[Update 2015/1/14: I consolidate feedback from Twitter, comments, email, and real life into the main text (StackExchange-style), so that we get a good and improving answer. Thanks to @TonyLFreitas,Â @PhDefunct, @bahniks, @JoeHilgard, @_r_c_a, @richardmorey, @R__INDEX, the commenters at the end of this post and on the OSF mailing list, and many others for their feedback!]
In a recent lecture I talked about the replication crisis in psychology. After the lecture my students asked: âWe learn so much stuff in our lectures, and now you tell us that a considerable proportion of these âfactsâ probably are just false positives, or highly exaggerated? Then, what can we believe at all?â. A short discussion soon led to the crucial question:
In the era of #repligate: What are validÂ cues for the trustworthinessÂ of a study?
Valid cues for trustworthiness of a single study:
- Pre-registration. This might be one of the strongest cues for trustworthiness. Pre-registration makes p-hacking and HARKing unlikely (Wagenmakers, Wetzels, Borsboom, Maas, & Kievit, 2012), and takes care for a sufficient amount of statistical power (At least, some sort of sample size planning has been done. Of course, this depends on the correctness of the a-priori effect size estimate).
- Sample size / Statistical Power. Larger samples mean higher power, higher precision, and less false positives (Bakker, van Dijk, & Wicherts, 2012; Maxwell, Kelley, & Rausch, 2008; SchĂ¶nbrodt & Perugini, 2013). Of course sample size alone is not a panacea. As always, the garbage in/garbage out principle holds, and a well designed lab study with n=40 can be much more trustworthy than a sloppy mTurk study with n=800. But allÂ other things being equal, I put more trust in larger studies.
- Independent high-power replications. If a study has been independently replicated from another lab with high power and preferably pre-registered, this probably is the strongest evidence for the trustworthiness of a study (How to conduct a replication? See the Replication Recipe by Brandt et al., 2014).
- I guess that studies with Open Data and Open Material have a higher replication rate
- “Willingness to Share Research Data Is Related to the Strength of the Evidence and the Quality of Reporting of Statistical Resultsâ (Wicherts, Bakker, & Molenaar, 2011) â> this is not exactly Open Data, because here authors only shared data upon request (orÂ not). But it points into the same direction.
- Beyond publishing Open Data at all, the neatness of the data set and the quality of the analysis script is an indicator (see also comment by Richard Morey). The journal “Quarterly Journal of Political Science” demands to publish raw data and analysis code that generates all the results reported in the paper. Of these submissions, 54% “had results in the paper that differed from those generated by the authorâs own code”! My fear is that analytical code that has not been refined and polished for publishing contains even more errors (not to speak of unreproducible point-and-click analyses). Therefore, a well prepared data set and analysis code should be a valid indicator.
- Open Material could be an indicator that people are not afraid of replications and further scrutiny
- An abstract with reasonable conclusions that stick close to the data – see also below: “Red flags”. This includes visible efforts of the authors to explain how they could be wrong and what precautions were/were not taken.
- A sensitivity analysis, which shows that conclusions do not depend on specific analytical choices. For Bayesian analyses this means to explore how the conclusions depend on the choice of the prior. But you could also show how your results change when you do not exlude the outliers, or do not apply that debatable transformation to your data (see also comment)
- Using the “21 Word Solution” of Simmons, Nelson, & Simonsohn (2012) leads to a better replication index.
Valid cues for trustworthiness of a research programme/ multiple studies:
Valid cues for UNtrustworthiness of a single study/ red flags:
In a comment below, Dr. R introduced the idea of “red flags”, which I really like. These red flags aren’t a prove of the untrustworthiness of a study – but definitely a warning sign to look closer and to be more sceptical.
- Sweeping claims, counterintuitive, and shocking results (that don’t connect to the actual data)
- Most p values are in the range of .03 – .05 (or, equivalently, most t-values in the 2-3 range, or most F-values are in the 4-9 range; see comment by Dr. R below).
- How does a distribution of p values look like when there’s an effect? See DaniĂ«l Lakens blog. With large samples, p-values just below .05 even indicate support for the null!
- It’s a highly cited result, but no direct replications have been published so far. That could be an indicator that many unsuccessful replication attempts went into the file-drawer (see comment by Ruben below).
- Too good to be true: If several low-power studies are combined in a paper, it can be very unlikely that all of them produce significant results. The “Test of Excess Significance” has been used to formally test for “too many significant results”. Although this formal test has been criticized (e.g., see The Etz-Files, and especially the long thread of comments, or this special issue on the test), I still think excess significance can be used as a red flag indicator to look closer.
Possibly invalid cuesÂ (cues which are often used, but only seemingly are indicators for a studyâs trustworthiness):
- The journalâs impact factor. Impact factors correlate with retractions (Fang & Casadevall, 2011), but do not correlate with a single paper’s citation count (see here).
- Iâm not really sure whether that is a valid or invalid cue for a study’s quality. The higher retraction rate might due to the stronger public interest and a tougher post-publication review of papers in high-impact journals. The IF seems not to be predictive of a single paper’s citation count; but I’m not sure either whether the citation count is an index of a study’s quality. Furthermore, “Impact factors should have no place in grant-giving, tenure or appointment committees.” (ibid.), see also a reccent article by @deevybee in Times Higher Education.
- On the other hand, the current replicability estimate of a full volume of JPSP is only at 20-30% (see Reproducibility Project: Psychology). A weak performance for one of our “best journals”.
- The authorâs publication record in high-impact journals or h-index. This might be a less valid cue as expected, or even an invalid cue.
- Meta-analyses. Garbage-in, garbage-out: Meta-analyses of a biased literature produce biased results. Typical correction methods do not work well. When looking at meta-analyses, at least one has to check whether and how it was corrected for publication bias.
“If a study has a large sample size, Open Data, and maybe even has been pre-registered, I would put quite some trust into the results.Â If the study has been independently replicated, even better. In contrast to common practice, I do not care so much whether this paper has been published in a high-impact journal or whether the author has a long publication record.Â The next step, of course, is: Read the paper, and judge itâs validity and the quality of its arguments!”
This list certainly is not complete, and I would be interested in your ideas, additions, and links to relevant literature!