Slashdot Mirror
Scientists Propose To Raise the Standards For Statistical Significance In Research Studies (sciencemag.org)
sciencehabit shares a report from Science Magazine: A megateam of reproducibility-minded scientists is renewing a controversial proposal to raise the standard for statistical significance in research studies. They want researchers to dump the long-standing use of a probability value (p-value) of less than 0.05 as the gold standard for significant results, and replace it with the much stiffer p-value threshold of 0.005. Backers of the change, which has been floated before, say it could dramatically reduce the reporting of false-positive results -- studies that claim to find an effect when there is none -- and so make more studies reproducible. And they note that researchers in some fields, including genome analysis, have already made a similar switch with beneficial results.
"If we're going to be in a world where the research community expects some strict cutoff ... it's better that that threshold be .005 than .05. That's an improvement over the status quo," says behavioral economist Daniel Benjamin of the University of Southern California in Los Angeles, first author on the new paper, which was posted 22 July as a preprint article on PsyArXiv and is slated for an upcoming issue of Nature Human Behavior. "It seemed like this was something that was doable and easy, and had worked in other fields."
"If we're going to be in a world where the research community expects some strict cutoff ... it's better that that threshold be .005 than .05. That's an improvement over the status quo," says behavioral economist Daniel Benjamin of the University of Southern California in Los Angeles, first author on the new paper, which was posted 22 July as a preprint article on PsyArXiv and is slated for an upcoming issue of Nature Human Behavior. "It seemed like this was something that was doable and easy, and had worked in other fields."
And if you ask me, that's not statistically significant enough.
Props to GNAA.
Make it Six Sigma, which is really 4.78 sigma (or something like that, I forget the actual number), because they allow a fudge factor to accommodate the fact that 6 sigma isn't realistic.
"National Security is the chief cause of national insecurity." - Celine's First Law
your a doosh
Global Climate Chance doesn't exclude cooling so you should get onboard.
A behavioral economist wanting stricter significance?
Career suicide.
That doesn't conform to the narrative so we shouldn't talk about that.
Just like they told us by 1990 that you wouldn't be able to go outside without a gas mask. Their point is still valid even if wrong.
There's a trade-off between sensitivity and specificity. If you increase the threshold for "significance", you reduce the power to discover a significant effect when it truly does exist.
And a major part of the problem with scientific studies is that they are already underpowered. According to conventional wisdom, ideally, scientists should strive for a power of about 80% (i.e., an 80% chance of detecting an effect if it truly exists), but very few studies actually achieve power of this level. In many fields, the power is less than 50% and sometimes much less.
Underpowered studies result in two major problems:
1) Most obviously, an underpowered study results in a greater number of FALSE NEGATIVES. You fail to find a true effect. You will either publish your incorrect result of no effect. (And why should we consider published false positives to be any worse than false negatives?) Alternatively, perhaps you don't publish your study because you couldn't reach significance. This exacerbates the "file-drawer effect" and also results in wasted research dollars because the results aren't published.
2) Somewhat counterintuitively, underpowered studies are often also more likely to result in FALSE POSITIVES. This is because, when your power to detect a true effect is low, and if you test a large number of effects that are unlikely to be null, most of the hypotheses that you say are "significantly" non-null will actually be false positives. We would say that the "false discovery rate" tends to be very high when the power is low.
Reducing the level of significance will do little to address these problems, and in some cases may even exacerbate the problem.
The key is *to move away from the binary concept of "significance" altogether*. It's obviously artificial to have an arbitrary numerical cutoff for "matters" vs. "doesn't matter", and this is not what Ronald Fisher intended when he popularized the p-value or developed the concept of "significance".
What we should be doing is measuring and reporting effect sizes along with their credible intervals. While using priors that are based on our real state of knowledge. In other words, we should be doing Bayesian statistics.
or is this more attempts to discredit climate change science by setting an impossible bar for proof? I honestly don't know, it's the first I've heard of this. Still, it's hard to imagine it being controversial otherwise.
Hi! I make Firefox Plug-ins. Check 'em out @ https://addons.mozilla.org/en-US/firefox/addon/youtube-mp3-podcaster/
I'm not convinced this will help. There are a couple of issues here. Often, the experimental design can be changed, like how certain variables are controlled for, to get a p-value that's below the threshold. The other problem is that p-value is sensitive to the sample size. If you want a lower p-value, increase the sample size. In many cases, p-values aren't a good way to show whether a result is useful or not.
I'm a meteorologist and I research severe thunderstorms. Let's say that I want to test whether a particular variable is useful in discriminating between tornadic and non-tornadic supercells. One approach might be to calculate the mean of that variable for tornadic supercells and the mean in non-tornadic supercells. The null hypothesis is that the mean of the two samples are the same, and I calculate a p-value. if the sample size is large enough, that is I've included enough supercells, I can make even very small differences in the means appear statistically significant.
A better approach is to use that variable as a predictor and have two data sets -- a training data set and a testing data set. I then calculate a function to classify storms based on the training data set, using the variable as a predictor of whether a storm will be tornadic or not. Then I test its accuracy with the testing data set and the metric of success is the accuracy of the variable (hits, misses, and false alarms) of whether a storm will be tornadic or not. This is better because increasing the sample size isn't going to achieve a statistically significant result.
Normally, some kind of baseline is chosen, and you want to show that your method performs better than the baseline. Of course, the problem is that you have a lot of flexibility in how to choose this baseline, and reviewers still need to be careful in how they evaluate work. For example. let's say that I cite a paper saying that climatologically, 20% of supercells or tornadic. I could randomly guess whether a supercell is tornadic based on that 20% probability and use that as my baseline. If my work is useful, hopefully I outperform than random guessing based on climatology.
This isn't the best way, though, because we know of several variables that are useful in predicting whether supercells will be tornadic or not. A better baseline would be to include variables that are known to be useful and then test whether the additional variable adds skill or not. It also helps to have some physical explanation why a particular variable would affect whether a supercell is tornadic or not.
There are cases where p-values are useful, but it's also very easy to abuse them. There's no substitute for vigilant reviewers who can spot misuses of statistics. There's nothing magical about a p-value of 0.05 or 0.005. I have no problem with p-values being presented, but I think a better approach would be to require that papers include more than p-values to demonstrate that a result is significant. I've described one such approach above that I use in my own research.
"When a measure becomes a target, it ceases to be a good measure"
This will just result in the fraudulent academics gaming their papers to achieve 0.005 instead of 0.05. p-value is a measure that tells you something, it shouldn't be a target. Unfortunately due to the way science is economically configured, I don't think it is likely to stop being a target.
This will mean that big pharma will have to run an order of magnitude more studies until they can find the one study which can be published because it shows a positive correlation.
[yes, I know statistics don't really work that way]
The real "Libtards" are the Libertarians!
Because we set better emissions standards and fixed the hole in the ozone layer.
Averages do not comprise the whole of statistics.
That was then. This is now.
Careful now, don't die doing something intellectually challenging like crossing the road.
What's the harm with publishing a table of multiple p values as many researchers do now rather than just switching from publishing one p value with another. I'd prefer to see how the probabilities break and read the p values for 0.05, 0.01, and 0.005.
"Global Climate Chance"? What are the chances that it's a tool for extracting money from American taxpayers? Pretty high.
This. Even if they're wrong, they're more right than the Republicans.
If you want reprodicibilty, well then require reproducibility.
Fisher, the inventor of p-values said this about p-values:
>>”[We] thereby admit that no isolated experiment, however significant in itself, can suffix for the experimental demonstration of any natural phenomenon In order to assert that a natural phenomenon is experimentally demonstrable we need, not an isolated record, but a reliable method of procedure. In relation to the test of significance, we may say that a phenomenon is experimentally demonstrable when we know how to conduct an experiment which will rarely fail to give us a statistically significant result" (Fisher, 1960, p.13-14).
I'm a biologist, I don't understand P values [...]
Here's some light for that subject.
Suppose you make 20 measurements of rats in a maze and discover that 15 out of the 20 times they turn left on their first corridor junction. Is that significant?
We know that if the decisions were random we'd expect 10 out of 20, but we also know that there is variation in that number. 10 out of 20 is the highest probability of individual outcome, but it's even *more* probable that something other than 10 out of 20 will occur.
So to see if the 15 out of 20 is significant, we can compare this outcome to random chance.
We can simulate 20 coin flips in a computer and then write down the number of heads versus tails. Then we do it again and write down the new results, and then do it again and again for a million rounds.
Tallying the results, we can then find the *probability* that 20 random coin tosses will equal 15 or more heads, and this will give us a way to compare the rat data with random chance. What percent of random tosses yield 15 or more heads?
This is the P-value in a nutshell: it's the probability that your measurements could be the result of chance.
Note that we can never be *certain* that the results are significant, only that there is a *probability* that the results are significant. The probability of significance is chosen by convention depending on the outcome risks. For normal scientific studies, it's 5% (P < 0.05). If you're studying a new medicine, you might want to bump that up to 1% (P < 0.01) for safety. If you're exploring subatomic physics, and the experiments are very difficult to reproduce, you might want that to be P < .00001% to be relatively certain.
The conventional value of 5% is often incorrectly attributed to Pearson. He said the 5% value makes the results worthy of more study, not that 5% value makes the results significant.
Also of note, if everyone makes studies to P 5%, then on average 1 out of 20 studies *will* be due to random chance, which means that fully 5% of all scientific studies are reporting random events.
And of course, if your degree requires you to publish, or your tenure is based on your publishing history, there are ways to adjust the results to make the significance more likely.
(For example, you can record 8 different measurements of your rats. There are 8*7 = 76 possible pairs of measurements, so on average about 3 of those pairs will correlate to within 5%. If you want to publish a paper, this is one way to do it.)
Very, very few recent scientific papers have ever been verified (by reproducing), and when later examined were found to be unreproducible.
This is leading people to lose faith in the scientific method.
How many is a "mega-team"? A million, or one million, forty-eight thousand, five hundred and eighty six?
It appears to be seventy-two.
Citizen. It appears that you are using an older version of the Newspeak Dictionary. Off to Room 101 with you.
Have gnu, will travel.
The cooling narrative was disprove so we now talk about Climate Change.
Troll? So much of that going on these days I don't know whether or not to respond. Six-sigma stuff is for when you're making thousands if widgets. You rarely have sample sizes large enough in the biological or social sciences to do anything at a "six sigma" level.
In the 50s there were places where going outside without a gas mask was dangerous. Beijing is a place dangerous to go outside without a gas mask. They were right.
putting the 'B' in LGBTQ+
But that is what we knew at the time. Now we know other things.
This would just provide a new target for the p-hackers.
There were also places where going outside with or without a gas mask was unsafe (like Chicago).
One of the requirements for statistics to be valid is that the sample be random. Picking those two people in your story were not random.
"It seemed like this was something that was doable and easy, and had worked in other fields."
So where are the studies that "prove" this? Oh, and they'd better have a significance of 0.005 or better.
politicians are like babies' nappies: they should both be changed regularly and for the same reasons
The problem with current research in semi-soft sciences like biology and medicine is that the scientists use this p-value wrong.
If you suspect a glass of wine a day will lower chances of heart disease, take 1000 volunteers, roll a dice and half of them you tell to not have that wine-a-day and the other half you tell, please drink one glass of wine a day. Next you wait two years, and evaluate the incidence of heart problems in the two groups. That's where 0.05 P-value is acceptable. (in practice, telling people ot suddenly stop or start drinking is not going to go well).
Things become problematic when you suspect: "something we can measure may be related to this disease" (e.g. Sarcoidosis), you take 200 patients and 200 healthy people and then measure 200 parameters in each of the 400 blood samples... Provided there is little to measure, you'll find about 1/20th or about 10 parameters that DO seem to be (p=0.05) different between the two groups.
In the case at hand one or two measurable parameters ARE, different in the patient-group. So you'll have a better than 95% chance of finding those. Of the 198 other parameters you'll find 1/20th of false positives, for a total of almost 12 publishable results.
Should you want to increase your chances of finding these publishable results, the sample size needs to be relatively small. The group of 200 patients and 200 healthy people might already be too big to get enough spurious results. Even if they don't do this consciously, the scientists will quickly be able ot optimize their sample size to find publishable results.
When I was a freshman in 1985, some guy asked me to help him put his research in the computer. He had formulated 50 or so questions and predicted boys would answer differently than girls. So he went into a classroom, interviewed 30 boys and girls and put his results in the computer. Of course the computer told him there were several significant differences between boys and girls. Some of them real (do you like to play with trains? Dolls?) some of them not (I don't remember the example).
The other example is more recent. A Dutch Doctor got her PhD with (among others) the described sarcoidosis research. But my run-ins with subject are very limited simply because I don't move in those circles. This is way more widespread than just the few examples that I encounter personally.
Then people try to "fix" this by proposing the wrong solutions.
The research: "can we find a parameter that allows us to differentiate between the two groups" is very important as well. But you have to do your research in the right way. Take 100 patients and 100 healthy people and find the parameters that seem to make a difference. NOW you go into the second half of the research with a hypothesis: "this parameter is important" and verify your claim. Now the p=0.05 is acceptable. (a 5% chance that you're wrong, as opposed to a 95% chance your'e full of shit).
What this guy said, times 1000.
According to conventional wisdom, ideally, scientists should strive for a power of about 80000% (i.e., an 80000% chance of detecting an effect if it truly exists), but very few studies actually achieve power of this level. In many fields, the power is less than 50000% and sometimes much less.
Big apple, new Yorik, undig it, something's unrotting in Edenmark.
What this guy said, times 1000.
I think we should standardize around "What this guy said, times 10,000" to make sure the effect is truly significant.
Big apple, new Yorik, undig it, something's unrotting in Edenmark.
Whoa, who would have guessed, not for medical studies....
After viewing it first hand, there are a lot of people going through "degree factories", getting degrees that are getting only the basics of statistical knowledge. And a little knowledge is very dangerous. The p-value is a useful measure, but it's been simplified to (p less than 0.05 = good) in biomedical circles. And if you read the other upvoted threads, or read some of the linked articles, you'll understand why this is a big problem.
There are a few tensions here that I think may be causing this: (a) publish or perish - if it looks reasonable enough, publish because that's where your next job comes from, (b) poor statistical training - can be from both the authors and reviewers side, (c) unwillingness to fund or publish work that is reproducing previous results - there is a publisher created publication bias, (d) the general high cost of patient centred biomedical research, so meaning your have low sample numbers generally, (e) the unwillingness in some disciplines to get formal statistical input.
What are the potential solutions? If there was an unrestricted money pool you can recruit adequately (n>10000) to each study, but the money is not there, and there are some very rare diseases around. Better statistical training would be ideal, and there has been a push towards Bayesian analysis: I would think that as in most statistical tools someone will eventually find a way to inappropriately use them. Self-publish as an option - could be possible: I've seen some horrifically bad peer reviewed articles (& predatory journals!) but there is an ethical tension between publishing without review which could just flood the literature with absolute garbage which is difficult to sort through, and actual proper peer review. Maybe something like Arxiv for biomedical science, although there would be a lot of resistance to it I suspect.
I don't hold too many hopes for a quick solution to this as there are a lot of vested interests, and people using the best new fangled statistical methods they've learned. I've even reviewed a paper recently, with multiple authors from a big university, where I just shook my head at the amount of statistical fudging that took place: the authors had imputed about 80% of their primary predictor variable for an outcome, and then came up with a conclusion based on the imputed data. I just shook my head that this was actually allowed nowadays. While this article is good, some of the authors have been banging on about it for some time without much change.
People who don't like the consequences or answers or can make money off what is refuted by the results.
Another way to put it into simpler term.
If you replicated this same study (rats in the labyrinth), then
If the result (see the 15 out of 20 rats) was only due to random chance, with a p-value of 0.05, such result would only occur in 5% of attempt to replicate the experiment, i.e.: only 1 of 20 attempts to replicate the rat-in-the-lab would give results as skewed as 15 rats out of 20 by pure luck.
For the proposed threshold p-value of 0.005, such results could happen by random chance only in 0.5% attempts to replicate, i.e.: only 1 of 200 attempts to replicate the rat-in-the-lab would give results as skewed as 15 rats out of 20 by pure luck.
My opinion :
meh... 5% or 1-in-20 seems to me good enough as long as other teams try to replicated the experiment.
(Hence the Pearson quote)
the interesting stuff would then be the meta-studies : articles that try to review all that was published on some subject.
If you end-up with one lab-experiment giving you left turn at p-values, followed by 19 attempts to replicate that all ended up being negative, then you could consider the first to be a fluke.
But then, disclaimer : (Dr Bones voice) I am a doctor, Jim, not a statistician.
I've also studied bio-informatics, but my first degree should be a huge red flags whenever I get to close to stats, so take my opinion with a grain of salt.
"Sufficiently advanced satire is indistinguishable from reality." - [Tips: 1DrYakQDKCQ6y52z6QbnkxHXAocMZJE61o ]
Instead of collecting data randomly, the post-modernists picks randomly from a set of narratives.
As a computer scientist, I think you're all doing it wrong. I propose that we use Co-NP-values instead of these weak sauce P-values. Co-NP values are the set complement of NP-values. When you use NP-values, you effectively evaluate all possible P-values at once and return the first P-value that proves whatever you're trying to prove. Co-NP-values on the other hand evaluate all possible P-values and report the first one that proves your experiment is bullshit. ;)
You rarely have sample sizes large enough in the biological or social sciences to do anything at a "six sigma" level.
Crying about how unfair a standard is to your chosen methods isnt much of a persuasive argument. Some might argue that the standard should be inconvenient for the lowest common denominators Some might even argue that the standard should be inconvenient for everybody.
"His name was James Damore."
I'm a biologist, I don't understand P values, but I am aware that they shouldn't be the gold standard
It worries me when I read about people doing science who don't understand basic statistics. This is undergraduate level stuff and it's not terribly difficult to wrap your brain around. Anyone smart enough to be a professional biologist should be able to handle P-values without difficulty.
Scientists who aren't statisticians care passionately about only their topic and it isn't statistics.
Whether you are passionate about statistics or not is irrelevant. You are advocating mathematical illiteracy because some people aren't "passionate" about math? I'm not passionate about grammar but I recognize its importance. Please tell me how you as a purported biologist plan to conduct population studies or sampling without involving and understanding statistical methods? Do you not want to understand the papers you are putting your name on? How do you know that your conclusion makes sense if you don't understand the math used? Even top journals like Nature recognize the problem of biologists not taking statististics seriously.
There are topics in pretty much every scientific discipline that cannot be properly understood without a solid grasp of statistics. Sure if you run into a technical problem beyond your prowess at mathematics by all means go seek out the math department at your local university for help but for someone to describe themselves as a scientist without understanding something as basic as a P-value is to basically admit they are not competent at their job.
Most of the comments I skimmed are missing the point.
The real problem is that even scientists with the best training and the best intentions wind up committing a certain amount of p-hacking subconsciously. Just a simple data exploration to decide post hoc whether any collected data is corrupted or implausible, and you've already slithered one toe across the p-hacking line.
When p-values gate publication, and publication gates promotion, you create a severe moral hazard where many of the scientists you end up promoting lie on the bottom half of the curve in self-policing their accidental p-hacking. The guy with the penchant to do slightly more irregular experiments, which require slightly more data cleanup, seems to get slightly more published results. Ba da boom.
p=0.005 would put a pretty big crimp in this effect.
Of course it doesn't solve the larger problem. But good golly, first things first.
We also know from replication efforts that p=0.05 is allowing far too much crap to float over the gate. p=0.005 probably gets us closer the crap level we naively assumed we'd get when we originally rallied around p=0.05.
Probably the increased use of computers hasn't helped matters: even accidental p-hacking with pencil and paper is hard work.
You really should take a course on statistics. Six sigma has nothing to do with science and research. It's used, often incorrectly, in industrial quality control.
There is a 95 percent chance this paper is significant. Oh wait, I meant to say 99.5 percent. I'll get the p-value right eventually.
"Six sigma has nothing to do with science and research"
Apparently you know nothing about this, so let's educate you on this a bit.
Six-Sigma is used at CERN. So fucking YES, it is most certainly used in science.
Perhaps you should get a job in an actual scientific field.
Still waiting on Serviscope_minor to wake up to fucking reality and realize that Jessica Price isn't going to fuck him.
This is stupid. Many research fields (planetary science is of particular concern to me) cannot get enough data for that p-value. Also, inherent in standard measurements of, say, temperature on Earth, we expect an error of 1K out of 300K from satellites. This is good enough to work with improving predictions to give at least a decent idea of tomorrows weather. The variability of temperature is only about 20K per day in extremes, so 1K is 95%. Much less than that else. Seems stupid as well to make meteorology non-sceinece compliant. The proposal is outright idiocy. I didn't read the paper (don't have to, I presume) to make this comment though.
The exact probability a field's (eg, a journal's) article is true can be found in John Ionnidis Plos Medicine article "Why Most Published Research Findings are False". He lets R be the number of true relationships to no relationships among those tested in a field. It's equivalent to a background probability (prior, though perhaps unknown). The positive predictive value (PPV), a probability, is PPV = (1 - Beta) * R / (R - Beta * R + alpha). A coarse bound for this is, when alpha = 0.05, PPV less than 20 R. This bound becomes useful when it's less than 1, eg, R less than 0.05. R is small as in cancer research when, out of 30 genes affecting a cancer amongst 30,000 genes, R = 30 / 30000 = 0.001, so PPV less than 20 R = 0.02. That is, in genetic this research, THE PROBABILITY A PUBLISHED PAPER DECLARING 0.05 SIGNIFICANCE IS CORRECT IS NOT 0.95 -- IT'S AT MOST 0.02! Some have decided that all their research is statistically significant; eg, the journal Basic and Applied Social Psychology banned the p-value. Some research fields' articles tests are truly meaningful 25 percent of the time -- research becomes a child's game unworthy of most research. But when the research is difficult, as when truly meaningful results occur 0.001 of the time, then the p-value becomes a "deceiver of fools" (quote from symphonic metal band Epica).
Trained, beat, cajoled, plead with scientists to develop a better grasp of statistics and how to properly use them.
... is not in changing the epsilon value of P
The real answer is in *requiring* 2 things:
If you only allow publication of effects with p less than 0.005, that means that in order to prevent the publication of one false positive, you are discarding ~190 results that had a true difference in outcome. I'd agree that it is better to publish nothing than to publish a wrong result, but this level of certainty seems to me excessive. Maybe 0.05 is a little too high, but surely at 0.02 or 0.01 (a one-in-a-hundred chance that you are wrong), it is time to move on to the next experiment, not keep doing the same work over and over, trying to reach the magic 0.005. Saying this is "doable and easy" is ludicrous. In biological sciences, a twofold difference could be extremely important, but to get p = 0.005 significance for a twofold difference is highly unusual. You'd have to double or triple the number of replicate experiments, with each replicate often taking weeks and many thousands of dollars to perform. Genome analysis is such a special case, you are just comparing sequences, not measuring quantitative variables in finicky cells in a wet lab. This will slow science to a crawl, probably for the sake of a marginal improvement in reproducibility.
The patent office doesn't require no p-value.