictive model may provide little causal insight.

Using imprecise labels such as “to determine factors” can hide these differences. The researcher should state whether the purpose is to describe, compare, explain or predict and define the intended output.

Descriptive research

Descriptive research characterises the occurrence, distribution or features of a health phenomenon in a defined population, place and time. It answers questions such as what, how much, who, where and when.

Examples include:

• prevalence of uncontrolled hypertension among clinic attendees;
• incidence of catheter-associated infection in an intensive-care unit;
• distribution of waiting time by clinic and month;
• characteristics of patients enrolled in a registry; and
• themes describing how participants experienced a service.

Descriptive epidemiology requires a clearly defined target population, sampling or case-ascertainment pathway, outcome definition, denominator, period and relevant subgroups. Person, place and time are not decorative table headings; they define the population to which the description refers [3].

A descriptive objective should not be undervalued or disguised as hypothesis testing. Good description is often the first requirement for surveillance, planning, equity assessment and hypothesis generation.

Analytical research

Analytical research examines differences or relationships. It asks whether an outcome varies across exposures, groups, interventions, time periods or characteristics. The term is broad and does not itself establish causal intent.

Examples include:

• association between shift work and burnout;
• comparison of infection rates between wards;
• relationship between treatment adherence and glycated haemoglobin; and
• difference in patient satisfaction before and after a service redesign.

Analytical studies require attention to temporality, selection, measurement, confounding, clustering and missing data. An observed association may result from a causal effect, reverse causation, common causes, selection processes or measurement error.

Explanatory research

Explanatory research seeks to understand why or how an outcome occurs. In quantitative health research, this may involve estimating causal effects or evaluating mechanisms. In qualitative research, explanation may focus on processes, meanings and contextual mechanisms. The word “explanatory” should therefore be accompanied by a clear account of the intended explanation.

A causal question usually compares potential outcomes under different interventions or exposures. For example: among eligible patients, what would the 30-day readmission risk be if all received a structured discharge programme compared with if all received usual care? This requires a credible comparison, correct time order, control of confounding where relevant, consistent outcome measurement and assumptions about adherence and missing data.

An explanatory variable should not be selected solely because it was statistically significant in the available dataset. Variable roles should follow subject-matter knowledge and the causal structure. Mediators, confounders and consequences of the outcome should not be treated as interchangeable covariates.

Predictive research

Predictive research estimates an unknown or future outcome for a new person, setting or time point. The goal is accurate and useful prediction, not necessarily explanation of the causal process [1].

Examples include predicting:

• 30-day readmission at hospital discharge;
• malignancy from clinical and imaging features;
• deterioration in an inpatient within 24 hours; or
• treatment response for a new patient.

Prediction requires a clearly defined intended use, target population, prediction time, outcome horizon, available predictors and decision context. Model performance should be evaluated on data not used to fit the model, with attention to calibration, discrimination and clinical usefulness. Current reporting guidance covers prediction models based on regression or machine-learning methods [4].

A predictor need not be causal. Previous admission may improve readmission prediction even if changing the historical fact is impossible. Conversely, a causal factor may add little incremental predictive value if its information is already captured by other predictors.

Comparing the purposes

Aligning research purpose and output

illustrative

Table 1. Four research purposes using 30-day readmission

Purpose	Example question	Primary output	Key methodological emphasis
Descriptive	What proportion of eligible patients are readmitted within 30 days?	Prevalence or risk with uncertainty	Population, denominator, outcome ascertainment, time period
Analytical	Which measured characteristics are associated with readmission?	Crude or adjusted associations	Temporality, confounding, measurement, selection
Explanatory	What is the effect of a discharge programme on readmission?	Causal contrast under explicit assumptions	Counterfactual comparison, allocation, adherence, bias
Predictive	What is a new patient’s probability of readmission at discharge?	Individual predicted probability and performance	Target use, predictors available at prediction time, validation, calibration

Research questions and objective verbs

Verbs can help but are not sufficient. “To determine” and “to assess” are too vague unless the object and intended inference are clear.

More explicit formulations include:

• Describe: To estimate the 30-day readmission risk among adults discharged with heart failure.
• Analyse association: To estimate the association between discharge-day sodium concentration and 30-day readmission.
• Explain causally: To estimate the effect of a structured discharge programme on 30-day readmission compared with usual care.
• Predict: To develop and internally validate a model that predicts 30-day readmission using information available at discharge.

Design and measurement implications

Descriptive

Define the case or outcome, population denominator, place and time. Sampling should support the claimed population description. Standardisation across sites and periods is essential when comparing distributions.

Analytical

Define exposure and outcome roles, comparison groups and time order. Decide whether the aim is exploratory association or a pre-specified contrast. Anticipate confounding and selection mechanisms.

Explanatory

Specify the intervention or exposure contrast, target population, time zero, follow-up, outcome and causal estimand. Design decisions should emulate a credible comparison. Randomisation is powerful where feasible, but explanatory observational research also requires explicit causal assumptions.

Predictive

Define when prediction occurs, what information is available at that moment, the prediction horizon and intended decision. Avoid predictor information that would only become available after the outcome or after the decision point. Evaluate performance in new data.

DECISION GUIDE

Clarify the purpose before selecting the analysis

1. Is the main output a population description, association, causal contrast or prediction?
2. What is the target population and time frame?
3. Which variables are outcomes, exposures, interventions or predictors?
4. Is the goal interpretation of coefficients or performance on new cases?
5. What design and validation are required for the intended claim?

A study may have multiple purposes

A cohort may first describe incidence, then analyse associations and later support prediction modelling. Multiple purposes are acceptable when they are explicit and adequately planned. Problems arise when exploratory associations are retrospectively reframed as causal explanations or when a model developed for association is advertised as a clinical prediction tool without validation.

The primary purpose should determine the main design and sample-size rationale. Secondary purposes may require additional measurements, larger samples or separate validation datasets.

Interface with statistical analysis

Descriptive analysis emphasises valid denominators, suitable summaries and uncertainty. Analytical and explanatory work may use similar regression models, but covariate selection and interpretation differ. Prediction prioritises out-of-sample performance, calibration and validation rather than coefficient p-values alone.

A model with a high coefficient of determination in the development data may still predict poorly in a new setting. Likewise, a statistically significant adjusted association does not prove that intervening on the predictor will improve the outcome.

Estimand, mechanism and prediction targetMMed extension

An estimand is the precise quantity a study intends to estimate. In explanatory work, it may be an average treatment effect under specified strategies. In predictive work, the target may be an individual probability at a defined prediction time and horizon. Stating the estimand or prediction target exposes hidden differences that a generic objective such as “identify factors” would conceal.

Worked example

WORKED EXAMPLE

Four studies from one heart-failure dataset

Descriptive study: Estimate 30-day readmission risk among all eligible discharges in 2026, overall and by hospital. The main threats are incomplete follow-up, inconsistent readmission capture and unclear denominators.

Analytical study: Examine associations between discharge characteristics and readmission. The findings identify patterns but are interpreted as associations because residual confounding and treatment selection remain.

Explanatory study: Compare a structured discharge programme with usual care using a design that establishes a common time zero and credible comparison. The protocol pre-specifies the causal contrast, follow-up, adherence strategy and outcome ascertainment.

Predictive study: Develop a discharge-time risk model using only information available before the patient leaves hospital. The model is internally validated and later externally validated in other hospitals. Performance and clinical utility are assessed before implementation.

Although all four studies may use readmission as the outcome, they require different questions, variable roles and conclusions.

Suggested protocol wording

Appraising alignment

APPRAISAL CHECKLIST

Does the report match its purpose?

• Is the research purpose stated explicitly rather than implied by the analysis?
• Does the design support the claimed type of inference?
• Are variable roles appropriate for description, explanation or prediction?
• Are validation and performance reported for predictive claims?
• Do the conclusions remain within association when causal assumptions are not justified?

Common mistakes

Knowledge check

A study estimates the prevalence of burnout among medical officers by hospital and specialty. What is its primary purpose?

Answer and reasoning

It is descriptive. It estimates the distribution of burnout in a defined population. Comparisons across subgroups may be included, but the core objective remains description unless the study explicitly seeks an explanatory or causal inference.

A variable is a strong predictor of mortality. Does that prove it causes mortality?

Answer and reasoning

No. Prediction depends on information that forecasts the outcome, whereas causation concerns what would happen if the exposure were changed. A predictor may reflect disease severity, prior consequences or common causes without being a modifiable cause.

Why can a model with good fit in its development dataset still be unsuitable for clinical prediction?

Answer and reasoning

It may overfit the development data, be poorly calibrated, use unavailable predictors or fail in a new population. Prediction requires out-of-sample validation and evaluation in the intended use setting.

What additional information is needed beyond the objective 'to determine factors associated with readmission'?

Answer and reasoning

The study should specify whether the aim is exploratory association, causal explanation or prediction; define the population, outcome time horizon and variable roles; and state the intended estimate or performance measure.

Key take-home messages

SUMMARY

Key take-home messages

• Descriptive, analytical, explanatory and predictive purposes address different questions.
• The statistical model alone does not determine the purpose or interpretation.
• Association is not equivalent to causal explanation.
• Prediction requires a defined use, prediction time, validation and performance assessment.
• State the intended output before selecting design, variables and analysis.

References

1. Shmueli G. To explain or to predict? Stat Sci. 2010;25(3):289–310. https://doi.org/10.1214/10-STS330
2. Breiman L. Statistical modeling: the two cultures. Stat Sci. 2001;16(3):199–231. https://doi.org/10.1214/ss/1009213726
3. Lesko CR, Fox MP, Edwards JK. A framework for descriptive epidemiology. Am J Epidemiol. 2022;191(12):2063–2070. https://doi.org/10.1093/aje/kwac115
4. Collins GS, Moons KGM, Dhiman P, Riley RD, Beam AL, Van Calster B, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378. https://doi.org/10.1136/bmj-2023-078378