Paired T Test but with nested data

[John Norton]

Hi List,

I'm hoping someone with a better understanding of complex designs than I can help me with suggestions for how I can essentially do a paired samples t test, but where the cases are nested.

The data are patient based, which is to say that each record is identified by a unique patient ID. However, each record represents the responses given by a physician relative to that patient.  So, whereas a patient occurs in the database only once, a particular MD can enroll several patients to the study.  So, a physician can occur in the database multiple times - once for each patient he or she enrolls.

The data are collected at baseline and then again after an intervention.  Physicians are asked for their responses prior to the intervention and then again afterward.  On the face of it, it looks like a straight forward paired samples T test: we want to determine whether the MDs treatment recommendations were impacted (and how) as a result of the intervention.  However, because an MD can have several records (patients) in the database, the cases are not unique.  If I aggregate the responses within each physician, I lose the variation across his/her patients.

Can any on this list suggest a way to approach this analysis?  Some colleagues have suggested that this may be a job for HGLM (a procedure with which I am not familiar) while others have been quite skeptical about that approach.  So, I'm hoping that some here on this list can offer assistance with suggestions for the proper approach.

Thanks,


John Norton
Biostatistician
Oncology Institute
Loyola University Medical Center

[phone]
[hidden email]

"Absence of evidence
      is not evidence of absence"

[statisticsdoc]

John,

Have you considered Hierarchical Linear Modeling (viz. Bryk and Raudenbush)?
It strikes me that you have a three-level model, with time nested within
patient, and patients nested within doctors.

HTH,

Stephen Brand

For personalized and professional consultation in statistics and research
design, visit
www.statisticsdoc.com


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]]On Behalf Of
John Norton
Sent: Tuesday, December 19, 2006 9:35 AM
To: [hidden email]
Subject: Paired T Test but with nested data


Hi List,

I'm hoping someone with a better understanding of complex designs than I can
help me with suggestions for how I can essentially do a paired samples t
test, but where the cases are nested.

The data are patient based, which is to say that each record is identified
by a unique patient ID. However, each record represents the responses given
by a physician relative to that patient.  So, whereas a patient occurs in
the database only once, a particular MD can enroll several patients to the
study.  So, a physician can occur in the database multiple times - once for
each patient he or she enrolls.

The data are collected at baseline and then again after an intervention.
Physicians are asked for their responses prior to the intervention and then
again afterward.  On the face of it, it looks like a straight forward paired
samples T test: we want to determine whether the MDs treatment
recommendations were impacted (and how) as a result of the intervention.
However, because an MD can have several records (patients) in the database,
the cases are not unique.  If I aggregate the responses within each
physician, I lose the variation across his/her patients.

Can any on this list suggest a way to approach this analysis?  Some
colleagues have suggested that this may be a job for HGLM (a procedure with
which I am not familiar) while others have been quite skeptical about that
approach.  So, I'm hoping that some here on this list can offer assistance
with suggestions for the proper approach.

Thanks,


John Norton
Biostatistician
Oncology Institute
Loyola University Medical Center

[phone]
[hidden email]

"Absence of evidence
      is not evidence of absence"

[John Norton]

Hi Steve,

Thanks for the suggestion.  Tyhis is not a procedure with which I am familiar.  Can you direct me to where in SPSS I can find this?

Thanks,

JN

John Norton
Biostatistician
Oncology Institute
Loyola University Medical Center

[phone]
[hidden email]

"Absence of evidence
      is not evidence of absence"

>>> Statisticsdoc <[hidden email]> 12/19/06 9:39 AM >>>
John,

Have you considered Hierarchical Linear Modeling (viz. Bryk and Raudenbush)?
It strikes me that you have a three-level model, with time nested within
patient, and patients nested within doctors.

HTH,

Stephen Brand

For personalized and professional consultation in statistics and research
design, visit
www.statisticsdoc.com


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]]On Behalf Of
John Norton
Sent: Tuesday, December 19, 2006 9:35 AM
To: [hidden email]
Subject: Paired T Test but with nested data


Hi List,

I'm hoping someone with a better understanding of complex designs than I can
help me with suggestions for how I can essentially do a paired samples t
test, but where the cases are nested.

The data are patient based, which is to say that each record is identified
by a unique patient ID. However, each record represents the responses given
by a physician relative to that patient.  So, whereas a patient occurs in
the database only once, a particular MD can enroll several patients to the
study.  So, a physician can occur in the database multiple times - once for
each patient he or she enrolls.

The data are collected at baseline and then again after an intervention.
Physicians are asked for their responses prior to the intervention and then
again afterward.  On the face of it, it looks like a straight forward paired
samples T test: we want to determine whether the MDs treatment
recommendations were impacted (and how) as a result of the intervention.
However, because an MD can have several records (patients) in the database,
the cases are not unique.  If I aggregate the responses within each
physician, I lose the variation across his/her patients.

Can any on this list suggest a way to approach this analysis?  Some
colleagues have suggested that this may be a job for HGLM (a procedure with
which I am not familiar) while others have been quite skeptical about that
approach.  So, I'm hoping that some here on this list can offer assistance
with suggestions for the proper approach.

Thanks,


John Norton
Biostatistician
Oncology Institute
Loyola University Medical Center

[phone]
[hidden email]

"Absence of evidence
      is not evidence of absence"

[statisticsdoc]

John,

Within SPSS, you can carry out HLM by using the MIXED procedure in the
Advanced Module.  The documentation provides an example of a three-level
multi-level model with students nested in classrooms nested in schools.
This approach would apply to your data in which time is nested within
patients who are nested within doctors.  BTW, I should have noted earlier
that HLM is a special case of HGLM (Raudenbush & Bryk, 2002).

HTH,

Stephen Brand

P.S. There is also an advanced HLM software package produced by Raudenbush
and sold by SSI, but I expect that the Mixed procedure in the SPSS Advanced
Module will accomplish your goals.

For personalized and professional consultation in statistics and research
design, visit
www.statisticsdoc.com


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]]On Behalf Of
John Norton
Sent: Tuesday, December 19, 2006 10:57 AM
To: [hidden email]
Subject: Re: Paired T Test but with nested data


Hi Steve,

Thanks for the suggestion.  Tyhis is not a procedure with which I am
familiar.  Can you direct me to where in SPSS I can find this?

Thanks,

JN

John Norton
Biostatistician
Oncology Institute
Loyola University Medical Center

[phone]
[hidden email]

"Absence of evidence
      is not evidence of absence"

>>> Statisticsdoc <[hidden email]> 12/19/06 9:39 AM >>>
John,

Have you considered Hierarchical Linear Modeling (viz. Bryk and Raudenbush)?
It strikes me that you have a three-level model, with time nested within
patient, and patients nested within doctors.

HTH,

Stephen Brand

For personalized and professional consultation in statistics and research
design, visit
www.statisticsdoc.com


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]]On Behalf Of
John Norton
Sent: Tuesday, December 19, 2006 9:35 AM
To: [hidden email]
Subject: Paired T Test but with nested data


Hi List,

I'm hoping someone with a better understanding of complex designs than I can
help me with suggestions for how I can essentially do a paired samples t
test, but where the cases are nested.

The data are patient based, which is to say that each record is identified
by a unique patient ID. However, each record represents the responses given
by a physician relative to that patient.  So, whereas a patient occurs in
the database only once, a particular MD can enroll several patients to the
study.  So, a physician can occur in the database multiple times - once for
each patient he or she enrolls.

The data are collected at baseline and then again after an intervention.
Physicians are asked for their responses prior to the intervention and then
again afterward.  On the face of it, it looks like a straight forward paired
samples T test: we want to determine whether the MDs treatment
recommendations were impacted (and how) as a result of the intervention.
However, because an MD can have several records (patients) in the database,
the cases are not unique.  If I aggregate the responses within each
physician, I lose the variation across his/her patients.

Can any on this list suggest a way to approach this analysis?  Some
colleagues have suggested that this may be a job for HGLM (a procedure with
which I am not familiar) while others have been quite skeptical about that
approach.  So, I'm hoping that some here on this list can offer assistance
with suggestions for the proper approach.

Thanks,


John Norton
Biostatistician
Oncology Institute
Loyola University Medical Center

[phone]
[hidden email]

"Absence of evidence
      is not evidence of absence"