R help - Jun 2010 - GEE: estimate of predictor with high time dependency

Hi,

I'm analyzing my data using GEE, which looks like below:
> interact <- geeglm(L ~ O + A + O:A,
+ data = data1, id = id,
+ family = binomial, corstr = "ar1")
> summary(interact)
Call:
geeglm(formula = lateral ~ ontask + attachment + ontask:attachment,
    family = binomial, data = firstgroupnowalking, id = id, corstr =
"ar1")

 Coefficients:
                   Estimate  Std.err  Wald Pr(>|W|)
(Intercept)        -1.89133  0.30363 38.80  4.7e-10 ***
O                    0.00348  0.00100 12.03  0.00052 ***
A1                  -0.21729  0.37350  0.34  0.56073
A2                  -0.14151  0.43483  0.11  0.74486
O:A1               -0.37540  0.16596  5.12  0.02370 *
O:A2               -0.27626  0.16651  2.75  0.09708 .
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Estimated Scale Parameters:
            Estimate Std.err
(Intercept)     1.27   0.369

Correlation: Structure = ar1  Link = identity

Estimated Correlation Parameters:
      Estimate Std.err
alpha    0.979 0.00586
Number of clusters:   49   Maximum cluster size: 533



I decided to use auto-regression as the correlation structure because of the
high auto-correlation of the dependent variable, "L".  However,
because one
of the predictors, "O", also has high time dependency (high
autocorrelation), the estimate of "O" (0.00348) seems to be too small.
In
this case, how shall I interpret the parameter?  Should I be using another
analysis, such as loglm?

Thank you in advance for your help!

Sachi

	[[alternative HTML version deleted]]

On Tue, 8 Jun 2010, Sachi Ito wrote:
> Hi,
>
> I'm analyzing my data using GEE, which looks like below:
>
>> interact <- geeglm(L ~ O + A + O:A,
> + data = data1, id = id,
> + family = binomial, corstr = "ar1")
>
>> summary(interact)
>
> Call:
> geeglm(formula = lateral ~ ontask + attachment + ontask:attachment,
>    family = binomial, data = firstgroupnowalking, id = id, corstr =
"ar1")
>
> Coefficients:
>                   Estimate  Std.err  Wald Pr(>|W|)
> (Intercept)        -1.89133  0.30363 38.80  4.7e-10 ***
> O                    0.00348  0.00100 12.03  0.00052 ***
> A1                  -0.21729  0.37350  0.34  0.56073
> A2                  -0.14151  0.43483  0.11  0.74486
> O:A1               -0.37540  0.16596  5.12  0.02370 *
> O:A2               -0.27626  0.16651  2.75  0.09708 .
> ---
> Signif. codes:  0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1
>
> Estimated Scale Parameters:
>            Estimate Std.err
> (Intercept)     1.27   0.369
>
> Correlation: Structure = ar1  Link = identity
>
> Estimated Correlation Parameters:
>      Estimate Std.err
> alpha    0.979 0.00586
> Number of clusters:   49   Maximum cluster size: 533
>
>
>
> I decided to use auto-regression as the correlation structure because of
the
> high auto-correlation of the dependent variable, "L".  However,
because one
> of the predictors, "O", also has high time dependency (high
> autocorrelation), the estimate of "O" (0.00348) seems to be too
small.  In
> this case, how shall I interpret the parameter?
First off, do you know how to interpret main effects in the presence of an 
interaction involving them?? I suspect not, but feel free to offer 
evidence to the contrary and then tell us why discussing 'the estimate of 
"O"' is sensible.

Secondly, without much more detail on the data it is hard to know what to 
make of a question like this even if the business of main 
effects/interactions is handled. As suggested, providing a minimal, 
reproducible example of R code will go a long way.

Chuck
> Should I be using another analysis, such as loglm?
>
> Thank you in advance for your help!
>
> Sachi
>
> 	[[alternative HTML version deleted]]
>
>
Charles C. Berry                            (858) 534-2098
                                             Dept of Family/Preventive Medicine
E mailto:cberry at tajo.ucsd.edu	            UC San Diego
http://famprevmed.ucsd.edu/faculty/cberry/  La Jolla, San Diego 92093-0901