Contrasts in regression analysis

[Frank Furter]

I am working on a Cox regression analysis with two categorical covariates (predictors) each of which has only two categories. In this particular situation, shouldn't the results be the same no matter which of the several available contrast options I choose (which in fact they aren't), or am I making an error in reasoning?

My covariates are Treatment (A or B) and Risk (High or Low), and I wan to know whether treatment has an effect, and whether the effect is modified by risk. Which contrasts should I choose for my two covariates?

Thanks, Andreas

[David Marso]

OK, I'll throw you a bone:
--
NEW FILE.
DATASET CLOSE ALL.
MATRIX.
SAVE MAKE (1000,3,-1) / OUTFILE * /VARIABLES X C Y .
END MATRIX.
IF $CASENUM GT 500 X=1.
IF RANGE($CASENUM,251,500,751,1000)  C=1.
COMPUTE XC = X*C.
COMPUTE Y = 10 + X * .5 + C * .2  + XC * 8 + RV.NORMAL(0,1).
DATASET NAME raw.
FREQUENCIES X C.
CROSSTABS TABLES X BY C.

* OMS.
DATASET DECLARE  Regression.
OMS
  /SELECT TABLES  /IF COMMANDS=['Regression'] SUBTYPES=['Coefficients']  /DESTINATION FORMAT=SAV
  OUTFILE='Regression'.
REGRESSION DEPENDENT Y / METHOD ENTER X C XC.
OMSEND.
* OMS.
DATASET DECLARE  Means.
OMS
  /SELECT TABLES /IF COMMANDS=['Means'] SUBTYPES=['Report']  /DESTINATION FORMAT=SAV NUMBERED=TableNumber_
  OUTFILE='Means'.
MEANS Y BY X.
MEANS Y BY XC.
OMSEND.
MATCH FILES /FILE Regression / KEEP Var2  B.
DATASET NAME Coef.
EXECUTE.
MATCH FILES /FILE Means / KEEP TableNumber_ Var1 Mean.
DATASET NAME MEInt.
IF  TableNumber_  EQ LAG( TableNumber_ ) AND Var1 NE 'Total' Delta=(Mean-LAG(Mean))/2.
FORMATS Delta (F6.4).
EXECUTE.

[Bruce Weaver]

I suspect you want to use INDICATOR.  From the FM (under COXREG > CONTRAST Subcommand):

INDICATOR(refcat). Indicator variables. Contrasts indicate the presence or absence of category
membership. By default, refcat is the last category (represented in the contrast matrix as a row of zeros).
To omit a category other than the last, specify the sequence number of the category (which is not
necessarily the same as its value) in parentheses after the keyword INDICATOR.


Furthermore, I suspect you want INDICATOR(1) to make the first category of each variable the omitted reference category.  (But I'm just guessing.)  

HTH.

Andreas Voelp wrote

I am working on a Cox regression analysis with two categorical covariates (predictors) each of which has only two categories. In this particular situation, shouldn't the results be the same no matter which of the several available contrast options I choose (which in fact they aren't), or am I making an error in reasoning?

My covariates are Treatment (A or B) and Risk (High or Low), and I wan to know whether treatment has an effect, and whether the effect is modified by risk. Which contrasts should I choose for my two covariates?

Thanks, Andreas

[Frank Furter]

Thanks for the programming lesson, David ;) I'm actually not sure what this tells me about my contrast problem, though.

It appears that when the predictors are binary, then Simple, Difference, Helmet, Repeated and Indicator should have the same result (only the reference category may differ), right?

Andreas