COLLEGE STATISTICS • DATA, VARIABLES & STUDY DESIGN

Random Assignment & Causation — Random Assignment and Causation

Why randomly assigning treatments is the gold standard for establishing cause-and-effect relationships in research.

Historical Context & Motivation

For centuries, scientists and physicians attempted to determine whether a treatment, intervention, or exposure truly caused an observed effect, yet their methods were often plagued by hidden biases that corrupted their conclusions. Early medical trials, for instance, would assign the wealthier patients to a new treatment and the poorer patients to the standard remedy, making it impossible to separate the effect of the treatment from the effect of socioeconomic status on health. The core challenge was the problem of confounding variables — factors that are associated with both the treatment and the outcome, creating spurious associations that masquerade as causal effects. The development of random assignment emerged as the methodological breakthrough that would allow researchers to neutralize confounders and isolate genuine causal relationships with unprecedented rigor.

1747
Lind's Scurvy Trial
James Lind conducted one of the earliest controlled experiments aboard HMS Salisbury, assigning groups of sailors to different dietary supplements. Although not randomized, his comparative design planted the seed for controlled experimentation in medicine.
1923
Fisher's Agricultural Experiments
Sir Ronald A. Fisher, working at the Rothamsted Experimental Station, formalized the principle of randomization in experimental design. He demonstrated mathematically that random assignment of treatments to plots eliminates systematic bias and provides a valid basis for statistical inference.
1948
First Modern Randomized Controlled Trial
The British Medical Research Council conducted the landmark streptomycin trial for tuberculosis, widely regarded as the first properly randomized controlled trial (RCT) in clinical medicine. Patients were allocated to treatment or control groups using random numbers, setting the template for modern clinical research.
1960s–1970s
RCTs in Social Science
Randomized experiments expanded into education, economics, and public policy. The Perry Preschool Project and negative income tax experiments demonstrated that random assignment could be applied to evaluate social programs, not just pharmaceutical interventions.
2019
Nobel Prize for Experimental Approaches
Abhijit Banerjee, Esther Duflo, and Michael Kremer received the Nobel Memorial Prize in Economics for their pioneering use of randomized controlled trials to address global poverty, cementing the RCT as the gold standard across disciplines.

The central question that drove these developments remains the fundamental puzzle of causal inference: when we observe that a group receiving a treatment has a better outcome than a group that did not, how can we be confident the treatment itself — rather than some pre-existing difference between the groups — produced that outcome? Random assignment provides the answer, and understanding why it works is essential for any student of statistics or research methodology.

Core Principles & Definitions

To understand why random assignment is the linchpin of causal inference, we must first distinguish it from a closely related concept and then articulate the foundational principles that govern experimental design. Many students conflate random sampling with random assignment, but they serve entirely different purposes. Random sampling determines who is selected from a population to participate in a study, and it supports external validity — the ability to generalize findings to the broader population. Random assignment, by contrast, determines which treatment each participant receives once they are already in the study, and it supports internal validity — the ability to draw causal conclusions from the data.

1

Random Assignment

A procedure in which each participant in an experiment has an equal probability of being placed into any treatment group. This is the mechanism that balances confounders across groups, enabling causal claims.
2

Causation

A relationship in which a change in one variable (the explanatory variable) directly produces a change in another variable (the response variable). Establishing causation requires ruling out alternative explanations.
3

Confounding Variable

A variable that is associated with both the treatment and the outcome, creating a spurious association that can be mistaken for a causal effect. Confounders are the primary threat random assignment is designed to eliminate.
4

Internal Validity

The degree to which a study can establish a causal relationship between the treatment and the outcome for the subjects in the study. Random assignment is the strongest tool for maximizing internal validity.
5

Counterfactual Framework

The conceptual foundation of causal inference: the causal effect of a treatment on a unit is the difference between the outcome the unit would have under treatment and the outcome under control. We can never observe both, creating the fundamental problem of causal inference.
KEY TAKEAWAY
Think of random assignment like shuffling a deck of cards before dealing hands to players in a poker game. If you shuffled thoroughly, no player can systematically end up with better cards — any advantage one hand has over another is due purely to chance. Similarly, when participants are randomly assigned to treatment and control groups, any pre-existing differences between the groups (age, motivation, health, socioeconomic status) are distributed by chance alone. This means that if the treatment group later shows a better outcome, the most plausible explanation is the treatment itself, not some lurking difference between the groups.

Visual Explanation — How Random Assignment Works

The diagram illustrates the logic of random assignment. Participants enter the study as a mixed group with varying characteristics (age, health, motivation, socioeconomic status). Random assignment distributes these characteristics approximately equally across the treatment and control groups. Because the groups are comparable on all background variables, any observed difference in outcomes can be attributed to the treatment rather than to confounding factors.

The diagram above captures the essential mechanism: random assignment does not eliminate confounding variables from the study — those variables still exist in each participant. Instead, it distributes them roughly equally across groups so that their effects cancel out in the comparison. When the sample size is sufficiently large, the Law of Large Numbers ensures that the average values of all confounders will converge between groups. Even with smaller samples, the randomization provides a valid probability model for computing p-values and confidence intervals, because we know exactly how much group differences can arise by chance alone. This is why Fisher referred to randomization as the "reasoned basis" for inference in experiments.

Mathematical Framework of Causal Inference

The formal mathematics of causal inference builds on the Rubin Causal Model (also called the potential outcomes framework), introduced by Donald Rubin in 1974. For each individual i in a study, we conceptualize two potential outcomes: Yi(1), the outcome if unit i receives the treatment, and Yi(0), the outcome if unit i receives the control. The individual causal effect is defined as the difference between these two potential outcomes.

INDIVIDUAL CAUSAL EFFECT
τᵢ = Yᵢ(1) − Yᵢ(0)
τi = causal effect for individual i; Yi(1) = potential outcome under treatment; Yi(0) = potential outcome under control. We can never observe both for the same individual — this is the fundamental problem of causal inference.

Since we can never observe both Yi(1) and Yi(0) for the same individual, we shift our focus to the Average Treatment Effect (ATE), which represents the expected causal effect averaged over the entire population of interest.

AVERAGE TREATMENT EFFECT
ATE = E[Yᵢ(1) − Yᵢ(0)] = E[Yᵢ(1)] − E[Yᵢ(0)]
E[·] denotes the expected value (population average). The ATE is the average difference between the potential outcomes under treatment and control across all units in the population.

In an experiment with random assignment, we estimate the ATE using the difference in sample means. Random assignment guarantees that the treatment indicator Ti is statistically independent of the potential outcomes, written as (Yi(1), Yi(0)) ⊥ Ti. This independence condition is what eliminates selection bias.

UNBIASED ESTIMATOR UNDER RANDOM ASSIGNMENT
ATE estimate = Ȳ_treatment − Ȳ_control
Ȳtreatment = sample mean of the treatment group; Ȳcontrol = sample mean of the control group. Under random assignment, E[Ȳtreatment] = E[Yi(1)] and E[Ȳcontrol] = E[Yi(0)], so this estimator is unbiased for the ATE.
SELECTION BIAS (WITHOUT RANDOM ASSIGNMENT)
E[Ȳ_T − Ȳ_C] = ATE + {E[Yᵢ(0)|T=1] − E[Yᵢ(0)|T=0]}
The bracketed term is the selection bias: it represents the difference in baseline outcomes between the treated and untreated groups that would exist even without the treatment. Random assignment makes this term zero because treatment status is independent of potential outcomes.

Experiments vs. Observational Studies — A Classification

The distinction between experiments and observational studies is perhaps the most consequential classification in statistics, because it determines whether the researcher can make causal claims. In an experiment, the researcher actively imposes a treatment on subjects through random assignment. In an observational study, the researcher merely records data on subjects who have self-selected into their conditions, or whose conditions were determined by circumstances beyond the researcher's control. The absence of random assignment in observational studies means that confounding variables may differ systematically between groups, precluding causal conclusions without strong additional assumptions.

Side-by-side comparison of experimental and observational designs. In the experiment (left), the researcher uses random assignment (symbolized by the die) to allocate subjects to treatment and control, ensuring confounders are balanced and enabling causal conclusions. In the observational study (right), subjects self-select or are naturally sorted into exposed and unexposed groups, leaving confounders potentially imbalanced and limiting conclusions to associations.
Comparison of experiments and observational studies
FeatureExperiment (RCT)Observational Study
Random assignment?Yes — researcher assigns treatmentNo — subjects self-select or nature assigns
ConfoundersBalanced across groups (in expectation)May differ systematically between groups
Conclusion typeCause-and-effectAssociation only
ExampleRandomly assigning patients to a drug vs. placeboComparing health outcomes of smokers vs. non-smokers
Key strengthHigh internal validityFeasible when experiments are unethical or impractical

Worked Example — Evaluating a Study Design

Consider the following scenario: A university wants to determine whether a new tutoring program causes improvements in students' exam scores. They recruit 200 volunteer students and randomly assign 100 to receive the tutoring program and 100 to a control group that receives no additional tutoring. After eight weeks, the mean exam score for the tutoring group is 82 and the mean for the control group is 75. Can the researchers conclude that the tutoring program caused higher scores?

Can We Claim Causation?
1
Step 1 — Identify the Study TypeThe researchers randomly assigned participants to the tutoring or control condition. Because the researcher actively imposed the treatment and used a random mechanism for group allocation, this is a randomized experiment.
Study type: Randomized Experiment
2
Step 2 — Identify Explanatory and Response VariablesThe explanatory variable (treatment) is whether or not a student received the tutoring program (two levels: tutoring vs. no tutoring). The response variable is the exam score measured after eight weeks.
Explanatory: tutoring (yes/no); Response: exam score
3
Step 3 — Assess the Role of Random AssignmentBecause participants were randomly assigned, we expect potential confounders — prior academic ability, motivation, study habits, socioeconomic background — to be approximately balanced across the two groups. This means any observed difference in exam scores is unlikely to be driven by pre-existing group differences.
Confounders balanced in expectation ✓
4
Step 4 — Compute the Observed Treatment EffectThe estimated average treatment effect is the difference in sample means: Ȳtreatment − Ȳcontrol = 82 − 75 = 7 points. Under random assignment, this is an unbiased estimate of the causal effect.
Estimated ATE = 7 points
5
Step 5 — State the ConclusionBecause this is a randomized experiment, we can conclude that the tutoring program caused an average improvement of approximately 7 points on exam scores (assuming the difference is statistically significant, which would be verified by a hypothesis test). Had this been an observational study — for example, if students had self-selected into tutoring — we could only say the tutoring was associated with higher scores, because motivated students might be more likely to seek tutoring and more likely to score well regardless.
Yes — the tutoring program caused higher exam scores (causal claim justified).

Strengths and Limitations of Random Assignment

While random assignment is the gold standard for establishing causation, it is not without limitations. Understanding both the strengths and the constraints of randomized experiments is essential for designing studies and critically evaluating published research. The following table summarizes the key trade-offs.

Strengths and Limitations of Random Assignment
StrengthsLimitations
Eliminates confounding in expectation, providing the strongest basis for causal inference.May be unethical — you cannot randomly assign people to smoke cigarettes, experience poverty, or forgo medical treatment.
Provides a probability model for statistical inference (p-values, confidence intervals) that does not rely on distributional assumptions.May be impractical or prohibitively expensive for certain research questions (e.g., long-term effects of diet over decades).
Results are straightforward to interpret — the difference in group means directly estimates the average causal effect.Small sample sizes can lead to chance imbalances in confounders despite randomization.
Controls for both measured and unmeasured confounders, unlike statistical adjustments in observational studies.Non-compliance (participants not following their assigned treatment) and attrition can undermine the experiment.
Widely accepted across disciplines as the highest level of evidence for treatment effects.Volunteer participants may not represent the broader population, limiting generalizability (external validity).
⚖️ KEY TAKEAWAY
Random assignment is like a well-calibrated scale in a laboratory: it is the most precise tool available for measuring the weight (causal effect) of a treatment, but it requires controlled conditions that aren't always feasible in the real world. When researchers cannot run an experiment — because of ethical, practical, or financial constraints — they rely on observational studies, sometimes using techniques like matching, instrumental variables, or regression discontinuity to approximate the conditions that random assignment would create. These quasi-experimental methods are valuable but always rest on assumptions that randomization would have made unnecessary.

Connection to Advanced Causal Inference

The principles of random assignment form the foundation upon which the entire field of causal inference is built. In more advanced coursework, you will encounter methods designed to approximate the conditions of random assignment when true experiments are not possible. These methods are central to modern applied statistics, econometrics, epidemiology, and data science. The table below previews how the ideas in this lesson connect to more advanced techniques.

From random assignment to advanced causal inference methods
Concept from This LessonAdvanced Extension
Random assignment balances confoundersPropensity Score Matching — estimates causal effects in observational data by matching treated and control units with similar probabilities of receiving treatment
Potential outcomes framework (Yᵢ(1), Yᵢ(0))Rubin Causal Model & SUTVA — formalizes assumptions like stable unit treatment value, enabling rigorous definition of causal effects in complex settings
Selection bias = 0 under randomizationInstrumental Variables (IV) — uses a variable correlated with the treatment but uncorrelated with confounders to estimate causal effects when assignment is not random
Comparison of treatment vs. control meansRegression Discontinuity Design (RDD) — exploits a cutoff in a continuous variable that quasi-randomly assigns subjects to treatment, allowing causal inference near the threshold
Ethical limitations of experimentsNatural Experiments — leverages events (policy changes, lotteries, weather shocks) that create as-if-random variation in treatment, enabling causal analysis from observational data

As you progress in statistics, you will see that the logic of random assignment serves as the benchmark against which all other causal methods are evaluated. Every quasi-experimental technique is judged by how well it mimics the conditions that a well-executed randomized experiment would provide. Understanding random assignment deeply — not just as a procedural step, but as a theoretical guarantee of independence between treatment and potential outcomes — equips you with the conceptual lens needed to evaluate causal claims in any context, from clinical trials to policy evaluations to A/B testing in technology.

Practice Problems

PROBLEM 1CONCEPTUAL
A researcher finds that people who drink more coffee tend to live longer. She concludes that coffee consumption causes increased longevity. Identify the flaw in her reasoning and explain what type of study would be needed to support a causal claim.
PROBLEM 2BASIC CALCULATION
In a randomized experiment testing a new fertilizer, 50 plots are randomly assigned to receive the fertilizer and 50 receive a placebo. The mean yield for the fertilizer group is 142 bushels per acre and the mean yield for the placebo group is 128 bushels per acre. Calculate the estimated average treatment effect (ATE) and explain why this estimate is unbiased.
PROBLEM 3INTERMEDIATE
A hospital wants to test whether a new surgical technique reduces recovery time compared to the standard procedure. Surgeons at the hospital allow each patient to choose which technique they prefer. After collecting data, the hospital finds that patients who chose the new technique recovered 3 days faster on average. (a) What type of study is this? (b) Identify at least two potential confounding variables. (c) Could the hospital justify a causal conclusion? Explain why or why not.
PROBLEM 4APPLIED
A tech company wants to determine whether a redesigned checkout page increases the proportion of users who complete a purchase. They randomly assign 10,000 website visitors to see the new design (treatment) and 10,000 to see the original design (control). The purchase rate is 4.8% for the treatment group and 4.2% for the control group. (a) What is the estimated causal effect? (b) Explain why this design supports a causal claim. (c) Does this experiment also support generalizing the result to all potential website visitors? What additional design feature would be needed?
PROBLEM 5CRITICAL THINKING
Consider the selection bias formula: E[Ȳ_T − Ȳ_C] = ATE + {E[Yᵢ(0)|T=1] − E[Yᵢ(0)|T=0]}. (a) Interpret the selection bias term in plain language. (b) Provide a concrete example of a study in which the selection bias term would be positive, making the naive comparison overestimate the true ATE. (c) Explain, using the independence condition (Yᵢ(1), Yᵢ(0)) ⊥ Tᵢ, why random assignment sets the selection bias term to zero.

Lesson Summary

Random assignment is the defining feature of a true experiment and the most powerful tool available for establishing causation. By using a chance mechanism to allocate participants to treatment and control groups, random assignment ensures that confounding variables — both measured and unmeasured — are balanced across groups in expectation. This balance eliminates selection bias and allows researchers to attribute observed differences in outcomes to the treatment itself, rather than to pre-existing group differences. The potential outcomes framework formalizes this logic: under randomization, the treatment indicator is independent of potential outcomes, making the simple difference in group means an unbiased estimator of the Average Treatment Effect (ATE).

It is critical to distinguish random assignment from random sampling: the former supports internal validity (the ability to make causal claims), while the latter supports external validity (the ability to generalize results). Without random assignment, as in observational studies, researchers can identify associations but cannot rule out confounding. When ethical or practical constraints prevent randomization, advanced techniques such as propensity score matching, instrumental variables, and regression discontinuity designs attempt to approximate the conditions of random assignment, but they always rely on assumptions that randomization would render unnecessary.

Varsity Tutors • College Statistics • Random Assignment & Causation