Lower confidence bound for the prediction-outcome correlation
rho_lower_bound.RdPlanning with the point estimate of \(\rho_{Yf}\) from a small pilot gives an underpowered study about half of the time, because the estimate is as likely to be too high as too low. This function returns the one-sided Fisher \(z\) lower bound $$\tanh\{\operatorname{atanh}(|\hat\rho|) - z_{c}/\sqrt{m - 3}\},$$ truncated at 0 and carrying the sign of \(\hat\rho\), where \(m\) is the number of labeled pairs in the pilot and \(c\) is the confidence level. Square it to get a lower bound for \(R^2\).
Details
The power functions power_ppi_mean(), power_ppi_paired(), and
power_ppi_ttest() apply this bound when a pilot size is given.