Confidence Interval Calculations

Why This Matters

Confidence intervals are the bridge between your sample data and the population you actually care about. Every time you calculate a CI, you're quantifying uncertainty: "here's my best estimate, and here's how much I trust it." This connects directly to core inference concepts like sampling distributions, standard error, degrees of freedom, and the trade-off between precision and confidence level.

The thing exams really test isn't whether you can plug numbers into formulas. It's whether you understand which formula to use and why. Different scenarios call for different distributions (Z, t, chi-squared, F) based on what you know about your population and what parameter you're estimating. So beyond memorizing formulas, focus on what assumptions each interval requires and what happens when those assumptions break down.

---

Estimating Single Population Means

When you're estimating a population mean from sample data, your choice of distribution hinges on one question: do you know the true population standard deviation, or are you estimating it from your sample?

Confidence Interval for Population Mean (Known $\sigma$)

This is the rare case where you actually know the population standard deviation, so the sampling distribution of $\bar{x}$ is exactly normal. You use the Z-distribution.

• Formula: $\bar{x} \pm Z_{\alpha/2} \left(\frac{\sigma}{\sqrt{n}}\right)$, where $\bar{x}$ is the sample mean and $n$ is sample size
• Interval width shrinks with larger $n$. The $\sqrt{n}$ in the denominator means you'd need to quadruple your sample size to cut the margin of error in half.

Confidence Interval for Population Mean (Unknown $\sigma$)

This is the far more common scenario. Substituting the sample standard deviation $s$ for $\sigma$ introduces extra uncertainty, and the t-distribution accounts for that with fatter tails.

• Formula: $\bar{x} \pm t_{\alpha/2, \, n-1} \left(\frac{s}{\sqrt{n}}\right)$, with degrees of freedom $df = n - 1$
• Converges to Z as $n$ increases. For large samples (roughly $n > 30$), the t and Z distributions become nearly identical, so the distinction matters most for small samples.

---

Estimating Proportions

Proportion intervals rely on the normal approximation to the binomial distribution. This works when your sample is large enough that the sampling distribution of $\hat{p}$ is approximately normal.

Confidence Interval for Population Proportion

• Formula: $\hat{p} \pm Z_{\alpha/2} \sqrt{\frac{\hat{p}(1 - \hat{p})}{n}}$, where $\hat{p}$ is your sample proportion
• Requires large sample verification. Check that $n\hat{p} \geq 10$ and $n(1-\hat{p}) \geq 10$ before using this formula. If either condition fails, the normal approximation isn't reliable.
• Standard error depends on $\hat{p}$. The SE is maximized when $\hat{p} = 0.5$, which is why opinion polls often assume a 50/50 split when planning sample sizes. It gives the most conservative (widest) interval.

---

Comparing Two Groups

Two-sample intervals answer the question: is there a real difference between these groups, or could sampling variability explain what we see? These show up constantly in clinical trials, A/B testing, and experimental design.

Confidence Interval for Difference Between Two Means

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

  When variances are unknown (the usual case), replace $\sigma^2$ with $s^2$ and use the t-distribution.

• Independence assumption is critical. The two samples must be drawn independently. If the same subjects appear in both groups (e.g., before/after measurements), you need a paired approach instead.

• If the interval contains zero, you cannot conclude the means differ at that confidence level. This directly parallels a two-sided hypothesis test: containing zero is equivalent to failing to reject $H_0$.

Confidence Interval for Difference Between Two Proportions

• Formula: $(\hat{p}_1 - \hat{p}_2) \pm Z_{\alpha/2} \sqrt{\frac{\hat{p}_1(1 - \hat{p}_1)}{n_1} + \frac{\hat{p}_2(1 - \hat{p}_2)}{n_2}}$
• Each group needs sufficient successes and failures. Verify the normal approximation conditions ($n\hat{p} \geq 10$ and $n(1-\hat{p}) \geq 10$) for both samples separately.
• Common in A/B testing scenarios such as comparing conversion rates, treatment success rates, or any binary outcome across two groups.

---

Estimating Variability

Sometimes you care about spread rather than center. Variance intervals use distributions specifically designed for squared quantities.

Confidence Interval for Population Variance

The quantity $(n-1)s^2/\sigma^2$ follows a chi-squared distribution with $n-1$ degrees of freedom. That's the theoretical basis for this interval.

• Formula: $\left(\frac{(n-1)s^2}{\chi^2_{\alpha/2, \, n-1}}, \quad \frac{(n-1)s^2}{\chi^2_{1-\alpha/2, \, n-1}}\right)$
• Highly sensitive to normality. This interval is much less robust than mean-based intervals. Non-normal data can severely distort results, so always check that assumption.
• Notice the interval is asymmetric around $s^2$ because the chi-squared distribution is right-skewed.

Confidence Interval for Ratio of Two Variances

The ratio of two independent chi-squared variables (each divided by their degrees of freedom) follows an F-distribution. This lets you compare variability across two groups.

• Formula: $\left(\frac{s_1^2}{s_2^2} \cdot \frac{1}{F_{\alpha/2, \, n_1-1, \, n_2-1}}, \quad \frac{s_1^2}{s_2^2} \cdot F_{1-\alpha/2, \, n_1-1, \, n_2-1}\right)$
• Tests equality of variances. If the interval contains 1, you cannot conclude the variances differ. This matters practically because it helps you decide between pooled and unpooled (Welch's) t-tests.

---

Relationship and Model Parameters

When you move beyond location and spread into relationships between variables, you need intervals for correlation and regression coefficients.

Confidence Interval for Correlation Coefficient

The sampling distribution of $r$ is skewed, especially when the true correlation is near $\pm 1$. Fisher's z-transformation fixes this by mapping $r$ onto a scale where the sampling distribution is approximately normal.

Here's the process:

1. Transform: $z' = \frac{1}{2} \ln\left(\frac{1+r}{1-r}\right)$
2. Build the CI on the z' scale: $z' \pm Z_{\alpha/2} \cdot \frac{1}{\sqrt{n-3}}$
3. Back-transform the endpoints to get the CI on the original $r$ scale (using the inverse formula).

The transformation stabilizes the variance so that the standard error is approximately $\frac{1}{\sqrt{n-3}}$ regardless of the true correlation value.

Confidence Interval for Regression Coefficients

• Formula: $\hat{\beta} \pm t_{\alpha/2, \, n-k} \cdot SE(\hat{\beta})$, where $k$ is the number of parameters estimated (including the intercept)
• Directly tests predictor significance. If the interval for a slope excludes zero, that predictor has a statistically significant linear relationship with the response at that confidence level.
• Assumes standard regression conditions: linearity, independence, homoscedasticity, and normally distributed errors (often remembered as the LINE assumptions).

---

Non-Parametric Approaches

When your data violates distributional assumptions or you're estimating a statistic without a known theoretical distribution, traditional formulas may not apply. That's where resampling methods come in.

Bootstrap Confidence Intervals

The bootstrap works by resampling your observed data with replacement thousands of times. Each resample gives you a new estimate of your statistic, and together they build an empirical sampling distribution.

• No distributional assumptions required. This makes bootstrap useful for medians, ratios, or any statistic where theoretical distributions are unknown or intractable.
• Multiple methods exist. The percentile method (simplest: just take the 2.5th and 97.5th percentiles of your bootstrap estimates for a 95% CI), BCa (bias-corrected and accelerated), and basic bootstrap each handle bias and skewness differently.

---

Quick Reference Table

---

Self-Check Questions

1. You're estimating a population mean from a sample of 25 observations, and you calculated the standard deviation from your data. Which distribution do you use, and why does it matter?

2. Compare and contrast the confidence interval for a single proportion versus the difference between two proportions. What's similar about their structures, and what additional consideration applies to the two-sample case?

3. Both chi-squared and F-distributions are used for variance-related intervals. When would you use each, and what assumption do they share that makes them sensitive to violations?

4. A colleague builds a 95% CI for a regression slope and finds it includes zero. Another builds a 95% CI for the correlation between the same two variables and finds it excludes zero. Is this possible? What might explain this?

5. When would you choose bootstrap confidence intervals over traditional parametric methods? Give two specific scenarios where bootstrap would be the better choice.