Timeline for Non-parametric bootstrap p-values vs confidence intervals
Current License: CC BY-SA 4.0
14 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| S Jun 22, 2019 at 17:02 | history | bounty ended | CommunityBot | ||
| S Jun 22, 2019 at 17:02 | history | notice removed | CommunityBot | ||
| Jun 14, 2019 at 22:25 | answer | added | EdM | timeline score: 3 | |
| Jun 14, 2019 at 19:36 | comment | added | Xavier Bourret Sicotte | Sampling under the null is clear when the test is a difference of means, but in many cases it is not obvious how to reproduce the null... for example the null is that the 75th percentile of two ratios is the same... how do I shift the numerators and denominators of the ratios in each sample to get that ? Also, how can I be sure that shifting the components of the ratio is actually reproducing the null ? | |
| Jun 14, 2019 at 17:37 | comment | added | jsk | I should clarify that the empirical distribution is only an approximation of the true data generating mechanism. The extent to which it's not representative of the truth will negatively impact the bootstrapped CI in unknown directions leading to less than 95% coverage. | |
| Jun 14, 2019 at 17:22 | comment | added | jsk | Isn't the simple answer just that it's clear how to sample under the null hypothesis, so there's an alternative method that is clearly better? Sampling under the bootstrap generally occurs under the empirical distribution, so the true data generating mechanism, so that clearly shouldn't be used instead of just sampling under the null. The bootstrapped CI is found from inverting the sampling distribution under the true data generating mechanism. It's true this CI may not work well, but like Dalgaard said, it's not necessarily obvious how to fix it. | |
| Jun 14, 2019 at 16:08 | comment | added | Xavier Bourret Sicotte | I think that most people will agree that when the following assumptions apply then using the CI for hypothesis test is OK: symmetrical distribution of test statistic, pivotal test statistic, CLT applying, no or few nuisance parameters etc.. but what happens when the statistic is weird or is not proven to be pivotal. Here is a real example I am working on: e.g. two sample difference between the 75th percentiles of a ratio statistic (ratio of two sums) | |
| Jun 14, 2019 at 16:03 | history | edited | Xavier Bourret Sicotte | CC BY-SA 4.0 |
Added section titles... rephrasing and simplifying this question would be great as I think part of the reason it has no answer is that it is too verbose
|
| Jun 14, 2019 at 16:01 | comment | added | Xavier Bourret Sicotte | I am starting a bounty on this question as I am very interested in gaining clarity on how and when bootstrap CIs can be used to accept / reject a hypothesis. Perhaps you could rephrase / reformat your question to make it more concise and appealing ? Thanks ! | |
| S Jun 14, 2019 at 15:59 | history | bounty started | Xavier Bourret Sicotte | ||
| S Jun 14, 2019 at 15:59 | history | notice added | Xavier Bourret Sicotte | Draw attention | |
| Mar 19, 2018 at 20:22 | history | tweeted | twitter.com/StackStats/status/975829861616021505 | ||
| Apr 13, 2017 at 12:44 | history | edited | CommunityBot |
replaced http://stats.stackexchange.com/ with https://stats.stackexchange.com/
|
|
| Oct 6, 2015 at 8:27 | history | asked | Erik | CC BY-SA 3.0 |