GLM assumptions and post hoc comparisons

[lken]

Ok, I have changed my analysis around a little bit.

I am trying to find differences in the amount of rot in trees (given that rot>0) in three different zones.

I am conducting a GLM in SPSS, with zone as a factor with diameter of tree and years since death as covariates.

I asked for post hoc comparisons and it gives me the adjusted means. I would like to report these means, but I don't know what assumptions I need to meet for this test and how I would test them in SPSS? Does anyone have any ideas?

If I find a significant interaction term, does that effect my results?

[Poes, Matthew Joseph-2]

A couple thoughts.  A GLM model based on ordinal DV data does not require normally distributed IV's (obviously) but I believe the post hoc means are still normal t-tests with normal distribution of the subgroup as assumption.  A better option in this situation would be to use resampling like bootstrapping for the post-hoc comparisons, which is not available without a plug in.  I think your post-hoc tests here  are unfortunately bogus because they violate the assumption (quite severely from what you say?).

Significant interactions are always meaningful, and they always change the final interpretation of your model.  First thing to remember, the coefficient of each variable (including the interaction) is that when every other variable in the model is equal to 0.  For instance, Diameter and Years since death are the effect (for each one) when your zone is equal to its reference level, the interaction is equal to zero, etc.

Also worth noting, interactions have normality assumptions, both variables should be normally distributed.  If you only have <5 to <10 levels, and its seriously non-normal, given enough data points, I would dummy code it and treat it as nominal.

Do you know how to test normality assumptions and have you done this?  I got the impression you had, but wasn't sure what all the rest of the variables looked like, and how bad things are across the board.  You might be able to violate assumptions and not worry about it if they aren't too bad.

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 lken
Sent: Monday, April 02, 2012 2:42 PM
To: [hidden email]
Subject: GLM assumptions and post hoc comparisons

Ok, I have changed my analysis around a little bit.

I am trying to find differences in the amount of rot in trees (given that
rot>0) in three different zones.

I am conducting a GLM in SPSS, with zone as a factor with diameter of tree and years since death as covariates.

I asked for post hoc comparisons and it gives me the adjusted means. I would like to report these means, but I don't know what assumptions I need to meet for this test and how I would test them in SPSS? Does anyone have any ideas?

If I find a significant interaction term, does that effect my results?

--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/GLM-assumptions-and-post-hoc-comparisons-tp5613465p5613465.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

[lken]

Thank you for bringing that up, I didn't know my covariates needed to be normally distributed, which they are not. My sample size is very small (12 for group1, 12 for group2, and 30 for group3).

I have not come across one but is there a non-parametric test while controlling for covariates?

Running out of ideas here.

[lken]

Here is an idea...

What if I just coded the presence of rot in trees (1 for presence of rot, 0 for no rot), then did a binary logistic regression. It would take away the exact amount of rot, but would still give me the frequencies in each area, which is still important and may be more relevant.

If I analyzed my data this way I would have a larger sample size because I could include my zero values.

I have heard that you can ignore the assumtion of normality in covariates if they are approximately normal. If diameter and years are approximately normal, would I be able to ask for post hoc comparisons for the proportion of trees with rot in them even though proportions are inherently not normal?

Your help is very much appreciated

[Poes, Matthew Joseph-2]

Yes sorry, when I say normality assumptions, I always mean approximately normal.  You just want to test that and see if they approximate normal.  Many people run these tests when they don't even approximate normal, and it's a noted problem that many of us don't report that we tested normality, and if the data met or didn't meet the standard assumptions.  It's been shown to be a relatively minor problem for a lot of methods when the skewness and kurtosis fall within a relatively small range (say about 3-4 times the standard error), and I've seen people do worse than that.  Didn't mean to create a needless panic.

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 lken
Sent: Monday, April 02, 2012 4:28 PM
To: [hidden email]
Subject: Re: GLM assumptions and post hoc comparisons

Here is an idea...

What if I just coded the presence of rot in trees (1 for presence of rot, 0 for no rot), then did a binary logistic regression. It would take away the exact amount of rot, but would still give me the frequencies in each area, which is still important and may be more relevant.

If I analyzed my data this way I would have a larger sample size because I could include my zero values.

I have heard that you can ignore the assumtion of normality in covariates if they are approximately normal. If diameter and years are approximately normal, would I be able to ask for post hoc comparisons for the proportion of trees with rot in them even though proportions are inherently not normal?

Your help is very much appreciated

--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/GLM-assumptions-and-post-hoc-comparisons-tp5613465p5613766.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

[lken]

Haha that's ok.

I just graduated and am pretty new at this, so anything that scares me into taking all the proper steps is good.

So then, if I did my analysis using a binomial distribution, is it valid to get post hoc comparisons between the groups?

Of not, how else would I report the differences between groups?

[Poes, Matthew Joseph-2]

Post Hoc's are fine for binomial logistic models.  The mean of the groups would be, in that case, the same as the percent of 1's in that group.  If you take the mean of 0 and 1, its .5, if you have lots of 0's and 1's, but they aren't equal, then it won't be .5, it shifts depending on which group has more 1's in it.  Last time I did something like this was in SAS, so my memory is fuzzy, but I recall that in both programs you can get the odds ratio estimate for the subgroups.  I don't believe I had to do custom hypothesis contrasts either, but I just don't recall exactly what you do in SPSS either.

Anyway, to make a useful statistical analysis, that is what you want, the comparison of odd ratio point estimates.

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 lken
Sent: Monday, April 02, 2012 4:53 PM
To: [hidden email]
Subject: Re: GLM assumptions and post hoc comparisons

Haha that's ok.

I just graduated and am pretty new at this, so anything that scares me into taking all the proper steps is good.

So then, if I did my analysis using a binomial distribution, is it valid to get post hoc comparisons between the groups?

Of not, how else would I report the differences between groups?

--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/GLM-assumptions-and-post-hoc-comparisons-tp5613465p5613845.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

[Bruce Weaver]

Where did the notion that "amount of rot" has to be treated as ordinal come from?  How is it measured?

Poes, Matthew Joseph-2 wrote

A couple thoughts.  A GLM model based on ordinal DV data does not require normally distributed IV's (obviously) but I believe the post hoc means are still normal t-tests with normal distribution of the subgroup as assumption.  A better option in this situation would be to use resampling like bootstrapping for the post-hoc comparisons, which is not available without a plug in.  I think your post-hoc tests here  are unfortunately bogus because they violate the assumption (quite severely from what you say?).

Significant interactions are always meaningful, and they always change the final interpretation of your model.  First thing to remember, the coefficient of each variable (including the interaction) is that when every other variable in the model is equal to 0.  For instance, Diameter and Years since death are the effect (for each one) when your zone is equal to its reference level, the interaction is equal to zero, etc.

Also worth noting, interactions have normality assumptions, both variables should be normally distributed.  If you only have <5 to <10 levels, and its seriously non-normal, given enough data points, I would dummy code it and treat it as nominal.

Do you know how to test normality assumptions and have you done this?  I got the impression you had, but wasn't sure what all the rest of the variables looked like, and how bad things are across the board.  You might be able to violate assumptions and not worry about it if they aren't too bad.

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 lken
Sent: Monday, April 02, 2012 2:42 PM
To: [hidden email]
Subject: GLM assumptions and post hoc comparisons

Ok, I have changed my analysis around a little bit.

I am trying to find differences in the amount of rot in trees (given that
rot>0) in three different zones.

I am conducting a GLM in SPSS, with zone as a factor with diameter of tree and years since death as covariates.

I asked for post hoc comparisons and it gives me the adjusted means. I would like to report these means, but I don't know what assumptions I need to meet for this test and how I would test them in SPSS? Does anyone have any ideas?

If I find a significant interaction term, does that effect my results?

--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/GLM-assumptions-and-post-hoc-comparisons-tp5613465p5613465.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

[lken]

The amount of rot is measured in cm (scale data?).

Either way, I don't think doing the GLM I proposed will work. I would like to find differences between rot in three BEC zones and my data do not meet the assumptions of an ANOVA. I thought doing a GLM would make up for it, but if I want to find differences I need to preform t-tests.

Unless there is a GLM that allows to test differences betwen a non-normal, continuous depedant in three groups, with two non-normal covariates, I don't think I can use a GLM.

I think my second proposal of just coding the trees for extensive rot (1), and not extensive rot (0) might be my best option, even if I lose the actual measurements.