[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

From |
"Nick Cox" <n.j.cox@durham.ac.uk> |

To |
<statalist@hsphsun2.harvard.edu> |

Subject |
RE: st: Chi-square test with probabilities |

Date |
Tue, 11 Nov 2008 19:48:18 -0000 |

What makes you think this is a chi-square problem? If you can supply a standard error for each probability then at a stretch each probability can be represented as z = (observed - expected) / se and the sum of those z-squareds might seem like a chi-square statistic losing 1 d.f. for the constraint that probabilities add to 1 and more if anything is estimated from the data. So where are your standard errors coming from? My gut feeling, set down before Maarten Buis says something similar, is that you would be better setting up (a) one or more measures of discrepancy and (b) a model specifically for your problem and getting sampling distributions by simulation. Nick n.j.cox@durham.ac.uk Christoph Merkle I have to clarify this a bit: The probabilities indeed add up to one, participants were made aware of to obey this rule. And if there were still some errors I corrected for those. They actually had to estimate with which probability a value falls in a certain quartile/decile of a distribution. My claim is not 'humans overestimate the chances of rare events happening' but rather 'humans violate the rules of Bayesian updating". Therefore ttests for the mean guessed probability compared to expected probability for quartiles/deciles one by one are fine for a start (and I calculated them already), but better would be to analyze the whole distribution. That is why I proposed a chi square goodness of fit test. I'm not an expert in repeated measure design, but I don't think it helps me out here. Zitat von Ronan Conroy <rconroy@rcsi.ie>: > On 11 Nov 2008, at 17:19, Christoph Merkle wrote: > >> Actually I'm only interested if the mean of these peobabilities over >> participants is different from hyposized proportions. If I use a >> simple ttest I can only test each of the variables one by one. But I >> want to test the distribution over the ten > > If the events are mutually exclusive and collectively exhaustive, then > the probabilities ought to add up to 1, but I fear that they won't. > > I think you are better testing one-by-one, using a t-test to test the > hypothesis that the mean guessed probability is the same as the actual > value, unless you have a hypothesis that is independent of the > probability being guessed (such as 'humans overestimate the chances of > rare events happening') in which case, I would treat it as a repeated > measures design. * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/

**References**:**st: Chi-square test with probabilities***From:*Christoph Merkle <chmerkle@rumms.uni-mannheim.de>

**Re: st: Chi-square test with probabilities***From:*Ronan Conroy <rconroy@rcsi.ie>

**Re: st: Chi-square test with probabilities***From:*Christoph Merkle <chmerkle@rumms.uni-mannheim.de>

- Prev by Date:
**Re: st: RE: Multilevel models and multiple imputation** - Next by Date:
**Re: st: Sort on time** - Previous by thread:
**Re: st: Chi-square test with probabilities** - Next by thread:
**[no subject]** - Index(es):

© Copyright 1996–2014 StataCorp LP | Terms of use | Privacy | Contact us | What's new | Site index |