Preregistration in practice: A comparison of preregistered and non-preregistered studies in psychology

Preregistration has gained traction as one of the most promising solutions to improve the replicability of scientific effects. In this project, we compared 193 psychology studies that earned a Preregistration Challenge prize or preregistration badge to 193 related studies that were not preregistered. In contrast to our theoretical expectations and prior research, we did not find that preregistered studies had a lower proportion of positive results (Hypothesis 1), smaller effect sizes (Hypothesis 2), or fewer statistical errors (Hypothesis 3) than non-preregistered studies. Supporting our Hypotheses 4 and 5, we found that preregistered studies more often contained power analyses and typically had larger sample sizes than non-preregistered studies. Finally, concerns about the publishability and impact of preregistered studies seem unwarranted, as preregistered studies did not take longer to publish and scored better on several impact measures. Overall, our data indicate that preregistration has beneficial effects in the realm of statistical power and impact, but we did not find robust evidence that preregistration prevents p-hacking and HARKing (Hypothesizing After the Results are Known).

Publication bias impacts on effect size, statistical power, and magnitude (Type M) and sign (Type S) errors in ecology and evolutionary biology

Collaborative efforts to directly replicate empirical studies in the medical and social sciences have revealed alarmingly low rates of replicability, a phenomenon dubbed the 'replication crisis'. Poor replicability has spurred cultural changes targeted at improving reliability in these disciplines. Given the absence of equivalent replication projects in ecology and evolutionary biology, two inter-related indicators offer the opportunity to retrospectively assess replicability: publication bias and statistical power. This registered report assesses the prevalence and severity of small-study (i.e., smaller studies reporting larger effect sizes) and decline effects (i.e., effect sizes decreasing over time) across ecology and evolutionary biology using 87 meta-analyses comprising 4,250 primary studies and 17,638 effect sizes. Further, we estimate how publication bias might distort the estimation of effect sizes, statistical power, and errors in magnitude (Type M or exaggeration ratio) and sign (Type S). We show strong evidence for the pervasiveness of both small-study and decline effects in ecology and evolution. There was widespread prevalence of publication bias that resulted in meta-analytic means being over-estimated by (at least) 0.12 standard deviations. The prevalence of publication bias distorted confidence in meta-analytic results, with 66% of initially statistically significant meta-analytic means becoming non-significant after correcting for publication bias. Ecological and evolutionary studies consistently had low statistical power (15%) with a 4-fold exaggeration of effects on average (Type M error rates = 4.4). Notably, publication bias reduced power from 23% to 15% and increased type M error rates from 2.7 to 4.4 because it creates a non-random sample of effect size evidence. The sign errors of effect sizes (Type S error) increased from 5% to 8% because of publication bias. Our research provides clear evidence that many published ecological and evolutionary findings are inflated. Our results highlight the importance of designing high-power empirical studies (e.g., via collaborative team science), promoting and encouraging replication studies, testing and correcting for publication bias in meta-analyses, and adopting open and transparent research practices, such as (pre)registration, data- and code-sharing, and transparent reporting.

Reproducibility in the absence of selective reporting: An illustration from large-scale brain asymmetry research

The problem of poor reproducibility of scientific findings has received much attention over recent years, in a variety of fields including psychology and neuroscience. The problem has been partly attributed to publication bias and unwanted practices such as p-hacking. Low statistical power in individual studies is also understood to be an important factor. In a recent multisite collaborative study, we mapped brain anatomical left-right asymmetries for regional measures of surface area and cortical thickness, in 99 MRI datasets from around the world, for a total of over 17,000 participants. In the present study, we revisited these hemispheric effects from the perspective of reproducibility. Within each dataset, we considered that an effect had been reproduced when it matched the meta-analytic effect from the 98 other datasets, in terms of effect direction and significance threshold. In this sense, the results within each dataset were viewed as coming from separate studies in an "ideal publishing environment," that is, free from selective reporting and p hacking. We found an average reproducibility rate of 63.2% (SD = 22.9%, min = 22.2%, max = 97.0%). As expected, reproducibility was higher for larger effects and in larger datasets. Reproducibility was not obviously related to the age of participants, scanner field strength, FreeSurfer software version, cortical regional measurement reliability, or regional size. These findings constitute an empirical illustration of reproducibility in the absence of publication bias or p hacking, when assessing realistic biological effects in heterogeneous neuroscience data, and given typically-used sample sizes.

Analysis goals, error-cost sensitivity, and analysis hacking: Essential considerations in hypothesis testing and multiple comparisons

The "replication crisis" has been attributed to perverse incentives that lead to selective reporting and misinterpretations of P-values and confidence intervals. A crude fix offered for this problem is to lower testing cut-offs (α levels), either directly or in the form of null-biased multiple comparisons procedures such as naïve Bonferroni adjustments. Methodologists and statisticians have expressed positions that range from condemning all such procedures to demanding their application in almost all analyses. Navigating between these unjustifiable extremes requires defining analysis goals precisely enough to separate inappropriate from appropriate adjustments. To meet this need, I here review issues arising in single-parameter inference (such as error costs and loss functions) that are often skipped in basic statistics, yet are crucial to understanding controversies in testing and multiple comparisons. I also review considerations that should be made when examining arguments for and against modifications of decision cut-offs and adjustments for multiple comparisons. The goal is to provide researchers a better understanding of what is assumed by each side and to enable recognition of hidden assumptions. Basic issues of goal specification and error costs are illustrated with simple fixed cut-off hypothesis testing scenarios. These illustrations show how adjustment choices are extremely sensitive to implicit decision costs, making it inevitable that different stakeholders will vehemently disagree about what is necessary or appropriate. Because decisions cannot be justified without explicit costs, resolution of inference controversies is impossible without recognising this sensitivity. Pre-analysis statements of funding, scientific goals, and analysis plans can help counter demands for inappropriate adjustments, and can provide guidance as to what adjustments are advisable. Hierarchical (multilevel) regression methods (including Bayesian, semi-Bayes, and empirical-Bayes methods) provide preferable alternatives to conventional adjustments, insofar as they facilitate use of background information in the analysis model, and thus can provide better-informed estimates on which to base inferences and decisions.

Public availability and adherence to prespecified statistical analysis approaches was low in published randomized trials

BACKGROUND AND OBJECTIVE

Prespecification of statistical methods in clinical trial protocols and statistical analysis plans can help to deter bias from p-hacking but is only effective if the prespecified approach is made available.

STUDY DESIGN AND SETTING

For 100 randomized trials published in 2018 and indexed in PubMed, we evaluated how often a prespecified statistical analysis approach for the trial's primary outcome was publicly available. For each trial with an available prespecified analysis, we compared this with the trial publication to identify whether there were unexplained discrepancies.

RESULTS

Only 12 of 100 trials (12%) had a publicly available prespecified analysis approach for their primary outcome; this document was dated before recruitment began for only two trials. Of the 12 trials with an available prespecified analysis approach, 11 (92%) had one or more unexplained discrepancies. Only 4 of 100 trials (4%) stated that the statistician was blinded until the SAP was signed off, and only 10 of 100 (10%) stated the statistician was blinded until the database was locked.

CONCLUSION

For most published trials, there is insufficient information available to determine whether the results may be subject to p-hacking. Where information was available, there were often unexplained discrepancies between the prespecified and final analysis methods.

Evidence of unexplained discrepancies between planned and conducted statistical analyses: a review of randomised trials

BACKGROUND

Choosing or altering the planned statistical analysis approach after examination of trial data (often referred to as 'p-hacking') can bias the results of randomised trials. However, the extent of this issue in practice is currently unclear. We conducted a review of published randomised trials to evaluate how often a pre-specified analysis approach is publicly available, and how often the planned analysis is changed.

METHODS

A review of randomised trials published between January and April 2018 in six leading general medical journals. For each trial, we established whether a pre-specified analysis approach was publicly available in a protocol or statistical analysis plan and compared this to the trial publication.

RESULTS

Overall, 89 of 101 eligible trials (88%) had a publicly available pre-specified analysis approach. Only 22/89 trials (25%) had no unexplained discrepancies between the pre-specified and conducted analysis. Fifty-four trials (61%) had one or more unexplained discrepancies, and in 13 trials (15%), it was impossible to ascertain whether any unexplained discrepancies occurred due to incomplete reporting of the statistical methods. Unexplained discrepancies were most common for the analysis model (n = 31, 35%) and analysis population (n = 28, 31%), followed by the use of covariates (n = 23, 26%) and the approach for handling missing data (n = 16, 18%). Many protocols or statistical analysis plans were dated after the trial had begun, so earlier discrepancies may have been missed.

CONCLUSIONS

Unexplained discrepancies in the statistical methods of randomised trials are common. Increased transparency is required for proper evaluation of results.

Rigor and reproducibility for data analysis and design in the behavioral sciences

The rigor and reproducibility of science methods depends heavily on the appropriate use of statistical methods to answer research questions and make meaningful and accurate inferences based on data. The increasing analytic complexity and valuation of novel statistical and methodological approaches to data place greater emphasis on statistical review. We will outline the controversies within statistical sciences that threaten rigor and reproducibility of research published in the behavioral sciences and discuss ongoing approaches to generate reliable and valid inferences from data. We outline nine major areas to consider for generally evaluating the rigor and reproducibility of published articles and apply this framework to the 116 Behaviour Research and Therapy (BRAT) articles published in 2018. The results of our analysis highlight a pattern of missing rigor and reproducibility elements, especially pre-registration of study hypotheses, links to statistical code/output, and explicit archiving or sharing data used in analyses. We recommend reviewers consider these elements in their peer review and that journals consider publishing results of these rigor and reproducibility ratings with manuscripts to incentivize authors to publish these elements with their manuscript.

A Reckless Guide to P-values : Local Evidence, Global Errors

This chapter demystifies P-values, hypothesis tests and significance tests and introduces the concepts of local evidence and global error rates. The local evidence is embodied in this data and concerns the hypotheses of interest for this experiment, whereas the global error rate is a property of the statistical analysis and sampling procedure. It is shown using simple examples that local evidence and global error rates can be, and should be, considered together when making inferences. Power analysis for experimental design for hypothesis testing is explained, along with the more locally focussed expected P-values. Issues relating to multiple testing, HARKing and P-hacking are explained, and it is shown that, in many situations, their effects on local evidence and global error rates are in conflict, a conflict that can always be overcome by a fresh dataset from replication of key experiments. Statistics is complicated, and so is science. There is no singular right way to do either, and universally acceptable compromises may not exist. Statistics offers a wide array of tools for assisting with scientific inference by calibrating uncertainty, but statistical inference is not a substitute for scientific inference. P-values are useful indices of evidence and deserve their place in the statistical toolbox of basic pharmacologists.

A Bayesian nonparametric test of significance chasing biases

There is a growing concern that much of the published research literature is distorted by the pursuit of statistically significant results. In a seminal article, Ioannidis and Trikalinos (2007, Clinical Trials) proposed an omnibus (I&T) test for significance chasing (SC) biases. This test compares the observed number of studies that report statistically significant results, against their expected number based on study power, assuming a common effect size across studies. The current article extends this approach by developing a Bayesian nonparametric (BNP) meta-regression model and test of SC bias, which can diagnose bias at the individual study level. This new BNP test is based on a flexible model of the predictive distribution of study power, conditionally on study-level covariates which account for study diversity, including diversity due to heterogeneous effect sizes across studies. A test of SC bias proceeds by comparing each study's significant outcome report indicator against its estimated posterior predictive distribution of study power, conditionally on the study's covariates. The BNP model and SC bias test are illustrated through the analyses of 3 meta-analytic data sets and through a simulation study. Software code for the BNP model and test, and the data sets, are provided as Supporting Information.

Inadvertent P-hacking among trials and systematic reviews of the effect of progestogens in pregnancy? A systematic review and meta-analysis

BACKGROUND

Progestogens have been evaluated in numerous trials and meta-analyses, many of which concluded they were effective. However, two large trials PROMISE and OPPTIMUM have recently concluded that progesterone was ineffective. This raises the possibility that earlier studies and reviews had been biased by either selective publication or selective choice of outcomes, so called "P-hacking".

OBJECTIVES

To compare the findings all progestogen trials and systematic reviews with those of trials with pre-registered primary outcomes which avoided selective outcome reporting.

SEARCH STRATEGY

Search of PubMed, the Cochrane Library and trial registries. Registration PROSPERO CRD42016035303.

SELECTION CRITERIA

Systematic reviews of randomised trials comparing progestogen with placebo in pregnancy and the individual trials included in those reviews. The subset of trials reporting a pre-registered primary outcome were compared with the totality of trials and reviews.

DATA COLLECTION AND ANALYSIS

For reviews all outcomes were included. For individual trials all outcomes reported in the systematic reviews were included. For the comparison group we recorded the registered primary outcome from trials that were either registered before they started, or registered during the recruitment phase and also double blind.

MAIN RESULTS

Nineteen of twenty-nine meta-analyses concluded that progestogens were effective. Twenty-two trials reported their pre-registered primary outcomes. There was no effect of progesterone on primary registered dichotomous outcome RR 1.00 (95% CI 0.94-1.07). Only one of the 22 showed a nominally statistically significant benefit.

AUTHOR'S CONCLUSIONS

When evaluated in registered double-blind trials with analysis restricted to predefined primary outcomes, progestational agents in pregnancy are ineffective.

TWEETABLE ABSTRACT

Progestogens to prevent pregnancy loss, an example of P-hacking.

Reanalyzing Head et al. (2015): investigating the robustness of widespread p-hacking

Head et al. (2015) provided a large collection of p-values that, from their perspective, indicates widespread statistical significance seeking (i.e., p-hacking). This paper inspects this result for robustness. Theoretically, the p-value distribution should be a smooth, decreasing function, but the distribution of reported p-values shows systematically more reported p-values for .01, .02, .03, .04, and .05 than p-values reported to three decimal places, due to apparent tendencies to round p-values to two decimal places. Head et al. (2015) correctly argue that an aggregate p-value distribution could show a bump below .05 when left-skew p-hacking occurs frequently. Moreover, the elimination of p = .045 and p = .05, as done in the original paper, is debatable. Given that eliminating p = .045 is a result of the need for symmetric bins and systematically more p-values are reported to two decimal places than to three decimal places, I did not exclude p = .045 and p = .05. I conducted Fisher's method .04 < p < .05 and reanalyzed the data by adjusting the bin selection to .03875 < p ≤ .04 versus .04875 < p ≤ .05. Results of the reanalysis indicate that no evidence for left-skew p-hacking remains when we look at the entire range between .04 < p < .05 or when we inspect the second-decimal. Taking into account reporting tendencies when selecting the bins to compare is especially important because this dataset does not allow for the recalculation of the p-values. Moreover, inspecting the bins that include two-decimal reported p-values potentially increases sensitivity if strategic rounding down of p-values as a form of p-hacking is widespread. Given the far-reaching implications of supposed widespread p-hacking throughout the sciences Head et al. (2015), it is important that these findings are robust to data analysis choices if the conclusion is to be considered unequivocal. Although no evidence of widespread left-skew p-hacking is found in this reanalysis, this does not mean that there is no p-hacking at all. These results nuance the conclusion by Head et al. (2015), indicating that the results are not robust and that the evidence for widespread left-skew p-hacking is ambiguous at best.

In Defense of P Values

P values have become the scapegoat for a wide variety of problems in science. P values are generally over-emphasized, often incorrectly applied, and in some cases even abused. However, alternative methods of hypothesis testing will likely fall victim to the same criticisms currently leveled at P values if more fundamental changes are not made in the research process. Increasing the general level of statistical literacy and enhancing training in statistical methods provide a potential avenue for identifying, correcting, and preventing erroneous conclusions from entering the academic literature and for improving the general quality of patient care.

Transparency in Ecology and Evolution: Real Problems, Real Solutions

To make progress scientists need to know what other researchers have found and how they found it. However, transparency is often insufficient across much of ecology and evolution. Researchers often fail to report results and methods in detail sufficient to permit interpretation and meta-analysis, and many results go entirely unreported. Further, these unreported results are often a biased subset. Thus the conclusions we can draw from the published literature are themselves often biased and sometimes might be entirely incorrect. Fortunately there is a movement across empirical disciplines, and now within ecology and evolution, to shape editorial policies to better promote transparency. This can be done by either requiring more disclosure by scientists or by developing incentives to encourage disclosure.

The signed Kolmogorov-Smirnov test: why it should not be used

The two-sample Kolmogorov-Smirnov (KS) test is often used to decide whether two random samples have the same statistical distribution. A popular modification of the KS test is to use a signed version of the KS statistic to infer whether the values of one sample are statistically larger than the values of the other. The underlying hypotheses of the KS test are intrinsically incompatible with this approach and the test can produce false positives supported by extremely low p-values. This potentially makes the signed KS test a tool of p-hacking, which should be discouraged by replacing it with standard tests such as the t-test and by providing confidence intervals instead of p-values.