scholarly journals The earth is flat (p>0.05): Significance thresholds and the crisis of unreplicable research

2017 ◽  
Author(s):  
Valentin Amrhein ◽  
Fränzi Korner-Nievergelt ◽  
Tobias Roth
Keyword(s):  
Publication Bias ◽  
Null Hypothesis ◽  
Statistical Power ◽  
Alternative Hypothesis ◽  
Statistical Significance ◽  
Practical Importance ◽  
Decision Rules ◽  
Effect Sizes ◽  
P Values ◽  
True Effect

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 (American Statistical Association, Wasserstein & Lazar 2016). 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 can 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 realistic statistical power of 40%, given that there is a true effect, only one in six studies will significantly replicate the significant result of another study. Even at a good power of 80%, results from two studies will be conflicting, in terms of significance, in one third of the cases if there is a true effect. This means that a replication cannot be interpreted as having failed only because it is nonsignificant. Many apparent replication failures may thus reflect faulty judgement based on significance thresholds rather than a crisis of unreplicable research. Reliable conclusions on replicability and practical importance of a finding can only be drawn using cumulative evidence from multiple independent studies. However, applying significance thresholds makes cumulative knowledge unreliable. One reason is that with anything but ideal statistical power, significant effect sizes will be biased upwards. Interpreting inflated significant results while ignoring nonsignificant results will thus lead to wrong conclusions. But current incentives to hunt for significance lead to publication bias against nonsignificant findings. Data dredging, p-hacking and publication bias should be addressed by removing fixed significance thresholds. Consistent with the recommendations of the late Ronald Fisher, p-values should be interpreted as graded measures of the strength of evidence against the null hypothesis. Also larger p-values offer some evidence against the null hypothesis, and they cannot be interpreted as supporting the null hypothesis, falsely concluding that 'there is no effect'. Information on possible true effect sizes that are compatible with the data must be obtained from the observed effect size, e.g., from a sample average, and from a measure of uncertainty, such as a confidence interval. We review how confusion about interpretation of larger p-values can be traced back to historical disputes among the founders of modern statistics. We further discuss potential arguments against removing significance thresholds, such as 'we need more stringent decision rules', 'sample sizes will decrease' or 'we need to get rid of p-values'.

scholarly journals The earth is flat (p>0.05): Significance thresholds and the crisis of unreplicable research

2017 ◽  
Author(s):  
Valentin Amrhein ◽  
Fränzi Korner-Nievergelt ◽  
Tobias Roth
Keyword(s):  
Publication Bias ◽  
Null Hypothesis ◽  
Statistical Power ◽  
Statistical Significance ◽  
Practical Importance ◽  
Effect Sizes ◽  
Selective Reporting ◽  
Interval Estimate ◽  
P Values ◽  
True Effect

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. A replication can therefore not be interpreted as having failed only because it is nonsignificant. Many apparent replication failures may thus reflect faulty judgment based on significance thresholds rather than a crisis of unreplicable research. Reliable conclusions on replicability and practical importance of a finding can only be drawn using cumulative evidence from multiple independent studies. However, applying significance thresholds makes cumulative knowledge unreliable. One reason is that with anything but ideal statistical power, significant effect sizes will be biased upwards. Interpreting inflated significant results while ignoring nonsignificant results will thus lead to wrong conclusions. But current incentives to hunt for significance lead to selective reporting and to publication bias against nonsignificant findings. Data dredging, p-hacking, and publication bias should be addressed by removing fixed significance thresholds. Consistent with the recommendations of the late Ronald Fisher, p-values should be interpreted as graded measures of the strength of evidence against the null hypothesis. Also larger p-values offer some evidence against the null hypothesis, and they cannot be interpreted as supporting the null hypothesis, falsely concluding that 'there is no effect'. Information on possible true effect sizes that are compatible with the data must be obtained from the point estimate, e.g., from a sample average, and from the interval estimate, such as a confidence interval. We review how confusion about interpretation of larger p-values can be traced back to historical disputes among the founders of modern statistics. We further discuss potential arguments against removing significance thresholds, for example that decision rules should rather be more stringent, that sample sizes could decrease, or that p-values should better be completely abandoned. We conclude that whatever method of statistical inference we use, dichotomous threshold thinking must give way to non-automated informed judgment.

scholarly journals The earth is flat (p>0.05): Significance thresholds and the crisis of unreplicable research

2017 ◽  
Author(s):  
Valentin Amrhein ◽  
Fränzi Korner-Nievergelt ◽  
Tobias Roth
Keyword(s):  
Publication Bias ◽  
Null Hypothesis ◽  
Statistical Power ◽  
Statistical Significance ◽  
Practical Importance ◽  
Effect Sizes ◽  
Selective Reporting ◽  
Interval Estimate ◽  
P Values ◽  
True Effect

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. A replication can therefore not be interpreted as having failed only because it is nonsignificant. Many apparent replication failures may thus reflect faulty judgment based on significance thresholds rather than a crisis of unreplicable research. Reliable conclusions on replicability and practical importance of a finding can only be drawn using cumulative evidence from multiple independent studies. However, applying significance thresholds makes cumulative knowledge unreliable. One reason is that with anything but ideal statistical power, significant effect sizes will be biased upwards. Interpreting inflated significant results while ignoring nonsignificant results will thus lead to wrong conclusions. But current incentives to hunt for significance lead to selective reporting and to publication bias against nonsignificant findings. Data dredging, p-hacking, and publication bias should be addressed by removing fixed significance thresholds. Consistent with the recommendations of the late Ronald Fisher, p-values should be interpreted as graded measures of the strength of evidence against the null hypothesis. Also larger p-values offer some evidence against the null hypothesis, and they cannot be interpreted as supporting the null hypothesis, falsely concluding that 'there is no effect'. Information on possible true effect sizes that are compatible with the data must be obtained from the point estimate, e.g., from a sample average, and from the interval estimate, such as a confidence interval. We review how confusion about interpretation of larger p-values can be traced back to historical disputes among the founders of modern statistics. We further discuss potential arguments against removing significance thresholds, for example that decision rules should rather be more stringent, that sample sizes could decrease, or that p-values should better be completely abandoned. We conclude that whatever method of statistical inference we use, dichotomous threshold thinking must give way to non-automated informed judgment.

scholarly journals The earth is flat (p > 0.05): significance thresholds and the crisis of unreplicable research

PeerJ ◽  
2017 ◽  
Vol 5 ◽  
pp. e3544 ◽  
Author(s):  
Valentin Amrhein ◽  
Fränzi Korner-Nievergelt ◽  
Tobias Roth
Keyword(s):  
Publication Bias ◽  
Null Hypothesis ◽  
Statistical Power ◽  
Statistical Significance ◽  
Practical Importance ◽  
Effect Sizes ◽  
Point Estimate ◽  
Selective Reporting ◽  
Interval Estimate ◽  
True Effect

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 degradingp-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 smallp-values at face value, but mistrust results with largerp-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. A replication can therefore not be interpreted as having failed only because it is nonsignificant. Many apparent replication failures may thus reflect faulty judgment based on significance thresholds rather than a crisis of unreplicable research. Reliable conclusions on replicability and practical importance of a finding can only be drawn using cumulative evidence from multiple independent studies. However, applying significance thresholds makes cumulative knowledge unreliable. One reason is that with anything but ideal statistical power, significant effect sizes will be biased upwards. Interpreting inflated significant results while ignoring nonsignificant results will thus lead to wrong conclusions. But current incentives to hunt for significance lead to selective reporting and to publication bias against nonsignificant findings. Data dredging,p-hacking, and publication bias should be addressed by removing fixed significance thresholds. Consistent with the recommendations of the late Ronald Fisher,p-values should be interpreted as graded measures of the strength of evidence against the null hypothesis. Also largerp-values offer some evidence against the null hypothesis, and they cannot be interpreted as supporting the null hypothesis, falsely concluding that ‘there is no effect’. Information on possible true effect sizes that are compatible with the data must be obtained from the point estimate, e.g., from a sample average, and from the interval estimate, such as a confidence interval. We review how confusion about interpretation of largerp-values can be traced back to historical disputes among the founders of modern statistics. We further discuss potential arguments against removing significance thresholds, for example that decision rules should rather be more stringent, that sample sizes could decrease, or thatp-values should better be completely abandoned. We conclude that whatever method of statistical inference we use, dichotomous threshold thinking must give way to non-automated informed judgment.

scholarly journals Are Most Published Research Findings False In A Continuous Universe?

2021 ◽  
Author(s):  
Kleber Neves ◽  
Pedro Batista Tan ◽  
Olavo Bohrer Amaral
Keyword(s):  
Publication Bias ◽  
Effect Size ◽  
Statistical Power ◽  
Statistical Significance ◽  
Significance Test ◽  
Effect Sizes ◽  
Diagnostic Screening ◽  
Research Findings ◽  
Published Research ◽  
Effect Size Distribution

Diagnostic screening models for the interpretation of null hypothesis significance test (NHST) results have been influential in highlighting the effect of selective publication on the reproducibility of the published literature, leading to John Ioannidis’ much-cited claim that most published research findings are false. These models, however, are typically based on the assumption that hypotheses are dichotomously true or false, without considering that effect sizes for different hypotheses are not the same. To address this limitation, we develop a simulation model that overcomes this by modeling effect sizes explicitly using different continuous distributions, while retaining other aspects of previous models such as publication bias and the pursuit of statistical significance. Our results show that the combination of selective publication, bias, low statistical power and unlikely hypotheses consistently leads to high proportions of false positives, irrespective of the effect size distribution assumed. Using continuous effect sizes also allows us to evaluate the degree of effect size overestimation and prevalence of estimates with the wrong signal in the literature, showing that the same factors that drive false-positive results also lead to errors in estimating effect size direction and magnitude. Nevertheless, the relative influence of these factors on different metrics varies depending on the distribution assumed for effect sizes. The model is made available as an R ShinyApp interface, allowing one to explore features of the literature in various scenarios.

scholarly journals Why and How We Should Join the Shift From Significance Testing to Estimation

2021 ◽  
Author(s):  
Daniel Berner ◽  
Valentin Amrhein
Keyword(s):  
Null Hypothesis ◽  
Evolutionary Biology ◽  
Statistical Power ◽  
Alternative Hypothesis ◽  
Effect Sizes ◽  
Significance Testing ◽  
Significance Threshold ◽  
Beta Error ◽  
Statistical Hypotheses ◽  
Zero Effect

A paradigm shift away from null hypothesis significance testing seems in progress. Based on simulations, we illustrate some of the underlying motivations. First, P-values vary strongly from study to study, hence dichotomous inference using significance thresholds is usually unjustified. Second, statistically significant results have overestimated effect sizes, a bias declining with increasing statistical power. Third, statistically non-significant results have underestimated effect sizes, and this bias gets stronger with higher statistical power. Fourth, the tested statistical hypotheses generally lack biological justification and are often uninformative. Despite these problems, a screen of 48 papers from the 2020 volume of the Journal of Evolutionary Biology exemplifies that significance testing is still used almost universally in evolutionary biology. All screened studies tested the default null hypothesis of zero effect with the default significance threshold of p = 0.05, none presented a pre-planned alternative hypothesis, and none calculated statistical power and the probability of ‘false negatives’ (beta error). The papers reported 49 significance tests on average. Of 41 papers that contained verbal descriptions of a ‘statistically non-significant’ result, 26 (63%) falsely claimed the absence of an effect. We conclude that our studies in ecology and evolutionary biology are mostly exploratory and descriptive. We should thus shift from claiming to “test” specific hypotheses statistically to describing and discussing many hypotheses (effect sizes) that are most compatible with our data, given our statistical model. We already have the means for doing so, because we routinely present compatibility (“confidence”) intervals covering these hypotheses.

Précis of Statistical significance: Rationale, validity, and utility

1998 ◽  
Vol 21 (2) ◽  
pp. 169-194 ◽  
Author(s):  
Siu L. Chow
Keyword(s):  
Effect Size ◽  
Null Hypothesis ◽  
Statistical Power ◽  
Statistical Significance ◽  
Research Result ◽  
Practical Importance ◽  
Test Procedure ◽  
Statistical Hypothesis ◽  
Interval Estimate ◽  
Test Statistic

The null-hypothesis significance-test procedure (NHSTP) is defended in the context of the theory-corroboration experiment, as well as the following contrasts: (a) substantive hypotheses versus statistical hypotheses, (b) theory corroboration versus statistical hypothesis testing, (c) theoretical inference versus statistical decision, (d) experiments versus nonexperimental studies, and (e) theory corroboration versus treatment assessment. The null hypothesis can be true because it is the hypothesis that errors are randomly distributed in data. Moreover, the null hypothesis is never used as a categorical proposition. Statistical significance means only that chance influences can be excluded as an explanation of data; it does not identify the nonchance factor responsible. The experimental conclusion is drawn with the inductive principle underlying the experimental design. A chain of deductive arguments gives rise to the theoretical conclusion via the experimental conclusion. The anomalous relationship between statistical significance and the effect size often used to criticize NHSTP is more apparent than real. The absolute size of the effect is not an index of evidential support for the substantive hypothesis. Nor is the effect size, by itself, informative as to the practical importance of the research result. Being a conditional probability, statistical power cannot be the a priori probability of statistical significance. The validity of statistical power is debatable because statistical significance is determined with a single sampling distribution of the test statistic based on H0, whereas it takes two distributions to represent statistical power or effect size. Sample size should not be determined in the mechanical manner envisaged in power analysis. It is inappropriate to criticize NHSTP for nonstatistical reasons. At the same time, neither effect size, nor confidence interval estimate, nor posterior probability can be used to exclude chance as an explanation of data. Neither can any of them fulfill the nonstatistical functions expected of them by critics.

How to Detect Publication Bias in Psychological Research

2019 ◽  
Vol 227 (4) ◽  
pp. 261-279 ◽  
Author(s):  
Frank Renkewitz ◽  
Melanie Keiner
Keyword(s):  
Publication Bias ◽  
Effect Size ◽  
Statistical Power ◽  
Type I Error ◽  
Psychological Research ◽  
Type I ◽  
True Effect Size ◽  
Questionable Research Practices ◽  
True Effect ◽  
Meta Analyses

Abstract. Publication biases and questionable research practices are assumed to be two of the main causes of low replication rates. Both of these problems lead to severely inflated effect size estimates in meta-analyses. Methodologists have proposed a number of statistical tools to detect such bias in meta-analytic results. We present an evaluation of the performance of six of these tools. To assess the Type I error rate and the statistical power of these methods, we simulated a large variety of literatures that differed with regard to true effect size, heterogeneity, number of available primary studies, and sample sizes of these primary studies; furthermore, simulated studies were subjected to different degrees of publication bias. Our results show that across all simulated conditions, no method consistently outperformed the others. Additionally, all methods performed poorly when true effect sizes were heterogeneous or primary studies had a small chance of being published, irrespective of their results. This suggests that in many actual meta-analyses in psychology, bias will remain undiscovered no matter which detection method is used.

scholarly journals Statistical Reliability of 10 Years of Cyber Security User Studies

2021 ◽  
pp. 171-190
Author(s):  
Thomas Groß
Keyword(s):  
Publication Bias ◽  
Cyber Security ◽  
Power Distribution ◽  
Statistical Power ◽  
Total Sample ◽  
User Studies ◽  
Effect Sizes ◽  
Standard Errors ◽  
Statistical Reliability ◽  
Statistical Inferences

AbstractBackground. In recent years, cyber security user studies have been appraised in meta-research, mostly focusing on the completeness of their statistical inferences and the fidelity of their statistical reporting. However, estimates of the field’s distribution of statistical power and its publication bias have not received much attention.Aim. In this study, we aim to estimate the effect sizes and their standard errors present as well as the implications on statistical power and publication bias.Method. We built upon a published systematic literature review of 146 user studies in cyber security (2006–2016). We took into account 431 statistical inferences including t-, $$\chi ^2$$ χ 2 -, r-, one-way F-tests, and Z-tests. In addition, we coded the corresponding total sample sizes, group sizes and test families. Given these data, we established the observed effect sizes and evaluated the overall publication bias. We further computed the statistical power vis-à-vis of parametrized population thresholds to gain unbiased estimates of the power distribution.Results. We obtained a distribution of effect sizes and their conversion into comparable log odds ratios together with their standard errors. We, further, gained funnel-plot estimates of the publication bias present in the sample as well as insights into the power distribution and its consequences.Conclusions. Through the lenses of power and publication bias, we shed light on the statistical reliability of the studies in the field. The upshot of this introspection is practical recommendations on conducting and evaluating studies to advance the field.

The null-hypothesis significance-test procedure is still warranted

1998 ◽  
Vol 21 (2) ◽  
pp. 228-235 ◽  
Author(s):  
Siu L. Chow
Keyword(s):  
Null Hypothesis ◽  
Statistical Power ◽  
Meta Analysis ◽  
Statistical Significance ◽  
Test Procedure ◽  
Significance Test ◽  
Null Hypothesis Significance Test ◽  
Wide Range ◽  
Statistical Hypotheses ◽  
Alternative Hypotheses

Entertaining diverse assumptions about empirical research, commentators give a wide range of verdicts on the NHSTP defence in Statistical significance. The null-hypothesis significance-test procedure (NHSTP) is defended in a framework in which deductive and inductive rules are deployed in theory corroboration in the spirit of Popper's Conjectures and refutations (1968b). The defensible hypothetico-deductive structure of the framework is used to make explicit the distinctions between (1) substantive and statistical hypotheses, (2) statistical alternative and conceptual alternative hypotheses, and (3) making statistical decisions and drawing theoretical conclusions. These distinctions make it easier to show that (1) H0 can be true, (2) the effect size is irrelevant to theory corroboration, and (3) “strong” hypotheses make no difference to NHSTP. Reservations about statistical power, meta-analysis, and the Bayesian approach are still warranted.

Sign in / Sign up

Export Citation Format

Share Document