Increased Modularity/Composability

[ParadaCarleton]

Perhaps a long-term goal since it would require quite a bit of work and might involve some breaking changes, but I'd like it if bootstrap estimates had greater modularity and composability. For instance, there's no particular reason that variance reduction techniques, like antithetic sampling, have to exclude techniques like the maximum entropy bootstrap for time series. I think you could break these features down into four separate questions:

1. What you're resampling (is this a maximum entropy bootstrap, block bootstrap, standard bootstrap, wild bootstrap, etc.)
2. Variance reduction techniques (control variates, antithetic. Maybe in the future things like drawing QMC rather than Monte Carlo samples or importance sampling?)
3. Residual/percentile (are you trying to generate the sampling distribution or the error distribution)?
4. Bootstrap distribution transformations -- e.g. BCa, studentizing, double bootstrap.
   So you could have something like:

bootstrap(data, BasicSampling, MaxEnt, BCa)

(The reason I mention 4 is that sometimes users may want to see a full bootstrap distribution, which might include a desire for corrections. I think this is very good practice and should be encouraged, since only calculating a 95% confidence interval can lead people into the trap of discounting outcomes outside the 95% interval as "basically impossible.")