Re: testing for homoscedasticity in SPSS?

Posted by Rich Ulrich on Feb 24, 2018; 10:30pm
URL: http://spssx-discussion.165.s1.nabble.com/testing-for-homoscedasticity-in-SPSS-tp5735562p5735571.html

Like Bruce, I'm not a fan of tests of assumptions, but I do pay attention to the

shape of distributions. In my experience - which is heavily biased towards

using rating scales - 90% or 99% of apparent heteroscedasticity is the fault

of "wrong scaling" rather than underlying lumpiness. Scale items can /usually/ be

analyzed as they are; scale totals occasionally benefit from transformation. Item

Response Theory uses logistic, though the complication may seem like over-kill; on

the cruder side, square root is most common, after deciding which end should

represent "zero".

Is there big skewness? Are there big outliers? Do these features represent scores

that you would consider at "equal intervals"?  Does taking a transformation give

something that is more Normal? If there is an outlier that represents a "real interval",

that raises the question of whether /that/ case actually belongs in a least-squares

analysis of these data; or if it should be removed and discussed as a special case.

If the transformation made no difference in the subsequent analyses and inferences,
PIs often liked to present the unmodified analysis along with the comment that doing
the analyses using XX-transformation to meet the variance assumptions made no difference.

--
Rich Ulrich

---

From: SPSSX(r) Discussion <[hidden email]> on behalf of Bruce Weaver <[hidden email]>
Sent: Saturday, February 24, 2018 4:55 PM
To: [hidden email]
Subject: Re: testing for homoscedasticity in SPSS?

 

I am not a fan of statistical tests of the assumptions for another test or
procedure.  Such tests often have too little power when n is small and too
much power when n is large. 

Rather than testing, you could just estimate your model via UNIANOVA and
allow for heteroscedasticity via the ROBUST sub-command (assuming your SPSS
version is recent enough).  See the link below for details.

https://www.ibm.com/support/knowledgecenter/en/SSLVMB_25.0.0/statistics_reference_project_ddita/spss/base/syn_unianova_robust.html

HTH.

[snip, previous]

===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD