# Are sampling distributions legitimate for inference?

Some Bayesians attack frequentist inference stating that "there is no unique sampling distribution" because it depends on the intentions of the researcher (Kruschke, Aguinis, & Joo, 2012, p. 733).

For instance, say a researcher starts data collection, but his funding was unexpectedly cut after 40 participants. How would the sampling distributions (and subsequent CIs and p-values) even be defined here? Would we just assume each constituent sample has N = 40? Or would it consist of samples with different N, with each size determined by other random times his funding may have been cut?

The t, F, chi-square (etc.), null distributions found in textbooks all assume that the N is fixed and constant for all the constituent samples, but this may not be true in practice. With every different stopping procedure (e.g., after a certain time interval or until my assistant gets tired) there seems to be a different sampling distribution, and using these 'tried and true' fixed-N distributions is inappropriate.

How damaging is this criticism to the legitimacy of frequentist CIs and p-values? Are there theoretical rebuttals? It seems that by attacking the concept of the sampling distribution, the entire edifice of frequentist inference is tenuous.

Any scholarly references are greatly appreciated.

• The citation is for: Kruschke, J. K., Aguinis, H., & Joo, H. (2012). The time has come: Bayesian methods for data analysis in the organizational sciences. But Kruschke has used it before in: (2010) Bayesian data analysis and (2010) What to believe: Bayesian methods for data analysis.
– ATJ
Jan 30 '14 at 23:27

Typically you'd carry out inference conditional on the actual sample size $n$, because it's ancillary to the parameters of interest; i.e. it contains no information about their true values, only affecting the precision with which you can measure them. Cox (1958), "Some Problems Connected with Statistical Inference", Ann. Math. Statist. 29, 2 is usually cited as first explicating what's sometimes known as the Conditionality Principle, though it was implicit in much earlier work, harking back to Fisher's idea of "relevant subsets".

If your researcher's funding was cut off because results so far were disappointing, then of course $n$ isn't ancillary. Perhaps the simplest illustration of the problem is estimation of a Bernoulli probability from either a binomial (fixed no. of trials) or negative binomial (fixed no. successes) sampling scheme. The sufficient statistic is the same under either, but its distribution differs. How would you analyze an experiment where you didn't know which was followed? Berger & Wolpert (1988), The Likelihood Principle discuss the implications of this & other stopping rules for inference.

You might want to think about what happens if you don't take any sampling distribution into account. Armitage (1961), "Comment on 'Consistency in Statistical Inference and Decision' by Smith", JRSS B, 23,1 pointed out that if you sample $x$ from a normal distribution until $\sqrt{n} \bar{x} \leq k$, the likelihood ratio for testing that the mean $\mu=0$ vs $\mu\neq0$ is $\frac{L(0)}{L(\bar{x})}\leq \mathrm{e}^{-k^2/2}$, so the researcher can set a bound on this in advance by an appropriate choice of $k$. Only a frequentist analysis can take the distribution of the likelihood ratio under this rather unfair-seeming sampling scheme into account. See the responses of Kerridge (1963), "Bounds for the frequency of misleading Bayes inferences", Ann. Math. Stat., 34, Cornfield (1966), "Sequential trials, sequential analysis, and the likelihood principle", The American Statistician, 20, 2, & Kadane (1996), "Reasoning to a foregone conclusion", JASA, 91, 435

Pointing out the dependence of frequentist inference on a researcher's intentions is a handy dig at people (if there still are any) who get on their high horse about the "subjectivity" of Bayesian inference. Personally, I can live with it; the performance of a procedure over a long series of repetitions is always going to be something more or less notional, which doesn't detract from its being a useful thing to consider ("a calibration of the likelihood" was how Cox described p-values). From the dates of the references you might have noticed that these issues aren't very new; attempts to settle them by a priori argumentation have largely died down (except on the Internet, always behind the times except in trivial matters) & been replaced by acknowledgement that neither Bayesian nor frequentist statistics are going to collapse under the weight of their internal contradictions, & that there's more than one useful way to apply probability theory to inference from noisy data.

PS: Thinking to add a counter-balance to Berger & Wolpert I happened upon Cox & Mayo (2010), "Objectivity and Conditionality in Frequentist Inference" in Error and Inference. There's quite likely an element of wishful thinking in my assertion that the debate has died down, but it's striking how little new there is to be said on the matter after half a century or so. (All the same, this is a concise & eloquent defence of frequentist ideas.)

• +1 (long time ago). I am wondering if Armitage's reasoning can be adapted to the well-known example of binomial vs neg-binomial sampling; e.g. observing TTTTTH sequence of coin tosses yields either p=0.03 or p=0.1 depending on the stopping rule. So, if we now consider yet another stopping rule, e.g. "Keep tossing until binomial p<0.05 and there was at least one H and at least one T", then it becomes rather intuitive that one should rather not ignore this stopping rule for inference (despite violating the Likelihood Principle). Does this make sense? Mar 17 '17 at 11:09

The short answer to your question is: it depends who you ask ;-) Die-hard Bayesians will declare victory over, or at least parity with, frequentist methodology. Die-hard frequentists will default to "This can't be answered". The other 99% of statisticians will use whatever methods have been shown to be reliable under inerrupted experiments.

I know that the sensitivity of the sampling distribution to the intentions of the researcher can be troubling, and there is really no good solution to that problem. Bayesians and frequentists alike must use some subjectivity and judgement in deciding how to form an inference. However, I think you are taking an example from an area that is generally controversial and laying the problems solely at the feet of frequentist inference. The sequential and/or stopped experiments are classic examples of the subjective nature of inference ... and to which there is no absolutely objective and agreed upon answer.

What about regular inference, where you actually collect the sample you intended to get? Here, I think the frequentists have the upper hand, as CI's and p-values are well calibrated wrt their repeated sampling properties, whereas Bayesian inference retains its personal and subjective nature.

If you want a more theoretical exposition of the Bayesian response, I would read about "conditional inference" with key researchers being Nancy Reid and Lehmann.