Experimental research is a quantitative method in which the researcher deliberately manipulates one variable (the independent variable, or IV), measures its effect on another (the dependent variable, or DV), and controls extraneous variables so that any change in the DV can be attributed to the IV. It is the only design that lets you establish a genuine cause-and-effect relationship rather than a mere association.
Use experimental research when your research question is causal — “Does X cause a change in Y?” — and when you can ethically and practically assign participants to conditions and intervene. If you can only observe variables as they naturally occur, you need correlational research instead.
What is experimental research?
Experimental research sits at the top of the evidence hierarchy precisely because of one feature: manipulation under control. In a correlational study you watch two variables move together and infer nothing about direction or cause. In an experiment you actively change the independent variable, hold everything else as constant as you can, and observe what happens to the dependent variable. Because you — not nature — decided who received the treatment, a difference in outcomes between conditions can reasonably be credited to the treatment itself.
Three logical conditions must be met before you can claim causation, and a well-built experiment is engineered to satisfy all three: (1) covariation — the IV and DV change together; (2) temporal precedence — the cause comes before the effect, which manipulation guarantees; and (3) no plausible alternative explanation — rival causes (confounders) are ruled out through control and randomisation. The third condition is the hardest to meet and is what separates a strong experiment from a weak one.
This is exactly why experimental research outranks observational designs such as correlational research when the goal is causal explanation. A correlation between, say, hours of revision and exam performance is consistent with several stories: revision raises marks, more able students simply revise more, or a third factor (motivation) drives both. An experiment that randomly assigns students to different amounts of structured revision dissolves that ambiguity, because the only systematic difference between the groups is the one the researcher created.
Key elements of an experiment
Whatever the discipline — psychology, business, education, health or sociology — every true experiment is assembled from the same building blocks:
- Independent variable (IV): the factor you deliberately manipulate (e.g. a new teaching method, a drug dose, an advertising message). Its different values are the conditions or levels.
- Dependent variable (DV): the outcome you measure and expect to change (e.g. exam score, recovery time, purchase intention). For more on classifying these, see our guide to types of variables.
- Experimental group: the participants who receive the treatment (the active level of the IV).
- Control group: the participants who receive no treatment, a placebo, or business-as-usual — the baseline against which the effect is judged.
- Randomisation: allocating participants to groups by chance so that, on average, the groups are equivalent at the outset on every characteristic, measured or not.
- Control of confounders: the procedures (holding variables constant, counterbalancing, blinding) that stop confounding variables from offering a rival explanation for your results.
The single most powerful of these is randomisation. By distributing both known and unknown extraneous variables evenly across groups, random allocation neutralises confounders you have not even thought to measure — something statistical control alone can never do.
“Random assignment of subjects to treatment groups distributes the characteristics of subjects among the groups in such a way that systematic differences are eliminated, leaving only chance differences.” (Source: Campbell & Stanley, 1963)
Types of experimental research
Experiments are usually classified by how rigorously they meet the conditions for causal inference — specifically, whether they include a control group and whether participants are randomly assigned. This gives three families.
- True experimental research — has both a control group and random assignment. The gold standard for causal claims (e.g. a randomised controlled trial of a new drug).
- Quasi-experimental research — has a comparison group but no random assignment, often because groups are pre-existing (e.g. comparing two schools’ outcomes after one adopts a new curriculum). Useful when randomisation is impossible, but vulnerable to selection bias.
- Pre-experimental research — lacks a control group, random assignment, or both (e.g. a single group measured before and after an intervention). The weakest for causal inference and best treated as exploratory.
| Type | Control group? | Random assignment? | Causal strength | Typical use |
|---|---|---|---|---|
| True experimental | Yes | Yes | Strongest — supports causal claims | Drug trials, controlled lab studies |
| Quasi-experimental | Yes (comparison group) | No | Moderate — selection bias risk | Field/policy studies, intact classes |
| Pre-experimental | Often none | No | Weakest — exploratory only | Pilot studies, early-stage testing |
Common experimental designs
Within those families, the specific design describes how conditions are arranged and how participants flow through them.
Between-subjects vs within-subjects
- Between-subjects (independent groups): each participant experiences only one condition. Avoids carry-over effects but needs more participants and risks group differences (managed by randomisation).
- Within-subjects (repeated measures): the same participants experience every condition. Highly efficient and controls for individual differences, but vulnerable to order and practice effects — usually addressed by counterbalancing the order of conditions.
Named designs you will meet
- Randomised controlled trial (RCT): participants are randomly allocated to treatment or control and the DV is compared. The benchmark true-experimental design (illustrated below).
- Pretest–posttest control-group design: both groups are measured before and after; comparing the change in each group isolates the treatment effect from baseline differences.
- Factorial design: two or more IVs are manipulated simultaneously (e.g. a 2×2 design), letting you test both main effects and their interaction — whether the effect of one IV depends on the level of another.
- Solomon four-group design: combines pretested and non-pretested versions of the experimental and control groups to detect (and remove) any sensitising effect of the pretest itself.
Internal vs external validity (and their threats)
Two kinds of validity are in constant tension in experimental design. Internal validity is the degree to which you can be confident the IV — and nothing else — caused the change in the DV. External validity is the degree to which your findings generalise to other people, settings and times. Tightly controlled lab experiments maximise internal validity but can sacrifice external validity; messy field experiments do the reverse. For a fuller treatment, see reliability and validity.
| Threat to internal validity | What it means | Defence |
|---|---|---|
| History | An outside event occurs during the study and affects the DV. | Control group experiences the same period. |
| Maturation | Participants naturally change over time (older, tired, practised). | Control group ages/tires equally. |
| Selection bias | Groups differ systematically before treatment. | Random assignment. |
| Testing | The pretest itself changes later scores. | Solomon four-group design. |
| Attrition | Participants drop out non-randomly. | Track and report dropout; intention-to-treat analysis. |
| Demand/experimenter effects | Cues lead participants or testers to behave as expected. | Blinding (single or double). |
Threats to external validity include unrepresentative samples (e.g. only undergraduates), artificial settings, and time-bound effects. Strong sampling helps — see sampling methods of research — as does replicating the study in different contexts.
The steps of an experiment
Running an experiment is a disciplined, sequential process. Follow these steps:
- State a directional hypothesis. Translate your causal question into a testable prediction with a clear IV and DV, plus a null hypothesis. See hypothesis testing.
- Operationalise your variables. Define exactly how the IV is manipulated and how the DV is measured, using a valid, reliable instrument.
- Identify and plan to control confounders. List plausible extraneous variables and decide how each will be held constant, counterbalanced or randomised away.
- Choose a design and sample. Select between- or within-subjects, decide on the number of conditions, and calculate the sample size needed for adequate statistical power.
- Randomly assign participants. Use a random mechanism (random number generator, sealed envelopes) to allocate participants to conditions.
- Run the manipulation and collect data. Administer conditions identically except for the IV; include a manipulation check where possible.
- Analyse the results. Use the appropriate inferential test (t-test, ANOVA), report effect sizes and confidence intervals, not just p-values.
- Interpret and report. Relate findings to the hypothesis, acknowledge limitations and validity threats, and avoid over-claiming causality beyond your design.
Worked example: computing the effect of a treatment
Suppose an education researcher tests whether a new study-skills workshop (the IV) raises end-of-term exam scores (the DV). Sixty students are randomly assigned to two groups of 30. The experimental group attends the workshop; the control group does not. At the end of term, both groups sit the same exam. To judge how big the effect is — not merely whether it is statistically significant — we compute the mean difference and a standardised effect size, Cohen’s d.
Experimental group (workshop): mean M₁ = 78, SD₁ = 9, n₁ = 30
Control group (no workshop): mean M₂ = 70, SD₂ = 8, n₂ = 30
Step 1 — Mean difference.
M₁ − M₂ = 78 − 70 = 8 marks. The workshop group scored 8 marks higher on average.
Step 2 — Pooled standard deviation.
sp = √[ ((n₁−1)·SD₁² + (n₂−1)·SD₂²) / (n₁+n₂−2) ]
= √[ ((29×81) + (29×64)) / 58 ]
= √[ (2349 + 1856) / 58 ]
= √[ 4205 / 58 ] = √72.5 ≈ 8.51
Step 3 — Cohen’s d.
d = (M₁ − M₂) / sp = 8 / 8.51 ≈ 0.94
Step 4 — Interpret. By Cohen’s conventions, d ≈ 0.2 is small, 0.5 is medium and 0.8 is large. A value of 0.94 is a large effect: the workshop group outperformed the control group by almost one full standard deviation. Because students were randomly assigned, this difference can reasonably be attributed to the workshop rather than to pre-existing differences between the groups. You would then run an independent-samples t-test to confirm the difference is statistically significant before drawing firm conclusions.
Effect sizes like Cohen’s d are essential because a difference can be statistically significant yet trivially small, or practically large yet non-significant in an underpowered study. If you would like the heavy lifting done for you, our statisticians can run and interpret the full analysis.
Need your experiment’s data analysed?
Our expert statisticians run t-tests, ANOVA and effect sizes in SPSS or R — and explain every result in plain English for your methods chapter.
Strengths and limitations
Experimental research is uniquely powerful, but it is not always feasible or appropriate.
- It is the only design that can establish cause and effect with confidence.
- Tight control over variables yields high internal validity and replicable procedures.
- Randomisation neutralises both known and unknown confounders.
- Quantitative outcomes support objective statistical testing and effect-size estimation.
Its limitations, however, are real and must be acknowledged:
- Artificial laboratory conditions can lower external validity (the “is this real life?” problem).
- Many important questions cannot be studied experimentally for ethical or practical reasons — you cannot randomly assign people to smoke or to a traumatic event.
- Demand characteristics and experimenter expectancy can bias results if blinding is absent.
- True experiments can be costly, slow and resource-intensive to run well.
Common mistakes to avoid
- Confusing correlation with causation — only manipulation plus control licenses a causal claim.
- Skipping or faking randomisation — convenience allocation reintroduces selection bias and downgrades a true experiment to a quasi-experiment.
- Ignoring confounders — failing to list and control extraneous variables is the most common reason reviewers reject causal claims.
- Reporting only p-values — always accompany significance with effect sizes and confidence intervals.
- Over-generalising — a result from 30 undergraduates in one lab does not automatically apply to everyone, everywhere.
How to do it well
Strong experiments share a few habits: a single, clearly operationalised IV; a measurable, reliable DV; genuine random assignment; an explicit plan for every plausible confounder; adequate statistical power decided before data collection; blinding wherever feasible; and honest reporting of effect sizes and limitations. Pilot your procedure first, pre-register your hypotheses and analysis plan where possible, and document everything so the study can be replicated.
It is also worth being pragmatic about which family of experiment your question and your resources actually allow. In disciplines such as education, business and public health, true randomisation is frequently impossible — you cannot randomly reorganise a company’s departments or break up intact school classes. A well-designed quasi-experiment, with a carefully matched comparison group and statistical adjustment for known baseline differences, is then the honest choice. The key is to be transparent about the design you used and to temper your causal language accordingly: a quasi-experiment can build a strong case, but it cannot deliver the airtight causal certainty of a randomised controlled trial. Choosing the right design, and reporting its limits candidly, is itself a mark of methodological maturity. Do these things and your experiment will support exactly the causal conclusion it was designed to test — no more and no less.
