At 13/03/2006, Howard Mann wrote:
> Why should I bother with p-values ?
There are two kinds of research question:
1. What is the world like? and
2. How does the world work?
Confidence intervals surround descriptive statistics as a measure of
uncertainty, but the descriptive statistics answer question 1. In any
type 1 question, the size of the effect, the actual value of the
statistic, is important, and its confidence interval is vital to its
interpretation.
However, in type 2 questions, we are concerned with relationships
between things. Here the question is not the size of the relationship
but its existence. Indeed, the size of the relationship may not be
definable in meaningful terms - how do you assess the size of the
relationship between self efficacy and self worth? Attempts made
using 'percentage of variance explained' are misleading, as only a
dataset with no measurement error can produce 100% shared variance
('explained' is begging the question!).
So p-values have their place in scientific reasoning. But they are
too often used in medicine where we really need to see measures of
effect size and confidence intervals.
Ronán Conroy
[log in to unmask]
|