Thank you Art..those were all helpful comments.....here are some responses:
(1) 99 pairs of patients were matched on age, gender, and LOS and then randomly assigned to exp and control group
(2) I had discussed the feasiblity of collapsing the various measures (e.g, respiratory rate) on a daily basis, especially given there is tremendous variation on the number of measures per day (from 2 to 10)...I'm not quite sure why there is such variation though it has more to do with the clinician than the patient;however, the director of the project would very much like to analyze all of the daily measures.......so we have some patients that had LOS ranging from 2 to 23 so anywhere from 10 to over 100 measures!!!
(3) I was thinking of using HLM6.0 (i.e, mixed model approach) since it affords the flexibility of varying occasions and varying times of measures, though the rate of missingness was so prominent here I was unsure of the stability of the parameter estimates
(4) Art, I was unsure of how to model this between-group design with varying repeated measures in a time series context....I've done the simple univariate time series, but I was uncertain of how a data base is set up for ARIMA modeling when you have two groups (n = 99 each) and you want to collapse all the repeated measures for all the participants?
thank you very much for your feedback...Dale
Art Kendall <Art@DrKendall.org> wrote:
Ouch! Perhaps these thoughts will help.
The MIXED procedure will handle repeated measures with missing data.
check and see if it will handle 3 way crossed repeated measures -- pair
* day * measurement event. With measurement events 1 to the max for
Each measurement event would be a case. Predictors would be pair, and
which member of the pair. and a event variable.
overall, Is there a special interest in the measurement event repeats
since they were not built into the design?
Can you greatly reduce the proportion of data points that are missing?
First figure out how many data points you would have if there were no
missing data. Especially if this is a convenience sample to start with,
to reduce the messiness, see what happens if you drop a few pairs and
the later repeats in a day. Perhaps drop the sixth day. Perhaps your
client can live with almost all of the measurement.
Why are there missing measurement events? Subjects fatigued?
Inconvenient for subject? Person doing measurement too busy?
The trouble for me of treating it as as a simple time series would be
to ignore the effects of pair and day.
Social Research Consultants
Inside the Washington, DC beltway
Dale Glaser wrote:
>Hi all.....I will be working on a project whereby there are 99 matches (i.e., 198 participants in the study) with approximately one dozen variables measured on a varying basis over a varying number of days. For example from one matched pair one individual from the experimental group is measured on one variable (e.g., respiratory rate) 8 times on day 1, then 9 times on day 2, and then 4 times on day 6; their matched cohort (control group) may be in the study for five days and have a very different number of measures per day. Also, the time by which they were measured was not standardized (i.e., not all were measured at 8 am, 9 am, etc.). There can be from 10 to 80 measures per person on a given variable given the length of their participation. I was envisioning a multi-sample latent growth curve approach until I found out the # of measures (note: the client would indeed like to model ALL the measures as opposed to collapsing across time/days). Subsequently, a time series!
> was proposed (e.g., ARIMA) though I am having difficulty how to best pool the 198 particpants given the varying rates of measures across a varying rate of days.
>If anyone has had experience with such a design, I would be most appreciative if you would be wiling to shed insight about the optimal way to handle such a time-unstructured design. I have pondered a mixed-model approach or pooled time series, but the number of measures across time see to prohibit a mixed-model approach (I do have HLM6.0 and Mplus) and I"m not sure if pooled time series is appropriate (i.e, does each column represent an individual's 10- 80 or so measures, and are they collapsed?).
>Again, thank you for your time...Dale Glaser
>Dale Glaser, Ph.D.
>4003 Goldfinch St, Suite G
>San Diego, CA 92103
Dale Glaser, Ph.D.
4003 Goldfinch St, Suite G
San Diego, CA 92103