Longitudinal study with different time points for each patient

[Omar Beidas]

Dear all,

I have collected data for a study that involves linear measurements taken
from patients at different times points over the last 5 years. These
patients had surgery performed to correct a unilateral facial deformity
and we took measurements from both sides of the face of various facial
features to see if growth is equal on both sides, and if symmetry was
achieved. We do not have a "specific" time point or interval for which we
took measurements, and we have varying numbers of measurements for each
patient (some have 1 measurement, others up to 7). I believe it is still
a 'longitudinal' study of sorts but not sure how to treat and interpret
the data.

Of note, these measurements are taken from digital images and were
calibrated using previous normative data. This normative data was split by
gender and age groups: 0-5 months, 6-11 months, 1 yr, 2 yrs, then yearly
thereafter until age 6. That being said, would it make sense (or do I
have) to split my data using these time groups as well? I went back and
forth but the problem with this is some patients have more than one data
point within a given time interval (for example a measurement at 13, 18,
and 21 months of age would all fall in the "1 year" interval, and I am not
sure how to treat that).

Basically I am not sure what tests I need to run initially, but I have
many independent variables and many dependent variables so I believe I
will need to do a MANOVA or MANCOVA, then run correlation/regression
analyses to see if there is a difference in growth by age and for each
dependent variable. Essentially, my problems boil down to:

1. How do I divide my groups (if I need to do so)?
2. What test(s) should I run, and in what order? I need to determine which
variables have an effect on growth/symmetry, and then correlation
statistics.

I appreciate anyone's input and feel free to directly e-mail me if you
need any more clarification. I realize this is a bit tricky (hence why I
am having trouble with it myself!).

Many thanks!

Omar Beidas
Research Fellow
Section of Plastic & Reconstructive Surgery
University of Oklahoma

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

[Mike Speed]

I am out of the office until 10/20/2010

[Maguin, Eugene]

Omar,

Ahh, the southern high plains.

There is a tremendous amount that I don't know about your subject area that
my comments might be useless. That said, here is my comments.  Gene Maguin


>>I have collected data for a study that involves linear measurements taken
from patients at different times points over the last 5 years. These
patients had surgery performed to correct a unilateral facial deformity
and we took measurements from both sides of the face of various facial
features to see if growth is equal on both sides, and if symmetry was
achieved. We do not have a "specific" time point or interval for which we
took measurements, and we have varying numbers of measurements for each
patient (some have 1 measurement, others up to 7). I believe it is still
a 'longitudinal' study of sorts but not sure how to treat and interpret
the data.

A key piece of information of missing information is the sample size. How
many patients? I'll assume 'sufficient to support the analysis'

One way to describe your data is that it has repeated measurements at
individually-varying times of observations. I'd bet there are many examples
of this sort of measurement structure on medline. The mixed command can
handle this kind of data, computationally. However, I'd like someone more
knowledgeable to comment on whether time would be regarded as a fixed
factor, which will affect the statistical test computations. Normally, I'd
think of doing an analysis such as this in Mplus rather than spss. Pending
an answer on the time factor, let me assume that time is fixed.

Using mixed you could cycle through your facial features measurement points
data. One way to incorporate growth and symmetry is by modeling 1) the
left-right average and 2) the left-right difference in separate models. I
suspect that growth and symmetry could be combined in a single model via a
second order growth model BUT I'm sure that more knowledgeable and
experienced list members will have other, better ideas for modeling growth
and symmetry simultaneously.


>>Of note, these measurements are taken from digital images and were
calibrated using previous normative data. This normative data was split by
gender and age groups: 0-5 months, 6-11 months, 1 yr, 2 yrs, then yearly
thereafter until age 6. That being said, would it make sense (or do I
have) to split my data using these time groups as well? I went back and
forth but the problem with this is some patients have more than one data
point within a given time interval (for example a measurement at 13, 18,
and 21 months of age would all fall in the "1 year" interval, and I am not
sure how to treat that).

I don't think that I know enough to comment on digital image and normative
calibration aspects of your data. All my comments above assumed the use of
the raw measurements. That is, I assumed that you had accurate, precise, and
valid measurements of facial features expressed in standard length units.


Basically I am not sure what tests I need to run initially, but I have
many independent variables and many dependent variables so I believe I
will need to do a MANOVA or MANCOVA, then run correlation/regression
analyses to see if there is a difference in growth by age and for each
dependent variable. Essentially, my problems boil down to:

1. How do I divide my groups (if I need to do so)?
2. What test(s) should I run, and in what order? I need to determine which
variables have an effect on growth/symmetry, and then correlation
statistics.

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

[Omar Beidas]

Gene,
 
First of all, thank you very much for your input! Believe me, if I was an expert in statistics I would not be on this thread--so I appreciate any and all help.
 
The data I measured is in normal units (millimeters) but I calibrated those measurements based on someone else's data. For example, I measured the distance between the eyes in pixels using an imaging program then calibrated that distance in millimeters using normative data based on gender and age. To verify the accuracy of my subsequent measurements, I measured another facial landmark that was also in the normative data and then went back at the end to check if my measurements were different from the published data by age and gender (they were not, meaning my measurements should be accurate). So to answer your question, all the data is in millimeters.

So far what I've done is a MANOVA comparing my measurements (both raw data which is measurements from both sides of the face) as well as the difference (cleft side vs noncleft side) for each of my independent variables (I have 4). I also have one intervention which I'm not sure how to treat (I think as a covariate, but not sure). I also ran a correlation between all my IDs and the "difference" data. I'm not exactly sure how to run a second order growth model in SPSS, as you suggested...
 
I don't know if this helps you answer any more of my questions. You mentioned Mplus, which I never heard of but checked out just now. I'm quite sure we don't have this at my institution but can check if it is something we can obtain. That being said, is it easy to use? I ask because I am sure I will have to do the statistics myself so I don't want to waste another month trying to do these tests...
 
Could anyone else please chime in? I sent this project to the biostatistics department over a month ago and have not been able to get an answer from them (they are "too busy"). I would sincerely appreciate anyone's input that could help out with this!
 
Thanks again,

Omar

> Date: Sun, 17 Oct 2010 11:39:50 -0400

> From: [hidden email]
> Subject: Re: Longitudinal study with different time points for each patient
> To: [hidden email]
>
> Omar,
>
> Ahh, the southern high plains.
>
> There is a tremendous amount that I don't know about your subject area that
> my comments might be useless. That said, here is my comments. Gene Maguin
>
>
> >>I have collected data for a study that involves linear measurements taken
> from patients at different times points over the last 5 years. These
> patients had surgery performed to correct a unilateral facial deformity
> and we took measurements from both sides of the face of various facial
> features to see if growth is equal on both sides, and if symmetry was
> achieved. We do not have a "specific" time point or interval for which we
> took measurements, and we have varying numbers of measurements for each
> patient (some have 1 measurement, others up to 7). I believe it is still
> a 'longitudinal' study of sorts but not sure how to treat and interpret
> the data.
>
> A key piece of information of missing information is the sample size. How
> many patients? I'll assume 'sufficient to support the analysis'
>
> One way to describe your data is that it has repeated measurements at
> individually-varying times of observations. I'd bet there are many examples
> of this sort of measurement structure on medline. The mixed command can
> handle this kind of data, computationally. However, I'd like someone more
> knowledgeable to comment on whether time would be regarded as a fixed
> factor, which will affect the statistical test computations. Normally, I'd
> think of doing an analysis such as this in Mplus rather than spss. Pending
> an answer on the time factor, let me assume that time is fixed.
>
> Using mixed you could cycle through your facial features measurement points
> data. One way to incorporate growth and symmetry is by modeling 1) the
> left-right average and 2) the left-right difference in separate models. I
> suspect that growth and symmetry could be combined in a single model via a
> second order growth model BUT I'm sure that more knowledgeable and
> experienced list members will have other, better ideas for modeling growth
> and symmetry simultaneously.
>
>
> >>Of note, these measurements are taken from digital images and were
> calibrated using previous normative data. This normative data was split by
> gender and age groups: 0-5 months, 6-11 months, 1 yr, 2 yrs, then yearly
> thereafter until age 6. That being said, would it make sense (or do I
> have) to split my data using these time groups as well? I went back and
> forth but the problem with this is some patients have more than one data
> point within a given time interval (for example a measurement at 13, 18,
> and 21 months of age would all fall in the "1 year" interval, and I am not
> sure how to treat that).
>
> I don't think that I know enough to comment on digital image and normative
> calibration aspects of your data. All my comments above assumed the use of
> the raw measurements. That is, I assumed that you had accurate, precise, and
> valid measurements of facial features expressed in standard length units.
>
>
> Basically I am not sure what tests I need to run initially, but I have
> many independent variables and many dependent variables so I believe I
> will need to do a MANOVA or MANCOVA, then run correlation/regression
> analyses to see if there is a difference in growth by age and for each
> dependent variable. Essentially, my problems boil down to:
>
> 1. How do I divide my groups (if I need to do so)?
> 2. What test(s) should I run, and in what order? I need to determine which
> variables have an effect on growth/symmetry, and then correlation
> statistics.
>
> =====================
> 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