COLLEGE STATISTICS • PROBLEM-SOLVING & STATISTICAL REASONING

Choosing Statistical Procedures — Choosing an Appropriate Procedure (CI vs Test; 1-sample vs 2-sample)

Learn to match every research question to the right inference tool by distinguishing estimation from testing and one-sample from two-sample designs.

Historical Context & Motivation

The modern statistical toolkit did not emerge fully formed; it was assembled piece by piece over roughly a century as researchers confronted increasingly sophisticated questions about populations. Early work focused on estimation — determining the plausible range of a population parameter — while parallel developments produced formal hypothesis testing frameworks for deciding whether observed data were compatible with a specific claim. Understanding why these two modes of inference arose, and how one-sample versus two-sample designs became formalized, clarifies why choosing the right procedure is the essential first step of any statistical analysis.

1908
Student's t-Distribution
William Sealy Gosset, writing under the pseudonym "Student," derived the t-distribution for small-sample inference about a single mean — the first rigorous one-sample procedure for settings where σ is unknown.
1925
Fisher's Statistical Methods
R. A. Fisher published Statistical Methods for Research Workers, formalizing significance testing and the p-value as tools for evaluating hypotheses about population parameters.
1933
Neyman–Pearson Framework
Jerzy Neyman and Egon Pearson introduced the concepts of Type I and Type II errors, power, and the formal alternative hypothesis, creating the decision-theoretic foundation for modern hypothesis tests.
1934
Confidence Intervals Formalized
Neyman published his theory of confidence intervals, establishing interval estimation as a complementary — and sometimes preferred — mode of inference distinct from binary hypothesis testing.
1940s–60s
Two-Sample and Multi-Sample Extensions
Procedures for comparing two population means (pooled and Welch t-tests) and two proportions became standard in textbooks, extending both confidence interval and hypothesis testing frameworks to comparative studies.

This timeline reveals a crucial insight: the statistical community developed estimation and testing as complementary tools, not interchangeable ones. A confidence interval answers "What is the parameter likely to be?" while a hypothesis test answers "Is this specific claim about the parameter plausible?" Simultaneously, the distinction between studying a single population and comparing two populations gives rise to fundamentally different designs, assumptions, and formulas. The question this lesson addresses is deceptively simple: given a real-world research scenario, how do you select the correct procedure?

Core Principles & Decision Criteria

Selecting the appropriate statistical procedure requires you to answer a short sequence of diagnostic questions about the research scenario. These questions revolve around two orthogonal axes: the goal of inference (estimation vs. testing) and the number of populations (one vs. two). Crossing these two axes produces a 2 × 2 decision matrix that covers the majority of introductory procedures. The four foundational principles below guide you through this classification.

1

Identify the Goal: Estimate or Decide?

If the research question asks "how much" or "what is the value," you need a confidence interval. If it asks "is there a difference" or "is the claim supported," you need a hypothesis test.
2

Count the Populations or Groups

A single sample drawn from one population calls for a one-sample procedure. When two independent samples (or matched pairs) from two populations are compared, a two-sample procedure is required.
3

Classify the Parameter

Determine whether you are making inferences about a population mean (μ) or a population proportion (p). This choice controls the test statistic formula and the shape of the sampling distribution you invoke.
4

Check Conditions Before Proceeding

Every procedure carries validity conditions — randomness of sampling, sample size thresholds, normality assumptions, or independence criteria. Verifying these conditions is part of choosing, not just performing, the procedure.
KEY TAKEAWAY
Think of your statistical toolkit like a physician's diagnostic protocol. A doctor first asks, "Am I measuring a quantity (like blood pressure) or making a binary decision (like whether to prescribe medication)?" — that is the CI vs. test axis. Then the doctor asks, "Am I looking at one patient relative to a reference value, or comparing two patients?" — that is the one-sample vs. two-sample axis. Just as a doctor would not prescribe treatment without first diagnosing the condition, you should never compute a statistic without first diagnosing which procedure the question demands.

Decision Flowchart — Mapping Questions to Procedures

The following flowchart translates the four principles from Section 2 into a step-by-step decision tree. Starting from the top, you classify the research question along two axes and arrive at one of four procedure families. Each terminal node names the specific procedure and references the formula you would use. Study this diagram carefully — it is the single most useful tool for choosing procedures on exams and in practice.

The flowchart begins with the research goal (estimate vs. test), then branches by the number of groups, and terminates at four procedure families. Dashed lines connect each procedure to the shared conditions box at the bottom.

Notice that every path through the flowchart requires exactly two decisions — goal and number of groups — before arriving at a procedure. The parameter type (mean vs. proportion) then determines which formula you plug into. The conditions box at the bottom serves as a final checkpoint: no matter which procedure you select, you must verify that the sample was collected randomly, that observations are independent, and that the sampling distribution is approximately normal (either by the Central Limit Theorem or by explicit checks on np and nq for proportions).

Mathematical Framework — Key Formulas

Each of the four procedure families identified in the flowchart has a signature formula. Understanding these formulas side by side reveals that they share a common architecture: every confidence interval is of the form point estimate ± (critical value)(standard error), and every test statistic is of the form (point estimate − null value) / standard error. Recognizing this shared architecture makes memorization unnecessary; you need only determine which point estimate, which standard error, and which null value to use.

ONE-SAMPLE CONFIDENCE INTERVAL FOR μ
x̄ ± t* × (s / √n)
x̄ = sample mean; t* = critical value from the t-distribution with df = n − 1; s = sample standard deviation; n = sample size. Use when σ is unknown and data are approximately normal or n ≥ 30.
ONE-SAMPLE HYPOTHESIS TEST FOR μ
t = (x̄ − μ₀) / (s / √n)
μ₀ = hypothesized population mean (from H₀). The test statistic t follows a t-distribution with df = n − 1. Compare the p-value to α to reach a conclusion.
TWO-SAMPLE CONFIDENCE INTERVAL FOR μ₁ − μ₂
(x̄₁ − x̄₂) ± t* × √(s₁²/n₁ + s₂²/n₂)
The standard error combines variability from both samples. Degrees of freedom are computed via the Welch approximation (or the conservative min(n₁ − 1, n₂ − 1)). The interval estimates the difference in population means.
TWO-SAMPLE HYPOTHESIS TEST FOR μ₁ − μ₂
t = (x̄₁ − x̄₂ − 0) / √(s₁²/n₁ + s₂²/n₂)
Under H₀: μ₁ − μ₂ = 0. The numerator subtracts zero (the hypothesized difference), and the denominator is the pooled standard error. The resulting t is compared against the t-distribution to compute a p-value.
📐 Proportions Follow the Same Pattern
For proportions, replace x̄ with p̂ and use the z-distribution instead of t. The one-sample CI uses SE = √(p̂q̂/n), while the one-sample test uses SE = √(p₀q₀/n) with the null proportion. For two-sample proportion tests, the standard error is computed under the pooled proportion p̂c = (x₁ + x₂)/(n₁ + n₂). The key architectural insight — point estimate ± critical value × SE for CIs, and (estimate − null)/SE for tests — remains identical.

Detailed Classification — Matching Scenarios to Procedures

In practice, research questions arrive as prose, not as tidy mathematical statements. The challenge lies in translating verbal cues into the correct procedure. The table below catalogs common phrasings and maps each to the appropriate procedure. After the table, a second diagram organizes all eight procedures (four for means, four for proportions) into a comprehensive reference grid.

Common problem phrasings mapped to procedures
Key Phrase in the ProblemGoalSamplesProcedure
"Estimate the average…"CI11-sample t interval for μ
"What proportion of…"CI11-sample z interval for p
"Is the mean different from 50?"Test11-sample t test for μ
"Is the pass rate above 0.70?"Test11-sample z test for p
"Estimate the difference in means between Group A and Group B."CI22-sample t interval for μ₁ − μ₂
"Is there a difference in proportions between men and women?"Test22-sample z test for p₁ − p₂
"Do students score higher after the intervention than before?"TestPairedPaired t test (1-sample t on differences)
"Estimate by how much Drug A lowers blood pressure more than Drug B."CI22-sample t interval for μ₁ − μ₂
The 2 × 2 × 2 grid organizes all eight introductory inference procedures. The left half covers means (t-based), the right half covers proportions (z-based). The top row is one-sample, the bottom row is two-sample. Gold condition boxes summarize the requirements for each quadrant. The note at the bottom reminds you that paired data reduce to a one-sample procedure on the differences.
⚠️ Paired Data — A Common Trap
Paired or matched-pairs designs (before/after, twin studies, repeated measures) appear to involve two samples, but they should not be analyzed with a two-sample procedure. Instead, compute the difference d = x₁ − x₂ for each pair and apply a one-sample t procedure on d̄. The null hypothesis is typically H₀: μd = 0. Treating paired data as independent two-sample data inflates the standard error and reduces power.

Worked Example — From Research Question to Procedure

Consider the following scenario: A university dining services director wants to know whether the average daily caloric intake of students living on campus differs from the recommended 2,000 calories. She collects a random sample of 45 students and records their average daily caloric intake over one week. The sample mean is x̄ = 2,140 calories with a sample standard deviation of s = 310 calories. She wants to test at the α = 0.05 significance level.

Choosing and Executing the Correct Procedure
1
Step 1 — Identify the GoalThe question asks whether the mean differs from a specific value (2,000). The phrase "differs from" signals a comparison against a benchmark, not an estimation. Therefore the goal is a hypothesis test, not a confidence interval.
Goal → Hypothesis test
2
Step 2 — Count the PopulationsThere is only one sample of 45 students. The comparison is between this sample's mean and a fixed reference value (2,000 calories), not between two groups of students. This is a one-sample scenario.
Design → One-sample
3
Step 3 — Classify the ParameterCaloric intake is a quantitative variable, and the parameter of interest is the population mean μ. Since σ is unknown (we have s = 310), we use the t-distribution.
Parameter → μ (mean); Distribution → t
4
Step 4 — State the HypothesesH₀: μ = 2,000 (the population mean caloric intake equals the recommended value). Hₐ: μ ≠ 2,000 (two-sided test because the director asked whether intake "differs," not whether it exceeds or falls below the recommendation).
H₀: μ = 2000; Hₐ: μ ≠ 2000
5
Step 5 — Verify ConditionsRandom: the problem states a random sample. Independence: 45 students is likely less than 10% of the campus population. Normality: n = 45 ≥ 30, so by the Central Limit Theorem the sampling distribution of x̄ is approximately normal. All conditions are satisfied.
Conditions ✓ — proceed with 1-sample t test
6
Step 6 — Compute the Test Statistict = (x̄ − μ₀) / (s / √n) = (2140 − 2000) / (310 / √45) = 140 / (310 / 6.708) = 140 / 46.21 ≈ 3.03. This test statistic follows a t-distribution with df = 45 − 1 = 44.
t = 3.03, df = 44
7
Step 7 — Find the p-value and ConcludeFor a two-sided test with t = 3.03 and df = 44, the p-value ≈ 0.004. Since 0.004 < 0.05 = α, we reject H₀. There is statistically significant evidence that the average caloric intake of on-campus students differs from the recommended 2,000 calories.
p ≈ 0.004 — Reject H₀
💡 What if She Had Asked a Different Question?
If the director had instead asked, "What is the average daily caloric intake of on-campus students?" the goal would shift to estimation, and the correct procedure would be a one-sample t confidence interval for μ: x̄ ± t* × (s / √n) = 2140 ± 2.015 × 46.21 ≈ (2046.9, 2233.1). Notice that 2,000 falls outside this 95% CI, which is consistent with rejecting H₀ at α = 0.05 in the test above — the CI and the test always agree at the same confidence/significance level.

CI vs. Test — Strengths, Limitations, and When to Prefer Each

Confidence intervals and hypothesis tests are not competing methods — they are complementary lenses on the same data. Nevertheless, each has distinct strengths and limitations, and certain contexts favor one over the other. The American Statistical Association's 2016 statement on p-values, for instance, encouraged researchers to report confidence intervals alongside or instead of p-values to convey effect size and precision. Understanding these trade-offs deepens your ability to choose wisely.

Comparing confidence intervals and hypothesis tests across five dimensions
DimensionConfidence IntervalHypothesis Test
Primary outputA range of plausible values for the parameterA p-value and a binary decision (reject or fail to reject H₀)
Communicates effect size?Yes — the width and location of the interval show both magnitude and precisionNot directly — statistical significance ≠ practical significance
Addresses a specific claim?Indirectly — check whether the claimed value falls inside or outside the intervalDirectly — designed to evaluate H₀ against Hₐ
Best when…The question is exploratory or the audience needs to understand the parameter's magnitudeThe question is confirmatory and requires a decision (e.g., regulatory approval, quality control)
Common misinterpretation"There is a 95% probability the parameter is in this interval" (wrong — it either is or isn't; 95% refers to the method's long-run coverage)"The p-value is the probability H₀ is true" (wrong — it is the probability of data this extreme given H₀ is true)
KEY TAKEAWAY
Think of a confidence interval as a satellite image: it gives you the full landscape, showing where the parameter probably lies and how much uncertainty remains. A hypothesis test, by contrast, is a spotlight: it illuminates one specific claim and tells you whether that claim is consistent with the data. In research, you often want both views — the satellite image for understanding and the spotlight for decision-making. When forced to choose, remember that a CI at confidence level (1 − α) and a two-sided test at significance level α will always yield the same conclusion about whether the null value is plausible.

Connection to Advanced Procedures

The one-sample and two-sample inference procedures for means and proportions are the building blocks of a much larger edifice. As you progress through statistics coursework, you will encounter procedures that extend these ideas in natural ways. The table below previews how the introductory procedures connect to more advanced methods, reinforcing that the decision-making framework you are learning now — identify the goal, count the groups, classify the parameter, check conditions — remains applicable throughout.

How introductory procedures generalize to advanced methods
Introductory ProcedureAdvanced ExtensionWhat Changes
1-sample / 2-sample t test for meansOne-way ANOVA (F-test)Extends to k ≥ 3 groups; uses F-distribution instead of t
2-sample z test for proportionsChi-square test of homogeneityCompares proportions across k ≥ 2 groups; uses χ² distribution
1-sample t test (is μ = μ₀?)Simple linear regression (is β₁ = 0?)Tests a slope rather than a mean; same t-statistic structure
Confidence interval for μPrediction intervals and bootstrap CIsPrediction intervals account for individual variability; bootstrap CIs relax normality assumptions
Paired t testRepeated-measures ANOVA / mixed modelsExtends paired comparisons to multiple time points or conditions with random effects

The most important takeaway from this table is that every advanced procedure is a generalization of a simpler one. ANOVA reduces to a two-sample t test when k = 2 groups. A chi-square test with two categories is algebraically equivalent to a two-sample z test for proportions. The regression t test for a slope uses the same t-statistic logic as a one-sample t test applied to the estimated coefficient. By mastering the decision framework now — CI vs. test, one-sample vs. two-sample, mean vs. proportion — you are laying the foundation for every inferential method you will encounter in the future.

Practice Problems

PROBLEM 1CONCEPTUAL
A researcher asks: "What is the average commute time of employees at this company?" She plans to survey a random sample of 60 employees. Should she construct a confidence interval or perform a hypothesis test? Explain your reasoning, and state whether this is a one-sample or two-sample scenario.
PROBLEM 2BASIC CALCULATION
A quality control engineer claims that a machine fills bottles to a mean of exactly 500 mL. A random sample of n = 36 bottles has x̄ = 497.5 mL and s = 8.4 mL. Identify the correct procedure, state the hypotheses, and compute the test statistic.
PROBLEM 3INTERMEDIATE
A political analyst wants to estimate the difference in support for a ballot initiative between suburban and urban voters. She surveys 400 suburban voters (52% support) and 350 urban voters (61% support). Identify the correct procedure, justify your choice, and explain which standard error formula you would use.
PROBLEM 4APPLIED
A pharmaceutical company tests whether a new drug lowers systolic blood pressure more than a placebo. 80 patients are randomly assigned to the drug group (x̄₁ = 128 mmHg, s₁ = 14) and 80 to the placebo group (x̄₂ = 135 mmHg, s₂ = 16). The company needs FDA approval, which requires demonstrating that the drug effect is statistically significant. (a) Identify the correct procedure. (b) Would a confidence interval have been equally appropriate? Justify your answer.
PROBLEM 5CRITICAL THINKING
A researcher measures the resting heart rate of 50 athletes before and after a 12-week training program. She wants to determine whether the training changed heart rates. A classmate recommends a two-sample t test comparing the "before" group (n = 50) with the "after" group (n = 50). (a) Explain why this recommendation is incorrect. (b) Identify the correct procedure. (c) Describe how the standard error in the correct procedure differs from the one in the two-sample approach, and explain why this matters for statistical power.

Lesson Summary

Choosing the correct statistical procedure begins with two fundamental questions. First, determine the goal of inference: if the research question asks "what is the value" or "how large is the difference," use a confidence interval; if it asks "is there evidence that" or "does the parameter equal," use a hypothesis test. Second, count the number of populations or groups: a single sample compared to a reference value calls for a one-sample procedure, while two independent groups call for a two-sample procedure. Remember that paired data should be reduced to differences and analyzed with a one-sample t procedure.

All procedures share a common architecture: confidence intervals take the form point estimate ± critical value × standard error, and test statistics take the form (point estimate − null value) / standard error. The parameter type — mean (μ) or proportion (p) — determines which standard error formula and which sampling distribution (t or z) you invoke. Before computing anything, always verify the validity conditions: randomness, independence (10% rule), and normality or large-sample requirements. Mastering this decision framework ensures that you apply the right tool to every inferential question you encounter.

Varsity Tutors • College Statistics • Choosing Statistical Procedures — Choosing an Appropriate Procedure (CI vs Test; 1-sample vs 2-sample)