We observe a general trade-off: methods which more thoroughly account for autocorrelation often discard too much data, while methods which less thoroughly account for autocorrelation often discard too little data. We recommend a method which balances these extremes.

(4/6)

To quantitatively assess the heuristics, we create sets of synthetic data modelled on long alchemical absolute binding free energy calculations. Since we know the true unbiased mean of our synthetic data, we can calculate the errors each of the heuristics introduces.

(3/6)

We test a range of truncation point selection heuristics. Following White (doi.org/10.1177/0037...), these all work by minimising the marginal standard error, but differ in how they account for autocorrelation.

(2/6)