Timeline for Statistical test for increasing incidence of a rare event
Current License: CC BY-SA 3.0
9 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jul 12, 2015 at 15:21 | comment | added | Karl Ove Hufthammer | @Wrzlprmft The link is about the problem of using a correlation measure (which is useful only in the bivariate case) in a regression setting. (I understand that it’s difficult to understand, as it references a book.) Correlation measures make no sense in a regression setting (which this is). That is, you can’t properly interpret the magnitude of the correlation coefficients when one of the variables is non-random. The corresponding tests may work (I guess that depends on which correlation measure one uses). And no, the ties would not be a problem for, e.g., Poisson regression. | |
| Jul 12, 2015 at 15:12 | comment | added | Wrzlprmft | @KarlOveHufthammer: That link makes too much references to unknown stuff to be understandable to me, though it does not really seem to be about the same thing. Also, in my understanding (ignoring ties for simplicity’s sake), you can break down ranked correlation coefficients down to the question whether there is a significant trend in one a sequence of ranks, so one variable being random does not really pose a problem. Finally, the ties should be a problem for any analysis. | |
| Jul 12, 2015 at 9:40 | comment | added | Karl Ove Hufthammer | Correlations measures, like Kendalls’s τ or Spearman’s ρ, are not appropriate, as they are created for random variables, and here one of the variables (time) is obviously not random at all. See, e.g., Don’t Summarize Regression Sampling Schemes with Correlation. Besides that, the Kendalls’s τ or Spearman’s ρ tests won’t work very well, as there is a large amount of ties in the data. A regression approach would be better, e.g., a Poisson regression (with a suitable trend function) and a likelihood ratio test. | |
| Jul 3, 2015 at 18:39 | comment | added | Wrzlprmft | @rnso: Ah, point taken. | |
| Jul 3, 2015 at 13:43 | history | edited | Wrzlprmft | CC BY-SA 3.0 |
added 9 characters in body
|
| Jul 3, 2015 at 13:20 | comment | added | rnso | I made that comment since you used the word 'prevalence' in your answer. Prevalence will include previous years' cases also (unless they have died). en.wikipedia.org/wiki/… | |
| Jul 3, 2015 at 4:38 | comment | added | Wrzlprmft | @rnso: I actually meant incidence i.e. number_affected indicates new cases in that year. – that’s how I understood it and I see no contradiction. | |
| Jul 3, 2015 at 0:56 | comment | added | rnso | I actually meant incidence i.e. number_affected indicates new cases in that year. But your method of simple correlation should work for that also. | |
| Jul 2, 2015 at 19:42 | history | answered | Wrzlprmft | CC BY-SA 3.0 |