Re: advice for using right statistic method
- From: "justinkdavis@xxxxxxxxx" <justinkdavis@xxxxxxxxx>
- Date: Fri, 31 Aug 2007 00:32:01 -0000
It may be reasonable to use dummy variables and a reasonable
regression scheme. Until you've justified why this isn't plausible,
this is the most obvious method, especially since it will usually give
you confidence intervals for the coefficients. The workload and
disruptions may be traditional real variables. The different methods
may be encoded as different variables.
Now, I don't think ANOVA is what I should use, because one of its
assumptions is that data within a group must have a normal distribution.
(For example, the performance drops if the workload or the number of
disruptions increase, so they're certainly not normally distributed).
I'm not certain the objection is meaningful here. The requirement of
the normality of the data is about the data in each group itself, not
relationships among them. You would use something like Shapiro-Wilk to
test this. It's also worthwhile to note that ANOVA may be useful even
if normality does not hold exceptionally well.
Or should I look more closely to principal component analysis instead?
I do not think that principal component analyses will be useful if you
have not exhausted traditional regressions. They are meant usually to
reduce the dimensionality of the data - they will contain no more data
than a full regression and should not be invoked too quickly.
.
- References:
- advice for using right statistic method
- From: JeeBee
- advice for using right statistic method
- Prev by Date: Re: Distrbuted Computing - counting prime numbers in 0 - 10^15
- Next by Date: Re: Pi in non-euclidiean spaces
- Previous by thread: Re: advice for using right statistic method
- Next by thread: morfing graphs
- Index(es):
Relevant Pages
|
