Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • #16
    The t/z/chi2 doesn't change; just the critical value.

    I guess the first question is why do you want a one-sided test?

    Comment


    • #17
      Sorry George, I did not see your response before I wrote the previous post.

      Regarding why I want to one-tailed test: it is because in each regime we expect the relation to be either positive or negative in line with theory.

      I am not sure what you meant by the t or z stats do not change. In other words, when I report the results in the table for a one-sided t/z stats, do I use the ones from the output (which are two-tailed)?
      Last edited by Mike Kraft; 23 Aug 2023, 18:44.

      Comment


      • #18
        Each of these produces the desired results in the format of Stata's ttest.

        the "test" version is more flexible since it can handle more complex tests (as shown)

        Code:
        mswitch dr gdp, switch(inf) states(2) varswitch vce(oim)
        
        display "Ha: diff < 0: Pr(T < t) = " 1 - r(table)[4,1]/2
        display "Ha: diff < 0: Pr(|T| > |t|) = " r(table)[4,1]
        display "Ha: diff < 0: Pr(T > t) = " r(table)[4,1]/2
        
        display "Ha: diff < 0: Pr(T < t) = " normal(r(table)[3,1])
        display "Ha: diff < 0: Pr(|T| > |t|) = " normal(-r(table)[3,1])*2
        display "Ha: diff < 0: Pr(T > t) = " 1 - normal(r(table)[3,1])
        
        test _b[State1:inf]==0
        
        display "Ha: diff < 0: Pr(T < t) = " 1 - r(p)/2
        display "Ha: diff < 0: Pr(|T| > |t|) = " r(p)
        display "Ha: diff < 0: Pr(T > t) = " r(p)/2
        
        display "Ha: diff < 0: Pr(T < t) = " 1 - chi2tail(1,r(chi2))/2
        display "Ha: diff < 0: Pr(|T| > |t|) = " chi2tail(1,r(chi2))
        display "Ha: diff < 0: Pr(T > t) = " chi2tail(1,r(chi2))/2
        
        test _b[State1:inf]==_b[State2:inf]
        
        display "Ha: diff < 0: Pr(T < t) = " 1 - r(p)/2
        display "Ha: diff < 0: Pr(|T| > |t|) = " r(p)
        display "Ha: diff < 0: Pr(T > t) = " r(p)/2
        
        display "Ha: diff < 0: Pr(T < t) = " 1 - chi2tail(1,r(chi2))/2
        display "Ha: diff < 0: Pr(|T| > |t|) = " chi2tail(1,r(chi2))
        display "Ha: diff < 0: Pr(T > t) = " chi2tail(1,r(chi2))/2

        Comment


        • #19
          response to #17 - if your theory says "positive" but your results say "negative", what do you want to do? one option is to say that the theory is not confirmed - but if you want to potentially learn something you need a 2-sided test (note that for general use of the chi-squared distribution in testing, one is only using one tail even though the conclusion is 2-sided)

          Comment


          • #20
            This looks a lot like p-hacking.

            Comment


            • #21
              Thanks Rich Goldstein! The idea is that the two regimes are showing positive and negative relations and thus I thought it would make sense to have a one tailed t stats as long as the two regimes are identified successfully by the model. Makes sense?!

              Comment


              • #22
                not unless you have no interest in finding that the "opposite" is true; personally I would do, and argue for, a two-sided test - but I don't even know what your substantive context is and I am probably not an expert in whatever it is

                Comment


                • #23
                  unless you know ahead of time that the relationship will be positive or negative for some theoretical reason, then two-sided is better.

                  or, you can take the increasingly modern approach and not report p-values at all, and comment on the coefficient and its confidence interval.

                  more regressors are likely to change things and may be theoretically justified, which could prove helpful (or harmful, depending).

                  Comment

                  Working...
                  X