Against NHST

50 21 December 2012 04:45AM

A summary of standard non-Bayesian criticisms of common frequentist statistical practices, with pointers into the academic literature.

Frequentist statistics is a wide field, but in practice by innumerable psychologists, biologists, economists etc, frequentism tends to be a particular style called “Null Hypothesis Significance Testing” (NHST) descended from R.A. Fisher (as opposed to eg. Neyman-Pearson) which is focused on

1. setting up a null hypothesis and an alternative hypothesis
2. calculating a p-value (possibly via a t-test or more complex alternatives like ANOVA)
3. and rejecting the null if an arbitrary threshold is passed.

NHST became nearly universal between the 1940s & 1960s (see Gigerenzer 2004, pg18), and has been heavily criticized for as long. Frequentists criticize it for:

1. practitioners & statistics teachers misinterpret the meaning of a p-value (LessWrongers too); Cohen on this persistent illusion:

What’s wrong with NHST? Well, among other things, it does not tell us what we want to know, and we so much want to know what we want to know that, out of desperation, we nevertheless believe that it does! What we want to know is, “Given these data, what is the probability that H0 is true?” But as most of us know, what it tells us is “Given that H0 is true, what is the probability of these (or more extreme) data?” These are not the same…

2. cargo-culting the use of 0.05 as an accept/reject threshold based on historical accident & custom (rather than using a loss function chosen through decision theory to set the threshold based on the cost of false positives).

Similarly, the cargo-culting encourages misuse of two-tailed tests, avoidance of multiple correction, data dredging, and in general, “p-value hacking”.
3. failing to compare many possible hypotheses or models, and limiting themselves to one - sometimes ill-chosen or absurd - null hypothesis and one alternative
4. deprecating the value of exploratory data analysis and depicting data graphically (see, for example, Anscombe’s quartet)
5. ignoring the more important summary statistic of “effect size”
6. ignoring the more important summary statistic of confidence intervals; this is related to how use of p-values leads to ignorance of the statistical power of a study - a small study may have only a small chance of detecting an effect if it exists, but turn in misleadingly good-looking p-values
7. because null hypothesis tests cannot accept the alternative, but only reject a null, they inevitably cause false alarms upon repeated testing

(An example from my personal experience of the cost of ignoring effect size and confidence intervals: p-values cannot (easily) be used to compile a meta-analysis (pooling of multiple studies); hence, studies often do not include the necessary information about means, standard deviations, or effect sizes & confidence intervals which one could use directly. So authors must be contacted, and they may refuse to provide the information or they may no longer be available; both have happened to me in trying to do my dual n-back & iodine meta-analyses.)

Critics’ explanations for why a flawed paradigm is still so popular focus on the ease of use and its weakness; from Gigerenzer 2004:

Hays (1963) had a chapter on Bayesian statistics in the second edition of his widely read textbook but dropped it in the subsequent editions. As he explained to one of us (GG) he dropped the chapter upon pressure from his publisher to produce a statistical cookbook that did not hint at the existence of alternative tools for statistical inference. Furthermore, he believed that many researchers are not interested in statistical thinking in the first place but solely in getting their papers published (Gigerenzer, 2000)…When Loftus (1993) became the editor of Memory & Cognition, he made it clear in his editorial that he did not want authors to submit papers in which p-, t-, or F-values are mindlessly being calculated and reported. Rather, he asked researchers to keep it simple and report figures with error bars, following the proverb that “a picture is worth more than a thousand p-values.” We admire Loftus for having had the courage to take this step. Years after, one of us (GG) asked Loftus about the success of his crusade against thoughtless significance testing. Loftus bitterly complained that most researchers actually refused the opportunity to escape the ritual. Even when he asked in his editorial letter to get rid of dozens of p-values, the authors insisted on keeping them in. There is something deeply engrained in the minds of many researchers that makes them repeat the same action over and over again.

Shifts away from NHST have happened in some fields. Medical testing seems to have made such a shift (I suspect due to the rise of meta-analysis):

Fidler et al. (2004b, 626) explain the spread of the reform in part by a shift from testing to estimation that was facilitated by the medical literature, unlike psychology, using a common measurement scale, to “strictly enforced editorial policy, virtually simultaneous reforms in a number of leading journals, and the timely re-writing [of] textbooks to fit with policy recommendations.” But their description of the process suggests that an accidental factor, the coincidence of several strong-willed editors, also mattered. For the classic collection of papers criticizing significance tests in psychology see Morrison and Hankel (1970) [The Significance Test Controversy: A Reader], and for a more recent collection of papers see Harlow et al. (1997) [What If There Were No Significance Tests?]. Nickerson (2000) provides a comprehensive survey of this literature.

More on these topics:

Sort By: Best
Comment author: 21 December 2012 08:32:21AM 9 points [-]

I think (hope?) most people already realize NHST is terrible. I would be much more interested in hearing if there were an equally-easy-to-use alternative without any baggage (preferably not requiring priors?)

Comment author: 23 December 2012 09:08:50PM 2 points [-]

NHST has been taught as The Method Of Science to lots of students. I remember setting these up explicitly in science class. I expect it will remain in the fabric of any given quantitative field until removed with force.

Comment author: 26 December 2012 07:04:54AM 1 point [-]

If you're right that that's how science works then that should make you distrustful of science. If they deserve any credibility, scientists must have some process by which they drop bad truth-finding methods instead of repeating them out of blind tradition. Do you believe scientific results?

Comment author: 30 December 2012 05:33:12AM 3 points [-]

If they deserve any credibility, scientists must have some process by which they drop bad truth-finding methods instead of repeating them out of blind tradition.

Plenty of otherwise-good science is done based on poor statistics. Keep in mind, there are tons and tons of working scientists, and they're already pretty busy just trying to understand the content of their fields. Many are likely to view improved statistical methods as an unneeded step in getting a paper published. Others are likely to view overthrowing NHST as a good idea, but not something that they themselves have the time or energy to do. Some might repeat it out of "blind tradition" -- but keep in mind that the "blind tradition" is an expensive-to-move Schelling point in a very complex system.

I do expect that serious scientific fields will, eventually, throw out NHST in favor of more fundamentally-sound statistical analyses. But, like any social change, it'll probably take decades at least.

Do you believe scientific results?

Unconditionally? No, and neither should you. Beliefs don't work that way.

If a scientific paper gives a fundamentally-sound statistical analysis of the effect it purports to prove, I'll give it more credence than a paper rejecting the null hypothesis at p < 0.05. On the other hand, a study rejecting the null hypothesis at p < 0.05 is going to provide far more useful information than a small collection of anecdotes, and both are probably better than my personal intuition in a field I have no experience with.

Comment author: 06 January 2013 07:01:20PM *  0 points [-]

Unconditionally? No, and neither should you. Beliefs don't work that way.

I should have said, "do you believe any scientific results?"

If a scientific paper gives a fundamentally-sound statistical analysis of the effect it purports to prove, I'll give it more credence than a paper rejecting the null hypothesis at p < 0.05. On the other hand, a study rejecting the null hypothesis at p < 0.05 is going to provide far more useful information than a small collection of anecdotes, and both are probably better than my personal intuition in a field I have no experience with.

To clarify, I wasn't saying that maybe you shouldn't believe scientific results because they use NHST specifically. I meant that if you think that scientists tend to stick with bad methods for decades then NHST probably isn't the only bad method they're using.

As you say though, NHST is helpful in many cases even if other methods might be more helpful. So I guess it doesn't say anything that awful about the way science works.

Comment author: 21 December 2012 05:09:17PM *  2 points [-]

Confidence intervals.

p<.05 means that the null hypothesis is excluded from the 95% confidence interval. Thus there is no political cost and every p-value recipe is a fragment of an existing confidence interval recipe.

added: also, the maximum likelihood estimate is a single number that is closely related to confidence intervals, but I don't know if is sufficiently well-known among statistically-ignorant scientists to avoid controversy.

Comment author: 21 December 2012 04:07:37PM 8 points [-]

Can you give me a concrete course of action to take when I am writing a paper reporting my results? Suppose I have created two versions of a website, and timed 30 people completing a task on each web site. The people on the second website were faster. I want my readers to believe that this wasn't merely a statistical coincidence. Normally, I would do a t-test to show this. What are you proposing I do instead? I don't want a generalization like "use Bayesian statistics, " but a concrete example of how one would test the data and report it in a paper.

Comment author: 21 December 2012 09:28:48PM 4 points [-]

You could use Bayesian estimation to compute credible differences in mean task completion time between your groups.

Described in excruciating detail in this pdf.

Comment author: 21 December 2012 04:17:29PM *  1 point [-]

Perhaps you would suggest showing the histograms of completion times on each site, along with the 95% confidence error bars?

Comment author: 21 December 2012 05:06:28PM 1 point [-]

Presumably not actually 95%, but, as gwern said, a threshold based on the cost of false positives.

Comment author: 21 December 2012 05:34:11PM *  3 points [-]

Yes, in this case you could keep using p-values (if you really wanted to...), but with reference to the value of, say, each customer. (This is what I meant by setting the threshold with respect to decision theory.) If the goal is to use on a site making millions of dollars*, 0.01 may be too loose a threshold, but if he's just messing with his personal site to help readers, a p-value like 0.10 may be perfectly acceptable.

* If the results were that important, I think there'd be better approaches than a once-off a/b test. Adaptive multi-armed bandit algorithms sound really cool from what I've read of them.

Comment author: 21 December 2012 04:56:16PM 1 point [-]

I'd suggest more of a scattergram than a histogram; superimposing 95% CIs would then cover the exploratory data/visualization & confidence intervals. Combine that with an effect size and one has made a good start.

Comment author: 22 December 2012 02:55:08AM 5 points [-]

This might be a good place to not that full Bayesianism is getting easier to practice in statistics. Doing fully Bayesian analysis has been tough for many models because it's computationally difficult since standard MCMC methods often don't scale that well, so you can only fit models with few parameters.

However, there are at least two statistical libraries STAN and PyMC3 (which I help out with) which implement Hamiltonian Monte Carlo (which scales well) and provide an easy language for model building. This allows you to fit relatively complex models, without thinking too much about how to do it.

Join the revolution!

Comment author: 14 April 2013 02:21:34AM 4 points [-]

A study with low statistical power has a reduced chance of detecting a true effect, but it is less well appreciated that low power also reduces the likelihood that a statistically significant result reflects a true effect. Here, we show that the average statistical power of studies in the neurosciences is very low. The consequences of this include overestimates of effect size and low reproducibility of results. There are also ethical dimensions to this problem, as unreliable research is inefficient and wasteful. Improving reproducibility in neuroscience is a key priority and requires attention to well-established but often ignored methodological principles.

Learned a new term:

Proteus phenomenon: The Proteus phenomenon refers to the situation in which the first published study is often the most biased towards an extreme result (the winner’s curse). Subsequent replication studies tend to be less biased towards the extreme, often finding evidence of smaller effects or even contradicting the findings from the initial study.

One of the interesting, and still counter-intuitive to me, aspects of power/beta is how it also changes the number of fake findings; typically, people think that must be governed by the p-value or alpha ("an alpha of 0.05 means that of the positive findings only 1 in 20 will be falsely thrown up by chance!"), but no:

For example, suppose that we work in a scientific field in which one in five of the effects we test are expected to be truly non-null (that is, R = 1 / (5 – 1) = 0.25) and that we claim to have discovered an effect when we reach p < 0.05; if our studies have 20% power, then PPV = 0.20 × 0.25 / (0.20 × 0.25 + 0.05) = 0.05 / 0.10 = 0.50; that is, only half of our claims for discoveries will be correct. If our studies have 80% power, then PPV = 0.80 × 0.25 / (0.80 × 0.25 + 0.05) = 0.20 / 0.25 = 0.80; that is, 80% of our claims for discoveries will be correct.

Third, even when an underpowered study discovers a true effect, it is likely that the estimate of the magnitude of that effect provided by that study will be exaggerated. This effect inflation is often referred to as the ‘winner’s curse’13 and is likely to occur whenever claims of discovery are based on thresholds of statistical significance (for example, p < 0.05) or other selection filters (for example, a Bayes factor better than a given value or a false-discovery rate below a given value). Effect inflation is worst for small, low-powered studies, which can only detect effects that happen to be large. If, for example, the true effect is medium-sized, only those small studies that, by chance, overestimate the magnitude of the effect will pass the threshold for discovery. To illustrate the winner’s curse, suppose that an association truly exists with an effect size that is equivalent to an odds ratio of 1.20, and we are trying to discover it by performing a small (that is, under-powered) study. Suppose also that our study only has the power to detect an odds ratio of 1.20 on average 20% of the time. The results of any study are subject to sampling variation and random error in the measurements of the variables and outcomes of interest. Therefore, on average, our small study will find an odds ratio of 1.20 but, because of random errors, our study may in fact find an odds ratio smaller than 1.20 (for example, 1.00) or an odds ratio larger than 1.20 (for example, 1.60). Odds ratios of 1.00 or 1.20 will not reach statistical significance because of the small sample size. We can only claim the association as nominally significant in the third case, where random error creates an odds ratio of 1.60. The winner’s curse means, therefore, that the ‘lucky’ scientist who makes the discovery in a small study is cursed by finding an inflated effect.

Publication bias and selective reporting of outcomes and analyses are also more likely to affect smaller, under-powered studies17. Indeed, investigations into publication bias often examine whether small studies yield different results than larger ones18. Smaller studies more readily disappear into a file drawer than very large studies that are widely known and visible, and the results of which are eagerly anticipated (although this correlation is far from perfect). A ‘negative’ result in a high-powered study cannot be explained away as being due to low power 19,20, and thus reviewers and editors may be more willing to publish it, whereas they more easily reject a small ‘negative study as being inconclusive or uninformative21. The protocols of large studies are also more likely to have been registered or otherwise made publicly available, so that deviations in the analysis plans and choice of outcomes may become obvious more easily. Small studies, conversely, are often subject to a higher level of exploration of their results and selective reporting thereof.

The actual strategy is the usual trick in meta-analysis: you take effects which have been studied enough to be meta-analyzed, take the meta-analysis result as the 'true' ground result, and re-analyze other results with that as the baseline. (I mention this because in some of the blogs, this seemed to come as news to them, that you could do this, but as far as I knew it's a perfectly ordinary approach.) This usually turns in depressing results, but actually it's not that bad - it's worse:

Any attempt to establish the average statistical power in neuroscience is hampered by the problem that the true effect sizes are not known. One solution to this problem is to use data from meta-analyses. Meta-analysis provides the best estimate of the true effect size, albeit with limitations, including the limitation that the individual studies that contribute to a meta-analysis are themselves subject to the problems described above. If anything, summary effects from meta-analyses, including power estimates calculated from meta-analysis results, may also be modestly inflated22.

Our results indicate that the median statistical power in neuroscience is 21%. We also applied a test for an excess of statistical significance72. This test has recently been used to show that there is an excess significance bias in the literature of various fields, including in studies of brain volume abnormalities73, Alzheimer’s disease genetics70,74 and cancer biomarkers75. The test revealed that the actual number (349) of nominally significant studies in our analysis was significantly higher than the number expected (254; p < 0.0001). Importantly, these calculations assume that the summary effect size reported in each study is close to the true effect size, but it is likely that they are inflated owing to publication and other biases described above.

Previous analyses of studies using animal models have shown that small studies consistently give more favourable (that is, ‘positive’) results than larger studies78 and that study quality is inversely related to effect size79–82.

Not mentioned, amusingly, are the concerns about applying research to humans:

In order to achieve 80% power to detect, in a single study, the most probable true effects as indicated by the meta-analysis, a sample size of 134 animals would be required for the water maze experiment (assuming an effect size of d = 0.49) and 68 animals for the radial maze experiment (assuming an effect size of d = 0.69); to achieve 95% power, these sample sizes would need to increase to 220 and 112, respectively. What is particularly striking, however, is the inefficiency of a continued reliance on small sample sizes. Despite the apparently large numbers of animals required to achieve acceptable statistical power in these experiments, the total numbers of animals actually used in the studies contributing to the meta-analyses were even larger: 420 for the water maze experiments and 514 for the radial maze experiments.

There is ongoing debate regarding the appropriate balance to strike between using as few animals as possible in experiments and the need to obtain robust, reliable findings. We argue that it is important to appreciate the waste associated with an underpowered study — even a study that achieves only 80% power still presents a 20% possibility that the animals have been sacrificed without the study detecting the underlying true effect. If the average power in neuroscience animal model studies is between 20–30%, as we observed in our analysis above, the ethical implications are clear.

Comment author: 14 April 2013 03:02:25AM 2 points [-]

Learned a new term:

Proteus phenomenon: The Proteus phenomenon refers to the situation in which the first published study is often the most biased towards an extreme result (the winner’s curse). Subsequent replication studies tend to be less biased towards the extreme, often finding evidence of smaller effects or even contradicting the findings from the initial study.

Oh great, researchers are going to end up giving this all sorts of names. Joseph Banks Rhine called it the decline effect, while Yitzhak Rabin* calls it the Truth Wears Off effect (after the Jonah Lehrer article). And now we have the Proteus phenomenon. Clearly, I need to write a paper declaring my discovery of the It Was Here, I Swear! effect.

* Not that one.

Comment author: 14 April 2013 03:31:40AM 0 points [-]

Clearly, I need to write a paper declaring my discovery of the It Was Here, I Swear! effect.

Make sure you cite my paper "Selection Effects and Regression to the Mean In Published Scientific Studies"

Comment author: 09 January 2013 01:59:59AM *  3 points [-]

"Do We Really Need the S-word?" (American Scientist) covers many of the same points. I enjoyed one anecdote:

Curious about the impact a ban on the s-word might have, three years ago I began banning the word from my two-semester Methods of Data Analysis course, which is taken primarily by nonstatistics graduate students. My motivation was to force students to justify and defend the statements they used to summarize results of a statistical analysis. In previous semesters I had noticed students using the s-word as a mask, an easily inserted word to replace the justification of assumptions and difficult decisions, such as arbitrary cutoffs. My students were following the example dominant in published research—perpetuating the false dichotomy of calling statistical results either significant or not and, in doing so, failing to acknowledge the vast and important area between the two extremes. The ban on the s-word seems to have left my students with fewer ways to skirt the difficult task of effective justification, forcing them to confront the more subtle issues inherent in statistical inference.

An unexpected realization I had was just how ingrained the word already was in the brains of even first-year graduate students. At first I merely suggested—over and over again—that students avoid using the word. When suggestion proved not to be enough, I evinced more motivation by taking off precious points at the sight of the word. To my surprise, it still appears, and students later say they didn’t even realize they had used it! Even though using this s-word doesn’t carry the possible consequence of having one’s mouth washed out with soap, I continue to witness the clasp of hands over the mouth as the first syllable tries to sneak out—as if the speakers had caught themselves nearly swearing in front of a child or parent.

Comment author: 21 December 2012 07:54:05PM 3 points [-]

setting up a null hypothesis and an alternative hypothesis calculating a p-value (possibly via a t-test or more complex alternatives like ANOVA) and accepting the null if an arbitrary threshold is passed.

Typo?

Comment author: 21 December 2012 08:28:25PM 0 points [-]

If you run together an enumerated list like that, of course it looks weird...

Comment author: 21 December 2012 08:53:19PM 6 points [-]

"Accepting the null if the threshold is passed." Not rejecting?

Comment author: 21 December 2012 08:57:14PM 2 points [-]

Oh. Yeah, should be rejecting.

Comment author: 21 December 2012 06:44:34AM *  3 points [-]

deprecating the value of exploratory data analysis and depicting data graphically

Should this be:

(deprecating the value of exploratory data analysis) and (depicting data graphically)

or

deprecating [(the value of exploratory data analysis) and (depicting data graphically)]?

ETA: Also, very nice article! I'm glad that you point out that NHST is only a small part of frequentist statistics.

Comment author: 21 December 2012 01:10:30PM 2 points [-]

I guess the latter.

Comment author: 26 April 2013 09:57:10PM *  2 points [-]

Via http://www.scottbot.net/HIAL/?p=24697 I learned that Wikipedia actually has a good roundup of misunderstandings of p-values:

The p-value does not in itself allow reasoning about the probabilities of hypotheses; this requires multiple hypotheses or a range of hypotheses, with a [prior distribution][1] of likelihoods between them, as in [Bayesian statistics][2], in which case one uses a [likelihood function][3] for all possible values of the prior, instead of the p-value for a single null hypothesis.

The p-value refers only to a single hypothesis, called the null hypothesis, and does not make reference to or allow conclusions about any other hypotheses, such as the [alternative hypothesis][4] in Neyman–Pearson [statistical hypothesis testing][5]. In that approach one instead has a decision function between two alternatives, often based on a [test statistic][6], and one computes the rate of [Type I and type II errors][7] as α and β. However, the p-value of a test statistic cannot be directly compared to these error rates α and β – instead it is fed into a decision function.

There are several common misunderstandings about p-values.[[16]][8][[17]][9]

1. The p-value is not the probability that the null hypothesis is true, nor is it the probability that the alternative hypothesis is false – it is not connected to either of these.
In fact, [frequentist statistics][10] does not, and cannot, attach probabilities to hypotheses. Comparison of [Bayesian][11] and classical approaches shows that a p-value can be very close to zero while the [posterior probability][12] of the null is very close to unity (if there is no alternative hypothesis with a large enough a priori probability and which would explain the results more easily). This is [Lindley's paradox][13]. But there are also a priori probability distributions where the [posterior probability][12] and the p-value have similar or equal values.[[18]][14]
2. The p-value is not the probability that a finding is "merely a fluke."
As the calculation of a p-value is based on the assumption that a finding is the product of chance alone, it patently cannot also be used to gauge the probability of that assumption being true. This is different from the real meaning which is that the p-value is the chance of obtaining such results if the null hypothesis is true.
3. The p-value is not the probability of falsely rejecting the null hypothesis. This error is a version of the so-called [prosecutor's fallacy][15].
4. The p-value is not the probability that a replicating experiment would not yield the same conclusion. Quantifying the replicability of an experiment was attempted through the concept of [p-rep][16] (which is heavily [criticized][17])
5. The significance level, such as 0.05, is not determined by the p-value.
Rather, the significance level is decided before the data are viewed, and is compared against the p-value, which is calculated after the test has been performed. (However, reporting a p-value is more useful than simply saying that the results were or were not significant at a given level, and allows readers to decide for themselves whether to consider the results significant.)
6. The p-value does not indicate the size or importance of the observed effect (compare with [effect size][18]). The two do vary together however – the larger the effect, the smaller sample size will be required to get a significant p-value.
Comment author: 21 December 2012 12:43:36PM 2 points [-]

And yet even you who are more against frequentist statistics than most (Given that you are even writing this among other things on the topic) inevitably use the frequentist tools. What I'd be interested in is a good and short(as short as it can be) summary of what methods should be followed to remove as many of the problems of frequentist statistics with properly defined cut-offs for p-values and everything else, where can we fully adapt Bayes, where we can minimize the problems of the frequentist tools and so on. You know, something that I can use on its own to interpret the data if I am to conduct an experiment today the way that currently seems best.

Comment author: 21 December 2012 04:53:40PM 5 points [-]

inevitably use the frequentist tools.

No, I don't. My self-experiments have long focused on effect sizes (an emphasis which is very easy to do without disruptive changes), and I have been using BEST as a replacement for t-tests for a while, only including an occasional t-test as a safety blanket for my frequentist readers.

If non-NHST frequentism or even full Bayesianism were taught as much as NHST and as well supported by software like R, I don't think it would be much harder to use.

Comment author: 28 December 2012 07:54:18AM 1 point [-]

I can't find BEST (as a statistical test or similar...) on Google. What test do you refer to?

Comment author: 28 December 2012 04:38:37PM 2 points [-]
Comment author: 22 December 2012 02:32:19AM 0 points [-]

If non-NHST frequentism

That'd be essentially Bayesianism with the (uninformative improper) priors (uniform for location parameters and logarithms of scale parameters) swept under the rug, right?

Comment author: 25 December 2012 02:04:12AM 1 point [-]

Not at all (I wrote a post refuting this a couple months ago but can't link it from my phone)

Comment author: 25 December 2012 10:16:52PM 3 points [-]
Comment author: 26 December 2012 06:10:43AM 2 points [-]

Thanks!

Comment author: 22 December 2012 03:51:43AM 1 point [-]

I really couldn't presume to say.

Comment author: 21 December 2012 03:51:00PM -1 points [-]

'Frequentist tools' are common approximations, loaded with sometimes-applicable interpretations. A Bayesian can use the same approximation, even under the same name, and yet not be diving into Frequentism.

Comment author: 25 March 2013 10:41:35PM *  1 point [-]

Three historical episodes in which the application of null hypothesis significance testing (NHST) led to the mis-interpretation of data are described. It is argued that the pervasive use of this statistical ritual impedes the accumulation of knowledge and is unfit for use.

(These deadly examples obviously lend themselves to Bayesian critique, but could just as well be classified by a frequentist under several of the rubrics in OP: under failures to adjust thresholds based on decision theory, and failure to use meta-analysis or other techniques to pool data and turn a collection of non-significant results into a significant result.)

Comment author: 10 January 2013 05:38:04PM 1 point [-]
Comment author: 31 December 2012 11:51:26PM 1 point [-]

http://library.mpib-berlin.mpg.de/ft/gg/GG_Null_2004.pdf

In 1962, Jacob Cohen reported that the experiments published in a major psychology journal had, on average, only a 50 : 50 chance of detecting a medium-sized effect if there was one. That is, the statistical power was as low as 50%. This result was widely cited, but did it change researchers’ practice? Sedlmeier and Gigerenzer (1989) checked the studies in the same journal, 24 years later, a time period that should allow for change. Yet only 2 out of 64 researchers mentioned power, and it was never estimated. Unnoticed, the average power had decreased (researchers now used alpha adjustment, which shrinks power). Thus, if there had been an effect of a medium size, the researchers would have had a better chance of finding it by throwing a coin rather than conducting their experiments. When we checked the years 2000 to 2002, with some 220 empirical articles, we finally found 9 researchers who computed the power of their tests. Forty years after Cohen, there is a first sign of change.

Comment author: 23 February 2013 04:21:32AM 0 points [-]

If the papers in the OP and comments are not enough reading material, there's many links and citations in http://stats.stackexchange.com/questions/10510/what-are-good-references-containing-arguments-against-null-hypothesis-significan (which is only partially redundant with this page, skimming).

Comment author: 23 February 2013 04:19:07AM *  0 points [-]

"P Values are not Error Probabilities", Hubbard & Bayarri 2003

...researchers erroneously believe that the interpretation of such tests is prescribed by a single coherent theory of statistical inference. This is not the case: Classical statistical testing is an anonymous hybrid of the competing and frequently contradictory approaches formulated by R.A. Fisher on the one hand, and Jerzy Neyman and Egon Pearson on the other. In particular, there is a widespread failure to appreciate the incompatibility of Fisher’s evidential p value with the Type I error rate, α, of Neyman–Pearson statistical orthodoxy. The distinction between evidence (p’s) and error (α’s) is not trivial. Instead, it reflects the fundamental differences between Fisher’s ideas on significance testing and inductive inference, and Neyman–Pearson views of hypothesis testing and inductive behavior. Unfortunately, statistics textbooks tend to inadvertently cobble together elements from both of these schools of thought, thereby perpetuating the confusion. So complete is this misunderstanding over measures of evidence versus error that is not viewed as even being a problem among the vast majority of researchers.

An interesting bit:

Fisher was insistent that the significance level of a test had no ongoing sampling interpretation. With respect to the .05 level, for example, he emphasized that this does not indicate that the researcher “allows himself to be deceived once in every twenty experiments. The test of significance only tells him what to ignore, namely all experiments in which significant results are not obtained” (Fisher 1929, p. 191). For Fisher, the significance level provided a measure of evidence for the “objective” disbelief in the null hypothesis; it had no long-run frequentist characteristics.

Indeed, interpreting the significance level of a test in terms of a Neyman–Pearson Type I error rate, α, rather than via a p value, infuriated Fisher who complained:

“In recent times one often-repeated exposition of the tests of significance, by J. Neyman, a writer not closely associated with the development of these tests, seems liable to lead mathematical readers astray, through laying down axiomatically, what is not agreed or generally true, that the level of significance must be equal to the frequency with which the hypothesis is rejected in repeated sampling of any fixed population allowed by hypothesis. This intrusive axiom, which is foreign to the reasoning on which the tests of significance were in fact based seems to be a real bar to progress....” (Fisher 1945, p. 130).