Experimental Design: Confounding Variables, Control Groups, Random Assignment and Blinding

A confounding variable is a third factor that affects both things being compared, so it can make them move together even when one does not cause the other. Ice cream sales and drownings rise together every summer, but hot weather drives both. A well-designed experiment guards against confounders with a control group for comparison, random assignment to make the groups alike, and blinding so expectations cannot sway the results.

Definitions at a Glance

ConceptWhat It MeansWhy It Matters
Experimental DesignPlanning a study so the treatment is the only systematic difference between the groupsIt is how you test whether something causes an effect
Observational Study vs ExperimentAn observational study records what people already do; an experiment assigns the treatmentDecides whether a study can show cause or only association
Confounding VariableA third factor linked to both the suspected cause and the outcomeIt can create a false link or hide a real one
Control GroupThe comparison group that does not receive the treatmentShows what would have happened anyway
Random AssignmentUsing chance to decide which group each subject joinsBalances confounders, known and unknown, across the groups
Random SamplingUsing chance to decide which members of a population are studiedLets results generalize to the whole population
Placebo EffectA real change in how people feel caused by expecting to be treatedThe control group gets a look-alike placebo to cancel it out
BlindingHiding who got which treatment from the subjects, the researchers or bothStops expectations from shaping responses or scoring
Statistical SignificanceA result unlikely to arise by chance alone if there were no real effectRules out chance as a good explanation, but says nothing about size

Observational Study vs Experiment

In an observational study, researchers measure what people already do and change nothing. In an experiment, researchers decide who receives the treatment and then measure what happens. One question sorts any study: did the researchers assign the treatment, or only record it?

FeatureObservational StudyExperiment
Who decides the treatmentThe subjects themselves, or their circumstancesThe researchers, ideally by chance
ExampleComparing adults who already jog with adults who do notAssigning adults by chance to a jogging plan or to their usual routine
What it can showAn association between two variablesCause and effect, if the design is sound
Main weaknessThe groups may differ in other ways (confounding)Can be costly, slow, or unethical for some questions

Why Only Experiments Can Show Cause and Effect

When people choose their own group, the groups almost always differ in more than one way. Adults who jog by choice may also eat differently, sleep more and have fewer health problems to begin with, so a gap in health between joggers and non-joggers could come from any of those differences. A randomized experiment removes the choice. Chance decides who jogs, so the two groups start out alike, and a clear difference at the end can be credited to the jogging.

Observational studies still matter. Some questions cannot be tested by experiment, because no one may ethically assign people to smoke or to skip school. The case that smoking causes lung cancer was built from many observational studies that ruled out confounders one at a time, together with laboratory evidence. A single observational study, though, shows association, not cause. See observational vs experimental studies for more practice telling them apart.

What Is a Confounding Variable?

A confounding variable, or confounder, is a third variable that affects both the explanatory variable (the suspected cause) and the response variable (the outcome you measure). Because it moves both, it can produce an association between them even when one has no effect on the other.

A variable has to pass two tests to be a confounder: it is linked to the explanatory variable, and it affects the outcome on its own. A factor that touches only one of the two cannot create the association. Confounders can also work the other way and hide a real effect. If doctors give a new medicine mostly to their sickest patients, those patients may still do worse than others, and a medicine that truly helps can look useless.

There are two ways to deal with confounders. In an experiment, random assignment spreads them evenly across the groups. In an observational study, you compare like with like: look at the association separately within each level of the suspected confounder, for example comparing smokers only with smokers and non-smokers only with non-smokers.

Confounding Variable Examples

Each of these associations is real in the sense that the two variables do move together. In each case a third factor explains the pattern.

Observed AssociationConfounding VariableHow It Links to Both
Ice cream sales and drownings rise togetherHot weatherHot days send more people to buy ice cream and more people into the water. Ice cream does not cause drowning.
A study might find coffee drinkers have more heart diseaseSmokingIf coffee drinkers are more likely to smoke, and smoking damages the heart, coffee can look harmful even if it has no effect.
Children with larger shoe sizes read betterAgeOlder children have bigger feet and more years of reading practice. Bigger shoes will not improve anyone's reading.
Fires with more firefighters cause more damageSize of the fireBigger fires bring more firefighters and do more damage. The firefighters are not the cause.
People who exercise live longerHealth to begin withPeople who are already healthy find it easier to exercise. Exercise may well help, but this comparison alone cannot show it.

A quick test for any headline: ask what kind of person, place or time ends up in each group, and whether that same factor could affect the outcome by itself. For more practice, work through these confounding variable examples.

How to Design a Good Experiment

Good experimental design rests on four principles. Each one closes a different gap that could let something other than the treatment explain the result.

1. Compare Against a Control Group

Give the treatment to one group and not to another. Without a comparison, there is no way to tell the treatment's effect from what would have happened anyway.

2. Assign Subjects at Random

Let chance decide who goes into each group, so the groups start out alike in every respect, including the confounders nobody thought to measure.

3. Keep Everything Else the Same

Use the same conditions, timing and measurements for every group, give the control group a placebo when expectations matter, and use blinding so nobody's knowledge of the groups can tilt the results.

4. Replicate

Use enough subjects in each group that chance differences average out, and repeat the study to check that the result holds.

What Is a Control Group?

A control group is the comparison group that does not receive the treatment being tested. It shows what happens anyway, through natural recovery, practice, the passing of time, or simply being part of a study. The treatment's effect is the difference between the two groups, not the change in the treated group alone.

Suppose students in a tutoring program raise their test scores by 12 points, while a matching control group without tutoring improves by 8 points over the same weeks. The tutoring added about 4 points, not 12. The other 8 points would have come anyway from a term of ordinary teaching and practice.

Depending on the question, a control group may receive nothing, a placebo, or the current standard treatment. What matters is that it matches the treatment group in every other way, which is why both groups are formed by random assignment. More on the control group concept page.

What Is Random Assignment?

Random assignment means using chance, such as a coin flip, a shuffled list or a random number generator, to decide which group each subject joins. Neither the researchers nor the subjects choose.

It works because chance pays no attention to age, motivation, health or anything else, so each of those traits tends to be spread evenly across the groups. That includes confounders no one thought to measure, which no other method can promise. The balance holds on average rather than perfectly, so larger groups balance better than small ones.

Assigning by a rule that someone can predict or bend is not random assignment. Letting volunteers pick their group, putting the first 20 arrivals in the treatment group, or letting a teacher choose which students try a new method can each build a confounder straight into the design.

Random Assignment vs Random Sampling

Both use chance, which is why they are so often mixed up, but they happen at different stages and support different conclusions. Random sampling decides who from the population gets into the study. Random assignment decides which treatment each person in the study receives.

FeatureRandom SamplingRandom Assignment
What chance decidesWhich members of the population are studiedWhich group each subject is placed in
When it happensWhen choosing who takes partAfter subjects are chosen, before the treatment starts
What it protects againstA sample that does not represent the population (sampling bias)Groups that differ before the treatment begins (confounding)
Conclusion it allowsGeneralizing the results to the whole populationSaying the treatment caused the difference
ExamplePicking 200 students at random from the school roster for a surveyFlipping a coin to put each of 60 volunteers on a new or usual study schedule

What Each Combination Lets You Conclude

SamplingRandomly AssignedNot Randomly Assigned
Randomly sampledCause and effect, for the whole populationAssociation only, for the whole population
Not randomly sampledCause and effect, for subjects like those studiedAssociation only, for the subjects studied

Most medical and classroom experiments sit in the bottom-left cell. Volunteers are randomly assigned, but they were never a random sample of everyone, so the findings apply most directly to people like the volunteers.

What Is a Placebo?

A placebo is a look-alike treatment with no active ingredient, such as a sugar pill shaped and colored like the real medicine. The placebo effect is a real change in how people feel or behave that comes from believing they are being treated.

The placebo effect is strongest for outcomes people feel or report themselves, such as pain, energy, mood or sleep. That is why comparing treated patients with untreated ones is not enough in a medicine trial: the treated patients expect to improve, and some will. Giving the control group a placebo gives both groups the same expectations, so the treatment's real effect is whatever it adds beyond the placebo group's improvement.

A placebo only works as a control if subjects cannot tell it apart from the real treatment, which is where blinding comes in.

Single-Blind vs Double-Blind

Blinding means keeping people in a study from knowing who received which treatment, so that knowledge cannot shape the results.

DesignWho Does Not KnowWhat It PreventsExample
Single-blindThe subjectsSubjects' expectations changing how they respond or what they reportPatients do not know whether their pill is the medicine or a placebo, but their doctors do
Double-blindThe subjects and the researchers who treat or measure themSubjects' expectations, plus researchers treating or scoring one group differentlyNeither patients nor doctors know who got the medicine until the data are collected

Double-blind is the stronger design whenever someone's judgment enters the measurement, such as rating pain or grading an essay. Coded labels make it possible: the pills are marked A and B, and someone outside the study holds the key until the data are in. Some studies also blind the people who analyze the data, which is called triple-blind.

Blinding is not always possible. Subjects in an exercise experiment know whether they are exercising. In that case, a study can still blind the people who measure the results.

Replication and Sample Size

Replication has two meanings in experimental design, and both matter. First, each treatment should be given to many subjects, not one or two. Random assignment balances the groups only on average, so with five people per group one group can easily end up healthier or more motivated by luck. With hundreds per group, chance differences mostly cancel out.

Second, a finding should be repeated by other researchers with new subjects. A result that appears once may be a fluke. One that holds up across several independent studies is far more trustworthy.

What Does Statistically Significant Mean?

A result is statistically significant when it would be unlikely to happen by chance alone if the treatment had no real effect. Researchers check this with a p-value: assuming there is no real effect, the p-value is the probability of a difference at least as large as the one observed. If the p-value falls below a threshold chosen before the study, usually 0.05, the result is called significant. This decision is the last step of a hypothesis test.

In plain words, significance says chance is a poor explanation for the difference. It does not say the difference is large, important or certainly real.

Significant Is Not the Same as Important

With a very large sample, a tiny difference can be statistically significant. Suppose a study of 20,000 students finds that a study app raises average scores by half a point out of 100, with a p-value below 0.05. The result is significant, but half a point would hardly matter to any student.

Not Significant Is Not the Same as No Effect

A small study can miss a real effect because chance variation is large compared with the effect. "Not significant" means the data did not rule out chance, not that they proved the treatment does nothing.

Significance Does Not Fix a Weak Design

A significant difference between groups that chose themselves can still come from a confounder. Significance speaks to chance; only the design speaks to confounding.

For worked practice, see these statistical significance examples.

Want to check your understanding?

Our interaction checks test whether you truly understand a concept — not just whether you can repeat a procedure.

Try an interaction check

Worked Examples

Example 1: Spot the Confounder in a Headline

Headline: "Students who eat breakfast every day get better grades." The study behind it surveyed students about the habits they already had.

Study type: Observational. No one was assigned to eat breakfast; the researchers recorded what students already did.

Possible confounder: A steady home routine. Families with a regular morning routine may be more likely to serve breakfast and also more likely to make sure homework gets done and bedtimes are kept, and both of those can raise grades.

Conclusion: The headline shows an association, not that breakfast causes better grades. Testing the cause would need an experiment that assigns breakfast at random, or at least a comparison of students from homes with similar routines.

Example 2: Design a Fair Experiment

Question: Does listening to music while studying change scores on a vocabulary quiz? You have 60 student volunteers.

Treatment and control: One group studies a list of 30 new words for 20 minutes with music playing; the control group studies the same list for the same time in silence.

Random assignment: Number the volunteers 1 to 60 and use a random number generator to pick 30 for the music group. The other 30 form the control group.

Keep everything else the same: Same word list, same room, same time of day, same quiz. Students cannot be blind to whether music is playing, so blind the scorer instead: quizzes carry ID numbers, not group labels.

Conclusion: With 30 students per group, a few unusual students cannot decide the result. If the groups' average scores differ and the difference is statistically significant, music caused the difference for students like these volunteers. Because the volunteers were not a random sample of all students, applying the result more widely needs care.

Example 3: Can This Study Support a Causal Claim?

Study A: Researchers randomly sample 500 adults in a city and ask how often they meditate and how stressed they feel. Those who meditate report less stress.

Verdict: Observational, with random sampling. It supports an association for the city's adults, but not a causal claim: people who meditate may also have more free time or less stressful jobs.

Study B: 120 volunteers are randomly assigned to spend 10 minutes a day with either a meditation app or an audiobook app for eight weeks. The meditation group reports less stress, and the difference is statistically significant.

Verdict: A randomized experiment with a control group. It supports the claim that meditation lowered stress for people like these volunteers. Extending that to all adults would need a random sample as well.

Example 4: Find the Flaws in an Experiment

Setup: A company gives its new energy drink to 40 volunteers. Each volunteer knows what they are drinking, and an hour later rates their energy from 1 to 10. Most give a 7 or higher, and the company claims the drink boosts energy.

Flaws: There is no control group, so nothing shows how energetic people would have felt anyway. There is no placebo, and self-rated energy is exactly the kind of outcome the placebo effect changes. There is no blinding: everyone knows who drank what.

Fix: Randomly assign volunteers to the energy drink or a placebo drink that looks and tastes the same. Label the cans with codes so neither the volunteers nor the staff collecting ratings know which is which, making the study double-blind, and then compare the average ratings of the two groups.

Common Mistakes

Treating correlation as causation

Two variables moving together, even very closely, does not show that one causes the other. Before accepting a causal claim, ask whether the study was an experiment with random assignment. If it was observational, look for a third factor that could drive both.

Confusing random assignment with random sampling

Random sampling picks who is studied and supports generalizing to the population. Random assignment sorts the chosen subjects into groups and supports cause-and-effect conclusions. A survey of a random sample cannot show cause, and an experiment on volunteers cannot by itself speak for everyone.

Thinking "significant" means "important"

Statistically significant means the result is unlikely to be due to chance alone if there were no real effect. It says nothing about size. Always check how large the difference actually is and whether it would matter in practice.

Forgetting the control group

Measuring one group before and after a treatment cannot separate the treatment from everything else that happened in the meantime: practice, growing older, the season or the placebo effect. The control group is the baseline that makes the difference meaningful.

Next Steps: Explore Each Concept

Frequently Asked Questions

What is a confounding variable in simple terms?

A confounding variable is a hidden third factor that affects both things you are comparing, so it can make them look connected even when one does not cause the other. For example, hot weather increases both ice cream sales and drownings, so the two rise and fall together. In an experiment, random assignment spreads confounders evenly between the groups so they cannot create a false link.

What is the difference between random assignment and random sampling?

Random sampling uses chance to choose which members of a population are studied, which lets the results generalize to that population. Random assignment uses chance to decide which treatment each subject in the study receives, which lets researchers conclude that the treatment caused any difference between the groups. A study can use one, both or neither.

Can an observational study prove cause and effect?

A single observational study shows association, not cause, because the people in each group chose their own behavior and may differ in other ways. A randomized experiment is the direct way to show cause and effect. When an experiment would be unethical, such as assigning people to smoke, causal evidence is built from many observational studies that rule out confounders one by one, together with other kinds of evidence.

What is the placebo effect?

The placebo effect is a real change in how people feel that comes from believing they are being treated, even when the treatment has no active ingredient. It is strongest for outcomes people report themselves, such as pain, mood, energy and sleep. Experiments handle it by giving the control group a look-alike placebo, so the treatment's real effect is whatever it adds beyond the placebo group's improvement.

What is the difference between single-blind and double-blind?

In a single-blind study, the subjects do not know which treatment they received, but the researchers do. In a double-blind study, neither the subjects nor the researchers who treat and measure them know until the data are collected. Double-blinding also stops researchers from treating one group differently or scoring its results more kindly.

What does statistically significant mean?

A result is statistically significant when it would be unlikely to happen by chance alone if there were no real effect, usually judged by a p-value below 0.05. It means chance is a poor explanation for the difference. It does not mean the effect is large or important: with a big enough sample, even a tiny difference can be significant.

Why does an experiment need a control group?

A control group shows what would have happened without the treatment, through natural recovery, practice, the passing of time or the placebo effect. The effect of the treatment is the difference between the treatment group and the control group. Without a control group there is nothing to compare against, and any change might have happened anyway.

Related Guides

About Sense of Study

Sense of Study is a concept-first learning platform that helps students build deep understanding in math, physics, chemistry, statistics, and computational thinking. Our approach maps prerequisite relationships between concepts so students master foundations before moving forward — eliminating the gaps that cause confusion later.

With 800+ interconnected concepts and mastery tracking, we help students and parents see exactly where understanding breaks down and how to fix it.