turn on suggestions

Auto-suggest helps you quickly narrow down your search results by suggesting possible matches as you type.

Showing results for

Find a Community

- Home
- /
- Analytics
- /
- Stat Procs
- /
- Reapeated measures linear regression

Topic Options

- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

09-05-2012 03:39 AM

Hi everybody,

I would like to perform a linear regression with repeated measurements: 20 patients with 4 measurements for each.

Do you have an idea how to take into account the correlated data?

I thought to random coefficient/intercept models using PROC MIXED. What do you think about this approach?

Thanking you in advance for your answer.

Regards,

Accepted Solutions

Solution

05-28-2013
04:20 PM

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

05-28-2013 04:20 PM

Plots of results comparing two methods are good. Whether the R-square statistic is good is another story because, with only two variables, the R-squared statistic is only the square of the Pearson's correlation coefficient between the two variables. Sometimes, the value of the Pearson's correlation coefficient (and its partner, the R-squared statistic) can be misleading when comparing two methods because the values from the two methods can be strongly correlated but the values from each of the methods can differ quite markedly.

The simple Bland-Altman scatter plot of differences between paired measurements (on the Y-axis) vs. the mean of the paired measurements (on the X-axis) is one way to depict the relationship between the two methods to detect possible biases and outlying measurements (although it does not account for repeated within-individual paired measurements).

Various statistical methods and measures have been suggested to compare two methods--intraclass correlations, Deming regression (both methods are subject to measurement error so that neither method is a "gold standard"), structural equation modelling (for measurement error models), etc. Even though you should produce the graphs that your colleagues want, it would be worthwhile for you in the future to read up more on the statistical methods relating to method comparisons. Besides books and Google Scholar, you might also look at Lex Jansen's Internet site about SAS articles from conference proceedings that discuss methods comparisons to see how you might use SAS to implement these models.

All Replies

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

09-05-2012 08:36 AM

PROC GLIMMIX or PROC GLM is definitely the way to go.

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

09-17-2012 05:36 AM

Thanks a lot for your answer.

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

05-28-2013 06:35 AM

Hi,

I would need to estimate the R-square and the p-value for the slope for this model. As I have to quantitative data, linear regression using PROC GLM is quite simple omitting CLASS statement.

Example:

proc glm data=RAW;

model VALUE_A = VALUE_B;

run;

Now, I have some difficulties to add the REPEATED statement to specify that measures is repeated on patients (replicates).

Do you have any idea?

Thanks in advance

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

05-28-2013 08:30 AM

Go to the SAS Internet site, http://support.sas.com/documentation/, to read the PROC MIXED example on repeated measures. Aligning your study with this example should allow you to analyze your data appropriately.

Generally, with mixed models and repeated measures, the R-squared measure is less useful than with ordinary least-squares regression because the interest is in modelling correctly the variance-covariance matrix among the repeated measures and in estimating the direction and the size of the model independent variable parameters rather than in estimating the proportion of the variance of the dependent variable "explained" by the independent variables.

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

05-28-2013 09:15 AM

Thanks a lot 1zmm.

I completely agree with your advices. I used to use PROC mixed for similar problematics. In this particular case (method comparison using X subjects with Y replicates for each), my colleagues want to provide a plot with the results of the linear regression (method A vs method B and corresponding fit) and indicate the R-square.

I do not know if it is feasible/relevant.

Thanks

Solution

05-28-2013
04:20 PM

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

05-28-2013 04:20 PM

Plots of results comparing two methods are good. Whether the R-square statistic is good is another story because, with only two variables, the R-squared statistic is only the square of the Pearson's correlation coefficient between the two variables. Sometimes, the value of the Pearson's correlation coefficient (and its partner, the R-squared statistic) can be misleading when comparing two methods because the values from the two methods can be strongly correlated but the values from each of the methods can differ quite markedly.

The simple Bland-Altman scatter plot of differences between paired measurements (on the Y-axis) vs. the mean of the paired measurements (on the X-axis) is one way to depict the relationship between the two methods to detect possible biases and outlying measurements (although it does not account for repeated within-individual paired measurements).

Various statistical methods and measures have been suggested to compare two methods--intraclass correlations, Deming regression (both methods are subject to measurement error so that neither method is a "gold standard"), structural equation modelling (for measurement error models), etc. Even though you should produce the graphs that your colleagues want, it would be worthwhile for you in the future to read up more on the statistical methods relating to method comparisons. Besides books and Google Scholar, you might also look at Lex Jansen's Internet site about SAS articles from conference proceedings that discuss methods comparisons to see how you might use SAS to implement these models.

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

05-30-2013 08:00 AM

Thanks one more time 1zmm for all your detailed advices.

Best,