1

I want to test a linear contrast (trend test) in SPSS with more than two groups. Let's assume my assumption is B>A>D>C.
I did the coding of the contrasts manually and obtained a signficance level of let's say P=0.075.
The only question I got - and which is not answered in most tutorials and books on this topic - do I need divide the p-value by two, since it's a one-sided (directed) hypothesis or is there no such one-sided hypothesis for linear contrasts - in other words is this already accounted for in the calculation?
In my case, this would decide upon whether the test is significant (p=0.0375 if divided by two) or not.

Glen_b
  • 282,281
Chris
  • 11
  • I don't understand the sense in which your hypothesis might be "one-sided," because I see twenty-four "sides". Your hypothesis posits one particular order for the four variables out of the 24 possible orders. – whuber Aug 15 '17 at 16:32
  • Thanks for your comment. That's exactly right, we assume one particular order out of 24, with strict differences between each group. I'm still now sure however, whether the p-vakue issued by SPSS is the correct one or whether we need to divide it by 2 (the reasoning would be that we only test whether one group is larger than a particular other, not just different from the other) – Chris Aug 15 '17 at 18:14
  • I still cannot see where a factor of $2$ could possibly come from in this setting. Regardless, your question doesn't look answerable until you describe exactly how you "obtained a significance level" from SPSS--and at that point to answer it is might a matter of consulting the SPSS documentation, depending on the details of what you did. – whuber Aug 15 '17 at 18:19
  • I've already checked the SPSS manual - unfortunately they don't talk about one-sided vs two-sided test related to linear contrasts. That's also where "divided by 2" comes from, that I can use the 10% level for one-sided tests to obtain 5% significance. – Chris Aug 15 '17 at 18:23
  • This test relates to oneway ANOVA-> contrasts -> linear http://www-01.ibm.com/support/docview.wss?uid=swg21587764 – Chris Aug 15 '17 at 18:31
  • Your question isn't clear, especially that my assumption is B>A>D>C. One contrast is comparing two means (two original groups or two some combined groups). "B>A>D>C" contains 4 terms, it looks like you want to test some omnibus hypothesis. Which one? What is null and what is the alternative? – ttnphns Aug 16 '17 at 04:58
  • Also, did you consider a nonparametric group trend test? – ttnphns Aug 16 '17 at 05:01
  • Thanks for your comment ttnphns. With linear group test you've got the possibility to compare trends for more than 2 groups (see p. 2 here: https://www.discoveringstatistics.com/repository/contrasts.pdf). What kind of non-parametric group trend tests did you think of? – Chris Aug 16 '17 at 14:11

0 Answers0