Using a continuous covariate to control for a covariate in LMM or GLM?

[DrDrDick]

Hi all! I apologize if this is a re-post, I did a brief search but I didn't find what I was looking for.

Here's what's happening:

I have a lot of gait (locomotion) data from animals, and I eventually want to be able to compare a Control group to an Experimental group across time. The real goal is to determine when there are significant changes in particular gait variables and when those changes return to baseline. Most of these variables are continuous, the others are nominal (Animal ID, Control vs. Experimental, Time Point and so on). I'm actually unsure if they should be categorical...

I have determined that the speed at which these animals move is strongly tied to the other variables, that is I have defined the relationship between Speed and Variable X and I am able to transform these variables such that the relationship with speed is linear.  In my analyses, I want to be able to control for the effect of speed in some way.

I have worked with a statistical consultant to get started on this project, however my lab's budget is getting low and we can no longer afford the service.

The consultant suggested using a Linear Mixed Model with adding speed (continuous) as my covariate, Variable X (continuous) as my dependent, and Time Point and Control vs. Experimental as my factors.

1.) I am not sure how to set up the model, i.e. what the fixed effects should be, if there should be interactions or not, what should be random? Or really how to interpret the results of interactions i.e. translate them into English...

2.) When I did an initial check, I found that the slope of my Variable X vs Speed changed significantly between different time points.  This makes me think that it is not a very stable reference point. It could also be an artifact from the fact that I do not see every speed every day. Is there a work-around to this?

3.) Because of #2, I think it would be ideal to have a lot of data, average it per speed and use it as a baseline to compare all other data to - almost like a standard curve with 95% confidence intervals. I'm not sure if this is possible to automate, but a look-up table is not going to work.

4.) Any other suggestions or approaches? I'm open!

I have SPSS v 20 right now, but I also have access to SAS, R and Stata if any of you think I would be better served using those.


Thanks!!

[David Marso]

"I found that the slope of my Variable X vs Speed changed significantly between different time points. "
Please be more specific.
Have you plotted the data?
-

DrDrDick wrote

Hi all! I apologize if this is a re-post, I did a brief search but I didn't find what I was looking for.

Here's what's happening:

I have a lot of gait (locomotion) data from animals, and I eventually want to be able to compare a Control group to an Experimental group across time. The real goal is to determine when there are significant changes in particular gait variables and when those changes return to baseline. Most of these variables are continuous, the others are nominal (Animal ID, Control vs. Experimental, Time Point and so on). I'm actually unsure if they should be categorical...

I have determined that the speed at which these animals move is strongly tied to the other variables, that is I have defined the relationship between Speed and Variable X and I am able to transform these variables such that the relationship with speed is linear.  In my analyses, I want to be able to control for the effect of speed in some way.

I have worked with a statistical consultant to get started on this project, however my lab's budget is getting low and we can no longer afford the service.

The consultant suggested using a Linear Mixed Model with adding speed (continuous) as my covariate, Variable X (continuous) as my dependent, and Time Point and Control vs. Experimental as my factors.

1.) I am not sure how to set up the model, i.e. what the fixed effects should be, if there should be interactions or not, what should be random? Or really how to interpret the results of interactions i.e. translate them into English...

2.) When I did an initial check, I found that the slope of my Variable X vs Speed changed significantly between different time points.  This makes me think that it is not a very stable reference point. It could also be an artifact from the fact that I do not see every speed every day. Is there a work-around to this?

3.) Because of #2, I think it would be ideal to have a lot of data, average it per speed and use it as a baseline to compare all other data to - almost like a standard curve with 95% confidence intervals. I'm not sure if this is possible to automate, but a look-up table is not going to work.

4.) Any other suggestions or approaches? I'm open!

I have SPSS v 20 right now, but I also have access to SAS, R and Stata if any of you think I would be better served using those.


Thanks!!

[Art Kendall]

If you have 2 values for experimental vs control, then you you have a dichotomy. With a dichotomy every interval is perfectly equal to every other interval.  You can use it as interval in calculations. (with other variables operationally intervals are often "not severely discrepant from equal.)

it is hard to think of time values that are nominal. Please explain.
On 1/25/2013 7:14 PM, DrDrDick wrote:

Hi all! I apologize if this is a re-post, I did a brief search but I didn't
find what I was looking for.

Here's what's happening:

I have a lot of gait (locomotion) data from animals, and I eventually want
to be able to compare a Control group to an Experimental group across time.
The real goal is to determine when there are significant changes in
particular gait variables and when those changes return to baseline. Most of
these variables are continuous, the others are nominal (Animal ID, Control
vs. Experimental, Time Point and so on). I'm actually unsure if they should
be categorical...

I have determined that the speed at which these animals move is strongly
tied to the other variables, that is I have defined the relationship between
Speed and Variable X and I am able to transform these variables such that
the relationship with speed is linear.  In my analyses, I want to be able to
control for the effect of speed in some way.

I have worked with a statistical consultant to get started on this project,
however my lab's budget is getting low and we can no longer afford the
service.

The consultant suggested using a Linear Mixed Model with adding speed
(continuous) as my covariate, Variable X (continuous) as my dependent, and
Time Point and Control vs. Experimental as my factors.

1.) I am not sure how to set up the model, i.e. what the fixed effects
should be, if there should be interactions or not, what should be random? Or
really how to interpret the results of interactions i.e. translate them into
English...

2.) When I did an initial check, I found that the slope of my Variable X vs
Speed changed significantly between different time points.  This makes me
think that it is not a very stable reference point. It could also be an
artifact from the fact that I do not see every speed every day. Is there a
work-around to this?

3.) Because of #2, I think it would be ideal to have a lot of data, average
it per speed and use it as a baseline to compare all other data to - almost
like a standard curve with 95% confidence intervals. I'm not sure if this is
possible to automate, but a look-up table is not going to work.

4.) Any other suggestions or approaches? I'm open!

I have SPSS v 20 right now, but I also have access to SAS, R and Stata if
any of you think I would be better served using those.


Thanks!!



--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Using-a-continuous-covariate-to-control-for-a-covariate-in-LMM-or-GLM-tp5717700.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

[DrDrDick]

Hi Art,

In SPSS, my data imported with experiment days as nominal. They are in the form of "Day 1, Day 2" etc. I have the ability to use actual dates, and I can also remove the word 'Day' from that variable if switching to a categorical variable will be helpful.  For my Group Type, I have the words "Experimental" and "Control." I can also switch this to 0 and 1 if and to Categorical or Interval if that makes more sense. This is my first foray into real statistics, so please forgive my level of knowledge.

Thanks again,

Dick

[DrDrDick]

David,

I forgot to mention that I have a lot of data: ~1,600 observations from which to make a baseline and around 120 a 160 observations for each group (Experimental vs control) per day of experimentation.

The 1,600 data points are spread across 12 days.  If I look at Variable X across the 12 days with Speed as a covariate in a LMM, the term for Variable X changes significantly with time, as does the slope of Variable X vs Speed for each day.When looking at a graph of all the data (1,600), the the points are very close together and best fit lines per day have very little separation. I suspect this could be due to the large # of observations per day, but it is a fact of the type of experiment I cannot get around. I also realize that dealing with animals will never be perfect, which is why I was thinking it could be helpful to generate a sort of standard curve form this data and compare both Experimental and Control data to it across time.

I can attach graphs if that will help.

Thanks,

Dick

[Maguin, Eugene]

Dick,

>> This is my first foray into real statistics, so please forgive my level of knowledge.

Crosstabs are real stats! But, to your point about this problem. It sounds like this is your first LMM adventure. If you haven't done so already, I strongly recommend some background reading. Bruce, I know, has a recommendation. My recommendation is Singer and Willett, "Applied longitudinal data analysis". I'm almost certain IU would have that book. God knows, some of you colleagues might have it as well. But there's many others as well. For instance, Joop Hox has or had the first edition of his book available online.

Gene Maguin


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of DrDrDick
Sent: Saturday, January 26, 2013 12:00 PM
To: [hidden email]
Subject: Re: Using a continuous covariate to control for a covariate in LMM or GLM?

Hi Art,

In SPSS, my data imported with experiment days as nominal. They are in the form of "Day 1, Day 2" etc. I have the ability to use actual dates, and I can also remove the word 'Day' from that variable if switching to a categorical variable will be helpful.  For my Group Type, I have the words "Experimental" and "Control." I can also switch this to 0 and 1 if and to Categorical or Interval if that makes more sense. This is my first foray into real statistics, so please forgive my level of knowledge.

Thanks again,

Dick



--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Using-a-continuous-covariate-to-control-for-a-covariate-in-LMM-or-GLM-tp5717700p5717722.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]

I agree with Gene that Singer & Willett's book is a great resource for analysis of longitudinal data via multilevel models.  I think the other resource he refers to is Applied Multilevel Analysis, by Jos Twisk.  I found it very helpful in introducing the basic concepts of multilevel models.  After reading it, I was better able to tackle other resources (e.g., Snijders & Bosker, Singer & Willett).  You can see a review of Twisk's book here:

  http://ije.oxfordjournals.org/content/36/4/934.full

HTH.

Maguin, Eugene wrote

Dick,

>> This is my first foray into real statistics, so please forgive my level of knowledge.

Crosstabs are real stats! But, to your point about this problem. It sounds like this is your first LMM adventure. If you haven't done so already, I strongly recommend some background reading. Bruce, I know, has a recommendation. My recommendation is Singer and Willett, "Applied longitudinal data analysis". I'm almost certain IU would have that book. God knows, some of you colleagues might have it as well. But there's many others as well. For instance, Joop Hox has or had the first edition of his book available online.

Gene Maguin


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of DrDrDick
Sent: Saturday, January 26, 2013 12:00 PM
To: [hidden email]
Subject: Re: Using a continuous covariate to control for a covariate in LMM or GLM?

Hi Art,

In SPSS, my data imported with experiment days as nominal. They are in the form of "Day 1, Day 2" etc. I have the ability to use actual dates, and I can also remove the word 'Day' from that variable if switching to a categorical variable will be helpful.  For my Group Type, I have the words "Experimental" and "Control." I can also switch this to 0 and 1 if and to Categorical or Interval if that makes more sense. This is my first foray into real statistics, so please forgive my level of knowledge.

Thanks again,

Dick



--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Using-a-continuous-covariate-to-control-for-a-covariate-in-LMM-or-GLM-tp5717700p5717722.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

[DrDrDick]

Gene and Bruce,

Thank you for the reading suggestions! I am definitely interested in understanding the foundations of these statistical models, but my current issue is lack of time and lack of fundamentals in statistics. I looked at both books and neither are written for the non-statistical audience. What I really need is a practical guide for linear mixed models of longitudinal data in SPSS or other software package. The real challenge here is handling the dependence of my variables on observed speed, and the fact that observed speed cannot be controlled; it is variable within animal, every day. This is something I don't even know how to describe in statistical terms, let alone how to search for a resource. If you have any other suggestions, I'd very much appreciate it.

Thanks again.

[David Greenberg]

David Garson has just edited a book published by Sage that gives an
introduction to mixed models. Special topics are covered by chapters
written by different offers. Julie Phillips and I have a chapter on
latent growth trajectory models, with detailed instructions on how to
estimate them using the HLM program. The chapters are not highly
demanding mathematically.  David Greenberg, Sociology Department, NYU

On Mon, Jan 28, 2013 at 2:23 PM, DrDrDick <[hidden email]> wrote:

> Gene and Bruce,
>
> Thank you for the reading suggestions! I am definitely interested in
> understanding the foundations of these statistical models, but my current
> issue is lack of time and lack of fundamentals in statistics. I looked at
> both books and neither are written for the non-statistical audience. What I
> really need is a practical guide for linear mixed models of longitudinal
> data in SPSS or other software package. The real challenge here is handling
> the dependence of my variables on observed speed, and the fact that observed
> speed cannot be controlled; it is variable within animal, every day. This is
> something I don't even know how to describe in statistical terms, let alone
> how to search for a resource. If you have any other suggestions, I'd very
> much appreciate it.
>
> Thanks again.
>
>
>
> --
> View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Using-a-continuous-covariate-to-control-for-a-covariate-in-LMM-or-GLM-tp5717700p5717767.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