SPSS Statistics

 View Only

  • 1.  Linear regression with only one variable produces odd residuals

    Posted Wed March 16, 2022 02:51 PM
    Reply
    Dear SPSS experts

    I have the volumes of 6 different brain structures that are confunded by other variables such as total brain volume, gender and age and would like to compare them between different patient groups using nonparametric statistics. Therefore I lineraly regressed out the influence of these 3 confounding variables from those 6 variables and saved their predicted values and fed them into nonparametric statistics. First, I regressed out all three variables togehter (total brain volume, gender and age) and got different between-group effects for the 6 different brain structures.
    However, when I regressed out only one of the 3 confounding variable (e.g. only the influence of age), saved these residuals and fed them into nonparametric statistics, I get identical (in terms of Kruskal-Wallis-H and p-value) between-group effects for all 6 volumes althought the residuals are different for the 6 volumes. In my opinion, this cannot be true.
    I guess something must be wrong, but cannot identify the problem.
    Any advise is hihgly appreciated.
    Thanks in advance
    Best Regards
    Jürgen Hänggi

    ------------------------------
    Jürgen Hänggi
    ------------------------------

    #SPSSStatistics


  • 2.  RE: Linear regression with only one variable produces odd residuals

    Posted Wed March 16, 2022 02:58 PM
    Reply
    Hi. If I were researching this, I think I might try using the RANK procedure on those dependent variables myself, to ensure that the ranks of the two kinds of residuals do indeed produce different rank orderings.

    ------------------------------
    Rick Marcantonio
    Quality Assurance
    IBM
    ------------------------------

    Original Message


  • 3.  RE: Linear regression with only one variable produces odd residuals

    Posted Thu March 17, 2022 03:09 AM
    Reply
    Dear Rick

    thank you a lot for your answer. I used the RANK procedure and indeed all 6 variables showed the same rank order after regressing out only one confounding variable. In my opinion, it is not plausibel that the effect of the confounding variable is exactly the same in all 6 variables of interest.
    If I regress out the influence of one confounding variable as a covariate in an ANCOVA model, I get different between-group effects for the 6 variables of interest, although the effect of the covariate is also linearly regressed out in an ANCOVA model.
    However, I also get identical between-group effects of 6 variables of interest when I first regress out the influence of the confounding variable using linear regression and then use the residuals in an ANOVA model, although the ANOVA model is not based on rank order and takes into account the real values of the variables, and should in my opinion result in different between-group effects.
    I tested more variables and found always the same behaviour, i.e. identical between-group effects of the residuals after regressing out one variable. I guess that the probability that different variables all are affected in the same exact way by another variable is rather low.
    For me, this is an unexpected behaviour and maybe I overlook something important.
    Would you expect this kind of behaviour for your data?
    Would it be possible to test whether you are also observing this kind of behaviour of the linear regression function in SPSS?

    Thank you a lot in advance for any advice
    Best Regards
    Jürgen


    Original Message


Related Content