Does Your Question Require Description, Comparison, Association, Prediction, Explanation, or Causal Evidence?

What Type of Evidence Does Your Research Question Actually Demand?

Two researchers may investigate the same educational intervention but need fundamentally different evidence. One wants to describe how frequently students use it. Another wants to compare achievement between users and nonusers. A third wants to determine whether the intervention actually causes an improvement in learning.

These questions share a topic, but they do not make the same intellectual demand. Each requires a different kind of inference, and evidence that answers one question may leave another unresolved.

The difficulty is that research questions often use similar language even when their underlying purposes differ. Words such as relationship, effect, influence, predict, and explain are sometimes treated as interchangeable, although their methodological implications can be substantial.

Before selecting a research design, you need to identify precisely what the question requires you to establish.

How Do You Identify the Evidence Type Your Question Requires?

Understanding Six Different Evidential Demands

1. Descriptive Evidence: What Exists, Happens, or Characterizes a Phenomenon?

Descriptive questions seek to characterize something without necessarily determining why it occurs. They may concern distributions, characteristics, experiences, practices, frequencies, or patterns.

Examples include:

• What proportion of university students use generative AI for academic writing?
• How frequently do teachers provide formative feedback?
• What challenges do students report when accessing online learning materials?

Descriptive evidence can be numerical, qualitative, or both. A survey may estimate the proportion of students engaging in a particular behavior, while interviews or observations may describe how that behavior manifests in a specific setting.

The appropriate evidence depends on the descriptive target. Estimating population prevalence requires attention to sampling, measurement, and denominators. Describing a set of experiences requires information sufficiently relevant and detailed to characterize those experiences credibly.

A descriptive finding does not, by itself, establish what caused the observed pattern. Knowing that 60% of surveyed students use an AI tool, for example, does not explain why they use it or whether it improves their academic performance.

Descriptive questions can nevertheless make substantial scientific contributions, especially when the phenomenon has not been adequately documented or when existing assumptions about its distribution lack empirical support.

2. Comparative Evidence: How Do Groups, Conditions, or Periods Differ?

Comparative questions ask whether characteristics or outcomes differ between two or more groups, conditions, settings, or periods.

Examples include:

• Do students enrolled in online courses report different levels of engagement from students in face-to-face courses?
• Are teachers' attitudes toward AI different across educational levels?
• How do assessment scores differ before and after a curriculum revision?

Answering these questions requires measurements or accounts that can be meaningfully compared. Researchers must consider whether the groups are defined appropriately, whether the same constructs are measured comparably, and whether differences could arise from sampling or measurement procedures.

In quantitative research, comparative evidence may involve differences in means, proportions, distributions, or other estimands. In qualitative research, comparison may involve identifying similarities and differences in experiences, practices, or interpretations across cases.

Comparison does not automatically establish causation. If students in one instructional condition outperform those in another, the difference may reflect the instruction, but it may also reflect prior achievement, selection processes, teacher characteristics, or other factors.

A comparative question becomes causal when it seeks to attribute the observed difference to a particular intervention or exposure rather than merely establish that the difference exists.

3. Associational Evidence: Are Variables Related?

Associational questions examine whether variation in one characteristic corresponds to variation in another.

Examples include:

• Is academic self-efficacy associated with students' use of digital learning resources?
• What is the relationship between teacher workload and perceived technology-related stress?
• Are patterns of AI tool usage associated with examination performance?

Such questions generally require observations of the relevant variables and an analytical strategy appropriate to their measurement scales, distributions, and study structure.

Researchers may use correlation, regression, contingency-table analysis, or other methods to examine relationships. The choice depends on the specific association being estimated, not simply on the presence of two variables.

An association can be positive, negative, nonlinear, conditional on other variables, or absent within the precision of the available evidence. Its magnitude and uncertainty are often more informative than a binary judgment based on statistical significance.

Associational evidence may also be affected by confounding, selection bias, measurement error, and reverse causation.

For example, a positive association between AI tutor usage and academic performance does not determine whether tutoring improves performance, stronger students use tutoring more often, or another factor influences both.

Association is therefore a legitimate research target in its own right, but it should not be casually converted into an explanation of causal influence.

4. Predictive Evidence: How Accurately Can an Outcome Be Anticipated?

Predictive questions concern the ability to estimate unknown or future outcomes from available information.

Examples include:

• Can early-semester learning management system activity predict which students will fail a course?
• How accurately can attendance patterns predict student dropout?
• Can prior achievement and study behaviors predict subsequent examination performance?

Predictive research requires evidence about the relationship between available predictors and the outcome being predicted, together with credible evaluation of model performance.

For quantitative prediction models, researchers may assess discrimination, calibration, prediction error, or other task-appropriate performance measures. A model intended to identify students at risk of failing a course should be evaluated on data that meaningfully represent its intended use.

Performance measured only on the data used to develop a model can be overly optimistic, particularly when the model is flexible or many candidate predictors have been considered. Appropriate validation, including temporal or external validation when feasible, helps assess whether performance generalizes beyond the development sample.

Prediction also introduces practical concerns. If a model is intended to guide educational decisions, its errors, subgroup performance, fairness, and consequences for students deserve scrutiny.

A variable may be an excellent predictor without being a useful intervention target. For instance, prior academic failure might help predict future failure, but the predictive relationship alone does not identify which educational intervention would prevent it.

5. Explanatory Evidence: Why or How Does a Phenomenon Occur?

Explanatory questions seek to account for patterns, processes, relationships, or events. They may ask why something happens, how it develops, or through what mechanisms particular conditions produce observed outcomes.

Examples include:

• Why do some teachers resist adopting AI-supported assessment systems?
• How do institutional policies shape teachers' decisions about educational technology?
• Through what processes do students develop trust in automated feedback?

Explanation is broader than a single statistical procedure. Researchers may develop explanatory accounts through qualitative process analysis, theory-informed statistical modeling, longitudinal evidence, experiments, comparative case analysis, or combinations of methods.

The evidence must be appropriate to the explanation being proposed. An account of teachers' decision-making might require interviews, observations, institutional documents, and careful analysis of how competing considerations influence practice. A proposed statistical mechanism may require evidence about temporal ordering, intermediate variables, and assumptions connecting those variables to the outcome.

Explanatory research should also distinguish participants' reported reasons from independently established causal mechanisms. When a teacher says that workload prevented adoption, the statement provides evidence about the teacher's understanding of the decision. Establishing workload as a causal determinant may require additional evidence.

Likewise, a statistical model that explains a substantial proportion of outcome variance does not necessarily establish why the outcome occurs. Statistical explanatory power and causal explanation are related in some contexts, but they are not equivalent.

Research questions concerning mechanisms and underlying processes require attention to the particular explanatory claim being advanced.

6. Causal Evidence: What Difference Does an Intervention or Exposure Make?

Causal questions ask whether an intervention, exposure, policy, or other factor produces a change in an outcome.

Examples include:

• Does providing access to an AI tutoring system improve mathematics achievement?
• Does reducing class size increase student participation?
• What is the effect of a new feedback policy on assignment completion?

A causal question involves a counterfactual comparison. For a specified population and intervention, the researcher seeks to determine how outcomes would differ under alternative exposure or intervention conditions.

The central challenge is that the same unit cannot ordinarily be observed simultaneously under both potential outcomes for the same intervention decision. Researchers must therefore use designs and assumptions that make a causal comparison credible.

Randomized experiments can provide strong causal evidence when allocation, implementation, measurement, and analysis are appropriate. Quasi-experimental approaches may also support causal inference through strategies such as regression discontinuity, difference-in-differences, or instrumental variables, provided their identifying assumptions are plausible.

Observational analyses may support causal claims under sufficiently justified assumptions, but statistical adjustment alone does not guarantee that confounding or other biases have been adequately addressed.

The causal question itself must also be sufficiently specific. Asking whether AI improves learning may be too broad to identify a meaningful effect. A more precise question would specify the intervention, comparison condition, outcome, population, and relevant period.

Researchers investigating whether an intervention is effective should therefore distinguish evidence of improvement from evidence that the intervention caused the improvement.

How the Six Evidence Types Relate to One Another

The six categories are useful distinctions, but they do not form a universal hierarchy in which causal research is inherently superior to descriptive research. Their value depends on the question.

Some relationships between the categories are especially important:

These categories can overlap. A causal study usually reports descriptive information and may compare outcomes between groups. An explanatory study may investigate causal mechanisms. A prediction study may use variables that are also examined in associational research.

However, the presence of one form of analysis does not automatically satisfy the evidential requirements of another.

Why the Wording of Your Question Matters, but Is Not Enough

Question wording can reveal intended inference, although researchers should not classify a study solely by looking for particular verbs.

The word effect, for example, generally suggests causality in ordinary research usage. Yet some manuscripts use it loosely when reporting regression coefficients from observational data. Similarly, predict may describe a genuine out-of-sample forecasting objective or simply refer to predictors in an explanatory regression model.

Rather than treating verbs as rigid labels, ask what conclusion the study is intended to justify.

If the researcher wants to know whether students with higher motivation also report greater engagement, the question is associational. If the researcher wants to know whether increasing motivation would improve engagement, it becomes causal. If the objective is to forecast which students will become disengaged, predictive performance is central.

The distinction rests on the intended inference, not the statistical software or terminology selected afterward.

Can One Research Question Require More Than One Type of Evidence?

Yes. Some research questions genuinely contain multiple evidential demands.

Consider: How does an AI-supported feedback system influence student achievement, and why do its effects differ across learners?

The first component asks for causal evidence about achievement. The second may require evidence about effect heterogeneity, contextual conditions, or explanatory mechanisms.

One study might address both through a suitable design, but the researcher must establish how the different components contribute to the overall inference.

Collecting quantitative and qualitative data does not automatically produce a coherent mixed-methods study. Integration must be justified by the research purpose, and each component must be sufficiently developed.

Sometimes it is preferable to separate an overly ambitious question into distinct questions with clearly specified evidential requirements.

Match the Evidence Type to the Claim, Not to the Prestige of a Method

A randomized trial may be appropriate for estimating a particular intervention effect, but it may be poorly suited to understanding how teachers interpret institutional expectations. A large survey may estimate prevalence effectively without revealing the mechanisms behind observed patterns.

Similarly, qualitative interviews may provide detailed explanations of participants' experiences while being unsuitable for estimating population prevalence without additional assumptions and evidence.

These are not deficiencies in the methods themselves. They reflect differences in the questions those methods are designed to answer.

The central methodological task is to select an approach whose strengths correspond to the intended inference and whose limitations do not invalidate the principal claim.

Six Different Questions About the Same AI Learning Platform

Suppose the researcher has only platform usage logs and final examination scores. Those records may support certain descriptive, comparative, or associational analyses. They might also contribute to prediction if predictors are available before the outcome and appropriate validation is possible.

They would not, without further design features and assumptions, establish the causal effect of platform access. Nor would they necessarily reveal students' reasons for discontinuing use.

The researcher should therefore avoid selecting the most ambitious question merely because the topic is important. The question should reflect the evidence that can realistically be obtained and analyzed.

Common Confusions Between Different Types of Research Evidence

How to Choose the Evidential Purpose of Your Research Question

Begin by completing the sentence: After conducting this study, I want to be able to establish that...

The ending should reveal the kind of inference you intend to make. You may want to establish how common something is, whether two groups differ, how variables are related, whether an outcome can be predicted, why a process unfolds, or what effect an intervention produces.

Next, identify what would make that conclusion credible. The question should be refined until the required evidence and the intended inference are sufficiently clear to guide research design decisions.

After identifying the evidential purpose, consider whether the question requires numerical measurements, experiential accounts, or both. That decision should follow the intended inference rather than precede it.

Also consider whether the question requires evidence about the surrounding context or observations across time. These requirements may affect the feasibility and credibility of the eventual design.

Finally, test whether the planned data can support the conclusion. If they cannot, either strengthen the design or revise the question. A narrower, defensible answer is more valuable than a broad conclusion the evidence cannot sustain.

Before Choosing a Research Design, Identify Your Evidence Type

Frequently Asked Questions About Research Evidence Types

Let the Intended Inference Determine the Evidence

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.