Troubleshooting
Problem
I have a 1:1 Matched Case-Control Study that I wish to analyze using Conditional Logistic Regression. Can I do this in SPSS Statistics?
Resolving The Problem
Yes, using SPSS Statistics Multinomial Logistic Regression (NOMREG), which is found in the Regression Models module. Only 1:1 matches can be analyzed using NOMREG. See Technote 1477360 for information on how to use Cox Regression to analyze a 1:n match.
In a matched case-control study, each "case," or observation which displays some condition, is paired with one (or several) observations, or "controls," which do not. The matching is done on the basis of one or more "risk factors." Since the "cases" and "controls" in each pair have identical values for these variables, they are eliminated from further consideration. (The likelihood function is said to be conditional on these risk factors; thus the term Conditional Logistic Regression.) The Conditional Logistic Regression model can assess the risk of other factors for which the "case" and "control" do not have matched values.
There are three points to remember in setting up the analysis. In outline:
* First, the dependent variable must be a constant (have only one level).
* Second, the difference between each case and corresponding control must be constructed, and the difference must be used as a covariate. (More than one difference variable may be used.)
* Third, the intercept must be omitted from the model.
To obtain a suitable constant variable, click on Transform>Compute... from the Data Editor window, give the Target Variable a name such as DUMMY or CONST, and type 0 (or 1, or any other number you like) as the Numeric Expression; then click OK. If using SPSS command syntax:
COMPUTE dummy = 1.
To compute the difference variable, the values for each "case" and "control" must be on the same row of the data editor, in different variables. Suppose that those variables are actually named Case and Control, return to the Data->Transform->Compute dialog, enter a new variable name such as diff, and give Case - Control as the Numeric Expression. If using SPSS command syntax,
COMPUTE diff = case - control.
Repeat for other unmatched variables to be analyzed.
Now go to the Analyze->Regression->Multinomial... dialog. For the Dependent variable, choose Dummy. Enter the difference variable(s). Leave Factors empty.
Click on the Model... button. Click in the check box to remove the check mark in front of Include Intercept in model. Click Continue, then OK.
If using command syntax,
NOMREG dummy WITH diff
/INTERCEPT EXCLUDE
/PRINT PARAMETER SUMMARY LRT .
Additional Notes:
Never enter any Factors into the model. If there are categorical variables, construct indicator variables for them using dummy coding, construct the differences of the indicators by subtracting the indicator for the control from the corresponding indicator for each case; enter the differences between the indicators in the model.
For example, if there is a Race variable with three values Black, Other, White where White is to be coded as the reference category, the indicators may be constructed as
Race Race_1 Race_2
Black 1 0
Other 0 1
White 0 0
Creation of dummy or indicator values for factors can be automated by using the same factors in a General Linear Model; the dependent variable used is irrelevant. To the GLM (or UNIANOVA) syntax add /OUTFILE=DESIGN(filename.sav). For example:
GLM anynumericvariable BY Race
/OUTFILE=DESIGN('filename.sav').
The file filename.sav would contain the variable Race, plus variables P1 through P4. Variable P1 would have a 1 for each case. Variables P2-P4 would contain dummies for the three categories of Race. You would merge variables P2 and P3 into the original file and perhaps rename them to make them more useful. With the original dataset open:
MATCH FILES /FILE=*
/FILE='filename.sav'
/RENAME (Race P1 P2 P3 P4= d0 d1 Race1 Race2 d2)
/DROP= d0 d1 d2.
EXECUTE.
The differences will still need to be constructed manually.
For advanced users: although matched variables never enter directly into the model because their differences are always zero, they may be included in interactions with the difference variables. This may be done after clicking the Model... button and choosing Custom, or by adding interaction terms to the /MODEL subcommand in command syntax.
Related Information
Historical Number
26706
Was this topic helpful?
Document Information
Modified date:
16 April 2020
UID
swg21477151