ANOVA and regression are but special cases of GLM, both sharing the same

basic assumptions (ANOVA may add its own equal variance assumption).

Keep off transformations as long as they are not dictated by theory. For

instance, if your theory predicts a constant rate of growth in a variable,

you can transform it into its logarithm you can fit a linear regression

function to its logarithmic transformation, Or, if you think that something

is a function of the proportional increase in something else, rather than a

function of its absolute increase, you may use the logarithm of the

independent variable instead of its original raw value. But transformations

in order to make up for an unfulfilled assumption are equivalent to cheating

combined with wishful thinking, and equally pointless.

Hector

-----Mensaje original-----

De:

[hidden email] [mailto:

[hidden email]]

Enviado el: Tuesday, July 18, 2006 12:06 PM

Para:

[hidden email]
Asunto: RE: Standard error using GLM in SPSS/dealing with kurtosis

Thanks for your reply, Hector!

In actual fact, I am not using a regression procedure but analysis of

variance (I have a mixed model between-within subjects design) -- I am not

sure if the same assumptions of normality hold true for ANOVA as for

regression (don't have my stats text handy) -- but if so, that is good news,

as I would prefer to avoid transformation if at all possible!

Cathy

-----Original Message-----

From: Hector Maletta [mailto:

[hidden email]]

Sent: July 18, 2006 10:32 AM

To:

[hidden email];

[hidden email]
Subject: RE: Standard error using GLM in SPSS/dealing with kurtosis

Cathy,

Regarding your first question I defer to others more knowledgeable than

myself. On your second question, I think you believe that the frequency

distribution of variables in your sample must be normal in order to apply

least squares procedures. This is not true, although a very common mistake.

Regression does not assume normal distribution in variables, either in the

population or the sample. What it assumes is normality of errors or

residuals, quite a different matter altogether.

Asymmetrical distributions may enlarge your standard error (without

invalidating the procedure, though) especially when outliers are present,

but more normal non-normal distributions (if you permit me a mild pun) may

hardly create any problem, least of all because of being a bit flatter than

Gauss's bell. Hector

-----Mensaje original-----

De: SPSSX(r) Discussion [mailto:

[hidden email]] En nombre de Cathy

Underhill Enviado el: Tuesday, July 18, 2006 10:59 AM

Para:

[hidden email]
Asunto: Standard error using GLM in SPSS/dealing with kurtosis

Hi,

I am new to this Listserv, but wondered if someone could help me with a

couple of issues:

* I am using GLM in SPSS for a mixed model (between-within) analysis

of variance and was told by someone that when using GLM the standard error

presented by SPSS is a "pooled" SE and is incorrect for my between-group

variables -- does anyone know is this is correct, and how I would obtain the

correct values?

* Secondly, how serious is the impact of platykurtosis (one of my w-s

variables) on the final analysis (I have a relative small sample -- N = 144)

and does anyone have a recommendation on the best way to deal with this? I

have looked into transformations but from what I have read so far, there are

mixed opinions about this approach.

Thanks,

Cathy

Cathy Underhill

613-951-6023 | facsimile / télécopieur 613-951-4179

[hidden email] Statistics Canada | 170 Tunney's Pasture Driveway

Ottawa ON K1A 0T6 Statistique Canada | 170, promenade du Pré Tunney Ottawa

ON K1A 0T6 Government of Canada | Gouvernement du Canada