Confidence Intervals

Why This Matters

Confidence intervals let you quantify how uncertain a sample-based estimate really is. Every time you estimate a parameter from data (a mean strength, a failure rate, a proportion of defective parts), you need to communicate how much that estimate could vary. CIs do exactly that, connecting directly to concepts like sampling distributions, the Central Limit Theorem, and hypothesis testing.

Don't just memorize the formulas. Know when each interval type applies and why the mechanics work. Can you explain why larger samples shrink your interval? Why you switch from Z to t? Why a 99% CI is wider than a 95% CI? These conceptual questions show up constantly, and the formulas alone won't save you. Master the underlying logic: confidence level trade-offs, sample size effects, distributional assumptions, and the duality between CIs and hypothesis tests.

---

Foundational Concepts

Before diving into specific interval formulas, you need a solid understanding of what a confidence interval actually means and what controls its width. The confidence level determines how often the procedure captures the true parameter; the margin of error determines how precise your estimate is.

Definition and Interpretation

A confidence interval is a range of plausible values for an unknown population parameter, constructed from sample data. It is not a probability statement about where the parameter lies after you've computed it.

The confidence level (e.g., 95%) means that if you repeated the sampling procedure many times, about 95% of the resulting intervals would contain the true parameter. The width of the interval reflects your uncertainty: wider intervals mean less precision, narrower intervals mean your point estimate is more tightly pinned down.

Confidence Level and Significance Level

• Confidence level $1 - \alpha$ and significance level $\alpha$ are complementary. A 95% CI corresponds to $\alpha = 0.05$.
• Higher confidence levels produce wider intervals because you're demanding more certainty that you've captured the true value.
• The trade-off is precision vs. confidence. You can't have both a narrow interval and high confidence without increasing your sample size.

Margin of Error

The margin of error (ME) is the "±" portion of your estimate, equal to half the total interval width. It quantifies the maximum expected sampling error.

ME depends on three factors:

• Confidence level: higher confidence → larger ME
• Sample size: larger $n$ → smaller ME
• Data variability: more spread in the data → larger ME

In practice, you'll often specify a required ME first, then solve for the sample size needed to achieve it.

---

Choosing the Right Distribution

The critical decision in constructing any CI is selecting the appropriate sampling distribution. Your choice depends on what you know about the population and how large your sample is.

Z-Score Intervals

Use the Z-distribution when $\sigma$ (population standard deviation) is known and either the population is normal or $n > 30$ (so the CLT applies).

Z critical values are fixed for each confidence level:

• 90% confidence: $Z = 1.645$
• 95% confidence: $Z = 1.96$
• 99% confidence: $Z = 2.576$

This situation is common in quality control settings where historical process data provides a reliable $\sigma$ value.

T-Score Intervals

Use the t-distribution when $\sigma$ is unknown and you estimate it with the sample standard deviation $s$.

Degrees of freedom ($df = n - 1$) control the shape. Smaller samples produce heavier tails, which makes the interval wider to account for the extra uncertainty in estimating $\sigma$. As $n \to \infty$, the t-distribution converges to the Z-distribution. Using t is always valid when $\sigma$ is unknown, even for large samples.

Chi-Squared Distribution for Variance

When you need a CI for a population variance (not a mean), you use the $\chi^2$ distribution. This works because the quantity $\frac{(n-1)s^2}{\sigma^2}$ follows a chi-squared distribution when the data is normal.

The resulting interval is asymmetric:

$\left(\frac{(n-1)s^2}{\chi^2_{\alpha/2}}, \frac{(n-1)s^2}{\chi^2_{1-\alpha/2}}\right)$

Note the "flipped" critical values in the denominators. The normality assumption is critical here; variance intervals are much more sensitive to non-normality than mean intervals.

---

Intervals for Means

These are your workhorse formulas. The structure is always the same: point estimate ± (critical value) × (standard error).

CI for Mean (Known $\sigma$)

$\bar{x} \pm Z_{\alpha/2} \cdot \frac{\sigma}{\sqrt{n}}$

The term $\frac{\sigma}{\sqrt{n}}$ is the standard error of the mean. This formula requires the sampling distribution of $\bar{x}$ to be approximately normal, which holds when the population is normal or $n > 30$.

This scenario is rare in practice since $\sigma$ is seldom truly known, but it's common on exams to test your understanding of the Z framework.

CI for Mean (Unknown $\sigma$)

$\bar{x} \pm t_{\alpha/2, \, n-1} \cdot \frac{s}{\sqrt{n}}$

Here $s$ replaces $\sigma$, and the t critical value depends on $df = n - 1$. Always check your t-table or calculator with the correct degrees of freedom.

This is the default real-world scenario. When in doubt on an exam and $\sigma$ isn't explicitly given, use t.

CI for Difference Between Two Means

For independent samples with known variances:

$(\bar{x}_1 - \bar{x}_2) \pm Z_{\alpha/2} \sqrt{\frac{\sigma_1^2}{n_1} + \frac{\sigma_2^2}{n_2}}$

For unknown variances (the more common case), use a t-interval (pooled or Welch's):

$(\bar{x}_1 - \bar{x}_2) \pm t_{\alpha/2} \sqrt{\frac{s_1^2}{n_1} + \frac{s_2^2}{n_2}}$

If the resulting interval contains zero, you cannot conclude the means differ at that confidence level. This connects directly to two-sample hypothesis tests.

---

Intervals for Proportions

Proportion intervals apply when your data is categorical (success/failure, defective/non-defective). The normal approximation requires checking sample size conditions before you use these formulas.

CI for Single Proportion

$\hat{p} \pm Z_{\alpha/2} \sqrt{\frac{\hat{p}(1-\hat{p})}{n}}$

where $\hat{p} = x/n$ is the sample proportion.

Check conditions before using this formula: $n\hat{p} \geq 10$ and $n(1-\hat{p}) \geq 10$. These ensure the normal approximation to the binomial is reasonable. This interval is used extensively to estimate defect rates, failure probabilities, and compliance percentages.

CI for Difference Between Two Proportions

$(\hat{p}_1 - \hat{p}_2) \pm Z_{\alpha/2} \sqrt{\frac{\hat{p}_1(1-\hat{p}_1)}{n_1} + \frac{\hat{p}_2(1-\hat{p}_2)}{n_2}}$

Both samples must independently satisfy the normal approximation conditions ($n_i\hat{p}_i \geq 10$ and $n_i(1-\hat{p}_i) \geq 10$ for each group). If zero is in the interval, you cannot conclude the proportions differ.

---

Sample Size and Interval Design

Often you work backwards: given a desired precision, how much data do you need? Sample size determination is about controlling the margin of error before collecting data.

Sample Size Effects

Larger $n$ shrinks the interval because standard error has $\sqrt{n}$ in the denominator. But the relationship is a square root, so the returns diminish quickly:

• To cut ME in half, you need $4\times$ the sample size.
• To cut ME to one-third, you need $9\times$ the sample size.

This means there are real resource constraints. You must balance statistical precision against time, cost, and feasibility.

One-Sided vs. Two-Sided Intervals

Two-sided intervals (the most common type) bound the parameter above and below: "we're 95% confident $\mu$ is between A and B."

One-sided intervals provide only an upper or lower bound: "we're 95% confident $\mu$ is at least A" or "we're 95% confident $\mu$ is at most B." Use one-sided intervals when the research question is directional, for example, "does the new process reduce defects?" only cares about the lower bound on improvement.

---

Assumptions and Connections

Valid inference requires meeting assumptions, and CIs connect directly to hypothesis testing. Understanding these relationships separates strong exam performance from formula memorization.

Assumptions for Valid CIs

Random sampling and independence are non-negotiable. Biased samples produce meaningless intervals regardless of formula correctness.

• For means: population normality or $n > 30$ (CLT kicks in)
• For proportions: $n\hat{p} \geq 10$ and $n(1-\hat{p}) \geq 10$
• For variance: the chi-squared method requires strict normality; variance intervals are the most sensitive to this assumption

CI and Hypothesis Test Duality

A two-sided $(1-\alpha)$ CI corresponds to a two-tailed test at significance level $\alpha$. If the null value falls outside the CI, you reject $H_0$. If it falls inside, you fail to reject.

This gives you a visual decision rule: construct the CI, then check whether the hypothesized value is inside or outside. CIs also give more information than tests alone because they show the full range of plausible values, not just a binary "reject" or "fail to reject."

Interpreting and Reporting CIs

Never say "there's a 95% probability the parameter is in this interval." The parameter is fixed; the interval is what's random across repeated samples.

Correct interpretation: "We are 95% confident that [parameter] lies between [lower bound] and [upper bound]."

When reporting, state the confidence level, sample size, assumptions you checked, and what the result means in context.

---

Quick Reference Table

---

Self-Check Questions

1. You're estimating the mean tensile strength of a new alloy. You have $n = 25$ specimens and calculate $s = 12$ MPa. Should you use a Z or t interval, and why does this choice affect your interval width?

2. Compare the CI for a single proportion to the CI for the difference between two proportions. What additional complexity arises in the two-sample case, and what conditions must both samples satisfy?

3. An engineer wants to estimate a defect rate with a margin of error of ±2% at 95% confidence. If a pilot study suggests $\hat{p} \approx 0.10$, approximately what sample size is needed? How would this change if $\hat{p} \approx 0.50$?

4. You construct a 95% CI for $\mu_1 - \mu_2$ and obtain $(-3.2, 1.8)$. What can you conclude about a hypothesis test of $H_0: \mu_1 = \mu_2$ at $\alpha = 0.05$? What if the interval were $(-3.2, -0.4)$?

5. Explain why a 99% confidence interval is wider than a 95% confidence interval for the same data. If you needed to maintain the same width while increasing confidence from 95% to 99%, what would you have to change?