The Central Idea
The central question is not merely whether two variables are related. It is whether changing one variable causes a change in another.
Suppose people who regularly eat breakfast report better concentration than people who skip breakfast. The data show an association:
\[ \text{breakfast habit}\quad\longleftrightarrow\quad\text{concentration}. \]The association alone does not prove that breakfast caused the difference. Sleep, work schedules, health, income, or other variables might help explain the pattern.
Study design tells us how far a conclusion may reasonably go.
Observational Studies
An observational study measures variables without assigning the explanatory condition to the participants or units being studied. Researchers observe, survey, record, or retrieve data, but they do not decide who receives which condition.
Examples include:
- surveying commuters about travel time and job satisfaction;
- comparing health records of smokers and nonsmokers;
- tracking rainfall and crop yields across several counties;
- examining the relationship between study time and course grades; and
- recording the feeding behavior of animals in their natural habitat.
In every case, the condition already exists or arises naturally. The researcher does not impose it.
Experiments
An experiment deliberately assigns one or more conditions to experimental units and measures the resulting outcomes. A condition imposed by the researcher is called a treatment.
The units might be people, plants, classrooms, machines, plots of land, or digital systems. For example, a researcher might randomly assign plants to receive either a new fertilizer or a standard fertilizer, then compare their growth after six weeks.
The explanatory variable is the assigned fertilizer, and the response variable is plant growth:
\[ \text{assigned fertilizer}\quad\longrightarrow\quad\text{measured growth}. \]The deliberate assignment is what makes the study an experiment. Merely comparing plants that happened to receive different fertilizers would be observational.
The Decisive Difference
| Question | Observational Study | Experiment |
|---|---|---|
| Who determines the condition? | Nature, circumstance, or the participants | The researcher |
| Is a treatment imposed? | No | Yes |
| What can a well-designed study establish? | Association and prediction | Potentially cause and effect |
| Main threat | Confounding and selection differences | Weak assignment, poor control, bias, or noncompliance |
The fastest classification test is therefore:
\[ \boxed{\text{Did the researcher assign the condition?}} \]If the answer is yes, the study is an experiment. If the answer is no, the study is observational.
Association Is Not Automatically Causation
An association exists when values of one variable tend to occur with particular values of another. A causal relationship means that changing the explanatory variable changes the response variable, all else being appropriately controlled.
Observational data may show a strong, consistent, and useful association. It may even contribute to a broader causal case. By itself, however, the observational design usually cannot rule out alternative explanations.
A strong relationship can still be noncausal. Conversely, a real causal effect can be small or difficult to detect. The size of an association and the validity of a causal claim are different questions.
Confounding
A confounding variable is associated with the explanatory variable and also affects the response. Its influence becomes mixed with the relationship under study.
Suppose an observational study finds that people who carry lighters have a higher rate of lung disease. Carrying a lighter does not plausibly explain the disease. Smoking is associated with carrying a lighter and affects lung health:
\[ \begin{aligned} \text{smoking}&\longrightarrow\text{carrying a lighter},\\ \text{smoking}&\longrightarrow\text{lung disease}. \end{aligned} \]Smoking confounds the observed association between lighter ownership and lung disease. The example is obvious, but real confounding can be subtle, multidimensional, and only partly measurable.
Why Random Assignment Matters
In a well-designed experiment, researchers often use random assignment to place experimental units into treatment groups. Each unit's assignment is determined by a chance process rather than by personal choice, researcher preference, or an existing difference.
Random assignment does not guarantee that the groups are identical. It makes systematic preexisting differences less likely and gives known and unknown characteristics an opportunity to balance across groups.
If the groups are treated similarly except for the assigned condition, a later difference in outcomes can more credibly be attributed to the treatment:
\[ \text{random assignment} \;\longrightarrow\; \text{comparable groups} \;\longrightarrow\; \text{stronger causal evidence}. \]Random assignment is not the same as random sampling. Random assignment supports causal comparison within a study. Random sampling supports generalizing from a sample to a population. A study may use one, both, or neither.
One Question, Two Study Designs
Consider the question: Does listening to instrumental music improve performance on a proofreading task?
Observational Version
Ask participants whether they normally listen to instrumental music while working. Compare the proofreading scores of those who answer yes with the scores of those who answer no.
This design can reveal an association. It is vulnerable to confounding because the groups may differ in concentration habits, reading experience, work environment, or other relevant characteristics.
Experimental Version
Randomly assign participants to complete the same proofreading task either with instrumental music or in silence. Keep the instructions, time limit, room, and scoring method the same. Compare the group results.
This design provides stronger evidence about whether the music itself affects performance because the researcher assigns the listening condition.
Example: Classify a Health Study
Researchers use medical records to identify adults who chose to take a daily vitamin supplement and adults who did not. They compare the number of respiratory infections recorded in each group during the following year.
Step 1: Identify the Explanatory Condition
The explanatory condition is whether a person took the supplement.
Step 2: Ask Who Assigned It
The participants chose whether to take the supplement. The researchers did not assign the condition.
Step 3: Classify the Study
This is an observational study.
Step 4: State an Appropriate Conclusion
The researchers may compare infection rates and report an association, but they should not claim from this design alone that the supplement caused a change in infections. Diet, age, health awareness, access to care, or baseline health could confound the result.
Example: Classify an Agricultural Study
A farmer divides a field into 40 similar plots. A random process assigns 20 plots to receive a new soil treatment and 20 plots to receive the standard treatment. At harvest, the yield from every plot is recorded.
Step 1: Identify the Treatment
The treatment is the soil preparation assigned to each plot.
Step 2: Ask Who Assigned It
The researcher assigned the treatment through a random process.
Step 3: Classify the Study
This is a randomized experiment.
Step 4: State the Scope of the Evidence
If the plots are managed alike except for the soil treatment, a meaningful difference in average yield provides evidence that the treatment caused the difference for plots like those studied. Whether the result generalizes to other fields, climates, or crops is a separate question.
Study Design and the Scope of a Conclusion
Two distinct issues govern what a study can conclude:
- Causal validity: Does the design support the claim that the treatment caused the observed difference?
- Generalizability: Does the sample represent the larger population to which the conclusion is being applied?
Random assignment primarily addresses causal validity. Representative random sampling primarily addresses generalizability. These goals should not be confused.
| Design Feature | Primary Benefit |
|---|---|
| Random assignment to treatments | Supports cause-and-effect reasoning |
| Random sampling from a population | Supports generalization to that population |
| Both | Can support both causal reasoning and population generalization |
| Neither | Requires especially cautious, limited interpretation |
When an Experiment Is Not Possible
Experiments are powerful, but they are not always ethical, practical, or physically possible. Researchers cannot ethically assign people to smoke, experience hazardous pollution, or be denied necessary treatment merely to observe long-term effects. Researchers also cannot assign a person's age, birthplace, or past experiences.
In such settings, carefully designed observational studies are essential. Researchers may match similar individuals, measure possible confounders, follow participants over time, use natural experiments, or combine evidence from several sources. These methods can strengthen an analysis, but the limits of the design must remain visible in the conclusion.
How to Read a Study Description
Use the following sequence whenever a study appears in a problem, report, or news article.
- Identify the units. Who or what was studied?
- Identify the variables. What condition and outcome were measured?
- Look for assignment. Did the researcher impose a condition?
- Look for randomization. If assigned, how were groups formed?
- Look for comparison. What treatment or control groups were compared?
- Check competing explanations. What confounders or biases remain?
- Match the conclusion to the design. Is the claim associative or causal?
The verbs often provide clues. Phrases such as surveyed, recorded, tracked, and reviewed records often describe observation. Phrases such as assigned, administered, and allocated to receive often describe an experiment. The decisive issue is still the design, not the vocabulary alone.
Common Errors
Calling Every Data Collection an Experiment
In everyday speech, people may call any investigation an experiment. In statistics, an experiment requires the deliberate assignment of a treatment or condition.
Treating a Large Sample as Proof of Causation
A large sample can reduce random sampling variability, but it does not remove confounding or repair a weak design. A very large observational study remains observational.
Confusing Random Sampling with Random Assignment
Random sampling concerns who enters the study. Random assignment concerns which treatment an experimental unit receives. They serve different goals.
Assuming Random Assignment Solves Every Problem
Experiments can still be weakened by attrition, noncompliance, poor measurement, unequal treatment beyond the intended condition, small samples, or conclusions that extend beyond the studied setting.
Guided Practice
Problem 1
A school records the amount of nightly sleep reported by students and compares it with their mathematics test scores. Is this observational or experimental? What kind of conclusion is appropriate?
Show the solution
The study is observational because the school did not assign students to sleep for different amounts of time. The data may show an association between reported sleep and test scores, but they do not establish that additional sleep caused higher scores. Study habits, health, stress, and home environment are possible confounders.
Problem 2
A manufacturer randomly assigns 100 batteries from one production run to operate at either a standard temperature or a high temperature. Battery life is then measured. Is this observational or experimental? What kind of conclusion is appropriate?
Show the solution
The study is a randomized experiment because the manufacturer assigns the operating temperature. If the groups are otherwise treated alike, a difference in battery life supports a causal conclusion about temperature for batteries and conditions like those studied.
Problem 3
Researchers compare the annual medical costs of people who voluntarily join a fitness program with the costs of people who do not join. Identify one possible confounding variable.
Show the solution
Possible confounders include baseline health, age, income, access to care, motivation, diet, and prior exercise habits. For example, people who join may already be more health-conscious, and that characteristic may affect both program participation and medical costs.
Problem 4
A researcher randomly samples adults from a city and asks whether they use a standing desk and whether they experience back pain. The sample is random. Does that make the study an experiment?
Show the solution
No. Random sampling does not make a study experimental. The researcher did not assign desk type, so the study is observational. The random sample may improve generalizability to the city's adult population, but it does not by itself justify a causal claim.
A Final Classification Check
Before interpreting any result, complete this sentence:
\[ \boxed{\text{The researcher did—or did not—assign the explanatory condition.}} \]If the researcher did not assign it, describe association and remain alert to confounding. If the researcher did assign it, inspect how the assignment, comparison, measurement, and implementation were carried out before making a causal claim.
Summary
- An observational study measures conditions without assigning them.
- An experiment deliberately assigns treatments or conditions and measures outcomes.
- Observational studies can establish association, but confounding usually limits causal interpretation.
- A well-designed randomized experiment can provide strong evidence of cause and effect.
- Random assignment supports causal reasoning; random sampling supports generalization to a population.
- Ethical and practical limits often make observational research necessary.
- Every conclusion must remain within the limits of the study design.
Next Step
The next lesson develops the basic parts of an experiment: treatments, controls, and outcomes. These elements turn a causal question into a comparison that can be measured and interpreted.