[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

From |
Maarten buis <[email protected]> |

To |
stata list <[email protected]> |

Subject |
Re: st: RE: Quantile question |

Date |
Mon, 3 Mar 2008 14:42:36 +0000 (GMT) |

--- Ronan Conroy wrote: > I was involved in looking at the effect of a number of measures of > the patient's response to illness as predictors of depression - work > and social adjustment, symptom frequency, symptom bother, and so on. > These were measured using standardised scales, but each scale had > its own theoretical range, its own empirical score distribution and, > most important, each scale was measured in arbitrary units. > > It was useful for the reader to be able to see the odds ratios (and > confidence intervals) associated with a 1-decile increase in each of > these predictors, as it gave them a way of comparing their effects > and of judging their practical importance as well as their > statistical significance. > > There are times, then, when quantiles do not lose information but > increase it, by converting unfamiliar and arbitrary measurement > scales to a definable measurement unit. I think the issue is terminology here. Say we have a variable with an arbitrary unit, say, symptom bother. The the 9th percentile is the value of the arbitrarily scaled variable sympton bother which has 9% of the observations below it. (http://en.wikipedia.org/wiki/Quantile) What you seem to mean is the percentile rank, which in the example above would be the number 9, and is a useful way of standardizing variables: you can create a new variable with now range (0 to 100), mean (50), standard deviation (approx. 28.6, depending on the number of knots) and distribution (uniform distribution). If that is the case than I agree with you. The more common alternative (in my discipline anyhow) is z-scores (the variable minus the mean devided by the standard deviation), and is often implicitly interpreted in terms of these percentile ranks: a value of 1.96 is large because if the variable is from a normal (Gaussian) distribution, 97.5 % of the respondents will have a value less than that. Notice that you now need the additional assumption about the distribution, which percentile scores don't make. However, Nick is right when he points out that this is a non-linear transformation. In particular, you will only retain information about the ordering of values and loose information about the distances between values. -- Maarten ----------------------------------------- Maarten L. Buis Department of Social Research Methodology Vrije Universiteit Amsterdam Boelelaan 1081 1081 HV Amsterdam The Netherlands visiting address: Buitenveldertselaan 3 (Metropolitan), room Z434 +31 20 5986715 http://home.fsw.vu.nl/m.buis/ ----------------------------------------- __________________________________________________________ Sent from Yahoo! Mail. A Smarter Inbox. http://uk.docs.yahoo.com/nowyoucan.html * * For searches and help try: * http://www.stata.com/support/faqs/res/findit.html * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/

- Prev by Date:
**RIF: st: chow test for more than 2 coef** - Next by Date:
**RE: st: evaluating probabilities using kdensity** - Previous by thread:
**Re: st: RE: Quantile question** - Next by thread:
**RE: st: -reshape- with more than one j()?** - Index(es):

© Copyright 1996–2024 StataCorp LLC | Terms of use | Privacy | Contact us | What's new | Site index |