The Statistical Crisis in Science

Total Page:16

File Type:pdf, Size:1020Kb

The Statistical Crisis in Science The Statistical Crisis in Science Data-dependent analysis— a "garden of forking paths"— explains why many statistically significant comparisons don't hold up. Andrew Gelman and Eric Loken here is a growing realization a short mathematics test when it is This multiple comparisons issue is that reported "statistically sig expressed in two different contexts, well known in statistics and has been nificant" claims in scientific involving either healthcare or the called "p-hacking" in an influential publications are routinely mis military. The question may be framed 2011 paper by the psychology re Ttaken. Researchers typically expressnonspecifically as an investigation of searchers Joseph Simmons, Leif Nel the confidence in their data in terms possible associations between party son, and Uri Simonsohn. Our main of p-value: the probability that a per affiliation and mathematical reasoning point in the present article is that it ceived result is actually the result of across contexts. The null hypothesis is is possible to have multiple potential random variation. The value of p (for that the political context is irrelevant comparisons (that is, a data analysis "probability") is a way of measuring to the task, and the alternative hypoth whose details are highly contingent the extent to which a data set provides esis is that context matters and the dif on data, invalidating published p-val- evidence against a so-called null hy ference in performance between the ues) without the researcher perform pothesis. By convention, a p-value be two parties would be different in the ing any conscious procedure of fishing low 0.05 is considered a meaningful military and healthcare contexts. through the data or explicitly examin refutation of the null hypothesis; how At this point a huge number of pos ing multiple comparisons. ever, such conclusions are less solid sible comparisons could be performed, than they appear. all consistent with the researcher's the How to Test a Hypothesis The idea is that when p is less than ory. for example, the null hypothesis In general, we could think of four some prespecified value such as 0.05, could be rejected (with statistical sig classes of procedures for hypothesis the null hypothesis is rejected by the nificance) among men and not among testing: (1) a simple classical test based data, allowing researchers to claim women—explicable under the theory on a unique test statistic, T, which strong evidence in favor of the alterna that men are more ideological than when applied to the observed data tive. The concept of p-values was origi women. The pattern could be found yields T(y), where y represents the nally developed by statistician Ronald among women but not among men— data; (2) a classical test prechosen from Fisher in the 1920s in the context of his explicable under the theory that wom a set of possible tests, yielding T(y;cp), research on crop variance in Hertford en are more sensitive to context than with preregistered (p (for example, <p shire, England. Fisher offered the idea men. Or the pattern could be statisti might correspond to choices of con of p-values as a means of protecting cally significant for neither group, but trol variables in a regression, transfor researchers from declaring truth based the difference could be significant (still mations, the decision of which main on patterns in noise. In an ironic twist, fitting the theory, as described above). effect or interaction to focus on); (3) p-values are now often manipulated to Or the effect might only appear among researcher degrees of freedom without lend credence to noisy claims based on men who are being questioned by fe fishing, which consists of computing small samples. male interviewers. a single test based on the data, but in In general, p-values are based on We might see a difference between an environment where a different test what would have happened under the sexes in the healthcare context but would have been performed given dif other possible data sets. As a hypo not the military context; this would ferent data; the result of such a course thetical example, suppose a researcher make sense given that health care is is T(y;(p(y)), where the function (p(») is interested in how Democrats and currently a highly politically salient is observed in the observed case. It is Republicans perform differently in issue and the military is less so. And generally considered unethical to (4) how are independents and nonparti commit outright fishing, computing sans handled? They could be exclud T(y;<p;) for j = 1,.../. This would be a Andrew Gelman is a professor in the depart ed entirely, depending on how many matter of performing / tests and then ments of statistics and political science at were in the sample. And so on: A sin reporting the best result given the Columbia University and the author o/Red gle overarching research hypothesis— data, thus T(y; cpbest(y)). State, Blue State, Rich State, Poor State: in this case, the idea that issue context It would take a highly unscrupulous Why Americans Vote the Way They Do (2008). Eric Loken is a research associate interacts with political partisanship to researcher to perform test after test in a professor of human development at Pennsyl affect mathematical problem-solving search for statistical significance (which vania State University. E-mail: gelman@stat. skills—corresponds to many possible could almost certainly be found at the columbia.edu choices of a decision variable. 0.05 or even the 0.01 level, given all 460 American Scientist, Volume 102 the options above and economic status (SES) the many more that will oppose wealth would be possible in redistribution, and that a real study). The diffi stronger men with low SES cult challenge lies else will support redistribution. where: Given a par These researchers had ticular data set, it can enough degrees of freedom for seem entirely appropri them to be able to find any number ate to look at the data and of apparent needles in the haystack construct reasonable rules of their data—and, again, it would be for data exclusion, coding, and easy enough to come across the statisti analysis that can lead to statistic, cally significant comparisons without significance. In such a case, resell Lit "fishing" by simply looking at the data ers need to perform only one test, data-based claim is more plausible to and noticing large differences that are but that test is conditional on the data; the extent it is a priori more likely, and consistent with their substantive theory. hence, T(y;cp(y)), with the same effect as any claim is less plausible to the extent Most notably, the authors report a if they had deliberately fished for those that it is estimated with more error. statistically significant interaction with results. As political scientists Macartan There are many roads to statistical no statistically significant main effect— Humphreys, Raul Sanchez de la Sierra, significance; if data are gathered with that is, they did not find that men with and Peter van der Windt wrote in 2013, no preconceptions at all, statistical sig bigger arm circumference had more a researcher faced with multiple rea nificance can obviously be obtained conservative positions on economic re sonable measures can think—perhaps even from pure noise by the simple distribution. What they found was that correctly—that the one that produces a means of repeatedly performing com the correlation of arm circumference significant result is more likely to be the parisons, excluding data in different with opposition to redistribution of least noisy measure, but then decide— ways, examining different interactions, wealth was higher among men of high incorrectly—to draw inferences based controlling for different predictors, socioeconomic status. Had they seen on that one measure alone. In the hy and so forth. Realistically, though, a the main effect (in either direction), pothetical example presented earlier, researcher will come into a study with they could have come up with a theo finding a difference in the healthcare strong substantive hypotheses, to the retically justified explanation for that, context might be taken as evidence that extent that, for any given data set, the too. And if there had been no main ef that is the most important context in appropriate analysis can seem evident fect and no interaction, they could have which to explore differences. ly clear. But even if the chosen data looked for other interactions. Perhaps, This error carries particular risks in analysis is a deterministic function of for example, the correlations could the context of small effect sizes, small the observed data, this does not elimi have differed when comparing stu sample sizes, large measurement er nate the problem posed by multiple dents with or without older siblings? rors, and high variation (which com comparisons. As we wrote in a 2013 critique for bine to give low power, hence less reli Slate, nothing in this report suggests able results even when they happen Arm Strength and Economic Status that fishing or p-hacking—which to be statistically significant, as dis In 2013, a research group led by Mi would imply an active pursuit of sta cussed by Katherine Button and her chael Petersen of Aarhus University tistical significance—was involved at coauthors in a 2013 paper in Nature published a study that claimed to find all. Of course, it is reasonable for scien Reviews: Neuroscience). Multiplicity is an association between men's upper- tists to refine their hypotheses in light of less consequential in settings with large body strength, interacted with socio the data. When the desired pattern does real differences, large samples, small economic status, and their attitudes not show up as a main effect, it makes measurement errors, and low variation.
Recommended publications
  • Downloading of a Human Consciousness Into a Digital Computer Would Involve ‘A Certain Loss of Our Finer Feelings and Qualities’89
    When We Can Trust Computers (and When We Can’t) Peter V. Coveney1,2* Roger R. Highfield3 1Centre for Computational Science, University College London, Gordon Street, London WC1H 0AJ, UK, https://orcid.org/0000-0002-8787-7256 2Institute for Informatics, Science Park 904, University of Amsterdam, 1098 XH Amsterdam, Netherlands 3Science Museum, Exhibition Road, London SW7 2DD, UK, https://orcid.org/0000-0003-2507-5458 Keywords: validation; verification; uncertainty quantification; big data; machine learning; artificial intelligence Abstract With the relentless rise of computer power, there is a widespread expectation that computers can solve the most pressing problems of science, and even more besides. We explore the limits of computational modelling and conclude that, in the domains of science and engineering that are relatively simple and firmly grounded in theory, these methods are indeed powerful. Even so, the availability of code, data and documentation, along with a range of techniques for validation, verification and uncertainty quantification, are essential for building trust in computer generated findings. When it comes to complex systems in domains of science that are less firmly grounded in theory, notably biology and medicine, to say nothing of the social sciences and humanities, computers can create the illusion of objectivity, not least because the rise of big data and machine learning pose new challenges to reproducibility, while lacking true explanatory power. We also discuss important aspects of the natural world which cannot be solved by digital means. In the long-term, renewed emphasis on analogue methods will be necessary to temper the excessive faith currently placed in digital computation.
    [Show full text]
  • Caveat Emptor: the Risks of Using Big Data for Human Development
    1 Caveat emptor: the risks of using big data for human development Siddique Latif1,2, Adnan Qayyum1, Muhammad Usama1, Junaid Qadir1, Andrej Zwitter3, and Muhammad Shahzad4 1Information Technology University (ITU)-Punjab, Pakistan 2University of Southern Queensland, Australia 3University of Groningen, Netherlands 4National University of Sciences and Technology (NUST), Pakistan Abstract Big data revolution promises to be instrumental in facilitating sustainable development in many sectors of life such as education, health, agriculture, and in combating humanitarian crises and violent conflicts. However, lurking beneath the immense promises of big data are some significant risks such as (1) the potential use of big data for unethical ends; (2) its ability to mislead through reliance on unrepresentative and biased data; and (3) the various privacy and security challenges associated with data (including the danger of an adversary tampering with the data to harm people). These risks can have severe consequences and a better understanding of these risks is the first step towards mitigation of these risks. In this paper, we highlight the potential dangers associated with using big data, particularly for human development. Index Terms Human development, big data analytics, risks and challenges, artificial intelligence, and machine learning. I. INTRODUCTION Over the last decades, widespread adoption of digital applications has moved all aspects of human lives into the digital sphere. The commoditization of the data collection process due to increased digitization has resulted in the “data deluge” that continues to intensify with a number of Internet companies dealing with petabytes of data on a daily basis. The term “big data” has been coined to refer to our emerging ability to collect, process, and analyze the massive amount of data being generated from multiple sources in order to obtain previously inaccessible insights.
    [Show full text]
  • Reproducibility: Promoting Scientific Rigor and Transparency
    Reproducibility: Promoting scientific rigor and transparency Roma Konecky, PhD Editorial Quality Advisor What does reproducibility mean? • Reproducibility is the ability to generate similar results each time an experiment is duplicated. • Data reproducibility enables us to validate experimental results. • Reproducibility is a key part of the scientific process; however, many scientific findings are not replicable. The Reproducibility Crisis • ~2010 as part of a growing awareness that many scientific studies are not replicable, the phrase “Reproducibility Crisis” was coined. • An initiative of the Center for Open Science conducted replications of 100 psychology experiments published in prominent journals. (Science, 349 (6251), 28 Aug 2015) - Out of 100 replication attempts, only 39 were successful. The Reproducibility Crisis • According to a poll of over 1,500 scientists, 70% had failed to reproduce at least one other scientist's experiment or their own. (Nature 533 (437), 26 May 2016) • Irreproducible research is a major concern because in valid claims: - slow scientific progress - waste time and resources - contribute to the public’s mistrust of science Factors contributing to Over 80% of respondents irreproducibility Nature | News Feature 25 May 2016 Underspecified methods Factors Data dredging/ Low statistical contributing to p-hacking power irreproducibility Technical Bias - omitting errors null results Weak experimental design Underspecified methods Factors Data dredging/ Low statistical contributing to p-hacking power irreproducibility Technical Bias - omitting errors null results Weak experimental design Underspecified methods • When experimental details are omitted, the procedure needed to reproduce a study isn’t clear. • Underspecified methods are like providing only part of a recipe. ? = Underspecified methods Underspecified methods Underspecified methods Underspecified methods • Like baking a loaf of bread, a “scientific recipe” should include all the details needed to reproduce the study.
    [Show full text]
  • Rejecting Statistical Significance Tests: Defanging the Arguments
    Rejecting Statistical Significance Tests: Defanging the Arguments D. G. Mayo Philosophy Department, Virginia Tech, 235 Major Williams Hall, Blacksburg VA 24060 Abstract I critically analyze three groups of arguments for rejecting statistical significance tests (don’t say ‘significance’, don’t use P-value thresholds), as espoused in the 2019 Editorial of The American Statistician (Wasserstein, Schirm and Lazar 2019). The strongest argument supposes that banning P-value thresholds would diminish P-hacking and data dredging. I argue that it is the opposite. In a world without thresholds, it would be harder to hold accountable those who fail to meet a predesignated threshold by dint of data dredging. Forgoing predesignated thresholds obstructs error control. If an account cannot say about any outcomes that they will not count as evidence for a claim—if all thresholds are abandoned—then there is no a test of that claim. Giving up on tests means forgoing statistical falsification. The second group of arguments constitutes a series of strawperson fallacies in which statistical significance tests are too readily identified with classic abuses of tests. The logical principle of charity is violated. The third group rests on implicit arguments. The first in this group presupposes, without argument, a different philosophy of statistics from the one underlying statistical significance tests; the second group—appeals to popularity and fear—only exacerbate the ‘perverse’ incentives underlying today’s replication crisis. Key Words: Fisher, Neyman and Pearson, replication crisis, statistical significance tests, strawperson fallacy, psychological appeals, 2016 ASA Statement on P-values 1. Introduction and Background Today’s crisis of replication gives a new urgency to critically appraising proposed statistical reforms intended to ameliorate the situation.
    [Show full text]
  • Replicability Problems in Science: It's Not the P-Values' Fault
    The replicability problems in Science: It’s not the p-values’ fault Yoav Benjamini Tel Aviv University NISS Webinar May 6, 2020 1. The Reproducibility and Replicability Crisis Replicability with 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 statistically significant results.” Fisher (1935) “The Design of Experiments”. Reproducibility/Replicability • Reproduce the study: from the original data, through analysis, to get same figures and conclusions • Replicability of results: replicate the entire study, from enlisting subjects through collecting data, and analyzing the results, in a similar but not necessarily identical way, yet get essentially the same results. (Biostatistics, Editorial 2010, Nature Editorial 2013, NSA 2019) “ reproducibilty is the ability to replicate the results…” in a paper on “reproducibility is not replicability” We can therefore assure reproducibility of a single study but only enhance its replicability Opinion shared by 2019 report of National Academies on R&R Outline 1. The misguided attack 2. Selective inference: The silent killer of replicability 3. The status of addressing evident selective inference 6 2. The misguided attack Psychological Science “… we have published a tutorial by Cumming (‘14), a leader in the new-statistics movement…” • 9. Do not trust any p value. • 10. Whenever possible, avoid using statistical significance or p- values; simply omit any mention of null hypothesis significance testing (NHST). • 14. …Routinely report 95% CIs… Editorial by Trafimow & Marks (2015) in Basic and Applied Social Psychology: From now on, BASP is banning the NHSTP… 7 Is it the p-values’ fault? ASA Board’s statement about p-values (Lazar & Wasserstein Am.
    [Show full text]
  • Just How Easy Is It to Cheat a Linear Regression? Philip Pham A
    Just How Easy is it to Cheat a Linear Regression? Philip Pham A THESIS in Mathematics Presented to the Faculties of the University of Pennsylvania in Partial Fulfillment of the Requirements for the Degree of Master of Arts Spring 2016 Robin Pemantle, Supervisor of Thesis David Harbater, Graduate Group Chairman Abstract As of late, the validity of much academic research has come into question. While many studies have been retracted for outright falsification of data, perhaps more common is inappropriate statistical methodology. In particular, this paper focuses on data dredging in the case of balanced design with two groups and classical linear regression. While it is well-known data dredging has pernicious e↵ects, few have attempted to quantify these e↵ects, and little is known about both the number of covariates needed to induce statistical significance and data dredging’s e↵ect on statistical power and e↵ect size. I have explored its e↵ect mathematically and through computer simulation. First, I prove that in the extreme case that the researcher can obtain any desired result by collecting nonsense data if there is no limit on how much data he or she collects. In practice, there are limits, so secondly, by computer simulation, I demonstrate that with a modest amount of e↵ort a researcher can find a small number of covariates to achieve statistical significance both when the treatment and response are independent as well as when they are weakly correlated. Moreover, I show that such practices lead not only to Type I errors but also result in an exaggerated e↵ect size.
    [Show full text]
  • Distributional Null Hypothesis Testing with the T Distribution Arxiv
    Distributional Null Hypothesis Testing with the T distribution Fintan Costello School of Computer Science and Informatics, University College Dublin and Paul Watts Department of Theoretical Physics, National University of Ireland Maynooth arXiv:2010.07813v1 [stat.ME] 15 Oct 2020 Distributional Nulls 2 Abstract Null Hypothesis Significance Testing (NHST) has long been central to the scien- tific project, guiding theory development and supporting evidence-based intervention and decision-making. Recent years, however, have seen growing awareness of serious problems with NHST as it is typically used, and hence to proposals to limit the use of NHST techniques, to abandon these techniques and move to alternative statistical approaches, or even to ban the use of NHST entirely. These proposals are premature, because the observed problems with NHST all arise as a consequence of a contingent and in many cases incorrect choice: that of NHST testing against point-form nulls. We show that testing against distributional, rather than point-form, nulls is bet- ter motivated mathematically and experimentally, and that the use of distributional nulls addresses many problems with the standard point-form NHST approach. We also show that use of distributional nulls allows a form of null hypothesis testing that takes into account both the statistical significance of a given result and the probability of replication of that result in a new experiment. Rather than abandoning NHST, we should use the NHST approach in its more general form, with distributional rather than point-form nulls. Keywords: NHST; Replication; Generalisation Distributional Nulls 3 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.
    [Show full text]
  • Responsible Use of Statistical Methods Larry Nelson North Carolina State University at Raleigh
    University of Massachusetts Amherst ScholarWorks@UMass Amherst Ethics in Science and Engineering National Science, Technology and Society Initiative Clearinghouse 4-1-2000 Responsible Use of Statistical Methods Larry Nelson North Carolina State University at Raleigh Charles Proctor North Carolina State University at Raleigh Cavell Brownie North Carolina State University at Raleigh Follow this and additional works at: https://scholarworks.umass.edu/esence Part of the Engineering Commons, Life Sciences Commons, Medicine and Health Sciences Commons, Physical Sciences and Mathematics Commons, and the Social and Behavioral Sciences Commons Recommended Citation Nelson, Larry; Proctor, Charles; and Brownie, Cavell, "Responsible Use of Statistical Methods" (2000). Ethics in Science and Engineering National Clearinghouse. 301. Retrieved from https://scholarworks.umass.edu/esence/301 This Teaching Module is brought to you for free and open access by the Science, Technology and Society Initiative at ScholarWorks@UMass Amherst. It has been accepted for inclusion in Ethics in Science and Engineering National Clearinghouse by an authorized administrator of ScholarWorks@UMass Amherst. For more information, please contact [email protected]. Responsible Use of Statistical Methods focuses on good statistical practices. In the Introduction we distinguish between two types of activities; one, those involving the study design and protocol (a priori) and two, those actions taken with the results (post hoc.) We note that right practice is right ethics, the distinction between a mistake and misconduct and emphasize the importance of how the central hypothesis is stated. The Central Essay, I dentification of Outliers in a Set of Precision Agriculture Experimental Data by Larry A. Nelson, Charles H. Proctor and Cavell Brownie, is a good paper to study.
    [Show full text]
  • P-Hacking' Now an Insiders' Term for Scientific Malpractice Has Worked Its Way Into Pop Culture
    CHRISTIE ASCHWANDEN 11.26.2019 09:00 AM We're All 'P-Hacking' Now An insiders' term for scientific malpractice has worked its way into pop culture. Is that a good thing? PHOTOGRAPH: YULIA REZNIKOV/GETTY IMAGES IT’S GOT AN entry in the Urban Dictionary, been discussed on Last Week Tonight with John Oliver, scored a wink from Cards Against Humanity, and now it’s been featured in a clue on the TV game show Jeopardy. Metascience nerds rejoice! The term p-hacking has gone mainstream. https://www.wired.com/stor/were-all-p-hacking-now/ Results from a study can be analyzed in a variety of ways, and p-hacking refers to a practice where researchers select the analysis that yields a pleasing result. The p refers to the p-value, a ridiculously complicated statistical entity that’s essentially a measure of how surprising the results of a study would be if the effect you’re looking for wasn’t there. Suppose you’re testing a pill for high blood pressure, and you find that blood pressures did indeed drop among people who took the medicine. The p-value is the probability that you’d find blood pressure reductions at least as big as the ones you measured, even if the drug was a dud and didn’t work. A p-value of 0.05 means there’s only a 5 percent chance of that scenario. By convention, a p-value of less than 0.05 gives the researcher license to say that the drug produced “statistically significant” reductions in blood pressure.
    [Show full text]
  • Investigating the Effect of the Multiple Comparisons Problem in Visual Analysis
    Investigating the Effect of the Multiple Comparisons Problem in Visual Analysis Emanuel Zgraggen1 Zheguang Zhao1 Robert Zeleznik1 Tim Kraska1, 2 Brown University1 Massachusetts Institute of Technology2 Providence, RI, United States Cambridge, MA, United States {ez, sam, bcz}@cs.brown.edu [email protected] ABSTRACT Similar to others [9], we argue that the same thing happens The goal of a visualization system is to facilitate data-driven when performing visual comparisons, but instead of winning, insight discovery. But what if the insights are spurious? Fea- an analyst “loses” when they observe an interesting-looking tures or patterns in visualizations can be perceived as relevant random event (e.g., two sixes). Instead of being rewarded for insights, even though they may arise from noise. We often persistence, an analyst increases their chances of losing by compare visualizations to a mental image of what we are inter- viewing more data visualizations. This concept is formally ested in: a particular trend, distribution or an unusual pattern. known as the multiple comparisons problem (MCP) [5]. Con- As more visualizations are examined and more comparisons sider an analyst looking at a completely random dataset. As are made, the probability of discovering spurious insights more comparisons are made, the probability rapidly increases increases. This problem is well-known in Statistics as the mul- of encountering interesting-looking (e.g., data trend, unex- tiple comparisons problem (MCP) but overlooked in visual pected distribution, etc.), but still random events. Treating analysis. We present a way to evaluate MCP in visualization such inevitable patterns as insights is a false discovery (Type I tools by measuring the accuracy of user reported insights on error) and the analyst “loses” if they act on such false insights.
    [Show full text]
  • Reproducibility and Privacy: Research Data Use, Reuse, and Abuse
    Reproducibility and Privacy: Research Data Use, Reuse, and Abuse Daniel L. Goroff Opinions here are his own. But based on the work of grantees. Alfred P. Sloan • Organized and ran GM • Sloan School, Sloan Kettering, too • Foundation (public goods business) • Emphasized role of data in decision making Privacy Protecting Research How can work on private data be reproducible? Protocols can impose obfuscation at 3 stages: input, computation, or output. Data Enclave • Work there • Data stays • Release write up if approved • Irreproducible results! • Same problem for NDA’s. Privacy Protecting Protocols De-Identification? • William Weld, while Governor of Massachusetts, approved the release of de-identified medical records of state employees. • Latania Sweeney, then a graduate student at MIT, re-identified Weld’s records and delivered a list of his diagnoses and prescriptions to his office. • Try stripping out names, SSNs, addresses, etc. • But 3 facts–gender, birthday, and zip code–are enough to uniquely identify over 85% of U.S. citizens. • Including my assistant, a male from 60035 born on 9/30/1988 as above. • See www.aboutmyinfo.org Netflix Challenge • Offered $1m prize for improving prediction algorithm. • Release “anonymized” training set of >100m records. • In 2007, researchers began identifying video renters by linking with public databases like IMDB. • Suite settled in 2010. NYC Taxi Records • Last year, NYC’s 187m cab trips were compiled and released, complete with GPS start and finish data, distance traveled, fares, tips, etc. • But the dataset also included hashed but poorly anonymized driver info, including license and medallion numbers, making it possible to determine driver names, salaries, tips, and embarrassing facts about some riders.
    [Show full text]
  • (P>0.05): Significance Thresholds and the Crisis of Unreplicable Research
    A peer-reviewed version of this preprint was published in PeerJ on 7 July 2017. View the peer-reviewed version (peerj.com/articles/3544), which is the preferred citable publication unless you specifically need to cite this preprint. Amrhein V, Korner-Nievergelt F, Roth T. 2017. The earth is flat (p > 0.05): significance thresholds and the crisis of unreplicable research. PeerJ 5:e3544 https://doi.org/10.7717/peerj.3544 The earth is flat (p>0.05): Significance thresholds and the crisis of unreplicable research Valentin Amrhein1,2,3, Fränzi Korner-Nievergelt3,4, Tobias Roth1,2 1Zoological Institute, University of Basel, Basel, Switzerland 2Research Station Petite Camargue Alsacienne, Saint-Louis, France 3Swiss Ornithological Institute, Sempach, Switzerland 4Oikostat GmbH, Ettiswil, Switzerland Email address: [email protected] Abstract The widespread use of 'statistical significance' as a license for making a claim of a scientific finding leads to considerable distortion of the scientific process (according to the American Statistical Association). We review why degrading p-values into 'significant' and 'nonsignificant' contributes to making studies irreproducible, or to making them seem irreproducible. A major problem is that we tend to take small p-values at face value, but mistrust results with larger p-values. In either case, p-values tell little about reliability of research, because they are hardly replicable even if an alternative hypothesis is true. Also significance (p≤0.05) is hardly replicable: at a good statistical power of 80%, two studies will be 'conflicting', meaning that one is significant and the other is not, in one third of the cases if there is a true effect.
    [Show full text]