COLLEGE STATISTICS • ESTIMATION AND CONFIDENCE INTERVALS

Conditions for Inference

The essential assumptions that must hold before any confidence interval or hypothesis test can be trusted.

Historical Context & Motivation

Statistical inference — the art of drawing conclusions about a population from a sample — is one of the most powerful tools in the quantitative sciences, but it rests on a set of assumptions that, when violated, can render conclusions meaningless or misleading. The development of formal conditions for inference evolved alongside the major breakthroughs of mathematical statistics in the late nineteenth and early twentieth centuries. Understanding these conditions is not merely a procedural checklist; it reflects a deeper understanding of why inferential methods work and when they fail.

1733
De Moivre's Normal Approximation
Abraham de Moivre demonstrated that the binomial distribution approaches a bell-shaped curve as n grows large, planting the seed for the normality assumption central to many inferential procedures.
1908
Student's t-Distribution
William Sealy Gosset, writing under the pen name 'Student,' derived the t-distribution for small samples, explicitly highlighting the requirement that the underlying population be approximately normal when sample sizes are small.
1922
Fisher Formalizes Random Sampling
Ronald Fisher published foundational work on estimation theory, emphasizing that valid inference requires data obtained through proper random sampling mechanisms — a condition we now call the randomness or independence condition.
1934
Neyman's Confidence Intervals
Jerzy Neyman introduced the modern framework of confidence intervals with repeated-sampling interpretation, making explicit that coverage probabilities are only guaranteed when all underlying conditions — randomness, independence, and distributional assumptions — are satisfied.
1960s–present
Robustness & Diagnostics
The work of George Box, John Tukey, and others explored how sensitive inferential procedures are to condition violations, spawning diagnostic tools such as residual plots, QQ-plots, and formal tests for normality and equal variances.

The central question this lesson addresses is deceptively simple: Under what conditions can we trust the output of an inferential procedure? Whether you are constructing a confidence interval for a population mean or proportion, or conducting a hypothesis test, you must verify a core set of conditions before interpreting the results. Failure to do so is one of the most common sources of error in applied statistics, and understanding these conditions will sharpen both your analytical rigor and your ability to critique published research.

Core Conditions for Inference

Although the precise conditions vary slightly depending on the parameter being estimated (a mean, a proportion, a difference), most inferential procedures in introductory statistics share three fundamental requirements. These conditions are sometimes organized under the mnemonic R-I-N (Randomness, Independence, Normality) or, equivalently, R-N-S (Random, Normal, Sufficient sample size). Regardless of the mnemonic, the substantive ideas are the same.

1

Randomness (Random Sampling or Assignment)

The data must come from a random sample of the population of interest, or from a randomized experiment. Without randomization, the sample may be systematically unrepresentative, and no amount of mathematical machinery can fix a biased sampling scheme.
2

Independence

Individual observations must be independent of one another. In practice, this is often checked via the 10% condition: sampling without replacement is approximately independent when the sample is no more than 10% of the population (n ≤ 0.10N).
3

Normality (Nearly Normal Condition)

The sampling distribution of the statistic must be approximately normal. For means, this is verified via the Central Limit Theorem when n ≥ 30 or when the population itself is roughly normal. For proportions, both np ≥ 10 and n(1 − p) ≥ 10 must hold (the success–failure condition).
KEY TAKEAWAY
Think of the conditions for inference like the pre-flight checklist for a commercial aircraft. No competent pilot would skip it simply because the plane 'looks fine.' Each condition — randomness, independence, normality — is a safety check that ensures the mathematical engine of inference is operating within its design specifications. Violating even one condition is like flying with an unchecked hydraulic system: you might arrive safely, but you have no guarantee that the outcome is reliable.

Visualizing the Conditions

The following diagram illustrates the decision flow a statistician should follow before performing inference. Each branch checks one of the three core conditions and indicates what remedial action may be available when a condition is not met.

The inference decision flowchart. Start at the top and work downward. A 'NO' at any stage diverts you to an alternative procedure or signals that standard inference is inappropriate. The randomness check comes first because no statistical fix can overcome a biased sampling design, followed by independence and normality.

Notice the hierarchical nature of the flowchart: the randomness condition is checked first because it is the most fundamental and the least forgivable when violated. A non-random sample introduces selection bias that cannot be corrected post hoc. The independence condition is checked next; violations here alter the variance of the sampling distribution, leading to confidence intervals that are either too narrow (undercoverage) or too wide (overcoverage). Finally, the normality condition is the most flexible — the Central Limit Theorem provides a powerful safety net for means when sample sizes are sufficiently large, and the success–failure condition provides an analogous safeguard for proportions.

Mathematical Framework

The conditions for inference exist because the derivations of confidence intervals and test statistics rely on specific mathematical properties of the sampling distribution. Let us trace exactly where each condition enters the machinery for the two most common parameters: a population mean μ and a population proportion p.

Inference for a Population Mean

CONFIDENCE INTERVAL FOR μ
x̄ ± t* × (s / √n)
where is the sample mean, t* is the critical value from the t-distribution with n − 1 degrees of freedom, s is the sample standard deviation, and n is the sample size. The formula assumes that x̄ follows a t-distribution, which requires independence and approximate normality of the population (or large n).

Inference for a Population Proportion

CONFIDENCE INTERVAL FOR p
p̂ ± z* × √(p̂(1 − p̂) / n)
where is the sample proportion, z* is the critical value from the standard normal distribution, and the expression under the radical is the estimated standard error of p̂. The z* multiplier is valid only when the sampling distribution of p̂ is approximately normal, which requires the success–failure condition.

Where Each Condition Enters

STANDARD ERROR DERIVATION (INDEPENDENCE)
SE(x̄) = σ / √n ⟵ requires Var(X₁ + X₂ + ⋯ + Xₙ) = n × σ²
The variance of a sum equals the sum of the variances only when the observations are independent. If observations are positively correlated, the true variance of the sum exceeds nσ², and the standard error formula underestimates the actual variability — producing confidence intervals that are too narrow.
SUCCESS–FAILURE CONDITION
np ≥ 10 and n(1 − p) ≥ 10
This condition ensures that the binomial distribution B(n, p) is well-approximated by a normal distribution. When either expected count is too small, the sampling distribution of p̂ is noticeably skewed, and symmetric z*-based intervals misrepresent the true coverage probability.

Checking the Conditions in Practice

Knowing the conditions conceptually is only half the battle; you must also learn how to verify them empirically using the information available in any given problem. Below is a detailed breakdown of how each condition is checked, along with common pitfalls.

A side-by-side reference for the three diagnostic panels. The Randomness panel focuses on study design, the Independence panel emphasizes the 10% rule and structural dependencies, and the Normality panel distinguishes between the criteria for means (CLT-based) and proportions (success–failure).

A few nuances deserve emphasis. First, the 10% condition is a practical approximation. When you sample without replacement from a finite population, successive draws are technically dependent — drawing one individual changes the composition of the remaining pool. However, when the sample constitutes only a small fraction of the population, this dependence is negligible and the observations can be treated as effectively independent. Second, the normality condition is not about the shape of the population itself (except for very small n), but about the shape of the sampling distribution of the statistic. The Central Limit Theorem guarantees that this distribution approaches normality for large n regardless of the population shape, which is why the n ≥ 30 guideline is so commonly cited. For smaller samples, you must scrutinize the data directly for strong skewness or outliers using histograms, boxplots, or normal probability plots.

⚠️ When n < 30 for Means
If your sample size is smaller than 30, the CLT may not provide sufficient protection. In this regime, you should examine a histogram or QQ-plot of the sample data. If the data appear roughly symmetric with no extreme outliers, the t-procedure remains reasonable. However, if the distribution is heavily skewed or has outliers, consider a nonparametric alternative such as the Wilcoxon signed-rank test or a bootstrap confidence interval.

Worked Example: Checking Conditions and Building a Confidence Interval

A university health center wants to estimate the average number of hours of sleep per night for undergraduates. Researchers take a simple random sample of 45 students from the registrar's database of 12,000 undergraduates. The sample yields x̄ = 6.8 hours and s = 1.3 hours. Construct a 95% confidence interval for the population mean μ, after first verifying all conditions.

95% Confidence Interval for Mean Hours of Sleep
1
Step 1 — Check RandomnessThe problem states that the sample was a simple random sample (SRS) drawn from the registrar's database. As long as the database covers the entire population of interest (all undergraduates at this university), the randomness condition is satisfied. We can generalize to the population of all undergraduates at this institution.
✓ Random condition met (SRS from population).
2
Step 2 — Check Independence (10% Condition)The sample size is n = 45 and the population size is N = 12,000. We compute n/N = 45/12,000 = 0.00375, which is 0.375%. Since 0.375% < 10%, the sample is a small fraction of the population, and observations can be treated as independent.
✓ Independence condition met (45 ≤ 10% × 12,000 = 1,200).
3
Step 3 — Check NormalitySince the sample size n = 45 is greater than 30, the Central Limit Theorem ensures that the sampling distribution of x̄ is approximately normal, even if the population of sleep hours is somewhat skewed. Additionally, suppose a histogram of the data shows a roughly unimodal, symmetric shape with no extreme outliers. The normality condition is satisfied.
✓ Normality condition met (n = 45 ≥ 30; CLT applies).
4
Step 4 — Compute the Standard ErrorThe standard error of the mean is SE = s / √n = 1.3 / √45 ≈ 1.3 / 6.708 ≈ 0.1938.
SE ≈ 0.194 hours
5
Step 5 — Find the Critical ValueFor a 95% confidence interval using the t-distribution with df = n − 1 = 44, the critical value t* ≈ 2.015 (from a t-table or software).
t* ≈ 2.015
6
Step 6 — Construct the IntervalThe margin of error is ME = t* × SE = 2.015 × 0.194 ≈ 0.391. Thus, the 95% confidence interval is x̄ ± ME = 6.8 ± 0.391, which gives (6.409, 7.191).
95% CI: (6.41, 7.19) hours
7
Step 7 — Interpret in ContextWe are 95% confident that the true mean number of hours of sleep per night for undergraduates at this university lies between approximately 6.41 and 7.19 hours. Notice that the interpretation references the population parameter (the true mean), not the sample statistic, and includes the context of the study.
Interpretation: 95% confident μ is between 6.41 and 7.19 hours.

Consequences of Violating Conditions

Understanding what goes wrong when a condition is violated is just as important as knowing the conditions themselves. The table below summarizes the impact of each type of violation on the reliability of your inferential conclusions.

Impact of violating each condition for inference
Condition ViolatedEffect on InferencePossible Remedy
RandomnessSelection bias: the sample systematically misrepresents the population. Confidence intervals and p-values are not meaningful because they describe the wrong population.Redesign the study with proper random sampling or randomized assignment. Post-hoc statistical adjustments (e.g., weighting) are partial fixes at best.
IndependenceStandard errors are miscalculated. Positive correlation between observations leads to underestimated SEs, producing intervals that are too narrow and p-values that are too small (inflated Type I error rate).Use methods designed for dependent data: cluster-robust standard errors, hierarchical models, or time-series methods.
Normality (small n)The t or z critical values no longer correspond to the actual tail probabilities of the sampling distribution. Coverage of confidence intervals departs from the nominal level.Increase sample size to invoke CLT, use a data transformation (e.g., log), or apply a nonparametric procedure (bootstrap, rank-based tests).
Success–Failure (proportions)Normal approximation to the binomial breaks down. Intervals may have actual coverage probabilities well below the stated confidence level.Use exact binomial methods (Clopper–Pearson interval) or the Wilson score interval, both of which perform better when np̂ or n(1 − p̂) is small.
KEY TAKEAWAY
The three conditions are not equally forgiving. The randomness condition is the hardest to fix after the fact — think of it like the foundation of a building: if the foundation is cracked, no amount of interior decoration (mathematical sophistication) can make the structure safe. The independence condition is moderately forgiving when violations are mild, and the normality condition is the most forgiving, particularly when n is large, thanks to the Central Limit Theorem.

Connection to Advanced Methods

The conditions discussed in this lesson apply to the most common introductory procedures — one-sample t-intervals, one-sample z-intervals for proportions, and their two-sample extensions. As you advance in statistics, you will encounter procedures that relax, modify, or replace some of these conditions. Understanding the introductory conditions provides a conceptual anchor for evaluating these more sophisticated methods.

How advanced methods relax introductory conditions
Introductory MethodAdvanced ExtensionCondition Relaxed
t-interval (requires independence)Mixed-effects / hierarchical modelsHandles clustered (non-independent) observations by modeling within-group correlation.
z-interval for p (requires np ≥ 10)Exact binomial (Clopper–Pearson) intervalRemoves the normal-approximation requirement; valid for any n and any p.
t-interval (requires normality or large n)Bootstrap confidence intervalDoes not require a parametric distributional assumption; uses resampling from the observed data.
SRS assumption (requires randomness)Survey-weighted estimation (Horvitz–Thompson)Accommodates complex sampling designs (stratification, clustering) with unequal selection probabilities.

A particularly important advance is the bootstrap, which constructs a confidence interval by repeatedly resampling (with replacement) from the observed sample, building an empirical approximation to the sampling distribution. The bootstrap still requires randomness and independence but does not rely on any normality assumption, making it a powerful tool for small samples or non-normal populations. Similarly, permutation tests provide hypothesis testing without distributional assumptions, relying only on exchangeability under the null hypothesis. These methods illustrate a broader theme: as you learn more statistics, you gain tools to handle increasingly realistic data-generating scenarios, but the foundational logic of checking assumptions before trusting results remains unchanged.

Practice Problems

PROBLEM 1CONCEPTUAL
A researcher surveys 80 customers who voluntarily fill out a comment card at a restaurant and uses the results to construct a 95% confidence interval for the mean satisfaction rating of all customers. Which condition for inference is most clearly violated, and why does this matter?
PROBLEM 2BASIC CALCULATION
A political polling firm takes a simple random sample of n = 600 voters from a city of 25,000 registered voters. They find that p̂ = 0.38 support a ballot measure. Verify all three conditions for constructing a one-proportion z-interval.
PROBLEM 3INTERMEDIATE
A researcher randomly selects 18 light bulbs from a production run of 5,000 and measures their lifetimes. The sample data are roughly symmetric with one moderate outlier. The sample mean is 1,020 hours with s = 85 hours. Should the researcher proceed with a one-sample t-interval? Justify your answer by checking all conditions, and suggest an alternative if appropriate.
PROBLEM 4APPLIED
An epidemiologist wants to estimate the proportion of adults in a rural county (population 3,200) who have received a flu vaccine. She obtains a random sample of 400 adults and finds that p̂ = 0.54 have been vaccinated. (a) Verify all conditions for a 90% confidence interval for p. (b) Construct the interval. (c) Interpret the result in context.
PROBLEM 5CRITICAL THINKING
Consider two scenarios: (A) A sample of n = 25 from a strongly right-skewed population, and (B) a sample of n = 250 from the same population. For each, discuss whether a t-interval for the mean is appropriate. Then explain why the Central Limit Theorem does not 'fix' every problem — give a specific example of a situation where even a large sample size would not rescue inference.

Lesson Summary

Before performing any inferential procedure — whether a confidence interval or a hypothesis test — you must verify three core conditions. The randomness condition ensures that data come from a random sample or randomized experiment, enabling generalization to the population. The independence condition ensures that individual observations do not influence one another, typically checked via the 10% condition (n ≤ 0.10N), so that the standard error formula is valid. The normality condition ensures that the sampling distribution of the statistic is approximately bell-shaped, verified for means via the Central Limit Theorem (n ≥ 30) or visual inspection of the data, and for proportions via the success–failure condition (np ≥ 10 and n(1 − p) ≥ 10).

Violating the randomness condition introduces bias that no statistical method can fully repair. Violating independence corrupts standard error estimates, and violating normality undermines the accuracy of critical values. When conditions are not met, advanced tools such as bootstrap methods, exact binomial intervals, and hierarchical models offer more robust alternatives. Mastering these conditions is not just a procedural requirement — it reflects a deep understanding of why inferential procedures work and how to apply them responsibly.

Varsity Tutors • College Statistics • Conditions for Inference