In a study of $n=100$ independent Bernoulli trials, a Wald confidence interval for the proportion $p$ is calculated as $\hat{p} \pm z_{\alpha/2} \sqrt{\frac{\hat{p}(1-\hat{p})}{n}}$. Under what circumstances is this interval most likely to exhibit poor coverage properties?