= Symmetric confidence-interval consistency identities
{title2=$L+U=2\widehat\theta,\quad U-L=2c\operatorname{SE}(\widehat\theta)$}
For a symmetric interval of the form $[L,U]=[\widehat\theta-c\operatorname{SE}(\widehat\theta),\widehat\theta+c\operatorname{SE}(\widehat\theta)]$, where $c\ge0$ and the <standard error> is nonnegative, adding and subtracting the endpoints proves the displayed identities. Thus the endpoints must be ordered and their midpoint must equal the point estimate. Rounding should be assessed using intervals of possible unrounded values before declaring a numerical contradiction.
When this <confidence interval> is obtained by inverting a matching <two-sided test> of $\theta=\theta_0$, exclusion of $\theta_0$ agrees with rejection at the corresponding <significance level>, with the usual convention at boundary equality. This is a check for a specified symmetric interval procedure, such as the usual t interval for a mean difference; it does not require asymmetric intervals or intervals on transformed scales to have their midpoint at the estimate on the original scale.
Back to article page
