Re: testing statistical difference between medians of a sample and a subsample extracted from the sample

[Poes, Matthew Joseph]

Sounds like you want to do a subgroup analysis via moderation effects.  In other words, you want to examine if there are differential effects within a subgroup, as compared to the main effect.  You still won't be doing an analysis of the subgroup and comparing it to the entire group, again, that would be a mistake.

In terms of seeing if the effect of a subgroup is driving the main effect, again, what you would need to show is that this subgroup has the effect, and the remaining group (not entire group) has no effect.  If there is no effect in the remaining sample, but a stronger overall effect in the same direction for the subgroup, you could argue that it's driving the main effect.  It should be noted that you haven't proven this by any means.  This analysis might be suggestive of this possibility, but without further testing to eliminate other possibilities, you really couldn't say what's going on in this now non-significant "other" group.

For example, let's say you wanted to test if obesity was getting worse in the United States.  You tested the overall population and compared the obesity rate of 2005 and the obesity rate of 2012, and found it was significantly greater in 2012, you could conclude that obesity has become worse since 2005 (this is overly simplistic I know).  Now let's say you test to see if people with brown eye color are driving that effect, and you find that people with brown eye color have increased in their obesity rate 10 fold, and now the effect of obesity rate is not significant in the remaining group, that doesn't really mean that the remaining group isn't also seeing variation in obesity rate, just that the remaining net effect is null.  You could further subdivide the remaining group into people with blonde hair, and everyone else.  You might find that the remaining group of people who don't have brown eyes or blonde hair also have a significant increase in the obesity rate, but peopl!
 e with blonde hair have a reduction in obesity rate.  This now suggests that the reason you had a null effect was that you had competing effects.  Of course none of this is causal.  The next problem to contend with is that subgroups are not always normally distributed, so dividing the groups and testing in this fashion has been argued to be problematic, as it may be failing certain assumptions.  In addition, the group sizes shrink, and the coefficients can become less stable, as well as the error distribution.  The fix to this has been things such as monte-carlo analysis or bootstrapping.  This is my preference, which, assuming certain assumptions are met, works pretty well.

Matthew J Poes
Research Data Specialist
Center for Prevention Research and Development
University of Illinois
510 Devonshire Dr.
Champaign, IL 61820
Phone: [phone]
email: [hidden email]


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of vini
Sent: Wednesday, August 22, 2012 5:54 AM
To: [hidden email]
Subject: Re: testing statistical dfference between medians of a sample and a subsample extracted from the sample

Thanks for your reply. In the light of discussion above, it seems to me that to test the statistical difference between 2 samples' mean would be a better idea and in that case I can go for t-test ( As the data based on field survey, it can 'safely' be considered as normally distributed. ANY COMMENT ?).

And as far as the relevance of comparing a full sample and a sub sample is concerned, the idea is to analysie if a particular sub sample (extracted based on certain parameter e.g. age group, education etc.) has the influence on the full sample .
However, my question was how to go about it in SPSS(steps?) i.e. comparing full sample and sub sample of the full sample. Any suggestions?

regards,
vini



--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/testing-statistical-dfference-between-medians-of-a-sample-and-a-subsample-extracted-from-the-sample-tp5714777p5714792.html
Sent from the SPSSX Discussion mailing list archive at Nabble.com.

=====================
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

=====================
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

[vini]

Thanks Matthew ! for your valuable advise. To summarise, you suggested 2 things. One, I should go for test of significance between sub sample and the remaining sample instead of full sample and two, I should go for wilcoxon test.
I would like to know how to go about it in SPSS. As in most of the cases, the subsamples within full sample are more than two. If I go for analyze-->-non-parametric--> 2 independent samples, I may not get the remaining sample (that comprise of > 2 subgroups) and if I go for analyze--> non-parametric--> K independent sample, the options of the test are Kruskal wallis and median. That will again test the significance among the sub-groups and not between sub group 1 with combined sub groups 2 & 3.

Kindly suggest how to go about it ?

Regards,
vini

[David Marso]

Please look at the RECODE command!