Monday, March 19, 2018

The Journal of Personality and Social Psychology: The Good, The Bad, and The Ugly


The Journal of Personality and Social Psychology is one of the main outlets for social and personality psychologists. It publishes around 110 empirical articles a year (and a small number of other types of articles) and is considered a prestigious outlet. It is also often criticized. That’s to be expected. As the Dutch saying goes: “High trees catch a lot of wind”.

For example, Simonsohn, Nelson, and Simmons validated their p-curve technique on articles that reported ANCOVA’s in JPSP. They found that if you look at a random selection of articles in JPSP that report ANCOVA’s, the pattern of results is what you would expect if there is no effect whatsoever, and we selectively publish Type 1 errors. JPSP also published the famous article on pre-cognition by Daryl Bem, and then desk-rejected a set of studies failing to replicate the original findings because they did not want to be the "Journal of Bem replication”. This is not very good.

Since these events, I have developed the habit of talking about JPSP as a journal that is prestigious, but not high quality. After recently reading a particularly bad article in JPSP, I said on Twitter that I thought JPSP was a 'crap journal', but Will Gervais correctly pointed out I can hardly conclude this based on n = 1. And if I'm honest, I believe it can't all be bad. So the truth must be somewhere in the middle. But where in the middle? I decided to take a quick look at the last 4 issues of JPSP (November, December, January, and February 2017/2018) to see if I was being unreasonably critical. So let’s go through the Good, The Bad, and the Ugly.

The Good

JPSP publishes a very special type of empirical article, which we might call the Odyssey format. They rarely publish single study papers, and most articles have between 4 and 7 empirical studies. Because these articles have a lot of studies, they can cover a lot of ground (such as The Good, The Bad, and The Ugly of Moral Affect, for example). I like this. It has potential to publish a coherent set of studies that really test a hypothesis thoroughly, and provide convincing support for an idea. Regrettably, the journal rarely lives up to this potential, but the format is good. If JPSP would require more replication and extension studies, where all studies use similar and well-validated measures, and if it would move further away from publishing sets of tangentially related conceptual replications, it could work. It’s an accomplishment that a journal can convince researchers to combine 7 studies into a single paper, instead of publishing 3 different papers, in this day and age.

It should not be a surprise that getting researchers to publish all their studies together leads to an outlet with a high impact factor. There are more reasons to cite an article that discusses The Good, The Bad, and The Ugly of X, than an article that discusses The Good of X. I’d say that simply based on the quantity of studies, JPSP should have an impact factor that is at least twice that of any journal that publishes single study papers – there’s just way more to cite. Furthermore, almost all articles are cited after a few years, with median citation counts from 2012 to 2017 currently being 33, 20, 16, 11, 5, and 2 (based on data I downloaded from Scopus). Remember that you can’t make statements about single articles or single researchers based on the performance of a journal (see and sign the DORA declaration). For example, I published a paper on equivalence testing in SPPS in 2017, which according to SCOPUS has 20 citations so far. This is more than any article published in JPSP in 2017.  So when evaluating a job applicant, I see no empirical justification to consider any single paper in JPSP better than any single paper in another journal (especially given The Ugly section below).




The journal consists of 3 sections. It took me a while to figure out how to identify them (the information in which section an article was published is not part of the bibliometric info) but the journal starts with a section on attitudes and social cognition, follows with a section on interpersonal relations and group processes, and ends with a section on personality processes and individual differences. The first article in the section has a header with the section name in the PDF, but unless you look for it, there is no way to know which section a paper was published in. This matters because the last section (on personality processes and individual differences) is rock solid. The other two sections are not unequivocally good, even though there are nice articles in these sections as well. If I was working in the personality processes and individual differences section of JPSP, I would split off from the journal (and roll over to a free open access publisher), and publish 4 instead of 10 issues a year filled with only articles on personality processes and individual differences. But I guess that's me.

The Bad

As far as I can see, what sets JPSP apart from other journals is that they publish huge articles with way more studies per article than other outlets, and they are in a position to reject articles that are unlikely to appeal to a large audience (and thus, get cited less). But that’s it. The studies that are done are not better than those in other outlets – there are just more of them.

If we ignore the theoretical content, and focus on the methodological content, there is really nothing to write home about in JPSP. A depressingly large number of studies relies exclusively on Null-Hypothesis Significance Testing, performed badly. In a majority of papers (and sorry if you are the exception – I love you!) interpretations are guided by p < .05 or p > .05. A p = .13 is the absence of an effect, p = 0.04 is an effect (see Lakens, Scheel, & Isager, 2018, on how to prevent this mistake). Effect sizes are sometimes not even reported, but if they are, they are literally never interpreted. Measures that are used are often ad hoc, not validated, and manipulations are often created for a single study but not extensively pilot tested or validated. There are a lot of studies, but the relationship between studies is almost always weak. The preferred approach is ‘now we show X in a completely different way’, whereas the potential of the JPSP format lies in ‘here we use the same validated measures across a set of studies with carefully piloted manipulations to show X’. But this rarely happens. As far as I could see when going through the papers, raw data availability is not zero, but very low (it would be super useful if JPSP used badges to clearly communicate where data is available, as some researchers recently proposed in an open letter to the then incoming editor of the first section of JPSP).

There is no reason to expect authors to use better methods and statistics in their work that ends up in JPSP, unless the editors at JPSP would hold these authors to higher standards. They don’t. I’m an editor at other journals (which might be a COI except I really don't care about journals in psychology and mainly read preprints), and I saw a lot of things I wouldn’t let authors get away with. Sample size justifications are horribly bad, power analyses (in the rare cases they are performed) are done incorrectly, and people who publish in (the first two sections of) JPSP generally love MTurk.

You might say that it’s just a missed opportunity that JPSP does not meet its potential and just publishes ‘more’. But there is another Bad. If we value a 7 study article in JPSP more than a 2 study article in another outlet, we will reward researchers who have lots of resources more than researchers with less resources. Whether or not you can perform 7 studies depends on how much money you have, the size of your participant pool, whether you have student assistants to help you or not, etcetera. So when evaluating a job applicant, I see no empirical justification to consider any single paper in JPSP with 7 studies as a better accomplishment than any single paper in another journal with 2 studies, without taking into account the resources the applicant had. It will not be a perfect correlation, but I predict that if you have more money, you will get more JPSP articles.

The Ugly

So one important question for me was: Is JPSP still publishing papers in which researchers increase their Type 1 error rate through p-hacking? Or is there, beyond the rare bad apple, mainly high quality work in JPSP? I took a look at four recent issues (November, December, January, and February 2017/2018). I evaluated all articles in the February issue, and checked how many articles in the other 3 issues raised strong suspicions of p-hacking. Being accused of p-hacking is not nice. Trust me, I know. But people admit they selectively report and p-hack (Fiedler & Schwarz, 2015, John et al., 2012), and if you read the articles in JPSP that I think are p-hacked, I doubt there will be much disagreement. If you wrote one of these articles, and are unaware of the problems with p-hacking, I'd recommend enrolling in my MOOC.

The articles in the personality processes and individual differences looked much better. This is to a large extent because the studies rarely rely on the outcome of a single DV measured after an experimental manipulation in a too small sample. The work is more descriptive, datasets are typically larger, and thus there is no need to selectively report what ‘works’. I’m not an expert in this field, so there might be things wrong with these papers that I was oblivious to, but to me these papers looked good.

I also skipped one article about artificial intelligence tools can detect whether someone is gay – the study is under ethical review and so problematic I thought it was fair to ignore it. Although I guess the study would deserve to fall under The Ugly.

This leaves 6 articles in the first two sections of the February issue. At the end of this post you can see the main tests for each study (copy pasted from the HTML version of the articles) and some comments about sample sizes, and my evaluation, if you prefer more detail about the basis of my evaluation. It’s obviously best to read the articles yourself. Of these 6 articles, two examined hypotheses that were plausible, and did so in a convincing manner. For example, an article showed that participants judged targets who they know performed immoral actions (broadly defined) as less competent.

The four other papers revealed the pattern I had feared. A logical question is: How can you identify a set of studies that is p-hacked? If the p-values in JPSP were realistically distributed, the distribution should look something like the curve in the picture below. Some predicted p-values should fall above .05, some fall below (indicated by the red area). Papers can have a surprising number of of p-values just below .05, when we should expect much smaller p-values (e.g., p = 0.001) often when there are true effects. This is in essence what p-curve analysis tests (or see TIVA by Uli Schimmack, and this recent blog by Will Gervais for related ideas). I don't present a formal p-curve analysis (although I checked some papers statistically), but in essence, I believe the pattern of p-values is unrealistic enough to cause doubt in objective readers with sufficient knowledge about what p-values across studies should realistically look like, especially in combination with a lack of pre-registration, and when many different dependent variables are reported across studies (and not all DV's are transparently reported). I want to make it clear that it might be possible, although very rare, that a single paper shows this surprising pattern of p-values even had every analysis been pre-registered - but there are too many papers like this in JPSP. If you published one of the articles I think was p-hacked and want to argue you did nothing wrong, that's fine with me. If everyone wants to do this, that's just not possible.

 
The studies that concern me showed that all predictions worked out as planned, even though sample sizes were decided upon quite randomly, and for all tests (or mediation models) p-values were just below .05. For example, DelPriore, Proffitt Leyva, Ellis, & Hill (2018) examined the effects of paternal disengagement on women’s perceptions of male mating intent. Their pattern of results across studies is Study 1, p = .030, Study 2, p = .049, Study 3, p = .04 and p = .028, Study 4, p = .019 and p = .012, Study 5, b = .09 (SE = .05), percentile 95% CI [.002, .18] (see how close these CI are to 0). Now, it is possible that couple of reviewers and an editor can miss the fact that this pattern is not realistic if they have not educated themselves on these matters somewhere during the last 5 years, but they really should have if they want to publish high quality work. Or take Stellar, Gordon, Anderson, Piff, McNeil, & Keltner (2018) who studies awe and humility. Study 1: p = .04, and p = .01, Study 2: p < .001, when controlling for positive affect, p = .02, Study 3: p = .02, Study 4: p < .001 and p = .03, Study 5: “We found a significant path from the in vivo induction condition (neutral = 0, awe = 1) to humility, via awe and self-diminishment (95% CI [0.004, 0.22]; Figure 4). (It’s amazing how close these CI can get to 0).

I don’t want to generalize this to all JPSP articles. I don’t intend to state that two-thirds of JPSP articles in the first two sections are p-hacked. I (more quickly) went through the 3 issues published before the February issue, to see if the February issue was some kind of fluke, but it isn't. Below are examples of articles with unrealistic patterns of p-values in the November, December, and January issues.

Issue 114(1)
Hofer, M. K., Collins, H. K., Whillans, A. V., & Chen, F. S. (2018). Olfactory cues from romantic partners and strangers influence women’s responses to stress. Journal of Personality and Social Psychology, 114(1), 1-9. In study 1, 96  couples, main result: “There was a nonsignificant main effect of scent exposure, F(2, 93) = 1.15, p = .32, η2  = 0.02, which—of most relevance for our hypothesis—was qualified by a significant interaction between time and scent exposure, F(5.36, 249.44) = 2.26, p = .04, η2  = 0.05.” Other effects: During stress recovery, women exposed to their partner’s scent reported significantly lower perceived stress than both those exposed to a stranger’s or an unworn scent (M = 20.25, SD = 14.96 vs. M = 27.14, SD = 16.67 and M = 29.01, SD = 14.19; p = .038 and .015, respectively, Table 1). Cortisol: There was a nonsignificant main effect of scent exposure, F(2, 93) = 0.83, p = .44, η2  = 0.02, which—of most relevance for our hypotheses—was qualified by a significant interaction between time and scent exposure, F(2.83, 131.76) = 3.05, p = .03, η2  = 0.06.

Issue 113(4)
Cortland, C. I., Craig, M. A., Shapiro, J. R., Richeson, J. A., Neel, R., & Goldstein, N. J. (2017). Solidarity through shared disadvantage: Highlighting shared experiences of discrimination improves relations between stigmatized groups. Journal of Personality and Social Psychology, 113(4), 547-567. Study 1: p = .078, Study 2: p = .001, and p = .010, Study 3: p = .031, p = .019, and p = .070, p = .036, and p = .120, Study 4: p = .030, p = .017, p = .014, and p= .040, Study 5: p = .020, p = .045 and p = .016.

Savani, K., & Job, V. (2017). Reverse ego-depletion: Acts of self-control can improve subsequent performance in indian cultural contexts. Journal of Personality and Social Psychology, 113(4), 589-607. Study 1A: p = .018, Study 1B: p = .049, Study 1c: p = .031, Study 2: p = .047. and p = .051, Study 3: p = .002 and p = .006 (this looks good, the follow up analysis is p = .045). Study 4: p = .023; p = .018; and p = .003.

Issue 113(3)
Chou, E. Y., Halevy, N., Galinsky, A. D., & Murnighan, J. K. (2017). The goldilocks contract: The synergistic benefits of combining structure and autonomy for persistence, creativity, and cooperation. Journal of Personality and Social Psychology, 113(3), 393-412. Study 1: p = .01, p = .007, p = .05, Study 2: manipulation check: p = .05. main result: p = .01, p = .02, p = .01 and, p = .08, p = .06, and p = .03. Study 3a: manipulation check: p = .03. Main result: p = .05, p = .02, p = .06. Study 3b: p = .04, p = .02, p = .02. Study 3C: p = .08, Study 4: p < .001, p = .01, Study 5A: p = .03, p = .03, p = .03, Study 5B: p = .01, p = .05. Experiment 6. p = .03.

Conclusion

I would say that something is not going right at JPSP. The journal has potential, in that it has convinced researchers to submit a large number of studies, consisting of a line of research, instead of publishing single study papers. And even despite the fact that most sets of studies lack strong coherence, are weak in sample size justification, validation of manipulations, and the choice of measures, there are some good articles published in the first two sections (the last section is doing fine).

However, there is a real risk that if you encounter a single article from JPSP, it is p-hacked and might just be a collection of Type 1 errors. You can easily notice this if you simply look at the main hypothesis tests in each study (or all tests in a mediation model).

When evaluating a job candidate, you can not treat a JPSP article as a good article. The error rate of making such statements will be too high. Especially in the first two sections, there is (based on my rather limited sampling, but still) a much higher error rate than 5% (enter a huge confidence interval here – if anyone wants to go through all issues, be my guest!) if you would attempt such simplistic evaluations of the work people have published. Now you should never evaluate the work of researchers just based on the outlet they published in. I’m just saying that if you would use this as a heuristic, you’d also be quite often wrong when it comes to JPSP.

This is regrettable. I’d like a journal that many people in my field consider prestigious and a flagship journal to mean something more than it currently does. High standards when publishing papers should mean more than ‘As an editor I counted the number of studies and there are more than 4 and I think many people are interested in this’. I hope JPSP will work hard to improve their editorial practices, and I hope researchers who publish in JPSP will not believe their work is high quality (remember that regardless of p-hacking, most studies had weak methods and statistics), but critically evaluate how they can improve. If JPSP is serious about raising the bar, there are straightforward things to do. Require a good sample size justification. Focus more on the interpretation of effect sizes, and less on p-values. Look for articles that use the same validated manipulations and measures consistently across studies. Publish (preferably preregistered) studies with mixed results, because not every prediction should be significant even when examining a true effect. And make sure the p-value distribution for all key hypothesis tests looks realistic. There has been so much work on improving research practices in recent years, that I expected a flagship journal in my field would have done more by now. 



Thanks to Will Gervais for motivating me to write this blog, and to Will, Farid Anvari and Nick Coles for feedback on an earlier draft.




















If you are interested, below is a more detailed look at the articles I read while preparing this blog post.



Attitudes and social cognition



Stellar, J. E., & Willer, R. (2018). Unethical and inept? the influence of moral information on perceptions of competence. Journal of Personality and Social Psychology, 114(2), 195-210.



General idea: Across 6 studies (n = 1,567), including 2 preregistered experiments, participants judged targets who committed hypothetical transgressions (Studies 1 and 3), cheated on lab tasks (Study 2), acted selfishly in economic games (Study 4), and received low morality ratings from coworkers (Study 5 and 6) as less competent than control or moral targets.



All studies show this convincingly. There is an OSF project where all data and materials are shared, which is excellent: https://osf.io/va6bj/?view_only=1220367fb74e44a4a15c0d8ef3cdfbf4. I think the hypothesis is very plausible (and even unsurprising). I want to point out that this is a very good paper, because I am less positive about another paper by the same first author in the same issue. But this is a later paper, so overall, it seems we are seeing progress in ways of working, which makes me happy.



Olcaysoy Okten, I., & Moskowitz, G. B. (2018). Goal versus trait explanations: Causal attributions beyond the trait-situation dichotomy. Journal of Personality and Social Psychology, 114(2), 211-229.



From the abstract: Participants tended to attribute the cause of others’ behaviors to their goals (vs. traits and other characteristics) when behaviors were characterized by high distinctiveness (Study 1A & 1B) or low consistency (Study 2). On the other hand, traits were ascribed as predominant causal explanations when behaviors had low distinctiveness or high consistency. Study 3 investigated the combined effect of those behavioral dimensions on causal attributions and showed that behaviors with high distinctiveness and consistency as well as low distinctiveness and consistency trigger goal attributions.



Evaluation: Good. The effects are all very large, and I would say had a very high prior. It would have been nice if data and materials would have been shared (especially the materials, to evaluate how surprising the data were, given the materials used).



Woolley, K., & Risen, J. L. (2018). Closing your eyes to follow your heart: Avoiding information to protect a strong intuitive preference. Journal of Personality and Social Psychology, 114(2), 230-245.



Main hypothesis: We predict that people avoid information that could encourage a more thoughtful, deliberative decision to make it easier to enact their intuitive preference.



Study 1: Sample 300 MTurk workers. Post-hoc power analysis. Main result: “As predicted, a majority of participants (62.7%; n = 188) chose to avoid calorie information, z = 4.33, p < .001, 95% CI = [56.9%, 68.2%].” But this is a nonsensical test against 50%. Why would we want to test against 50%? It makes no sense.



Study 2A: Sample 150 guests at a museum. Main result: “We found the predicted effect of payment information, β = −.63, SE = .27, Wald = 5.45, p = .020, OR = .53, 95% CIExp(B) = [.32, .90].” Regrettably, the authors interpret a p = 0.183 as evidence for the absence of an effect: “As predicted, there was no interaction between choice of information and bonus amount, β = .36, SE = .27, Wald = 1.78, p = .183, OR = 1.43, 95% CIExp(B) = [.85, 2.42],” This is wrong.



Study 2B: Sample 300 MTurkers. Main result: “As predicted, the stronger participants’ intuitive preference for the cartoon task, the more they avoided the bonus information, β = .23, SE = .07, Wald = 9.61, p = .002, OR = 1.26, 95% CIExp(B) = [1.09, 1.45].” This seems quite expected.



Study 3: 200 guests at a museum. Main result: “ndeed, using a chi-square analysis we found that people avoid information more when offered an opportunity to bet that a student would do poorly (57.8%) than when offered a chance to bet that a student would do well (42.9%), χ2 (1, N = 200) = 4.49, p = .034, φ = .15, OR = 1.83, 95% CIExp(B) = [1.04, 3.21].”



Study 4a: Sample 200 Mturkers. Main result: “As predicted, there was greater information avoidance in the plan-choice condition when the information was relevant to the decision (58.4%)  [  8  ]  than in the plan-assigned condition (41.4%), χ2 (1, N = 200) = 5.78, p = .016, φ = .17, OR = 1.99, 95% CIExp(B) = [1.13, 3.49].”



Study 4b: Sample size is doubled from 4a. Main result: A chi-square analysis of condition (assigned to information or not) on choice (Plans A-C vs. Plan D) revealed the predicted effect. More participants selected Plan D, the financially rational option, when assigned to receive information (59.6%, n = 121), than when assigned no information (48.0%, n = 95), χ2 (1, N = 401) = 5.45, p = .020, φ = .12, OR = 1.60, 95% CIExp(B) = [1.08, 2.38].



Study 5: 200 guests at a museum. Main result: “We first tested our main prediction that information avoidance is greater when it can influence the decision. As predicted, more people chose to avoid information when offered an opportunity to accept or refuse a bet that a student would do poorly (61.2%) than when assigned to bet that a student would do poorly (43.4%), χ2 (1, N = 197) = 6.25, p = .012, φ = .18, OR = 2.06, 95% CIExp(B) = [1.17, 3.63].”



Evaluation: Some tests are nonsensical – such as the test against 50% in Study 1. Weird that passed peer review. Everything else works out way too nicely. All critical p-values are either good (but then the test is kind of trivial as in Study 1) or all between the .01-.05 range, which is not plausible. This does not look realistic. This pattern of p-values suggest massive selective reporting, and flexibility in the data analysis to yield p < .05.  



Interpersonal relations and group processes,



Stellar, J. E., Gordon, A., Anderson, C. L., Piff, P. K., McNeil, G. D., & Keltner, D. (2018). Awe and humility. Journal of Personality and Social Psychology, 114(2), 258-269.



Main idea: “We hypothesize that experiences of awe promote greater humility. Guided by an appraisal-tendency framework of emotion, we propose that when individuals encounter an entity that is vast and challenges their worldview, they feel awe, which leads to self-diminishment and subsequently humility.”



Study 1: 119 freshmen, no justification for sample size. Main result: “In keeping with Hypothesis 1, participants who reported frequent and intense experiences of awe were judged to be more humble by their friends controlling for both openness and positive affect, r(92) = .22, p = .04, as well as openness and a discrete positive emotion—joy, r(92) = .25, p = .01.  [  2  ]  These two analyses generally remained significant when we added liking as an additional control variable, positive affect: r(91) = .18, p = .09; joy: r(91) = .21, p = .05.”



Study 2: Sample: 106 same freshmen from Study 1. Authors do not report all measures that were collected (“Embedded among other self-report items not relevant to this study”). Main result: Feeling humble and awe is correlated. No doubt, but almost no care to control for confounds, and only including positive affect is almost enough to make the effect disappear: Participants reported feeling more humble on days when they experienced more awe, B = .18, t(176) = 6.56, p < .001). This effect held when controlling for positive affect, B = .07, t(1178) = 2.31, p = .02, and when controlling for the prosocial emotion of compassion, B = .13, t(197) = 4.71, p < .001.



Study 3: Sample 104 adults, 14 excluded for non-preregistered reasons. Main result: “Participants in the awe and neutral conditions had a different balance between disclosing their strengths and weaknesses, t(84) = 2.38, p = .02.”



Study 4: Sample 598 adults from MTurk. Main result: “Participants who recalled an awe experience reported a significantly larger amount of their success coming from external forces compared with the self (M = 55.28, SD = 25.89) than those who wrote about a neutral (M = 44.03, SD = 21.50), t(593) = 4.78, p < .001), or amusing experience (M = 50.27, SD = 23.79), t(593) = 2.12 p = .03.” Mediation model is similarly hanging on a thread.



Study 5: Sample: 93 undergraduates, no justification. Mediation model: “We found a significant path from the in vivo induction condition (neutral = 0, awe = 1) to humility, via awe and self-diminishment (95% CI [0.004, 0.22]; Figure 4).” It’s amazing how close these CI can get to 0.



Evaluation: Everything works out for these authors, and that without sample size planning. This is simply not realistic. I believe almost all critical tests in this paper are selectively reported, and there are clear signs of flexibility in the data analysis to yield p < .05.  



Webber, D., Babush, M., Schori-Eyal, N., Vazeou-Nieuwenhuis, A., Hettiarachchi, M., Bélanger, J. J., . . . Gelfand, M. J. (2018). The road to extremism: Field and experimental evidence that significance loss-induced need for closure fosters radicalization. Journal of Personality and Social Psychology, 114(2), 270-285.



Study 1: Sample is 74 members suspected of a terrorist organization. Main result: “Analyses on the full sample first revealed a nonsignificant total effect between the predictor (LoS) and the outcome (extremism); b = .15, SE = .12, p = .217. Results next revealed that LoS predicted NFC; b = .26, SE = .13, p = .050; and that NFC subsequently predicted extremism; b = .36, SE = .10, p < .001. The direct effect of LoS on extremism was not significant; b = .06, SE = .11, p = .622. To examine the significance of the indirect effect, we calculated bias corrected 95% confidence intervals of the indirect effects using 10,000 bootstrapped resamples. As “0” was not contained within the confidence intervals, the indirect effect was indeed significant; 95% CI [.024, .215].”

Study 2: Sample: 237 (male) former members of the LTTE. Main result: “Analyses on the full sample revealed a significant total effect of LoS on extremism; b = .27, SE = .05, p < .001. Analyses further revealed that LoS was related to increased NFC; b = .27, SE = .11, p = .012; and NFC was related to increased extremism; b = .06, SE = .03, p = .048. The direct effect of LoS on extremism remained significant; b = .25, SE = .05, p < .001. Ninety-five percent confidence intervals obtained with 10,000 bootstrapped resamples revealed that the indirect effect was significant; 95% CI [.0001, .044]. Analyses on the reduced sample and including covariates revealed an identical pattern of results, and levels of significance were unchanged; 95% CI [.0004, .051].”

Study 3: Sample: 196 people through online websites. Main result: “Power analysis is based on a medium effect (which is bad practice). Main result: “Only the main effect of LoS condition was significant, such that participants in the LoS condition (M = 5.19; SE = .17) expressed significantly greater extremism than participants in the control condition (M = 4.63; SE = .17); F(1, 192) = 5.52, p = .020, η2  = .03.”

Study 4: Sample: 344 participants from Amazon Turk. Main result: “The total effect of LoS condition on endorsement of extreme political beliefs was not significant (p = .828). Analyses further revealed that participants in the LoS (vs. control) condition reported higher NFC; b = .24, SE = .08, p = .003; and NFC was related to increased extremism; b = .17, SE = .08, p = .028. The direct effect of LoS condition on extremism remained nonsignificant (p = .892). Ninety-five percent confidence intervals obtained with 10,000 bootstrapped resamples revealed a significant indirect effect, 95% CI [.006, .100].”



Evaluation: First of all, major credits for collecting these samples or people (suspected to be) involved in terrorist organizations. This is really what social psychology can contribute to the world. Regrettably, the data is overall not convincing. The data is not messy enough (across the 4 studies, everything that needs to work works) but very often things are borderline significant (think about a bootstrapped CI (which will vary a bit every time) that has a lower limit of .0001, reported to 4 decimals!). Still, major credits for data collection.



DelPriore, D. J., Proffitt Leyva, R., Ellis, B. J., & Hill, S. E. (2018). The effects of paternal disengagement on women’s perceptions of male mating intent. Journal of Personality and Social Psychology, 114(2), 286-302.



Main conclusion from abstract: Together, this research suggests that low paternal investment (including primed paternal disengagement and harsh-deviant fathering) causes changes in daughters’ perceptions of men that may influence their subsequent mating behavior.



Study 1, n1 = 34, n2 = 41. Substantial data exclusions without clear reasons. Not pre-registered. Main finding “However, there was a significant main effect of priming condition, F(1, 73) = 4.91, p = .030, d = .52.” No corrections for multiple comparisons.

Study 2, n1 = 35, n2 = 33. Main result: “As predicted, there was a significant simple main effect of condition on women’s perceptions of male sexual arousal, F(1, 65) = 4.01, p = .049, d = .49.”

Study 3, similar sample sizes, main result: “There was, however, a significant three-way interaction between priming condition, target sex, and target emotion, F(3, 81) = 2.82, p = .04. This interaction reflected a significant simple main effect of priming condition on women’s perceptions of male sexual arousal, F(1, 83) = 5.02, p = .028, d = .49,”

Study 4, where I just looked at the main result: “The analysis revealed a significant main effect of priming condition on women’s perceptions of the male confederate’s dating intent, F(1, 60) = 5.82, p = .019, d = .61, and sexual intent, F(1, 60) = 6.69, p = .012, d = .66.”

Study 5: “Both indirect pathways remained statistically significant: paternal harshness → perceived sexual intent → unrestricted sociosexuality: b = .09 (SE = .05), percentile 95% CI [.002, .18]; paternal harshness → residual father-related pain → perceived sexual intent: b = .04 (SE = .03), bias corrected 95% CI [.004, .12].” You can see how close these CI are to 0.

Finally, the authors perform an internal meta-analysis.



Evaluation: This does not look realistic. This pattern of p-values suggest massive selective reporting, and flexibility in the data analysis to yield p < .05.  



Personality processes and individual differences



Humberg, S., Dufner, M., Schönbrodt, F. D., Geukes, K., Hutteman, R., van Zalk, Maarten H. W., . . . Back, M. D. (2018). Enhanced versus simply positive: A new condition-based regression analysis to disentangle effects of self-enhancement from effects of positivity of self-view. Journal of Personality and Social Psychology, 114(2), 303-322.



Main idea (from abstract): “We provide a new condition-based regression analysis (CRA) that unequivocally identifies effects of SE by testing intuitive and mathematically derived conditions on the coefficients in a bivariate linear regression. Using data from 3 studies on intellectual SE (total N = 566), we then illustrate that the CRA provides novel results as compared with traditional methods. Results suggest that many previously identified SE effects are in fact effects of PSV alone.”



Evaluation. The materials are on the OSF, and it is primarily a new statistical technique (no new data was collected). It does not fall in the empirical studies I am examining, but it was a good paper.



Dejonckheere, E., Mestdagh, M., Houben, M., Erbas, Y., Pe, M., Koval, P., . . . Kuppens, P. (2018). The bipolarity of affect and depressive symptoms. Journal of Personality and Social Psychology, 114(2), 323-341.



Main idea (from abstract): “these findings demonstrate that depressive symptoms involve stronger bipolarity between positive and negative affect, reflecting reduced emotional complexity and flexibility.”



Evaluation: This was too far out of my domain to confidently evaluate. There were nice things in the article, pretty close replications across 3 experience sampling studies, a multiverse analysis to explore all possible combinations of items.



Siddaway, A. P., Taylor, P. J., & Wood, A. M. (2018). Reconceptualizing anxiety as a continuum that ranges from high calmness to high anxiety: The joint importance of reducing distress and increasing well-being. Journal of Personality and Social Psychology, 114(2), e1-e11.



Evaluation: I’ll keep this short, because unless I’m mistaking, this seems to be an electronic only replication study: “We first replicate a study by Vautier and Pohl (2009), who used the State–Trait Anxiety Inventory (STAI) to reexamine the structure of anxiety. Using two large samples (N = 4,138 and 1,824), we also find that state and trait anxiety measure continua that range from high calmness to high anxiety.” It’s good and clear.



Vol 114 (1)

Hofer, M. K., Collins, H. K., Whillans, A. V., & Chen, F. S. (2018). Olfactory cues from romantic partners and strangers influence women’s responses to stress. Journal of Personality and Social Psychology, 114(1), 1-9.



Study 1: 96  couples. Main result: “There was a nonsignificant main effect of scent exposure, F(2, 93) = 1.15, p = .32, η2  = 0.02, which—of most relevance for our hypothesis—was qualified by a significant interaction between time and scent exposure, F(5.36, 249.44) = 2.26, p = .04, η2  = 0.05.” Other effects: During stress recovery, women exposed to their partner’s scent reported significantly lower perceived stress than both those exposed to a stranger’s or an unworn scent (M = 20.25, SD = 14.96 vs. M = 27.14, SD = 16.67 and M = 29.01, SD = 14.19; p = .038 and .015, respectively, Table 1). Cortisol: There was a nonsignificant main effect of scent exposure, F(2, 93) = 0.83, p = .44, η2  = 0.02, which—of most relevance for our hypotheses—was qualified by a significant interaction between time and scent exposure, F(2.83, 131.76) = 3.05, p = .03, η2  = 0.06.

Evaluation: This does not look realistic. This pattern of p-values suggest massive selective reporting, and flexibility in the data analysis to yield p < .05.  



Vol 113(4)



Cortland, C. I., Craig, M. A., Shapiro, J. R., Richeson, J. A., Neel, R., & Goldstein, N. J. (2017). Solidarity through shared disadvantage: Highlighting shared experiences of discrimination improves relations between stigmatized groups. Journal of Personality and Social Psychology, 113(4), 547-567.



Study 1: 47 participants, in 2 between subject groups (party like it’s1999). Main result: Black participants’ support for same-sex marriage was somewhat higher when it was framed as a civil rights issue and similar to the experiences of Black Americans (M = 5.74, SD = 1.06) compared with when it was framed as a gay rights issue (M = 4.82, SD = 2.09), Brown-Forsythe t(28.22) = 1.83, p = .078, d = 0.57.



Study 2: 35 participants (15 vs 20 in two between subject conditions). Main results: As predicted, and replicating results from Experiment 1, Black participants in the shared experience with discrimination condition (framing gay marriage as a “civil rights issue”) expressed more support for same-sex marriage (M = 6.01, SD = 0.66) compared with participants in the control condition (framing gay marriage as a “gay rights issue”; M = 4.13, SD = 2.14), Brown-Forsythe t(23.65) = 3.72, p = .001, d = 1.12. Consistent with predictions, Black participants in the shared experience with discrimination condition reported greater empathy for same-sex couples (M = 5.70, SD = 1.33) than did participants in the control condition (M = 4.03, SD = 2.28), Brown-Forsythe t(31.42) = 2.72, p = .010, d = 0.86.



Study 3: 63 participants across 3 (yes, three) conditions. An effect of experimental condition emerged for attitudes toward lesbians, Brown-Forsythe F(2, 46.63) = 3.76, p = .031, and gay men, Brown-Forsythe F(2, 50.20) = 4.29, p = .019. Consistent with predictions, Games-Howell post hoc analyses revealed that compared with participants in the control condition (M = 5.71, SD = 1.14), participants in the blatant shared experience condition expressed somewhat more positive attitudes toward lesbians (M = 6.34, SD = 0.53, p = .070, d = 0.69).  [  5  ]  Similarly, participants in the blatant shared experience condition expressed more positive attitudes toward gay men (M = 5.97, SD = 0.71) than participants in the control condition (M = 5.40, SD = 1.13), although this effect was unreliable (p = .140, d = 0.59). Further, compared with participants in the control condition, participants in the subtle shared experience condition expressed more positive attitudes toward gay men (M = 6.14, SD = 0.69, p = .036, d = 0.78) and more positive attitudes toward lesbians (M = 6.30, SD = 0.75), although this effect was unreliable (p = .120, d = 0.61).



Study 4: Power analysis (1) expecting a d = 0.76 (!). Completely unreasonable, but ok. 102 participants. The results are really lovely: As shown in Table 1 and consistent with predictions, Asian American participants in the shared experience with discrimination condition expressed more perceived similarity with gay/lesbian people compared with those in the control condition, t(100) = 2.20, p = .030, d = 0.43 (see Table 1). Furthermore, conceptually replicating Experiment 3, Asian American participants in the shared experience with discrimination condition expressed more positive attitudes toward lesbians compared with those in the control condition, Brown-Forsythe t(90.57) = 2.44, p = .017, d = 0.48. In addition, Asian American participants in the shared experience with discrimination condition expressed more positive attitudes toward gay men compared to those in the control condition, Brown-Forsythe t(92.53) = 2.52, p = .014, d = 0.50. Finally, Asian American participants in the shared experience with discrimination condition expressed more support for gay and lesbian civil rights compared to those in the control condition, Brown-Forsythe t(90.84) = 2.09, p = .040, d = 0.41.



Study 5: 201 participants. Results: We conducted a 2 (mindset: similarity-seeking, neutral) × 2 (pervasive sexism: sexism salient, control) between-subjects ANOVA on participants’ anti-Black bias scores, revealing the predicted Pervasive Sexism × Mindset interaction, F(1, 184) = 5.55, p = .020, ηp 2  = .03. No main effects emerged (mindset: F(1, 184) < 1, p = .365, ηp 2  = .00; pervasive sexism: F(1, 184) < 1, p = .751, ηp 2  = .00). As seen in Figure 1 and consistent with predictions and prior research (Craig et al., 2012), among participants who described the series of landscapes in the neutral mindset condition, salient sexism led to greater anti-Black bias compared with the control condition (salient sexism article: M = 3.63, SD = 1.09; control article: M = 3.21, SD = 1.08), F(1, 184) = 4.06, p = .045, d = 0.39, 95% CI [0.01, 0.78]; see Figure 1). Furthermore, consistent with our prediction that manipulating a similarity-seeking mindset in the context of salient ingroup discrimination should reduce bias, among White women for whom sexism was made salient, inducing a similarity-seeking mindset (M = 3.12, SD = 1.07) led to less expressed anti-Black bias compared with inducing a neutral mindset, F(1, 184) = 5.91, p = .016, d = 0.48, 95% CI [0.09, 0.87].



Evaluation: This does not look realistic. This pattern of p-values suggest massive selective reporting, and flexibility in the data analysis to yield p < .05.  





Savani, K., & Job, V. (2017). Reverse ego-depletion: Acts of self-control can improve subsequent performance in indian cultural contexts. Journal of Personality and Social Psychology, 113(4), 589-607.



Study 1A: 77 ppn, result: As predicted, the Condition × Incongruence interaction was significant, B = −.055, SE = .023, z = 2.36, p = .018

Study 1B: 57 ppn, result: We found a significant effect of condition, B = 0.17, z = 1.97, p = .049,

Study 1c: 500 Mturkers. Main result: We found a significant effect of condition, B = 0.031, SE = .014, incidence rate ratio = 1.03, z = 2.15, p = .031

Note how the sample sizes change wildly in these studies, but the p-values stay just below .05? That’s a desk-reject if any of my bachelor students had been reviewers, but ok. Let’s read on.

Study 2: 180 Indians and 193 Americans on MTurk. Results: For Indians, we again found a main effect of incongruent trials, B = .12, SE = .009, z = 13.58, p < .001, and a Condition × Incongruence interaction, B = −.027, SE = .014, z = 1.99, p = .047. For Americans, we found a main effect of trial incongruence, B = .17, SE = .008, z = 21.27, p < .001, and a Condition × incongruence interaction, B = .022, SE = .011, z = 1.95, p = .051.

Study 3: 143 students in the lab. Results: As predicted, the Condition × Incongruence interaction was significant, B = −.046, SE = .015, z = 3.04, p = .002. The Condition × Incongruence interaction was significant, B = .034, SE = .012, z = 2.77, p = .006.

Hey, this looks good! So they do a follow-up analysis: Among Indian participants, however, we found a significant three-way interaction, B = −.0088, SE = .0044, z = 2.00, p = .045.

Study 4: 400 Mturkers from Inda, 400 from US. Results: We also found three two-way interactions: Culture × Strenuous versus nonstrenuous task condition, B = .048, SE = .021, z = 2.27, p = .023; Culture × Belief condition, B = −.051, SE = .021, z = 2.37, p = .018; and strenuous versus nonstrenuous task Condition × Belief condition, B = .062, SE = .021, z = 2.93, p = .003. The three-way Culture × Task Condition × Belief Condition interaction was nonsignificant, B = .038, SE = .043, z = .89, p = .38.



Evaluation: This does not look realistic. This pattern of p-values suggest massive selective reporting, and flexibility in the data analysis to yield p < .05.  



Vol 113(3)



Chou, E. Y., Halevy, N., Galinsky, A. D., & Murnighan, J. K. (2017). The goldilocks contract: The synergistic benefits of combining structure and autonomy for persistence, creativity, and cooperation. Journal of Personality and Social Psychology, 113(3), 393-412.



Study 1: 124 Mturkers. Results: An analysis of variance (ANOVA) revealed a significant effect of contract type on task persistence, F(2, 118) = 4.25, p = .01, partial η2  = .06. As predicted, the general-contract workers (M = 554.30 s, SD = 225.65) worked significantly longer than both the no-contract (M = 417.91 s, SD = 205.92), t(82) = 2.77, p = .007, Cohen’s d = .63, and the specific-contract workers (M = 455.66 s, SD = 177.15), t(63) = 1.97, p = .05, Cohen’s d = .49; the specific and no-contract groups did not differ, t(91) = .91, p = .36



Study 2: 188 MTurkers. Manipulation check:The contract manipulation was effective: Workers rated the general contract as less specific (M = 4.14, SD = .88) than the specific contract (M = 4.50, SD = 1.18), t(121.04) = −1.95, p = .05. (Phew! It worked). As predicted, workers’ contracts influenced their feelings of autonomy, F(2, 185) = 4.06, p = .01, partial η2  = .04: Workers who received the general contract (M = 4.92, SD = 1.01) or no contract at all (M = 4.98, SD = 1.07) felt more autonomy than workers who received the specific contract (M = 4.50, SD = 1.06), t(124) = 2.24, p = .02; Cohen’s d = .41; and, t(127) = 2.58, p = .01; Cohen’s d = .45, respectively. The general and no-contract groups did not significantly differ, t(119) = −.36, p = .23, suggesting that specific contracts reduced people’s feelings of autonomy. Ok, the next measures are not significant, let’s report them and call it ‘discriminant validity’(!). Contract type had no effect on feelings of competence, F(2, 185) = .67, p = .51, or belongingness, F(2, 185) = 1.53, p = .21. These results provide discriminant validity and support the importance of autonomy needs in driving the effects of contract specificity. And then: Time 1’s contracts influenced Time 2’s task persistence. Workers with a general contract at Time 1 worked almost twice as long at Time 2 (M = 914.06 s, SD = 1365.61) as workers who had received either the specific (M = 519.17 s, SD = 220.69), t(40.41) = 1.78, p = .08; Cohen’s d = .40, or no contract (M = 496.08, SD = 205.96), t(40.66) = 1.88, p = .06; Cohen’s d = .43. A planned contrast showed that workers who received the general contract persisted longer than workers in the other two conditions combined, t(93) = −2.20, p = .03.

Study 3a: 175 MTurk workers. Manipulation check. The contract manipulation was effective: Workers rated the general contract as less specific (M = 3.08, SD = 1.03) than the specific contract (M = 3.45, SD = .78), t(114) = −2.14, p = .03, Cohen’s d = .40. Phew! The manipulation check worked again! Lucky us! Results: Replicating our findings from Experiment 2, workers’ contracts influenced their feelings of autonomy, F(2, 172) = 2.94, p = .05, partial η2  = .03: Workers who received the general contract (M = 3.20, SD = .96) felt greater autonomy than workers who received the specific contract (M = 2.80, SD = .90; t(114) = 2.29, p = .02, Cohen’s d = .42). Those who did not receive any contract (M = 3.12, SD = .97) felt marginally more autonomy than those who received the specific contract t(118) = 1.89, p = .06, Cohen’s d = .34.

Study 3b: 82 students. Results:  Those who thought that the lab’s code of conduct was more general felt more autonomy (M = 3.68, SD = .73) than those who thought the lab code of conduct was more specific (M = 3.35, SD = .73), t(80) = 2.01, p = .04, Cohen’s d = .45. As predicted, we found a significant interaction between contract condition and perceived structure on autonomy (B = −.37, SE = .16, t = −2.32, p = .02). Bootstrapping analysis verified that the effect of general contract on autonomy is significant only when people perceive a sense of structure (95% bias-corrected bootstrapped CI [−1.21, −.07]). Likewise, we found a significant interaction between contract condition and perceived structure on intrinsic motivation (B = −.42, SE = .18, t = −2.36, p = .02).

Study 3C: Single-indicator path modeling using nonparametric bootstrapping indicates that the proposed model fit the data well (comparative fit index [CFI] = 0.98, root-mean-square error of approximation [RMSEA] = 0.05), χ2 (3) = 6.67, p = .08,

Study 4: 149 undergraduates. Results: Participants who received the general legal clauses worked significantly longer (M = 590.08 s, SD = 332.44) than those who received the specific legal clauses (M = 415.12 s, SD = 217.01), F(1, 145) = 14.66, p < .001, Cohen’s d = .62. The opposite pattern emerged for the technical clause manipulation: Participants who received the general technical clause spent less time on the task (M = 443.29 s, SD = 184.38) as compared with those who received the specific technical clauses (M = 552.37 s, SD = 357.41), F(1, 145) = 6.27, p = .01, Cohen’s d = .38. [These results on their own would be ok, if the first test did not yield a slightly too large effect size].

Study 5A: 91 MTurkers. Results: As predicted, the general contract led workers to produce more original ideas (M = 4.02, SD = .89) than the specific contract did (M = 3.57, SD = 1.11), t(89) = 2.14, p = .03, Cohen’s d = .45. General contracts also led workers to produce more unique ideas (M = 8.29, SD = 3.71 vs. M = 6.55, SD = 3.85), t(89) = 2.18, p = .03, Cohen’s d = .46. We replicated the main effect of general contracts on idea generation with a separate sample (80 MTurk workers; mean age = 37.95, SD = 12.34; 67% female), using slightly different contracts. Workers who received the general contract generated more unique uses than those who received the specific contract did (M = 8.02, SD = 3.8 vs. M = 6.43, SD = 2.86), t(78) = 2.15, p = .03, d = .47.

Study 5B: 143 MTurkers. Results: As predicted, workers in the general contract condition solved significantly more problems correctly (M = 1.38, SD = .70) than workers in the specific contract condition (M = 1.07, SD = .73), t(141) = 2.61, p = .01, Cohen’s d = .43. The general contract (M = 4.51, SD = 1.09) also produced stronger intrinsic motivation than the specific contract (M = 4.15, SD = 1.04), t(141) = 1.95, p = .05, Cohen’s d = .34.

Experiment 6. As predicted, participants who received the general contract cooperated at a significantly higher rate (M = 84%, SD = 36%) than those who received the specific legal clauses (M = 70%, SD = 46%), χ2 (1) = 4.60, p = .03.



Evaluation: This does not look realistic. This pattern of p-values suggest massive selective reporting, and flexibility in the data analysis to yield p < .05.  

2 comments:

  1. I found this interesting, from DelPriore et al.'s Study 5: "Given the absence of statistically significant main effects or interactions following from the randomly assigned writing prime, we proceeded to analyze for the effects of the emotions about fathers expressed in the essays." That sounds to me rather like "We decided which analyses to perform based on what we found in the data". (I'm looking forward to seeing how many registered reports have, as an a priori hypothesis, the specific prediction of a full mediation effect.)

    Also, this, from Study 4: "Given the extant literature demonstrating reliable effects of paternal absence-disengagement on sexually proceptive behavior in women (as reviewed in the Introduction), a one-tailed statistical test (p = .031) could be justified here, supporting a causal effect of paternal disengagement on flirting." Now I know that Daniel is a big fan of one-tailed tests, but it seems to me that this is basically pleading for "something that we wish was 'true', but we can't say it's 'true' because its p value is >.05" to be turned into "something that we can say is 'true' because its p value is <.05", with the justification that *somebody else previously found a similar result*. Taking this to its logical conclusion, we could run absolutely any study, and whenever the p value doesn't pan out, just say "Oh well, we didn't quite get lucky today, but we know it's 'true' because these other people found a similar effect, so we'll pretend we had their results instead of ours".

    ReplyDelete
    Replies
    1. I agree with everything - maybe only should point out I believe one-sided tests should be used more often, but after pre-registering.

      Delete