Hi, after running some tests on the eFAST model I found odd behaviour of the generated Confidence Intervals.

I tested the model

$$f(x_1,x_2) = x_1 a_1 + x_2 a_2 $$

for varying a_1.

Currently, the $CI(S_i)$ show no correlation to the size of $S_i$

this is due to the random sampling done here as:

sample_idx = np.random.choice(T_data, replace=True, size=n_size)

This destroys the function incorporated in the model ouput.

Unfortunately, I do not know what the correct bootstrapping method would be. Two methods that could work would be:

1. sorting after the random sampling
2. using a smaller sample but retaining order and position (e.g. sample_idx = linspace(random_int,random_int + n_size) ... of course some overflow consideration is needed)

These Methods result in:

For comparison, I did the same analysis with Sobol and a kstest. My kstests implementation is consistent with the results of method 2 and the sobol implementation in this library suggests method 1. What are your Thoughts on this ?