.

ONE YEAR AGO: …and growing more relevant all the time. Rather than leak any of my new book*, I reblog some earlier posts, even if they’re a bit scruffy. This was first blogged here (with a slightly different title). It’s married to posts on “the P-values overstate the evidence against the null fallacy”, such as this, and is wedded to this one on “How to Tell What’s True About Power if You’re Practicing within the Frequentist Tribe”. 

In their “Comment: A Simple Alternative to p-values,” (on the ASA P-value document), Benjamin and Berger (2016) recommend researchers report a pre-data Rejection Ratio:

It is the probability of rejection when the alternative hypothesis is true, divided by the probability of rejection when the null hypothesis is true, i.e., the ratio of the power of the experiment to the Type I error of the experiment. The rejection ratio has a straightforward interpretation as quantifying the strength of evidence about the alternative hypothesis relative to the null hypothesis conveyed by the experimental result being statistically significant. (Benjamin and Berger 2016, p. 1)

The recommendation is much more fully fleshed out in a 2016 paper by Bayarri, Benjamin, Berger, and Sellke (BBBS 2016): Rejection Odds and Rejection Ratios: A Proposal for Statistical Practice in Testing Hypotheses. Their recommendation is:

…that researchers should report the ‘pre-experimental rejection ratio’ when presenting their experimental design and researchers should report the ‘post-experimental rejection ratio’ (or Bayes factor) when presenting their experimental results. (BBBS 2016, p. 3)….

The (pre-experimental) ‘rejection ratio’ R_pre , the ratio of statistical power to significance threshold (i.e., the ratio of the probability of rejecting under H₁ and H₀ respectively), is shown to capture the strength of evidence in the experiment for H₁ over H₀. (ibid., p. 2)

But it does no such thing! [See my post from the FUSION 2016 conference here.] J. Berger, and his co-authors, will tell you the rejection ratio (and a variety of other measures created over the years) are entirely frequentist because they are created out of frequentist error statistical measures. But a creation built on frequentist measures doesn’t mean the resulting animal captures frequentist error statistical reasoning. It might be a kind of Frequentstein monster! [1]

~~~~~~~~~~~~~~

The Law of Comparative Support

It comes from a comparativist support position which has intrinsic plausibility, although I do not hold to it. It is akin to what some likelihoodists call “the law of support”: if H₁ make the observed results probable, while H₀ make them improbable, then the results are strong (or at least better) evidence for H₁ compared to H₀ . It appears to be saying (sensibly) that you have better evidence for a hypothesis that best “explains” the data, only this is not a good measure of explanation. It is not generally required H₀ and H₁ be exhaustive. Even if you hold a comparative support position, the “ratio of statistical power to significance threshold” isn’t a plausible measure for this. Now BBBS also object to the Rejection Ratio, but only largely because it’s not sensitive to the actual outcome; so they recommend the Bayes Factor post data. My criticism is much, much deeper. To get around the data-dependent part, let’s assume throughout that we’re dealing with a result just statistically significant at the α level.

~~~~~~~~~~~~~~

Take a one-sided Normal test T+: with n iid samples:

H₀: µ ≤  0 against H₁: µ >  0

σ = 10,  n = 100,  σ/√n =σ_x= 1,  α = .025.

So the test would reject H₀ iff Z > c_.025 =1.96. (1.96. is the “cut-off”.)

People often talk of a test “having a power” but the test actually specifies a power function that varies with different point values in the alternative H₁ . The power of test T+ in relation to point alternative µ’ is

Pr(Z > 1.96; µ = µ’).

We can abbreviate this as POW(T+,µ’).

~~~~~~~~~~~~~~

Jacob Cohen’s slips

By the way, Jacob Cohen, a founder of power analysis, makes a few slips in introducing power, even though he correctly computes power through the book (so far as I know). [2] Someone recently reminded me of this, and given the confusion about power, maybe it’s had more of an ill effect than I assumed.

In the first sentence on p. 1 of Statistical Power Analysis for the Behavioral Sciences, Cohen says “The power of a statistical test is the probability it will yield statistically significant results.” Also faulty, and for two reasons, is what he says on p. 4: “The power of a statistical test of a null hypothesis is the probability that it will lead to the rejection of the null hypothesis, i.e., the probability that it will result in the conclusion that the phenomenon exists.”

Do you see the two mistakes? 

~~~~~~~~~~~~~~

Examples of alternatives against which T+ has high power:

• If we add σ_x (i.e.,σ/√n) to the cut-off  (1.96) we are at an alternative value for µ that test T+ has .84 power to detect. In this example, σ_x = 1.
• If we add 3σ_x to the cut-off we are at an alternative value for µ that test T+ has ~ .999 power to detect. This value, which we can write as µ^.999 = 4.96

Let the observed outcome just reach the cut-off to reject the null, z₀ = 1.96.

If we were to form a “rejection ratio” or a “likelihood ratio” of μ = 4.96 compared to μ₀ = 0 using

[POW(T+, 4.96)]/α,

it would be 40.  (.999/.025).

It is absurd to say the alternative 4.96 is supported 40 times as much as the null, even understanding support as comparative likelihood or something akin. The data 1.96 are even closer to 0 than to 4.96. The same point can be made with less extreme cases.) What is commonly done next is to assign priors of .5 to the two hypotheses, yielding

Pr(H₀|z₀) = 1/(1 + 40) = .024, so Pr(H₁|z₀) = .976.

Such an inference is highly unwarranted and would almost always be wrong. Back to our question:

Here’s my explanation for why some think it’s plausible to compute comparative evidence this way:

I presume it stems comes from the comparativist support position noted above. I’m guessing they’re reasoning as follows:

The probability is very high that z > 1.96 under the assumption that μ = 4.96.

The probability is low that z > 1.96 under the assumption that μ = μ₀ = 0.

We’ve observed z₀ = 1.96 (so you’ve observed z > 1.96).

Therefore, μ = 4.96 makes the observation more probable than does  μ = 0.

Therefore the outcome is (comparatively) better evidence for μ= 4.96 than for μ = 0.

But the “outcome” for a likelihood is to be the specific outcome, and the comparative appraisal of which hypothesis accords better with the data only makes sense when one keeps to this.

I can pick any far away alternative I like for purposes of getting high power, and we wouldn’t want to say that just reaching the cut-off (1.96) is good evidence for it! Power works in the reverse. That is,

If POW(T+,µ’) is high, then z₀ = 1.96 is poor evidence that μ  > μ’.

That’s because were μ as great as μ’, with high probability we would have observed a larger z value (smaller p-value) than we did. Power may, if one wishes, be seen as a kind of distance measure, but (just like α) it is inverted.

(Note that our inferences take the form μ > μ’, μ < μ’, etc. rather than to a point value.) 

In fact:

if Pr(Z > z₀;μ =μ’) = high , then Z = z₀ is strong evidence that  μ < μ’!

Rather than being evidence for μ’, the statistically significant result is evidence against μ being as high as μ’.
~~~~~~~~~~~~~~

A post by Stephen Senn:

In my favorite guest post by Stephen Senn here, Senn strengthens a point from his 2008 book (p. 201), namely, that the following is “nonsense”:

[U]pon rejecting the null hypothesis, not only may we conclude that the treatment is effective but also that it has a clinically relevant effect. (Senn 2008, p. 201)

Now the test is designed to have high power to detect a clinically relevant effect (usually .8 or .9). I happen to have chosen an extremely high power (.999) but the claim holds for any alternative that the test has high power to detect. The clinically relevant discrepancy, as he describes it, is one “we should not like to miss”, but obtaining a statistically significant result is not evidence we’ve found a discrepancy that big. 

Supposing that it is, is essentially  to treat the test as if it were:

H₀: μ < 0 vs H₁: μ  > 4.96

This, he says,  is “ludicrous”as it:

would imply that we knew, before conducting the trial, that the treatment effect is either zero or at least equal to the clinically relevant difference. But where we are unsure whether a drug works or not, it would be ludicrous to maintain that it cannot have an effect which, while greater than nothing, is less than the clinically relevant difference. (Senn, 2008, p. 201)

The same holds with H₀: μ = 0 as null.

If anything, it is the lower confidence limit that we would look at to see what discrepancies from 0 are warranted. The lower .975 limit (if one-sided) or .95 (if two-sided) would be 0 and .3, respectively. So we would be warranted in inferring from z:

μ  > 0 or μ  > .3.

~~~~~~~~~~~~~~

What does the severe tester say?

In sync with the confidence interval, she would say SEV(μ > 0) = .975 (if one-sided), and would also note some other benchmarks, e.g., SEV(μ > .96) = .84.

Equally important for her is a report of what is poorly warranted. In particular the claim that the data indicate

μ > 4.96

would be wrong over 99% of the time!

Of course, I would want to use the actual result, rather than the cut-off for rejection (as with power) but the reasoning is the same, and here I deliberately let the outcome just hit the cut-off for rejection.

~~~~~~~~~~~~~~

The (Type 1, 2 error probability) trade-off vanishes

Notice what happens if we consider the “real Type 1 error” as Pr(H₀|z₀)

Since Pr(H₀|z₀) decreases with increasing power, it decreases with decreasing Type 2 error. So we know that to identify “Type 1 error” and Pr(H₀|z₀) is to use language in a completely different way than the one in which power is defined. For there we must have a trade-off between Type 1 and 2 error probabilities.

Upshot

Using size/ power as a likelihood ratio, or even as a preregistrated estimate of expected strength of evidence (with which to accord a rejection) is problematic. The error statistician is not in the business of making inferences to point values, nor to comparative appraisals of different point hypotheses. It’s not unusual for criticisms to start out forming these ratios, and then blame the “tail areas” for exaggerating the evidence against the test hypothesis. We don’t form those ratios. But the pre-data Rejection Ratio is also misleading as an assessment alleged to be akin to a Bayes ratio or likelihood assessment. You can marry frequentist components and end up with something frequentsteinian.

REFERENCES

Bayarri, M., Benjamin, D., Berger, J., & Sellke, T. (2016, in press). “Rejection Odds and Rejection Ratios: A Proposal for Statistical Practice in Testing Hypotheses“, Journal of Mathematical Psychology

Benjamin, D. & Berger J. 2016. “Comment: A Simple Alternative to P-values,” The American Statistician (online March 7, 2016).

Cohen, J. 1988. Statistical Power Analysis for the Behavioral Sciences. 2nd ed. Hillsdale, NJ: Erlbaum.

Mayo, D. 2016. “Don’t throw out the Error Control Baby with the Error Statistical Bathwater“. (My comment on the ASA document)

Mayo, D. 2003. Comments on J. Berger’s, “Could Jeffreys, Fisher and Neyman have Agreed on Testing?  (pp. 19-24)

*Mayo, D. Statistical Inference as Severe Testing, forthcoming (2017) CUP.

Senn, S. 2008. Statistical Issues in Drug Development, 2^nd ed. Chichster, New Sussex: Wiley Interscience, John Wiley & Sons.

Wasserstein, R. & Lazar, N. 2016. “The ASA’s Statement on P-values: Context, Process and Purpose”, The American Statistician (online March 7, 2016).

[1] I don’t say there’s no context where the Rejection Ratio has a frequentist role. It may arise in a diagnostic screening or empirical Bayesian context where one has to deal with a dichotomy. See, for example, this post (“Beware of questionable front page articles telling you to beware…”)

[2] It may also be found in Neyman! (Search this blog under Neyman’s Nursery.) However, Cohen uniquely provides massive power computations, before it was all computerized.