Confidence Interval¶
Formula¶
\[
\hat{\theta}\ \pm\ z_{\alpha/2}\,\mathrm{SE}(\hat{\theta})
\]
Parameters¶
- \(\hat{\theta}\): estimator
- \(\mathrm{SE}(\hat{\theta})\): standard error
- \(z_{\alpha/2}\): critical value for target confidence level (normal approximation)
What it means¶
A confidence interval is a data-derived range produced by a procedure that captures the true parameter with a stated long-run frequency.
What it's used for¶
- Quantifying estimator uncertainty.
- Reporting effect sizes with uncertainty, not just point estimates.
Key properties¶
- Interpretation is about the procedure, not the realized interval containing a random parameter.
- Width depends on variance, sample size, and confidence level.
Common gotchas¶
- A 95% CI does not mean "95% probability the parameter is in this specific interval" (frequentist meaning).
- Approximation formulas may fail in small samples or non-normal settings.
Example¶
Estimate a mean and report \(\hat\mu \pm 1.96\cdot \mathrm{SE}(\hat\mu)\) for an approximate 95% CI.
How to Compute (Pseudocode)¶
Input: estimator theta_hat, standard error estimate SE(theta_hat), confidence level, critical-value method
Output: confidence interval
compute the critical value for the chosen confidence level/test distribution
compute margin <- critical_value * SE(theta_hat)
return [theta_hat - margin, theta_hat + margin] # or a one-sided variant
Complexity¶
- Time: \(O(1)\) once the estimator and standard error are available
- Space: \(O(1)\)
- Assumptions: The dominant cost is usually estimating \(\hat\theta\) and \(SE(\hat\theta)\), not the interval arithmetic itself