5
$\begingroup$

I am trying to understand the uniformly most powerful (UMP) equivalence tests presented by Welleck (2010) with respect to $p$-values for the tests. In the case of the paired $t$ test example concerning measures before and after an intervention (pp 94–96) Wellek presents a UMP one-sample $t$ test of mean equivalence with symmetric equivalence boundaries where $ -\theta_{1} = \theta_{2} = \varepsilon = 0.5$, for the negativist null hypothesis H$^{-}_{0}\text{: }\left|t\right| \ge \tilde{C}_{\alpha,\nu}(\varepsilon)$ and gives:

  1. $n=23$ (sample size)
  2. $t=\frac{0.16}{\frac{3.99}{\sqrt{23}}} = 0.1923$ (test statistic)
  3. $\tilde{C}_{.05, 22}(0.5) = \sqrt{F^{-1}_{\nu_{\text{n}},\nu_{\text{d}},\psi^{2}}(0.05)} = 0.7595$ (critical value),

    where $F^{-1}$ is the quantile function (inverse CDF) of the noncentral F distribution, $\nu_{\text{n}}= 1$, $\nu_{\text{d}}= n-1 = 22$, and $\psi^{2} = n\varepsilon^{2} = 5.75$.

    We would therefore reject H$^{-}_{0}$ in favor of H$^{-}_{\text{A}}$, and conclude that, for $\alpha=0.05$ and $\varepsilon = 0.5$, mean measurements before and after intervention were equivalent (we would also draw this conclusion in a combined test of H$^{-}_{0}$ and H$^{+}_{0}$, since we would reject the first, but not the second).

    What is the $p$-value of this test? My guess is that it is [edited to remove the "$\mathbf{1-}$" confusion I had... forgot which tail I was in]:

  4. $p = \text{P}(\left|T\right| < \left|t\right|) = F_{\nu_{\text{n}},\nu_{\text{d}},\psi^{2}}(t^2) = 0.00882$,

where $F$ is the CDF of the noncentral F distribution.

Is this the correct $p$-value for this specific example, and is this the way to approach $p$-values generally for t tests for equivalence?

In the non-symmetric equivalence interval case, Wellek describes an iterative process to numerically solve for the values $C_{1}$ and $C_{2}$, corresponding to the rejection regions H$^{-}_{0}\text{: } t \le C_{1}$ or $t \ge C_{2}$. Any suggestions for how to obtain the $p$-values for $t$ in this case?

Update: would: $\psi = n\left|\theta_{1}\theta_{2}\right|$ in (4) above in the asymmetric case?

Update 2: Bounty for explicitly describing the steps required to iteratively estimate $p$-values in the asymmetric case as hinted at in Horst's response, including how to solve $\min_{i \in l,u}\left|T−t_{i}\right|=0$.


References

Wellek, S. (2010). Testing Statistical Hypotheses of Equivalence and Noninferiority. Chapman and Hall/CRC Press, second edition.

$\endgroup$
2
+200
$\begingroup$

Calculating the $p$-values for asymmetric equivalence bounds is a nested process. First, one has to set up the procedure to find the critical values for each $\alpha^*$. Note that this $\alpha^*$ is not the intended type I error but a placeholder for any such level. Then, one has to find the $\alpha^*$ such that one of the critical values equals the test statistic.

Procedure for Critical Values

Remember that the power function (or rejection probability) $\beta_.(\delta)$ for a point hypothesis test with test statistic $T$ and critical region $\mathbb{R}\setminus[C_1(\alpha),C_2(\alpha)]$ is $$\beta_.(\delta)=P_{\delta}(T>C_2(\alpha))+P_{\delta}(T<C_1(\alpha)).$$ Turning to an equivalence test, we somehow change hypothesis and alternative. For point hypotheses, the rejection probability $\alpha$ had to be maintained at the point the hypothesis was formulated about. For equivalence tests, it is the border of the equivalence region $]\epsilon_1,\epsilon_2[$ where the rejection probability has to be equal to $\alpha$. Then it will be $\leq\alpha$ also for the rest of the hypothesis space. That means that we have to plug the $\epsilon$s into the noncentrality parameters. So in order to get the equivalence test's critical values for a $\alpha^*$, we have to solve both $$T_{df,\epsilon_1}(C_1) - T_{df,\epsilon_1}(C_2) = \alpha^*$$ $$T_{df,\epsilon_2}(C_1) - T_{df,\epsilon_2}(C_2) = \alpha^*,$$ where $T_{df,\delta}$ denotes the CDF of a $t$-distribution with $df$ degrees of freedom and noncentrality parameter $\delta$. Numerically this can be done by the Newton-(Raphson)-Algorithm for the mapping $$h_{\alpha^*}(C_1,C_2) = (T_{df,\epsilon_1}(C_1) - T_{df,\epsilon_1}(C_2) -\alpha^*, T_{df,\epsilon_2}(C_1) - T_{df,\epsilon_2}(C_2) - \alpha^*)$$ In R and with df as degree of freedom, e1 and e2 as lower and upper equivalence margin and library(rootSolve) loaded, this can be implemented by

h <- function(c1,c2){ c(pt(c2,df=df,ncp=e1) - pt(c1,df=df,ncp=e1) - alphastar pt(c2,df=df,ncp=e2) - pt(c1,df=df,ncp=e2) - alphastar) } c <- multiroot(h,c(e1,e2))\$root #please ignore the backslash

Note that this involves numerical derivation of pt. It will be computationally faster and more precise to program the Newton algorithm by hand here. Then the function dt, the density of the $t$-distribution, can be used directly in the Jacobian of h.

$p$-values

Now as the $p$-value is the particular $\alpha^*$ that makes test statistic T and either critical value coincide, one has to solve $C_1(\alpha_1^*)+T=0$ and $C_2(\alpha_2^*)+T=0$. Then the $p$-value is the minimum of both $\alpha^*$s. So you have to nest the critical value algorithm within the solving process of $T-C_1(\alpha^*)$. In R this can look like this (e1, e2, df, test statistic T fixed as before):

hC1 <- function(alphastar){ #previous code block goes here C <- multiroot(h,c(e1,e2))\$root T+C[1] } p1 <- multiroot(hC,1).root hC2 <- function(alphastar){#previous code block goes here C <- multiroot(h,c(e1,e2))\$root T+C[2] } p2 <- multiroot(hC2,1)\$root pval <- min(p1,p2) In the end, pval is the asymmetric $p$-value. Maybe it is possible with some heuristics to tell in advance if p1 od p2 is smaller. Then one need only to calculate the respective value. But since the equivalence region is asymmetric, I'm not sure about it.

$\endgroup$
  • $\begingroup$ +1 mega-kudos, Horst. Do you have thoughts about initial values for $\alpha^{*}$? $\endgroup$ – Alexis May 9 '14 at 17:20
  • 1
    $\begingroup$ The critical values are strictly monotonous in $\alpha^*$ so "good" initial values don't change the result, they would spare you only some iterations. Maybe 0.5 is faster than what I suggested, 1. $\endgroup$ – Horst Grünbusch May 11 '14 at 15:56
3
$\begingroup$

For all kinds of tests, the $p$-value can be imagined as the particular $\alpha$ for which the present data's test would be on the edge between significance and nonsignificance. In other words: The $p$-value is the $\alpha$ that makes the critical value coincide with the test statistic. So if the test statistic of your data is $T$, the $p$-value is $F_{1,n-1,\psi^{2}=n\epsilon^{2}}(T^2)$.

Your guess is inspired from point hypothesis testing: There you can calculate $P_{H_0}(|T|<|t|)$. Here, $H_0$ is no more just a point in the parameter space. Likewise, $P_{H_0}$ is a set of different probability measures. (Also, I could not reproduce your evaluation of the cdf-F function.)

For the case of assymetric equivalence regions, it's basically similar but since the critical values change asymetrically as $\alpha$ changes, calculation of the $p$-values would require nested iterative calculations. Namely, let $t_l(\epsilon_l, \epsilon_u, \alpha)$ be the lower critical value calculated iteratively, similarly $t_u$. A crude approach to find the $p$-value would be to solve $\min_{i\in {l,u}} |T-t_i| = 0$ numerically for $\alpha$ e.g. by Newton's method. This $\alpha$ would be the $p$-value, because $T$ is on the edge of significance iff it conicides with one bound of the critical region. That's an expensive approach, though.

Note that due to FDA/ICH guidelines, the easier, more flexible (but also more conservative) TOST should be preferred to this exact test in biostatistical applications.

$\endgroup$
  • $\begingroup$ Horst, do you mean $\delta = n\varepsilon^{2}$? (see page 96, second paragraph). And if you do mean $\varepsilon^{2}$, then why this difference, when the statistic itself uses $\psi$ as I have reproduced from the text? Also: could you (or has someone else) elaborate on the iterative process in the asymmetric case? I have played (extensively) with Wellek's code for asymmetric intervals, but he leaves out the $p$-values. $\endgroup$ – Alexis Apr 28 '14 at 22:18
  • $\begingroup$ I think $\delta$ must indeed be $\psi (= n\varepsilon^{2}$) and not merely $\varepsilon^{2}$, otherwise for Wellek's example (Rejected H$^{-}_{0}$) we get $p=0.133$ which cannot be for $\alpha=0.05$. $\endgroup$ – Alexis Apr 28 '14 at 22:32
  • $\begingroup$ I think I had p. 94 in mind. Unfortunately, he somehow changed notation. Just edit my answer and I'll check it. $\endgroup$ – Horst Grünbusch Apr 28 '14 at 23:03
  • $\begingroup$ In the asymmetric case, does $\psi = n\left|\theta_{1}\theta_{2}\right|$? For example, in the asymmetric example referenced at the bottom of page 96, and implemented in the tt1st file, would $\psi = 36\left|-0.4716\times 0.3853\right| = 6.542$? $\endgroup$ – Alexis Apr 28 '14 at 23:20
  • $\begingroup$ @All: The source code of tt1st can be found at crcpress.com/product/isbn/9781439808184. I did not see there why $psi=n|\theta_1 \theta_2|$. (Also I might unintentionally have rejected your edit as I wanted to look it first up in the book.) $\endgroup$ – Horst Grünbusch Apr 29 '14 at 10:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.