Independent groups vs. matched pairs (re-post)

[Staffan Lindberg]

Sorry for re-posting this but I am in dire need of some comments on this
(however short!)

Dear list!

I have clinical sample of 420 female patients I want to follow up in some
national registers after an approximately 20 year period. As a comparison
group I have a randomly drawn sample matched on age at admission for the
patient, place of residence, civil status (married, unmarried,
divorced,widowed), educational level and socioeconomic status. For each
patient I have 5 controls matched on the above criteria.

I remember from my statistics courses (years ago) that you gained
significantly more power (assuming  a correlation between the matched
variables and an outcome variable) using a significance tests for matched
pairs. My questions are:

1. Is there any analysis model when you have 5 matched controls for every
patient?

2. Would I be terribly wrong in treating the patient group and the control
group as independent using for example t-test other statistics for
independent groups? This seem to me simpler but is it correct?

best

Staffan Lindberg
National Institute of Public Health
Sweden

[Marta García-Granero]

Hi Staffan

I've been having some problems trying to post messages to the list.
Here it goes again:

SL> I have clinical sample of 420 female patients I want to follow up in some
SL> national registers after an approximately 20 year period. As a comparison
SL> group I have a randomly drawn sample matched on age at admission for the
SL> patient, place of residence, civil status (married, unmarried,
SL> divorced,widowed), educational level and socioeconomic status. For each
SL> patient I have 5 controls matched on the above criteria.

SL> I remember from my statistics courses (years ago) that you gained
SL> significantly more power (assuming  a correlation between the matched
SL> variables and an outcome variable) using a significance tests for matched
SL> pairs.

Correct!

SL> My questions are:

SL> 1. Is there any analysis model when you have 5 matched controls for every
SL> patient?

Conditional logistic regression. You can trick SPSS into it. See
Raynald's web page for a simple example I posted time ago

http://www.spsstools.net/Syntax/RegressionRepeatedMeasure/ConditionalLogisticRegression.txt

I normally add tho the table with Odds Ratio (with its 95%CI) extra
columns with cases and controls means (with their SD) for
queantitative variables and percentages for qualitative.

Contact me at the list if you need further assistance to get
everything running

SL> 2. Would I be terribly wrong in treating the patient group and the control
SL> group as independent using for example t-test other statistics for
SL> independent groups? This seem to me simpler but is it correct?

By ignoring the matched nature of your data, you will normally bias a
bit your results towards non-significance. Anything that is
significant (in spite of being analysed as independent data) should be
reliable.


(crossing my fingers I'm going to click "Send").


--
Regards,
Dr. Marta García-Granero,PhD           mailto:[hidden email]
Statistician

---
"It is unwise to use a statistical procedure whose use one does
not understand. SPSS syntax guide cannot supply this knowledge, and it
is certainly no substitute for the basic understanding of statistics
and statistical thinking that is essential for the wise choice of
methods and the correct interpretation of their results".

(Adapted from WinPepi manual - I'm sure Joe Abrahmson will not mind)

[Art Kendall-2]

As usual Marta is giving good advice. This will tell you about the main
effect of  cases with the diagnosis and cases that did not report the
condition.
If you have a later version of SPSS,  for your categorical dv's that
have more levels, you might take a look at Multiple Correspondence
Analysis.

Often studies have more than one dependent variable.  If so, and if some
of your dependent variables are not too discrepant from interval level,
a repeated measures 4 way anova with 4 between subjects factors (place
of residence, civil status (married, unmarried, divorced,widowed),
educational level and socioeconomic status) and 2  repeats might tease
out if  there are difference in the subpopulation groups.


Art Kendall
Social Research Consultants

Marta García-Granero wrote:

>Hi Staffan
>
>I've been having some problems trying to post messages to the list.
>Here it goes again:
>
>SL> I have clinical sample of 420 female patients I want to follow up in some
>SL> national registers after an approximately 20 year period. As a comparison
>SL> group I have a randomly drawn sample matched on age at admission for the
>SL> patient, place of residence, civil status (married, unmarried,
>SL> divorced,widowed), educational level and socioeconomic status. For each
>SL> patient I have 5 controls matched on the above criteria.
>
>SL> I remember from my statistics courses (years ago) that you gained
>SL> significantly more power (assuming  a correlation between the matched
>SL> variables and an outcome variable) using a significance tests for matched
>SL> pairs.
>
>Correct!
>
>SL> My questions are:
>
>SL> 1. Is there any analysis model when you have 5 matched controls for every
>SL> patient?
>
>Conditional logistic regression. You can trick SPSS into it. See
>Raynald's web page for a simple example I posted time ago
>
>http://www.spsstools.net/Syntax/RegressionRepeatedMeasure/ConditionalLogisticRegression.txt
>
>I normally add tho the table with Odds Ratio (with its 95%CI) extra
>columns with cases and controls means (with their SD) for
>queantitative variables and percentages for qualitative.
>
>Contact me at the list if you need further assistance to get
>everything running
>
>SL> 2. Would I be terribly wrong in treating the patient group and the control
>SL> group as independent using for example t-test other statistics for
>SL> independent groups? This seem to me simpler but is it correct?
>
>By ignoring the matched nature of your data, you will normally bias a
>bit your results towards non-significance. Anything that is
>significant (in spite of being analysed as independent data) should be
>reliable.
>
>
>(crossing my fingers I'm going to click "Send").
>
>
>--
>Regards,
>Dr. Marta García-Granero,PhD           mailto:[hidden email]
>Statistician
>
>---
>"It is unwise to use a statistical procedure whose use one does
>not understand. SPSS syntax guide cannot supply this knowledge, and it
>is certainly no substitute for the basic understanding of statistics
>and statistical thinking that is essential for the wise choice of
>methods and the correct interpretation of their results".
>
>(Adapted from WinPepi manual - I'm sure Joe Abrahmson will not mind)
>
>
>
>

[Staffan Lindberg]

Thank you ever so much Marta and Art for your input!

To Art: I'm afraid as far as I can foresee most of our dependent variables
(such as deaths, hospital admissions, disability pensions etc) are nominal
in nature. We don´t have Multiple Correspondence Analysis. To Marta: We will
look at Raynalds site, although we feel somewhat intimidated by the concept
"conditional logistic regression". We feel some comfort though, by your
words "a bit" if using methods for independent groups, thus staying somewhat
on the conservative side. Thank you for your kind offer to get back to you.
I'm afraid we maybe will have to do that.

best to all of you on the list

Staffan Lindberg
National Institute of Public Health
Sweden

-----Ursprungligt meddelande-----
Från: SPSSX(r) Discussion [mailto:[hidden email]] För Art Kendall
Skickat: den 22 januari 2007 15:56
Till: [hidden email]
Ämne: Re: Independent groups vs. matched pairs (re-post)


As usual Marta is giving good advice. This will tell you about the main
effect of  cases with the diagnosis and cases that did not report the
condition. If you have a later version of SPSS,  for your categorical dv's
that have more levels, you might take a look at Multiple Correspondence
Analysis.

Often studies have more than one dependent variable.  If so, and if some of
your dependent variables are not too discrepant from interval level, a
repeated measures 4 way anova with 4 between subjects factors (place of
residence, civil status (married, unmarried, divorced,widowed), educational
level and socioeconomic status) and 2  repeats might tease out if  there are
difference in the subpopulation groups.


Art Kendall
Social Research Consultants

Marta García-Granero wrote:

>Hi Staffan
>
>I've been having some problems trying to post messages to the list.
>Here it goes again:
>
>SL> I have clinical sample of 420 female patients I want to follow up
>SL> in some national registers after an approximately 20 year period.
>SL> As a comparison group I have a randomly drawn sample matched on age
>SL> at admission for the patient, place of residence, civil status
>SL> (married, unmarried, divorced,widowed), educational level and
>SL> socioeconomic status. For each patient I have 5 controls matched on
>SL> the above criteria.
>
>SL> I remember from my statistics courses (years ago) that you gained
>SL> significantly more power (assuming  a correlation between the
>SL> matched variables and an outcome variable) using a significance
>SL> tests for matched pairs.
>
>Correct!
>
>SL> My questions are:
>
>SL> 1. Is there any analysis model when you have 5 matched controls for
>SL> every patient?
>
>Conditional logistic regression. You can trick SPSS into it. See
>Raynald's web page for a simple example I posted time ago
>
>http://www.spsstools.net/Syntax/RegressionRepeatedMeasure/ConditionalLo
>gisticRegression.txt
>
>I normally add tho the table with Odds Ratio (with its 95%CI) extra
>columns with cases and controls means (with their SD) for queantitative
>variables and percentages for qualitative.
>
>Contact me at the list if you need further assistance to get everything
>running
>
>SL> 2. Would I be terribly wrong in treating the patient group and the
>SL> control group as independent using for example t-test other
>SL> statistics for independent groups? This seem to me simpler but is
>SL> it correct?
>
>By ignoring the matched nature of your data, you will normally bias a
>bit your results towards non-significance. Anything that is significant
>(in spite of being analysed as independent data) should be reliable.
>
>
>(crossing my fingers I'm going to click "Send").
>
>
>--
>Regards,
>Dr. Marta García-Granero,PhD           mailto:[hidden email]
>Statistician
>
>---
>"It is unwise to use a statistical procedure whose use one does not
>understand. SPSS syntax guide cannot supply this knowledge, and it is
>certainly no substitute for the basic understanding of statistics and
>statistical thinking that is essential for the wise choice of methods
>and the correct interpretation of their results".
>
>(Adapted from WinPepi manual - I'm sure Joe Abrahmson will not mind)
>
>
>
>

[Marta García-Granero]

Hi Staffan

SL> Thank you ever so much Marta and Art for your input!

Always wellcome!

SL> To Art: I'm afraid as far as I can foresee most of our dependent variables
SL> (such as deaths, hospital admissions, disability pensions etc) are nominal
SL> in nature. We don´t have Multiple Correspondence Analysis. To Marta: We will
SL> look at Raynalds site, although we feel somewhat intimidated by the concept
SL> "conditional logistic regression".

If you take a look at this page, you can find some useful info:

http://www2.chass.ncsu.edu/garson/pa765/logit.htm#conditional

SL> We feel some comfort though, by your words "a bit" if using
SL> methods for independent groups, thus staying somewhat on the
SL> conservative side.

The ammount of power lost depends on the strength of the matching. The
more important the degree of matching is, the more sensitive an
analysis that takes it into account will be. The underestimation of
the association can be important if an unmatched analysis is used for
stricktly matched data. See the following example (Shlech et al: "Risk
factors for development of Toxic Shock Syndrome", JAMA 1982; 248:
835-9). Each case was matched with 1 to 4 controls:

DATA LIST FREE/matching exposure disease (3 F8).
BEGIN DATA
 1 1 1  1 0 0  2 1 1  2 0 0  3 1 1  3 0 0  4 1 1  4 0 0
 5 0 1  5 0 0  6 0 1  6 0 0  7 1 1  7 1 0  7 1 0  8 1 1
 8 1 0  8 1 0  9 1 1  9 1 0  9 1 0 10 1 1 10 1 0 10 0 0
11 1 1 11 1 0 11 0 0 12 1 1 12 1 0 12 0 0 13 0 1 13 0 0
13 1 0 14 0 1 14 0 0 14 1 0 15 0 1 15 0 0 15 1 0 16 1 1
16 0 0 16 0 0 17 1 1 17 0 0 17 0 0 18 1 1 18 0 0 18 0 0
19 1 1 19 0 0 19 0 0 20 1 1 20 0 0 20 0 0 21 1 1 21 0 0
21 0 0 22 1 1 22 0 0 22 0 0 23 0 1 23 0 0 23 0 0 24 0 1
24 0 0 24 0 0 25 0 1 25 0 0 25 0 0 26 0 1 26 0 0 26 0 0
27 1 1 27 1 0 27 1 0 27 1 0 28 1 1 28 1 0 28 1 0 28 0 0
29 0 1 29 1 0 29 1 0 29 0 0 30 1 1 30 1 0 30 0 0 30 0 0
31 1 1 31 1 0 31 0 0 31 0 0 32 1 1 32 1 0 32 0 0 32 0 0
33 1 1 33 0 0 33 1 0 33 0 0 34 1 1 34 1 0 34 0 0 34 0 0
35 0 1 35 1 0 35 0 0 35 0 0 36 1 1 36 0 0 36 0 0 36 0 0
37 1 1 37 0 0 37 0 0 37 0 0 38 1 1 38 0 0 38 0 0 38 0 0
39 1 1 39 0 0 39 0 0 39 0 0 40 0 1 40 0 0 40 0 0 40 0 0
41 1 1 41 1 0 41 0 0 41 0 0 41 0 0
END DATA.
VAR WIDTH ALL (8).
VAL LAB exposure 0'No' 1'Yes'/
        disease 0'Control' 1'Case'.
VAR LEV ALL(NOMINAL).

* Unmatched analysis (OR is 6.1, 95%CI: 2.7 to 13.8) *.

LOGISTIC REGRESSION disease
  /METHOD = ENTER exposure
  /CONTRAST (exposure)=Indicator(1)
  /PRINT = CI(95).

* Matched analysis (OR is 7.7; 95%CI: 2.9 to 20.5) *.

COMPUTE vtime=1+(disease=0).
COXREG vtime
  /STATUS=disease(1)
  /STRATA=matching
  /CONTRAST (exposure)=Indicator(1)
  /METHOD=ENTER exposure
  /PRINT=CI(95).

--
Regards,
Dr. Marta García-Granero,PhD           mailto:[hidden email]
Statistician

---
"It is unwise to use a statistical procedure whose use one does
not understand. SPSS syntax guide cannot supply this knowledge, and it
is certainly no substitute for the basic understanding of statistics
and statistical thinking that is essential for the wise choice of
methods and the correct interpretation of their results".

(Adapted from WinPepi manual - I'm sure Joe Abrahmson will not mind)