Checking Conditions - College Statistics

SECTION 1

Historical Context & Motivation

Statistical inference has always rested on a compact set of mathematical assumptions, yet the formal practice of checking conditions before applying a procedure evolved over more than a century. Early pioneers such as Carl Friedrich Gauss and Pierre-Simon Laplace derived theoretical results under idealized assumptions—normality, independence, equal variance—without necessarily prescribing systematic verification steps for practitioners. As statistical methods migrated from mathematical journals into experimental laboratories and, eventually, into introductory coursework, the gap between theory and practice became a source of persistent errors. Researchers would apply a t-test or construct a confidence interval without asking whether the data actually satisfied the conditions that legitimized the calculation. The modern emphasis on condition checking therefore represents a pedagogical and methodological corrective—one that insists on verifying assumptions before trusting conclusions.

1809 • Gauss and the Normal Error Model

Carl Friedrich Gauss publishes Theoria Motus, formalizing the method of least squares under the assumption that errors are normally distributed and independent—conditions that would later become cornerstones of regression diagnostics.

1908 • Gosset's t-Distribution

William Sealy Gosset ("Student") derives the t-distribution for small samples, explicitly noting the condition that the parent population should be approximately normal. This marks one of the earliest instances where a specific condition was tied to a specific procedure.

1935 • Fisher's Design of Experiments

R. A. Fisher emphasizes randomization as a condition for valid causal inference, arguing that experimental design—not just analysis—must satisfy structural conditions like random assignment and replication.

1977 • Tukey's Exploratory Data Analysis

John Tukey popularizes graphical methods—boxplots, stem-and-leaf displays, residual plots—as tools for exploring data before formal analysis, institutionalizing the habit of visually checking conditions.

2000s • Modern Pedagogy of Condition Checking

Introductory statistics curricula (e.g., AP Statistics, college-level courses) formally embed condition-checking as a required step in inference problems, making it a graded component of every hypothesis test and confidence interval.

The central question that checking conditions addresses is deceptively simple: Can I trust the output of this statistical procedure when applied to these particular data? Without affirmative answers to the relevant conditions, confidence intervals may fail to achieve their nominal coverage rate, p-values may be systematically too small or too large, and entire lines of scientific reasoning may rest on a shaky foundation. Condition checking transforms statistics from rote formula application into genuine statistical reasoning.

SECTION 2

Core Principles & Definitions

Every inferential procedure in statistics—whether a one-sample z-test, a two-sample t-test, a chi-square test, or a regression model—is derived under a set of mathematical assumptions. When we translate those assumptions into checkable statements about real data, we obtain conditions. Some conditions are verifiable directly (e.g., the sample size exceeds a threshold), while others can only be assessed approximately (e.g., the population distribution is roughly normal). The discipline of checking conditions before proceeding with inference is what separates rigorous statistical practice from mechanical computation.

Randomness / Independence

Data must be collected via a random mechanism—random sampling or random assignment. Observations must be independent of one another, often verified by confirming n < 10% of the population (the 10% condition).

Normality / Distribution Shape

Many procedures require the sampling distribution to be approximately normal. For proportions, this is checked with np ≥ 10 and n(1 − p) ≥ 10. For means, the CLT applies when n ≥ 30 or the population is unimodal and symmetric.

Sample Size & Success/Failure

The Large Counts condition (also called the success/failure condition) ensures enough expected successes and failures exist for the normal approximation to hold in proportion-based inference.

Equal Variance (Homoscedasticity)

In two-sample comparisons and ANOVA, we often assume that the populations have equal variances. When this fails, adjusted procedures (e.g., Welch's t-test) should be employed.

Linearity & Constant Spread (Regression)

For linear regression inference, we check that the relationship is linear, residuals are independent, residuals are normally distributed, and variance of residuals is constant across fitted values (LINE conditions).

✦

KEY TAKEAWAY

Think of conditions as the structural integrity inspection of a bridge before you drive heavy traffic across it. The bridge (the statistical formula) was engineered to bear loads under specific design specs (conditions). If the soil composition (data characteristics) doesn't match the specs, the bridge might hold, or it might not—and the consequences of failure are conclusions that mislead. Checking conditions is your engineering report: it either greenlights the crossing or flags which remediation steps (transformations, alternative methods) are needed before proceeding.

SECTION 3

Visual Explanation: The Condition-Checking Flowchart

The following diagram illustrates the general decision pathway a statistician follows when preparing to conduct inference. The process begins with identifying the procedure (e.g., one-proportion z-test, two-sample t-test, chi-square goodness of fit), then moves through a sequence of condition checks. Each check is a gate: if a condition is met, you proceed; if it is violated, you must either choose an alternative procedure or acknowledge reduced reliability in your conclusions.

The flowchart traces the three primary gates—Randomness, Independence (10% Rule), and Normality / Large Counts—that must be passed sequentially before inference is valid. Each failed gate diverts to a remediation pathway.

Notice that the flowchart is sequential: you do not skip ahead to the normality check without first confirming randomness and independence. This ordering reflects a logical hierarchy—if the data were not collected randomly, the sampling distribution theory that underpins the normality condition is itself inapplicable, rendering any subsequent check moot. In practice, many textbooks present these conditions as a checklist, but thinking of them as ordered gates reinforces the idea that later conditions presuppose earlier ones.

SECTION 4

Mathematical Framework

While condition checking is largely a qualitative process, several of the conditions have precise quantitative expressions. Understanding the mathematics behind these thresholds deepens your ability to justify whether a condition is satisfied and helps you appreciate why certain numbers (10, 30, 10%) recur throughout statistics.

Conditions for Inference on a Proportion

LARGE COUNTS CONDITION (SUCCESS/FAILURE)

np₀ ≥ 10 and n(1 − p₀) ≥ 10

Here n is the sample size and p₀ is the hypothesized (or estimated) population proportion. When both products are at least 10, the binomial distribution is well-approximated by the normal distribution, justifying the use of a z-test statistic.

10% CONDITION (INDEPENDENCE)

n < 0.10 × N

Where N is the population size. When sampling without replacement, draws are technically dependent. However, if the sample comprises less than 10% of the population, the dependence is negligible and we can treat observations as independent. This condition is a practical relaxation of the strict independence assumption.

Conditions for Inference on a Mean

CENTRAL LIMIT THEOREM THRESHOLD

n ≥ 30 (rule of thumb for approximately normal sampling distribution of x̄)

By the Central Limit Theorem, the sampling distribution of x̄ becomes approximately N(μ, σ/√n) regardless of the population shape, provided n is large enough. For strongly skewed or heavy-tailed populations, n may need to be substantially larger than 30; for populations already near normal, even n = 15 may suffice.

STANDARD ERROR OF THE SAMPLE PROPORTION

SE(p̂) = √[ p̂(1 − p̂) / n ]

This formula is valid only when the independence and large counts conditions are met. The standard error quantifies the typical distance between the sample proportion p̂ and the true population proportion p. If conditions are violated, this formula under- or overestimates variability.

💡

Why These Specific Numbers?

The thresholds 10, 30, and 10% are not magic constants—they are practical guidelines derived from simulation studies and asymptotic theory. The Large Counts threshold of 10 ensures that the skewness of the binomial distribution is small enough that a normal curve provides a close approximation (Berry–Esseen bound). The n ≥ 30 guideline for the CLT is a rough heuristic; for symmetric populations, n = 15 often suffices, while for highly skewed populations, n > 50 may be needed. Always consult the shape of the data distribution alongside these numerical thresholds.

SECTION 5

Conditions by Procedure: A Detailed Breakdown

Different statistical procedures demand different condition sets. The table below provides a comprehensive reference, organized by the type of inference. When performing any inference problem, begin by identifying the procedure and then consult the corresponding conditions. The right-most column indicates how each condition is typically verified in practice.

Procedure	Conditions Required	How to Check
1-Proportion z-test / CI	Random sample; n < 10% of N; np₀ ≥ 10 and n(1 − p₀) ≥ 10	State sampling method; verify population size; compute both products
1-Sample t-test / CI	Random sample; n < 10% of N; population ~Normal or n ≥ 30 (no strong skew/outliers)	State sampling method; verify population size; examine dotplot/histogram/Normal QQ plot
2-Sample t-test / CI	Independent random samples from each population; each n < 10% of respective N; each sample ~Normal or n ≥ 30	Confirm separate random mechanisms; check 10% for each; check distribution shape for each
Paired t-test / CI	Random sample of pairs; differences ~Normal or n ≥ 30 (examine distribution of differences)	Verify pairing structure; plot differences; check for outliers in differences
Chi-Square Test	Random sample; independent observations; all expected counts ≥ 5	State sampling; verify independence; compute expected counts for every cell
Linear Regression Inference	Linear relationship; independent residuals; Normal residuals; Equal (constant) variance of residuals (LINE)	Scatterplot for linearity; context for independence; histogram/QQ of residuals; residual-vs-fitted plot for constant spread

Conditions organized by inferential procedure

Top row: three diagnostic plots showing conditions satisfied. Bottom left: a fan-shaped residual plot revealing heteroscedasticity (condition violated). Bottom right: a summary of which plot addresses which condition.

The graphical diagnostics above illustrate that condition checking is not purely arithmetic—it is a visual and interpretive skill. A histogram that is roughly bell-shaped supports the normality assumption, while a Normal QQ plot in which points cling to the reference line provides even stronger evidence. In regression contexts, a residual vs. fitted plot is indispensable: random scatter around zero indicates constant variance, while a fan or funnel shape signals heteroscedasticity—a violation that could distort confidence intervals and inflate Type I error rates.

SECTION 6

Worked Example: Checking Conditions for a One-Proportion z-Test

A university admissions office claims that 40% of applicants who are admitted choose to enroll. A student researcher suspects the true proportion is lower. She obtains a simple random sample of 120 admitted applicants from the current cycle and finds that 39 enrolled. Before performing a one-proportion z-test at α = 0.05, she must check the conditions.

Checking Conditions: One-Proportion z-Test

1
Step 1 — State Hypotheses

The null hypothesis is H₀: p = 0.40, and the alternative is Hₐ: p < 0.40. The parameter p represents the true proportion of admitted applicants who enroll. This is a one-sided (left-tailed) test.
2
Step 2 — Random Condition

The problem states that the researcher obtained a simple random sample of 120 admitted applicants. Because the data were collected via a random mechanism, the random condition is satisfied. This ensures the sample is representative of the population of admitted applicants and guards against selection bias.

✓ Random condition met (SRS stated)
3
Step 3 — 10% Condition (Independence)

We need n < 10% of N. The sample size is n = 120. It is reasonable to assume that the university admits at least 1,200 applicants per cycle (most universities admit thousands). Therefore, 120 < 0.10 × 1,200 = 120. In practice, the population is likely much larger, so this condition is comfortably met. The 10% condition ensures that sampling without replacement does not create meaningful dependence between observations.

✓ 10% condition met (120 < 10% of at least 1,200 admitted applicants)
4
Step 4 — Large Counts Condition

Under H₀, the hypothesized proportion is p₀ = 0.40. We compute: np₀ = 120 × 0.40 = 48 ≥ 10 ✓ n(1 − p₀) = 120 × 0.60 = 72 ≥ 10 ✓ Both expected counts exceed 10, so the sampling distribution of p̂ is approximately normal under H₀. This justifies using the standard normal (z) distribution to calculate the p-value.

✓ Large Counts condition met (48 ≥ 10 and 72 ≥ 10)
5
Step 5 — Proceed with Inference

All three conditions—Random, 10%, and Large Counts—are satisfied. The researcher may proceed with the one-proportion z-test. The sample proportion is p̂ = 39/120 = 0.325. The test statistic is: z = (p̂ − p₀) / √[p₀(1 − p₀)/n] = (0.325 − 0.40) / √[0.40 × 0.60 / 120] = −0.075 / √0.002 = −0.075 / 0.04472 ≈ −1.677 The p-value for a left-tailed test is P(Z < −1.677) ≈ 0.0468. Since 0.0468 < 0.05, we reject H₀ and conclude there is statistically significant evidence that the true enrollment proportion is less than 40%.

z ≈ −1.68, p-value ≈ 0.047; reject H₀ at α = 0.05

⚠

What If a Condition Failed?

Suppose the Large Counts condition had not been met—for example, if p₀ = 0.02 and n = 120, giving np₀ = 2.4 < 10. In that case, the normal approximation to the binomial would be poor, and you should use an exact binomial test or a simulation-based approach instead of the z-test. Always state which condition failed and what alternative method you would use.

SECTION 7

Strengths, Limitations & Common Pitfalls

While checking conditions is essential for valid inference, it is not a mechanical exercise with black-and-white answers. Some conditions are more consequential than others, and the impact of a violation depends on the degree to which the condition is breached. The table below contrasts the strengths of the condition-checking framework with its inherent limitations.

Strengths	Limitations / Pitfalls
Provides a systematic, reproducible protocol that prevents analysts from blindly applying formulas to inappropriate data.	Thresholds (10, 30, 10%) are approximations—not sharp boundaries. Conditions at the margin (e.g., np₀ = 9.8) create ambiguity.
Encourages data exploration (histograms, QQ plots, residual plots) before inference, fostering deeper understanding of the data.	Graphical checks are inherently subjective; two analysts may disagree on whether a QQ plot shows 'close enough' linearity.
Connects abstract mathematical assumptions to concrete, verifiable properties of the sample and study design.	Some conditions (e.g., independence) depend on the data collection process, which may not be fully described in the problem. Analysts must sometimes assume based on context.
Alerts the analyst when alternative methods (nonparametric tests, bootstrapping, exact tests) are more appropriate.	Students may treat condition checking as a ritualistic checklist rather than an exercise in critical reasoning, writing 'condition met' without genuine evaluation.
Helps control Type I and Type II error rates by ensuring the distributional assumptions underlying p-values and confidence levels are approximately correct.	Robustness varies: t-tests are fairly robust to mild non-normality, while chi-square tests with small expected counts are not. Blanket rules can oversimplify.

Strengths versus limitations of the condition-checking framework

✦

KEY TAKEAWAY

Condition checking is to statistics what a preflight checklist is to aviation. Pilots don't skip the checklist just because the sky looks clear—each item exists because a specific failure mode was identified through experience or engineering analysis. Similarly, each statistical condition exists because violating it can produce a specific kind of inferential error. The discipline lies in treating the checklist as essential, not optional, while also understanding why each item matters, rather than checking boxes mindlessly.

SECTION 8

Connection to Advanced Theory & Methods

The conditions encountered in introductory statistics are simplified versions of deeper assumptions that surface in advanced methodology. Understanding where introductory conditions lead helps you appreciate both their power and their limitations, and prepares you for more sophisticated tools in upper-division courses and graduate study.

Introductory Condition	Advanced Generalization
Random sample / random assignment	In causal inference (Rubin's potential outcomes framework), randomization enables identification of average treatment effects. Observational studies require propensity score methods or instrumental variables when randomization is absent.
10% condition for independence	Formally, sampling without replacement follows a hypergeometric distribution. The finite population correction factor √((N − n)/(N − 1)) explicitly adjusts standard errors when n/N is non-negligible, replacing the binary '10% rule' with a continuous correction.
Large Counts / CLT (n ≥ 30)	The Berry–Esseen theorem quantifies the rate of convergence to normality, showing that the required n depends on the third absolute moment of the population distribution. Bootstrap methods and permutation tests bypass normality assumptions entirely by resampling from the data.
Constant variance (homoscedasticity)	Generalized Least Squares (GLS) and Heteroscedasticity-Consistent (HC) standard errors (White's estimator) provide valid inference even when variance is non-constant, rendering the assumption less restrictive in advanced regression.
Normality of residuals (LINE)	Generalized Linear Models (GLMs) extend regression to non-normal response distributions (e.g., Poisson for count data, logistic for binary outcomes), removing the normality requirement by modeling the conditional distribution directly.

From introductory conditions to advanced statistical methodology

The trajectory is clear: as you move to advanced statistics, the rigid conditions of introductory inference are relaxed—not because they stop mattering, but because more flexible tools are developed to handle their violations. Bootstrapping, for instance, constructs empirical sampling distributions through resampling, sidestepping the normality condition entirely. Permutation tests generate null distributions by shuffling labels, requiring only the exchangeability assumption rather than a specific distributional form. Understanding why introductory conditions exist is the foundation for understanding when and how these advanced methods improve upon them.

SECTION 9

Practice Problems

Question 1: CONCEPTUAL

Explain why the Large Counts condition (np ≥ 10 and n(1 − p) ≥ 10) is necessary for a one-proportion z-test. What would happen to the accuracy of your p-value if this condition were not met?

Show answer

The one-proportion z-test uses the standard normal distribution as an approximation to the binomial sampling distribution of p̂. This approximation works well when the distribution is roughly symmetric and bell-shaped, which occurs when both np and n(1 − p) are sufficiently large (at least 10). When this condition is violated—for example, when p is very close to 0 or 1 and n is small—the binomial distribution is markedly skewed, and the normal curve is a poor fit. As a result, the z-score does not map accurately to tail probabilities, meaning the calculated p-value may be substantially different from the true probability. In practice, this often leads to either inflated Type I error (rejecting H₀ too frequently) or reduced power (failing to detect real effects).

Question 2: BASIC CALCULATION

A researcher surveys a random sample of 85 college students and finds that 22 report using a particular study app. She wants to construct a 95% confidence interval for the true proportion of students using the app. Verify all three conditions for constructing a one-proportion z-interval.

Show answer

Random condition: The problem states the sample is random, so this condition is met. 10% condition: We need n < 10% of N. There are likely tens of thousands of college students at the university (or in the broader population of interest), so 85 < 0.10 × N is easily satisfied—for example, if N ≥ 850. Large Counts condition: We use the sample proportion p̂ = 22/85 ≈ 0.259. Then np̂ = 85 × 0.259 ≈ 22 ≥ 10 ✓, and n(1 − p̂) = 85 × 0.741 ≈ 63 ≥ 10 ✓. (For confidence intervals, we use p̂ instead of p₀ since there is no hypothesized proportion.) All three conditions are satisfied, so a one-proportion z-interval is appropriate.

Question 3: INTERMEDIATE

A hospital administrator collects data on patient wait times from a random sample of 24 emergency room visits. A histogram of the data reveals moderate right skewness with one potential outlier. The administrator wants to perform a one-sample t-test to determine if the mean wait time exceeds 45 minutes. Evaluate the conditions. Should she proceed with the t-test? Justify your reasoning.

Show answer

Random condition: The sample is described as random, so this condition is met. 10% condition: The hospital presumably handles far more than 240 ER visits (likely thousands per year), so 24 < 10% of N is satisfied. Normality condition: This is where the analysis requires judgment. With n = 24, the sample is not large enough to rely solely on the Central Limit Theorem (the n ≥ 30 guideline is not met). The histogram shows moderate right skewness and a potential outlier. The t-test is moderately robust to non-normality, but the combination of small sample size, skewness, and an outlier is concerning. The outlier, in particular, could heavily influence the sample mean and inflate the standard deviation. The administrator should consider removing the outlier and reassessing the histogram, or she should use a nonparametric alternative such as the Wilcoxon signed-rank test. Proceeding with the t-test without addressing the outlier and skewness risks unreliable results.

Question 4: APPLIED

A political polling organization wants to estimate the proportion of registered voters in a city who support a proposed ballot measure. They randomly select 500 voters from the voter registration list (the city has 42,000 registered voters). Of those surveyed, 215 support the measure. The organization plans to report a 99% confidence interval. (a) Verify all conditions. (b) If the city had only 4,000 registered voters, would any conditions need to be revisited? Explain the impact on the standard error.

Show answer

(a) Random condition: The problem states the voters were randomly selected from the voter registration list—condition met. 10% condition: n = 500, and N = 42,000. Since 500 < 0.10 × 42,000 = 4,200, the condition is met. We can treat observations as independent. Large Counts condition: p̂ = 215/500 = 0.43. Then np̂ = 500 × 0.43 = 215 ≥ 10 ✓, and n(1 − p̂) = 500 × 0.57 = 285 ≥ 10 ✓. All conditions are met. (b) If N = 4,000, then 10% of N = 400, but n = 500 > 400. The 10% condition is violated. This means observations are not approximately independent because removing 500 from 4,000 materially changes the composition of the remaining pool. To correct for this, the standard error should be multiplied by the finite population correction factor: √((N − n)/(N − 1)) = √((4,000 − 500)/(3,999)) = √(3,500/3,999) ≈ √0.8752 ≈ 0.9355. This would narrow the confidence interval by approximately 6.5%, reflecting the fact that a sample comprising 12.5% of the population carries more information than the basic formula assumes.

Question 5: CRITICAL THINKING

A researcher fits a simple linear regression model predicting GPA from hours of weekly study, using a sample of 60 college students. The residual plot (residuals vs. fitted values) shows a clear curved pattern, the Normal QQ plot of residuals shows heavy tails, and the researcher suspects that some students in the sample are from the same study group (introducing dependence). Identify which LINE conditions are violated, explain the consequences of each violation for inference, and propose at least one remediation strategy for each.

Show answer

Three of the four LINE conditions appear to be violated. Linearity (L): The curved pattern in the residual plot indicates that the true relationship between study hours and GPA is not linear—perhaps it levels off at high hours (diminishing returns). Consequence: The regression coefficients are biased estimates of the true relationship, and predictions will be systematically wrong. Remediation: Try a polynomial term (e.g., add hours²) or apply a transformation to the predictor (e.g., log(hours)). Independence (I): Students from the same study group likely have correlated outcomes, violating the independence assumption. Consequence: Standard errors will be underestimated (because the effective sample size is smaller than 60), leading to p-values that are too small and confidence intervals that are too narrow. Remediation: Use clustered standard errors or a multilevel (mixed-effects) model that accounts for the group structure. Normality (N): Heavy tails in the QQ plot suggest the residual distribution has more extreme values than a normal distribution would predict. Consequence: While OLS coefficient estimates remain unbiased, the p-values from t-tests and F-tests may be inaccurate, particularly in the tails. Remediation: Use robust regression methods or bootstrap confidence intervals, which do not require normality. The Equal variance (E) condition should also be checked; if the spread of residuals increases with fitted values, heteroscedasticity-consistent standard errors (e.g., HC3) would be appropriate. Overall, proceeding with standard regression inference without addressing these violations would produce misleading conclusions.

SUMMARY

Lesson Summary

Summary

Checking conditions is the disciplined practice of verifying that the mathematical assumptions underlying a statistical procedure are approximately satisfied by your data before you trust the results of that procedure. Every major inferential method—from one-proportion z-tests to linear regression—has a specific set of conditions. The three most universal conditions are the Random condition (data come from a random process), the 10% condition (n < 10% of N to ensure approximate independence), and the Normality / Large Counts condition (ensuring the sampling distribution is approximately normal via np ≥ 10 and n(1 − p) ≥ 10 for proportions, or n ≥ 30 / population normality for means).

Condition checking involves both quantitative verification (computing products, comparing sample size to population) and graphical diagnostics (histograms, QQ plots, residual plots). When conditions are violated, the appropriate response is not to ignore the violation but to choose an alternative method—such as nonparametric tests, bootstrap procedures, or data transformations—that does not depend on the unmet assumption. Mastering condition checking transforms statistical practice from mechanical formula application into genuine statistical reasoning.