Chi-Square Test for Trend

Comments

• EJ

  June 18

  Do you have a concrete example, possible urls, and an existing R package in mind?

  EJ
• PerPalmgren

  June 18

  Yes, in R, (what I found) the Cochran-Armitage test for trend is often used as a Chi-square test for trend in proportions across ordered groups. The DescTools package provides the function CochranArmitageTest() for performing the Chi-square test for trend.
• andersony3k

  June 19

  I wonder if an adequate substitute would be a rank biserial correlation https://onlinelibrary.wiley.com/doi/10.1002/ejsp.2420020412 , which quantifies the relationship between an ordinal variable and a dichotomous variable. JASP calculates the rank biserial correlation as an effect size for the Mann-Whitney U test (within the independent samples t test analysis).

  R
• EJ

  June 19

  That is an interesting thought! Of course the ranks are consistent with any monotonic increase (not just linear) but that might actually be a good idea in this case -- it is clear that the trend cannot be linear on the original scale, so it makes sense to test linearity on the logit scale, but few people would be deeply committed to this choice.

  EJ
• andersony3k

  June 20

  This seems legit to me:

  

  
  

  R
• PerPalmgren

  June 21

  My understanding is that while rank biserial correlation is available, it is not a substitute. It measures association between binary and ordinal or continuous variables but does not test for trends in categorical proportions. Including the chi-square test of trend would fill this gap.

  Per the JASP lover🥰
• andersony3k

  June 21 edited June 21

  But my argument is that "association between binary and ordinal variable" is equivalent to "trend in categorical proportion." In my example above, there's a significant negative relationship between passing or failing, and the ordinal level of the group. How is that any different from saying there's a significant trend for the pass-proportion to decrease as the ordinal group-level decreases? Aren't they simply two linguistic options for describing the same result? Logically, shouldn't it be the case that one hypothesis is true when and only when the other is true?

  R