Open beallen opened 4 years ago
There is covariation among the estimates, so using the full bootstrap distribution is better (try predicting based on CIs of a linear model of y=a+b*x, it will be a funnel). It also gives a handle on uncertainty in the predictions by looking at quantiles of the 100 runs.
Adding either the complete set of bootstrapped coefficients or the upper and lower bounds to help scenario modeling.