Timeline for When to check model assumptions
Current License: CC BY-SA 4.0
20 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Oct 31, 2022 at 11:27 | history | edited | Christian Hennig | CC BY-SA 4.0 |
added 5 characters in body
|
| Oct 27, 2022 at 10:02 | history | edited | Christian Hennig | CC BY-SA 4.0 |
added 84 characters in body
|
| Oct 27, 2022 at 10:02 | comment | added | Christian Hennig | @Björn However I'll add a qualification to my original claim in point 5, so thanks! | |
| Oct 27, 2022 at 10:00 | comment | added | Christian Hennig | @Björn Fair enough. The thing in that example is that the Wilcoxon is not independent of the residuals, so this doesn't disprove what I had claimed. What I claimed is that the t-test is independent. Spanos's case is that if you only run the t-test in case residuals in some way don't reject the model, the t-test distribution conditionally on "passing the test" won't be affected. What you simulate is something more complex than that. For sure your simulation makes sense and is of interest, but it doesn't contradict the claim. | |
| Oct 27, 2022 at 9:03 | comment | added | Björn | @ChristianHennig I'm not sure what makes you think that. It's very easy to show that you get issues by doing a pre-test for normality of the residuals. We can argue about whether the issues are large enough to be truly important (but of course the size of the issue can vary substantially from situation to situation, so knowing that something can in theory cause issues should make us cautious as the issue could be large in some situations). I've now added a little example to the response below, but I'm most certainly not the first to show this. | |
| Oct 26, 2022 at 15:26 | comment | added | Christian Hennig | @Björn I'm only claiming this if independence holds. However, under normality, e.g., linear regression residuals are independent of the regression parameter estimators, and probably also of the corresponding t-statistics. Once more, I'm not saying this holds almost always or very often. | |
| Oct 26, 2022 at 15:24 | comment | added | Björn | @ChristianHennig Sure, but looking at e.g. whether QQ-plots of the residuals follow a nice line or some test of normality on them is not independent of the test statistic. In fact, most commonly done model checks are not. | |
| Oct 26, 2022 at 15:19 | comment | added | Christian Hennig | @Björn If the misspecification test is independent of the final test statistic, the distribution of the final test statistic does not change whether or not conditioned on the misspecification test result. | |
| Oct 26, 2022 at 15:17 | comment | added | Björn | @ChristianHennig Did you see some proofs or extensive simulations there? I have seen such statements from that author before in some other discussions (e.g. here: stats.stackexchange.com/questions/303887/…), but I was not clear on the basis of the statements. | |
| Oct 26, 2022 at 15:13 | comment | added | Christian Hennig | @Björn This is pretty easy to show; I have seen Aris Spanos mentioning it in some publications, I believe even the one cited in the question. | |
| Oct 26, 2022 at 15:11 | comment | added | Björn | Do you have a reference for this claim:"There are even some situations in which one can show that misspecification testing will not affect the characteristics of the later ANOVA in case the assumed model held before misspecification testing, namely where the misspecification test is independent of the finally used statistic. This happens for example when running misspecification tests on only the residuals in the linear model." I was under the impression that this is not the case. | |
| Oct 26, 2022 at 14:56 | comment | added | Christian Hennig | @NickCox Thanks. I edited #8 as the deleted "not" was in fact correct, but my original double negation there was hard to read. I'm looking at boxplots but of course also at means and other plots, so I'm not that worried about what boxplots omit. Of course people should know the limits of any plot. | |
| Oct 26, 2022 at 14:53 | history | edited | Christian Hennig | CC BY-SA 4.0 |
added 12 characters in body
|
| Oct 26, 2022 at 13:13 | comment | added | Nick Cox | Outstanding answer (+1). I made various edits that are mostly minor but please check my deletion of a "not" in #8. A small point of disagreement: I find (conventional) box plots oversold, especially in the context of ANOVA (1) especially if they don't show means too (2) because they often omit detail of importance (3) because people are over-confident about their interpretation. | |
| Oct 26, 2022 at 13:11 | history | edited | Nick Cox | CC BY-SA 4.0 |
deleted 3 characters in body
|
| Oct 26, 2022 at 11:07 | history | bounty ended | medium-dimensional | ||
| Oct 26, 2022 at 10:57 | history | edited | Christian Hennig | CC BY-SA 4.0 |
deleted 14 characters in body
|
| Oct 26, 2022 at 10:47 | history | edited | Christian Hennig | CC BY-SA 4.0 |
added 73 characters in body
|
| Oct 26, 2022 at 10:32 | history | edited | Christian Hennig | CC BY-SA 4.0 |
added 365 characters in body
|
| Oct 26, 2022 at 10:21 | history | answered | Christian Hennig | CC BY-SA 4.0 |