SPSSX Discussion

Using GEE for analysis of ranking task

Classic

List

Threaded

5 messages Options

Kirill Orlov

Using GEE for analysis of ranking task

Generalized Estimating Equations (GEE) can be used to perform repeated measures ordinal regression.

Can it be used in situation when the ordinal data are ranking list given by each subject? That is, the RM (within-subject) factor is some p items ranked. And so, for each subject the sum of ranks across the p items is the same constant value, and no two items can have the same rank value. (I.e. we have ordinal "compositional" data.)

A between-subject predictors is, say, a grouping factor.

My propose is that it would be correct to use GEE - after removal of any one item, to untie summing to the constant. Then the only "unusual" thing remaining about the data is the restriction mentioned above that items are not allowed to have same values (within same subject). I guess it is not a counter-indication for GEE. Is it?

And I would probably specify the correlation structure for the GEE as "Exchangeable" - for under the null the correlation between the item variables, which are ranking values, is one value equal to 1/(p-1). GEE will estimate then the effects of GROUP, ITEM, and their interaction.

What is your opinion? Is the approach right for you? What might be alternative approaches to analyze such data - to test, at minimum, the group differences?

Это сообщение проверено на вирусы антивирусом Avast.
www.avast.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

Ryan

Re: Using GEE for analysis of ranking task

Read this article:

http://www.statisticalhorizons.com/wp-content/uploads/AllisonChristakis.pdf

Ryan

Sent from my iPhone

On Aug 4, 2016, at 4:15 AM, Kirill Orlov <[hidden email]> wrote:

Generalized Estimating Equations (GEE) can be used to perform repeated measures ordinal regression.

Can it be used in situation when the ordinal data are ranking list given by each subject? That is, the RM (within-subject) factor is some p items ranked. And so, for each subject the sum of ranks across the p items is the same constant value, and no two items can have the same rank value. (I.e. we have ordinal "compositional" data.)

A between-subject predictors is, say, a grouping factor.

My propose is that it would be correct to use GEE - after removal of any one item, to untie summing to the constant. Then the only "unusual" thing remaining about the data is the restriction mentioned above that items are not allowed to have same values (within same subject). I guess it is not a counter-indication for GEE. Is it?

And I would probably specify the correlation structure for the GEE as "Exchangeable" - for under the null the correlation between the item variables, which are ranking values, is one value equal to 1/(p-1). GEE will estimate then the effects of GROUP, ITEM, and their interaction.

What is your opinion? Is the approach right for you? What might be alternative approaches to analyze such data - to test, at minimum, the group differences?

Это сообщение проверено на вирусы антивирусом Avast.
www.avast.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

Kirill Orlov

Re: Using GEE for analysis of ranking task

Ryan,
It is great, thank you.

I would love if somebody else want to share his/her suggestions/opinions as well.

04.08.2016 15:20, Ryan Black пишет:

Read this article:

http://www.statisticalhorizons.com/wp-content/uploads/AllisonChristakis.pdf

Ryan

Это сообщение проверено на вирусы антивирусом Avast.
www.avast.com

Rich Ulrich

Re: Using GEE for analysis of ranking task

The paper: I thought that it was very much worth noting that the testing is not
symmetrical. That is, the authors state that you get different results depending on
whether you rank from high-to-low or from low-to-high. (They also claim that it
won't make much difference, but that sounds like an accident waiting to happen --
natural rankings of "bests" have log(rank) as a good metric, so there is a big
differences between ranks 1 and 2 and a tiny difference between ranks n and n-1.)

So you want to put your more critical distinctions at the more important end,
whichever end that is (probably "1").

--
Rich Ulrich

Date: Thu, 4 Aug 2016 20:32:26 +0300
From: [hidden email]
Subject: Re: Using GEE for analysis of ranking task
To: [hidden email]

Ryan,
It is great, thank you.

I would love if somebody else want to share his/her suggestions/opinions as well.

04.08.2016 15:20, Ryan Black пишет:

Read this article:

http://www.statisticalhorizons.com/wp-content/uploads/AllisonChristakis.pdf

Ryan

Kirill Orlov

Re: Using GEE for analysis of ranking task

When I was segmenting people based on their ratings of some items I've often used Canberra distance (https://en.wikipedia.org/wiki/Canberra_distance) which I find intuitively very appealing in comparing rating results or sport spots (1st, 2nd, 3rd, 4th etc).

I wonder if group differences could be tested statistically based on a matrix of such Canberra distances between individuals. Any ideas?

Это сообщение проверено на вирусы антивирусом Avast.
www.avast.com