Social Influence in Donation Behavior: The Effect of Mean, Variance, and Psychopathy

Reviewer #1 (Public review):

This manuscript investigates how people use sequential social information when deciding how much to donate to charity. Across four preregistered experiments, participants first made baseline donations to a set of charities, then observed a sequence of donations from five other people whose mean and variability were experimentally manipulated, and finally made a second donation to the same charities. The authors ask whether the mean and variability of others' donations affect the mean and variability of participants' own donations, and whether individual differences in psychopathy and empathy are associated with responsiveness to social information.

The main behavioral finding is that participants shifted their second donations toward the mean of the donations they observed: generous social information increased donations, whereas stingy social information decreased donations. In contrast, the variability of observed donations had little effect on the mean donation shift, but did affect the variability of participants' subsequent donations, with more consistent social information producing stronger reductions in variability. The authors also fit several computational models and conclude that a hybrid model, in which second donations reflect both participants' initial donations and learned predictions of others' donations, best accounts for the data. Finally, they report that psychopathic traits are positively associated with donation change and with model-derived social-information use, and that this association generalizes to a perceptual social-influence task in Experiment 4.

The paper addresses an interesting question and has several strengths, especially the repeated experimental design, the direct manipulation of social-information statistics, and the attempt to connect descriptive behavior with computational modeling and individual-difference measures. However, several aspects of the design and analysis currently block some of the major conclusions. The behavioral results provide convincing evidence that observed donation levels affect later donation decisions. The current evidence is less decisive for the stronger claims that the winning computational model identifies the underlying mechanism, that individual-level model parameters are robust phenotypes, and that psychopathy specifically increases susceptibility to social information.

Strengths:

A major strength of the manuscript is that it investigates social influence in charitable giving across four preregistered experiments with relatively large samples. The core mean-effect result is replicated across different donation scales, across hypothetical and incentivized settings, and across student and more general online samples. This gives the descriptive behavioral finding substantially more credibility than would be available from a single experiment.

The experimental manipulation is also valuable. Rather than presenting only a single prior donation or a simple group average, the authors expose participants to sequences of donations and independently manipulate the mean and variability of this social information. This design allows the authors to ask not only whether social information changes donation levels, but also whether the distributional structure of that information changes the variability of participants' own responses.

Another strength is the combination of traditional statistical analyses with computational modeling. The hybrid model is a reasonable descriptive candidate because it formalizes the intuitive idea that second donations may depend both on participants' initial preferences and on learned expectations about others' donations. This modeling approach has the potential to clarify mechanisms of social-information use, especially if the validation of the model and its individual-level parameters is strengthened.

Experiment 4 is a sensible extension because it uses an incentivized design, includes a more diverse sample, examines transfer to novel charities, and adds a perceptual social-influence task. These features broaden the empirical scope of the manuscript and make the psychopathy-related findings more interesting, although the perceptual-task result should still be treated as requiring replication.

Weaknesses

The first limitation concerns causal interpretation of the phase effects. Participants always make baseline donations first, then observe social information, and then make second donations to the same charities. There is no non-social repeated-donation control condition. This type of design does support the conclusion that donation changes differ as a function of the observed social-information condition, especially the mean of others' donations. However, it does not by itself fully isolate social influence from other processes that could also occur between a first and second donation to the same item, such as repeated exposure to the charities, slider familiarity, memory of the first donation, regression to the mean, reduced uncertainty, fatigue, or "the experiment clearly wants me to update" demand effects. This issue is especially relevant for the claim that observing others' donations generally reduces the variability of individual donations. The variability effect may well be socially driven, but the absence of a non-social or irrelevant-information repeated-donation control means that this cannot be decisively demonstrated.

The second limitation concerns the trial-level mixed models. The primary mixed-effects models include random intercepts for participants and items, but do not appear to include random slopes for within-participant or within-item phase effects. Since phase is repeatedly manipulated within participants and items, random-intercept-only models may underestimate uncertainty for some phase interactions, resulting in anti-conservative p-values. The convergent participant-level ANOVA analyses are reassuring, but the trial-level inferential claims would be stronger if the authors reported additional analyses using fuller random-effects structures or other methods that better reflect the repeated-measures structure.

The third limitation concerns model comparison and model validation. The computational models are fit separately to each participant, and model comparison is based on summed information criteria and protected exceedance probabilities derived from those participant-level fits. This is informative about relative conditional fit within the tested sample and model set. However, the manuscript uses the winning model to support broader claims about latent computational mechanisms, individual computational phenotypes, psychopathy-related susceptibility, and potential intervention relevance. For these claims, the relevant prediction target is generalization to new participants, whose individual parameters are not known in advance. The current model-comparison approach is not well aligned with that target. Additionally, the loss appears to combine prediction trials and donation outcomes, so the selected model may more strongly reflect performance at predicting participants' guesses about others rather than specifically predicting their own donation decisions.

The fourth limitation concerns the model adequacy checks and recovery analyses. The analyses described as posterior predictive checks do not appear to be posterior predictive checks, because the models are not Bayesian and there consequently isn't a posterior to check. Instead, the analyses appear closer to some sort of in-sample fitted-value reconstruction checks. Such checks provide limited evidence of model adequacy, especially because the same second-donation data used to estimate individual parameters are then used to assess whether the fitted model reproduces the main behavioral patterns. In addition, the reported model and parameter recovery analyses use extremely favorable response-noise assumptions that are not expected to be met in real data. The analyses establish that the models and parameters are mathematically distinguishable in principle, but they do not establish that the individual-level parameters are reliably recoverable under realistic empirical noise levels to the extent required for the analyses performed in the manuscript.

The fifth limitation concerns the interpretation of the psychopathy results. The association between psychopathic traits and donation change is interesting and appears directionally consistent across experiments. However, the interpretation that psychopathy increases susceptibility to social information is vulnerable to biasing by baseline-distance. The manuscript reports that psychopathy is negatively associated with baseline donations in Experiments 1-3. Participants with lower baseline donations have more room to move toward generous social information, and absolute donation change is partly a function of the distance between the initial donation and the observed social mean for mechanical reasons. Thus, an association between psychopathy and absolute donation change could theoretically arise even if psychopathy does not directly increase social susceptibility.

A sixth limitation is that we could not find the links to the preregistration. The authors state when preregistered hypotheses were or were not supported, but it is unclear how these hypotheses were phrased. Most notably, it is unclear how variance in the observed donation choices was supposed to influence participants. As a side note, it was not quite clear if the variance in the observations was higher or lower across charities, across observed persons, or across both.

Several more minor suggestions can also be made regarding the modelling and the presentation of the task, etc.

Reviewer #2 (Public review):

Summary:

This manuscript examines how the statistical properties of others' charitable donations shape subsequent giving using four preregistered experiments and computational modelling. The authors find that both the average level and variability of observed donations influence donation behaviour, and that individual differences in social information use are associated with psychopathic traits.

Strengths:

This is a well-executed paper on the important question of how social information shapes charitable giving. In my view, the combination of preregistered experiments, large sample sizes, computational modelling, and a multi-paradigm approach makes for convincing evidence. The progression across experiments, the use of real donation data rather than deception, the incentivized experiment 4, and the generalization to a second paradigm are all notable strengths. The introduction is clearly written and well-motivated - an enjoyable read. The experimental paradigm is thoughtfully designed, and the methods and supplementary materials are described in considerable detail. The computational modelling provides useful additional insights beyond the behavioural analyses.

As far as I could tell, the manuscript also adheres closely to the preregistrations. The primary hypotheses, experimental designs, exclusion criteria, and key analyses are all consistent with the preregistered plans. Deviations seem to consist of methodological improvements (e.g., mixed-effects models replacing ANOVAs), additional computational and robustness analyses, and therefore strengthen rather than weaken the manuscript. (NB: for transparency, I would appreciate a clearer distinction between preregistered and post hoc analyses, as well as a brief explanation for why some preregistered secondary analyses are no longer reported; see minor comments below).

Overall, I enjoyed reading this paper. I believe it will make a valuable contribution. My comments below are intended to further strengthen an already solid manuscript.

Weaknesses:

(1) The rationale for the social-information phase could be clarified further. Given the research question, I wondered why participants observed the five donations sequentially (and only briefly) rather than simultaneously. In particular, variance is arguably more difficult than the mean to encode and remember, and a sequential presentation may both obscure distributional differences and introduce primacy or recency effects. It would be helpful if the authors could better motivate this design choice, and indicate whether they examined possible order effects.

Relatedly, I felt somewhat uncertain about the purpose of asking participants to predict each donation before observing it. The prediction phase appears to play an important role in the computational model, but its theoretical role is not clearly introduced. Is it intended as a measure of participants' evolving beliefs about the descriptive norm, or primarily as a modelling device? Finally, were these predictions incentivized (e.g., for accuracy), and if not, how should readers interpret them?

(2) I would appreciate having the full experimental materials reproduced in the Supplementary Information. This would make it easier to understand what participants experienced during the task, including what they were told about the "other participants" whose donations they observed.

Minor points:

(1) The interpretations around domain-generality would be strengthened by reporting the association between social information use in the charitable giving task and in the BEAST. Currently, both measures are shown to correlate with psychopathy, but it remains unclear whether individuals who rely strongly on social information in one task also do so in the other. Reporting this correlation (or explaining why it cannot be meaningfully computed) would provide a nice and direct test of a domain-general tendency to use social information.

(2) It would help to explain more explicitly why the standard deviation of donations is theoretically interesting in its own right. The motivation for studying the mean seems immediately intuitive, whereas the motivation for focusing on variability could be elaborated on further in the Introduction.

(3) As I said above, I think the manuscript follows the preregistrations closely. Maybe I missed it, but it seems that prediction accuracy and reaction-time analyses were omitted. It would improve transparency further if the authors would briefly mention the preregistered secondary analyses that are no longer reported (and explain why they were omitted).

Reviewer #3 (Public review):

Summary:

In this manuscript, the authors aimed to assess the mechanisms of social influence on charitable giving, particularly by separating the role of donation magnitude and variability in others' donations, and by examining the role of incremental social information in a learning framework. They additionally investigated individual differences in the magnitude effects in relation to self-reported psychopathy and empathy. The main findings suggest that magnitude and variability of others' donation impacted the magnitude and variability of the participants' donations, respectively, and that the weight of social information on individual decisions correlates positively with psychopathy, but not with empathy.

Strengths:

(1) The findings extend previous evidence for social influence on charitable giving to contexts where social information is provided incrementally, and to effects on the variability in social information (in addition to the mean).

(2) Individual differences suggest a role for psychopathy, but not empathy.

(3) Findings are replicated across all 4 (or for some findings 3 out of the 4) experiments, which helps strengthen the claims.

(4) Multiple experiments are a strength, especially Experiment 4, which helped address concerns/potential confounds in the previous experiments, increase representativeness of the sample, add incentive compatibility, and generalize to another task domain (perceptual).

(5) For modelling, strong model and parameter recovery was obtained, thus validating the modelling pipelines.

(6) The experiments were pre-registered, though it's unclear whether only planned analyses were pre-registered, or specific directional hypotheses. It would help if the manuscript took the reader through the pre-registration (and any deviation from it), instead of expecting the reader to do the comparison between the pre-registrations and actual manuscripts.

(7) The studies are appropriately powered, and power analyses are provided.

Weaknesses

(1) Lack of rationale and justification for the between-subjects design.

While this design may be appropriate in some cases (for example, for the generalization of donation to new charities or as a potential "intervention"), it would have been great to know if the findings related to social influence extend to a within-subjects design, especially given the weak results related to the effects of standard deviation in others' donations. It is possible that variability in others' responses would have a stronger effect if manipulated within individuals, since the same individual exposed to both high-SD and low-SD social information may weight low-SD information more, but this effect may lack when individuals are only exposed to the same variability across trials.

(2) Motivation for the RL framework.

The use of reinforcement learning (RL) isn't very well motivated, both in the introduction and methods/results (given the task). In particular, why is RL relevant to studying the problem of social influence, which isn't inherently a learning problem? This should be better motivated in the introduction. Second, when taking the task into account, it's unclear why RL is an appropriate model, given that from the perspective of the participant, the 5 others are different individuals, so the model shouldn't assume that predicting an individual's donation should be related to the previous individual's donation. Unless participants are informed that there is some dependency between the 5 donors they observe on each trial? If so, this should be made clear.

(3) Specifics of modelling analyses, and separability between prediction and second donation data.

Does the RL-based model (either prediction-only or hybrid) explain more variance in second donations than a simple linear regression model predicting second donation from initial donation and the mean of others' donations (or each individual other's donation)? It could be helpful to add some models that include social influence (i.e., integration of social and individual information) but no learning mechanisms per se. If this is not done, I do not believe that current results show that participants combine "their initial self-donation tendencies with their predictions of observed others' giving to guide their second individual donations". While participants may update their predictions, the authors should test multiple models of prediction update (fit only on the prediction data to understand the specific mechanisms of prediction update independently of second donation - for example, is it RL, or could it just be a running average, or some other heuristic? In parallel, it would be helpful to test whether it's the learned predictions (or whatever other prediction update mechanism was found to best explain the prediction data) or the actual others' donation information that best explains second donation - when combined with initial donation. These latter models would be fit on second donation data only in order to be comparable. If it's not possible to separate people's predictions from the actual social information (others' donations) then this should be acknowledged as a limitation. Ultimately, separating the modelling by data type (prediction only vs second donation data only) would help provide more insights into the learning mechanisms (if any) and whether it's learned prediction, or just social information, which influences second donation.

(4) Missing statistics in generalization to novel donation results.

On page 13, in the generalization effect, the authors mention that "Compared with participants exposed to High-SD social information, those exposed to Low-SD social information exhibited less variability in their novel donations, with this effect being especially pronounced in the Low-Mean condition." Was this supported by a significant interaction between SD and Mean condition? If so, please report the statistics of the interaction; if not, it's probably better to refrain from making this claim.

(5) Behavioral index of social influence individual differences.

For the first analysis reported on the association with psychopathy (Figure S9), as well as empathy (Figure S10), the absolute change between first and second donation does not seem like the appropriate marker of social influence. While I understand from Figure 2 that most participants changed their donation in a direction consistent with the social information, it would appear more appropriate to calculate an index of donation change consistent with influence, so calculated as D2 - D1 for the high mean groups and D1 - D2 for the low mean groups. This would be a better measure to interpret high values as an index of social influence.

(6) Interpretation of psychopathy effects.

a) The general idea that high psychopathy would be associated with increased social influence seems counterintuitive. While I appreciate that the authors controlled for additional variables such as age, gender, condition, and other model parameters, is it possible that this effect could be instead explained by the availability heuristic (the social information is more readily available to participants than their individual choice from the baseline trials), lower memory for their own choice, or lower IQ/cognitive abilities? These appear to be important confounds to address to be able to interpret the findings.

b) Related to this, and given that psychopathy/empathy were negatively/positively related to baseline donation amounts, it would be good to account for baseline mean donation amount in the individual difference analyses.

c) Finally, the authors interpret this association in line with other studies that have shown strategic social blending in psychopathy - while this seems possible in contexts where others are present, it doesn't really seem to be the case in this task. Did participants believe the other donors were watching them somehow? It also appears contradictory for the incentivized experiment, whereby if high psychopathy participants would no longer be able to "maintain a favorable social image while still pursuing their own self-interests" (p.23), since as soon as incentivization is added, participants' own self-interests are directly in conflict with the social image. Was participants' understanding of the incentive compatibility tested in Experiment 4?

(7) Asymmetry between generous vs stingy social influence and link with psychopathy.

a) Was such an asymmetry present - in other words, were people more strongly influenced by generous others or stingy others, or were the two comparable? I believe some analyses could be added to test this, and this is also where a within-subject design could help (e.g., different parameters for the two directions of social influence at the individual levels).

b) Related to that, does the correlation with psychopathy vary between conditions? It appears important to test if the increased social susceptibility is general or specific to increases (~high mean group, generous social influence) or decreases (~low mean group, stingy social influence) in donation. I understand that the main effect of psychopathy survived controlling for conditions, but it would still be interesting to test for an interaction between psychopathy and condition in predicting donation changes (calculated as suggested in point 5 above) or social influence weight.

(8) Perceptual task in Experiment 4.

a) While it is good to show that there was no correlation between psychopathy and initial estimate in the perceptual task, were there differences in initial estimate accuracy (i.e., difference between initial estimate and correct answer) along psychopathology? If so, this should be controlled for in the analyses. Given that social influence is always in the direction of the true value, the proportional deviations between initial estimate and social information could yield larger numerical differences and induce larger changes in estimate.

b) Even if previous studies have excluded rounds in which participants update their estimate in the opposite direction of the social information or move beyond it, I believe analyses that include those rounds should be included, especially in the context of individual difference analyses. Could it be that individuals who are high in psychopathy or low in empathy have a higher proportion of rounds where they go against the social influence? The same question applies to the main 4 experiments (in case this criterion was applied to) as well as the perceptual task.

c) Because the perceptual task was completed by the same participants as Experiment 4, were the two social influence measures correlated across tasks? Was psychopathy better predicted by a combination of predictors across the two tasks?

(9) Were individual difference measures examined in relation to the variability effect?

(10) Discussion.

The authors argue against a role for opportunistic conformity. While I tend to agree with their interpretation, I believe that it could be strengthened as follows:

a) First, it relies on a null result (the absence of a difference in decreases between low-mean low-SD and low-mean high-SD groups), which I do not believe was explicitly tested; and even if it was, it should ideally be corroborated by Bayesian statistics to provide strength of evidence for the null effect.

b) Second, this could be a great opportunity to dive into the mechanisms of social influence in the model, by testing the theory that only the lowest (or highest) donation from the group (rather than the mean, or the learned prediction) influences donation. Could a subset of participants be better fitted by such a model?

(11) Methods. Maybe I missed it, but it's unclear what participants were told about the other donors they are observing. It is mentioned that they were fully debriefed after the experiment, but what they were told in the instructions appears important. Was believability tested (this also relates to my comment #1 about the rationale for a between-subjects design, which creates fairly biased sets of social information from the perspective of a single participant)? And related to my comment #2, what participants were told about the donors could help justify the rationale for the RL framework.

Author response:

We thank the editors and reviewers for their thoughtful and constructive comments on our manuscript. We are pleased that they considered the core behavioral findings important and robust, especially the results showing that the magnitude and variability of others’ donations affected the magnitude and variability of participants' donations, respectively. We also appreciate their acknowledgement of the strengths of the experimental design, large sample sizes, the incentive-compatible and across-domain measures included in Experiment 4, and combined behavioral and computational approaches.

We agree that the manuscript would benefit from greater clarification in several areas, further analyses, and more cautious interpretations. In the revised manuscript, we plan to clarify the rationale for sequentially presenting social information, the role of prediction responses, the theoretical motivation of examining the variability of others’ donation, the use of the between-subjects design, and the motivation for the RL framework. We also agree that the lack of a non-social repeated-donation control condition limits the interpretation of the phase effects. Our design permits strong inferences about differences in donation changes across different conditions, but it cannot establish that the phase-related changes are exclusively attributable to social information exposure. We will revise the wording accordingly, moderate the causal language, and explicitly discuss the limitations of our design.

To strengthen the behavioral analyses, we plan to supplement the current mixed-effects linear models with models that reflect the repeated-measure structure of the task, including random slopes for the phase. We will also add statistics in the generalization results section and test the asymmetry between generous vs. stingy social influence. Moreover, the reviewers raised an important concern regarding the associations between psychopathy and the donation change. Because psychopathy is negatively correlated with initial donations in several experiments, absolute donation changes may partly reflect the distance between initial donation and the observed donation mean. We therefore plan to reanalyze the psychopathy effects by using signed donation changes and trial-level discrepancies between participants’ initial donations and the observed social information. These additional analyses will enable a more direct and precise assessment of whether psychopathy is associated with greater susceptibility to social influence.

We further agree that the comparison and validation of the computational models should be strengthened. In the revised manuscript, we plan to clarify that the learning models are intended to describe the updating beliefs about a group-level donation norm from sequential social information, rather than learning about a single donor. We will expand the candidate model set to include non-learning models, such as models based on the actual social mean, a running average. We will also model the prediction phase and the second donation phase separately. This will help identify the models that provide explanatory values for both predictions of others’ donations and individual donation behaviors. In addition, because the models were not estimated via a Bayesian framework, we agree that the term “posterior predictive checks” is inappropriate. We will rename these analyses. We will also rerun the parameter and model recovery analyses using empirically informed noise levels separately for the prediction and donation phases. In addition, we will implement model-evaluation processes, such as cross-validation, that better reflect prediction for new participants.

In Experiment 4, we plan to directly report the association between social-information-use measures in the perceptual and the donation task to strengthen the domain-generality effect. We will additionally examine whether social susceptibility in the perceptual task is associated with psychopathy by including all trials, including those in which participants moved away from or beyond the social value.

Finally, we will correct the reporting and presentation issues identified by the reviewers, including the social information use equation in the perceptual task, the pseudo-SD of individual donations formula, supplementary figure captions, and task duration. We will also provide fuller experimental materials and make the preregistration links more prominent. In addition, we intend to make the analysis code, model-fitting scripts, and data available during the revision process.

We greatly appreciate the editors’ and reviewers’ thoughtful suggestions, which will help us substantially strengthen the manuscript. We are grateful for the opportunity to address these important points and believe that the planned revisions will enhance the manuscript’s clarity, robustness, and its contribution to the understanding of social influence in donation behaviors.