Re: Which type of variance?

On 6 Aug 2005 05:35:13 -0700, jhartdc@xxxxxxxxx wrote:

> Is there a simple explanation for knowing which type of variance (equal
> versus unequal) to use when doing a t-test?

The naive approach is considered wrong by experts -
XXX "Use the test of equality to tell you which to use."

Where it matters is when Ns are unequal, too (as I just
posted to the other question).

Some people have recommended
"Always use the 'unequal' test, because it is more robust."
It is True, that test is slightly more robust to the most
generic unequal variances. But it is less robust to (say)
analyzing 0/1 scores or rating scales with limited ranges,
where the pooled variance does better.

For my usual data, I'm comfortable using the pooled test.

However, in general, neither test is very robust -- when you
look at the one-tailed results, either may be grossly wrong.


If you can normalize the data by a rational transformation
(if it is counts, use square root; if it is reasonably log-normal,
use log), the better test is to test after the transformation.
- When the shapes are from the same family, then the
rank-order transformation is also effective, so you can use
the 'nonparametric' test. But generally, the unequal variances
can be a bad indicator for rank tests, too.

If you have any doubt, consider the alternate results, and
report what is going on -- if there is any difference that
requires explanation.



http://www.pitt.edu/~wpilib/index.html
.

Relevant Pages

  • Re: Principle Components Analysis
    ... Can i do that by using stat of SAS or is there any software that can help? ... unit variance I believe is the default. ... You don't have to specifically tell SAS to do that transformation. ... dichotomous variables in this situation, you don't want each one of them to have a variance of 1. ...
    (sci.stat.math)
  • Re: Mean and Variance
    ... Actually what I want is the symbolic representation of the ... mean and the variance of the function? ... normal, with mean mux and variance vx, ... then we can think of it as a transformation ...
    (comp.soft-sys.matlab)