Reward learning on social media

We tested our hypothesis that online social behavior, in the form of posts, follows principles of reward learning theory in four independent social media datasets (see Methods) (total N_Obs = 1,046,857, N_Users = 4,168) with computational modeling. These datasets came from four distinct social media platforms, where people post pictures and, in response, receive social reward in the forms of “likes.” In Study 1 (N_Users = 2,039), we tested our hypothesis in a large dataset of Instagram posts³⁰ (average number of posts per individual = 418, see Supplementary Table 1 for additional descriptive statistics). Instagram exemplifies modern social media, with over 1 billion registered users, and its format—focused primarily on simple postings and the receipt of likes as feedback—makes it a unique case study. However, because there are significant economic motives on Instagram and similar platforms³¹, an assessment of reward learning on Instagram could be limited to some extent by the possibility of fraudulent accounts and “fake likes,” among other strategic uses³². We therefore replicated and extended Study 1 in Study 2 (N_Users = 2,127) with data from three different topic-focused social media sites (discussion forums focused on Men’s fashion, Women’s fashion, and Gardening, see Methods), where economic motives are less likely (average number of posts per individual = 91, see Supplementary Table 1 for descriptive statistics). Finally, we conducted an online experiment (Study 3, N_Participants = 176), designed to mimic key aspects of social media platforms, in which we manipulated the social reward rate to verify its causal impact on response latencies.

Social rewards predict social media posting

We conceptualized the act of posting on a social media platform (e.g., Instagram) as free-operant behavior in a Skinner box with one response option (e.g., a single lever), where responses are followed by reward (i.e., likes). As outlined, a key prediction from learning theory for such situations, in which the agent can decide when to respond, is that the latency between responses should be affected by the average rate of rewards^28,29. Before formally testing our computational hypothesis, we evaluated, in two complementary and model-independent ways, whether social media behavior was sensitive to social rewards.

First, we drew inspiration from classic work in animal learning theory, which established that response rates, an aggregate measure of response latency, typically follow a saturating positive (i.e., hyperbolic) function of reward rates²⁸. This relationship, known as the quantitative law of effect²⁸, is a signature of reward driven behavior (especially on interval schedules of reinforcement²⁸). To directly test whether social media behavior exhibits this pattern, we compared how well a hyperbolic function explained the relationship between likes and response rates relative to a linear function (Supplementary Note 1). We found that the “quantitative law of effect” explained behavior better than a linear relationship in all four social media datasets (mean R²: Study 1 = 0.43, Study 2: = 0.37, see Supplementary Note 1), demonstrating that an aggregate measure of response latencies on social media exhibits a classic signature of reward learning²⁸.

Second, we defined a high resolution measure of response latency (τ_Post) as the time between two successive social media posts (similar to the interval between responses in human laboratory tasks and in animal free-operant behavior, see Fig. 1), and tested whether τ_Post was predicted by the history of likes using Granger causality analysis (see Methods). Granger causality is established if a variable (e.g., likes) improves on the prediction of a second variable (e.g., τ_Post) over and above earlier (lagged) values of the second variable in itself. To ascertain the selectivity of this method, we first applied it to simulated data from generative models where the ground truth was known (causality or no causality). We then fine-tuned the analysis parameters (the lag number, see Supplementary Note 2) to reliably detect Granger causality in data simulated from our reward learning model, which we introduce next, but not from models without learning (in which likes are unrelated to behavior, see Supplementary Note 2). Applying this optimized analysis method to the empirical data showed that likes Granger-caused τ_Post in all four datasets (Study 1: Z̃ = 23.65, p < 0001; Study 2: Men’s Fashion: Z̃ = 3.94, p < 0001; Women’s Fashion: Z̃ = 14.16, p < 0001; Gardening: Z̃ = 6.78, p < 0001). Together, these results demonstrate that the history of social rewards (i.e., likes) influenced both the rate and the time distribution of social media posting. Such reward sensitivity is a minimal criterion for more formally testing the explanatory power of learning theory.

Fig. 1

Schematic illustration of the computational hypothesis.

a The $$ \overline{\mathrm{R}}\mathrm{L}$$ model describes how τ_Post, the latency to the next social media post (denoted by the “camera” icon), is shaped by social rewards. Each post is followed by social reward (denoted by the “heart” symbol), which varies in number. The model adjusts the response policy, or threshold, which determines τ_Post, to maximize the average net rate of reward. b The $$ \overline{\mathrm{R}}\mathrm{L}$$ model posits an effort cost to responding (e.g., taking pictures, uploading), which decreases as a function of τ_Post. The effort cost term penalizes posting in quick succession, because high effort reduces the average reward rate. c The opportunity cost of time increases as a function of the average reward rate $$ \overline{R}$$. The gradient of red lines indicates increasing values of $$ \overline{R}$$ (darker colors represent higher values), and thereby higher opportunity cost. d The optimal value of τ_Post, which maximizes the net reward δ, varies as function of $$ \overline{R}$$ (darker colors represent higher values). The δ is used to update average reward rate $$ \overline{\mathrm{R}}$$. Note that the optimum, indicated by the peak of the function, moves to shorter response latencies when $$ \overline{\mathrm{R}}$$ is higher because the opportunity cost of time increases with $$ \overline{\mathrm{R}}$$. The horizontal line denotes 0. The figure assumes a constant effort cost C. e Simulated model predictions. The $$ \overline{\mathrm{R}}\mathrm{L}$$ model predicts that τ_Post, the latency between successive social media posts, will be shorter with high compared to low average reward rate, $$ \overline{\mathrm{R}}$$. The simulation involved N = 1000 independent synthetic individuals. The prediction is presented as estimated mean ± 99% CI from mixed-effects regression.

Modeling the dynamics of social media behavior

Having established that social media behavior is sensitive to reward, we next developed a generative model, based on RL theory of free-operant behavior in non-human animals²⁹. The key principle of this theory is that agents should balance the effort costs of responding and the opportunity costs of passivity (i.e., the posting-related rewards one misses while not posting) to maximize the average net (i.e., gains minus losses) reward rate²⁹. The consequence is that average response latencies should be shorter when the average reward rate is higher. This prediction holds both when the amount of reward is a direct function of the number of responses (i.e., ratio schedules of reinforcement) and when rewards become available at specific time points (i.e., interval schedules of reinforcement).

Building directly on these principles, our $$ \overline{\mathrm{R}}\mathrm{L}$$ model specifies how agents adjust the latency of their responses to maximize the average net reward rate $$ \overline{R}$$ (see Methods and Supplementary Methods). Hence, the model provides a formal account of our hypothesis that behavior on social media conforms to basic principles of reward learning. Formally, the model conceptualizes social media use as a sequence of decisions regarding the latency between successive posts, τ_Post (Fig. 1a)²⁹, where the agent maximizes the reward rate by adaptively adjusting τ_Post after observing each accrued reward. Psychologically, τ_Post can be thought of as the accumulation of motivation towards a threshold for posting (in similarity to the boundary in evidence accumulation models of decision-making³³). The model policy, or threshold, which determines τ_Post,, is dynamically adjusted based on the net reward prediction error, δ, the difference between the experienced reward and the reference level. The reference level is determined both by the individual’s effort cost sensitivity (e.g., the subjective cost of taking pictures and uploading) and the subjective estimate of the average net reward rate $$ \overline{\mathrm{R}}$$^29,34, which determines the opportunity cost of slow responding (Fig. 1b, c). Both the effort cost and opportunity cost depend on the response latency, τ_Post. In other words, the optimal response latency balances these two costs to maximize the net reward δ (Fig. 1d). The subjective estimate of $$ \overline{\mathrm{R}}$$ is updated using the same reward prediction error, thereby reflecting the integration of prediction errors across time²⁹. In total, the model has three free parameters: learning rate, ɑ; initial policy, P; and effort cost sensitivity, C (see Methods). We verify with simulations that our $$ \overline{\mathrm{R}}\mathrm{L}$$ model accurately reproduces standard patterns of animal behavior in Skinner boxes (Supplementary Methods and Supplementary Fig. 1). This demonstrates the validity of the $$ \overline{\mathrm{R}}\mathrm{L}$$ model as an account of instrumental reward learning.

We simulated the model (~250,000 data points from 1000 simulated users, with random parameter values, see Supplementary Methods for details) to generate predictions for reward learning on social media. According to learning theory, τ_Post should be lower when the average reward rate is relatively higher. To verify this prediction in a simple manner, we rank-transformed and standardized $$ \overline{\mathrm{R}}$$ for each synthetic user and then dichotomized the variable at 0 to produce a qualitative “Low vs High $$ \overline{\mathrm{R}}$$” predictor (nearly identical results are observed with other definitions, see Supplementary Table 7). To facilitate subsequent comparison with empirical analyses, we summarized the simulated data using mixed-effects models. These analyses revealed a clear effect of low vs. high $$ \overline{\mathrm{R}}$$ on τ_Post (β = 0.18, SE = 0.007, t = 31, p < 0001), as expected. In other words, the model predicts (given the set of simulation parameters) that average response latencies should be ~18% longer when the average reward rate is low versus high (see Fig. 1e).

Our empirical analysis of the four social media platforms tested these model-based predictions with model estimation, statistical analyses, and generative model simulations. We optimized the parameters of the $$ \overline{\mathrm{R}}\mathrm{L}$$ model for each individual user and quantitatively compared the explanatory value of the $$ \overline{\mathrm{R}}\mathrm{L}$$ model to a null model without reward learning (see Supplementary Methods; model estimation and comparison procedure recovered the models with high probability, Supplementary Fig. 2). The null model assumes that posting on social media reflects a stable behavioral tendency (i.e., average response latency, one free parameter), which is not affected by reward. The model comparison provides a direct, quantitative test of reward learning as an explanation for social media use.

Study 1

We first modeled online behavior in the Instagram dataset of Study 1³⁰. Model comparison showed that the $$ \overline{\mathrm{R}}\mathrm{L}$$ model accounted better for the time distribution of responses (τ_Post) than the model without learning for ~70% of the users (mean individual-level Akaike Information Criterion weight (AIC_W) = 0.7, 99% CI [0.68, 0.81], one-sample t-test relative to equal AIC_W for the two models: t(2038) = 23.1, p < 0001, see Fig. 2a). The AIC_W expresses the relative likelihood of one model over another³⁵. Equivalently, Bayesian random effects model comparison³⁶ showed that the $$ \overline{\mathrm{R}}\mathrm{L}$$ model was more common than the model without learning (exceedance probability [xp] = 1), and classified ~70% of individuals as better explained by the $$ \overline{\mathrm{R}}\mathrm{L}$$ model.

Fig. 2

Behavior on Instagram is explained by reward learning (Study 1).

a Model comparison shows that the $$ \overline{\mathrm{R}}\mathrm{L}$$ model explained behavior on Instagram (N = 2,039 independent individuals) better than a model without learning. The AIC_W expresses the relative likelihood for each model, and are presented as means ± 99% CI. The horizontal line at 0.5 represents the chance level of no difference between models. The distribution of AIC_W is displayed in Supplementary Fig. 3. Source data are provided as a Source Data file. b The model-derived estimate of $$ \overline{\mathrm{R}}$$, the average reward rate, predicted the latency between posts (N = 2,039 independent individuals). As implicated by reward learning theory, the latency between posts was shorter with high compared to low $$ \overline{\mathrm{R}}$$. Points indicate the corresponding estimates from synthetic data, based on ten generative simulation runs of the $$ \overline{\mathrm{R}}\mathrm{L}$$ model (see text for details). The colored line denotes the average effect in the simulated data. Results are presented as means (fixed effects regression estimates) +/− 99% CI from mixed-effects regression. c, d Model fit for an example individual. c The posting history of an individual user over 673 days was well approximated by the $$ \overline{\mathrm{R}}\mathrm{L}$$ model. The model policy (or posting threshold) denotes the average response latency predicted by the model at a given time point. The faded purple lines show 100 simulations of τ_Post from the estimated model policy, which illustrate the expected degree of variability given that policy, and how the empirical τ_Post falls within this range. The yellow line indicates the model estimate of the net reward rate, $$ \overline{\mathrm{R}}$$. Note that a higher estimated $$ \overline{\mathrm{R}}$$ is associated with shorter response latencies (τ_Post). See Supplementary Fig. 4 for additional example individuals. Source data are provided as a Source Data file. d The distribution of τ_Post for the same individual. The faded purple line shows 100 simulations of τ_Post from the estimated model policy. Source data are provided as a Source Data file.

This conclusion was robust to the removal of individuals with especially short or long (outside the 20^th and 80^th deciles) average τ_Post, or with few (or many) posts (see Supplementary Note 3), which confirms that the fit of the $$ \overline{\mathrm{R}}\mathrm{L}$$ model was not driven by outliers. Similarly, splitting the dataset into four equally sized partitions showed that the $$ \overline{\mathrm{R}}\mathrm{L}$$ model was the most common in all four partitions (mean AIC_W: 0.68–0.73, t-test against equal AIC_W: t(508) = 9.63−13.9, ps < 0001), which indicates that our conclusion is robust to sample idiosyncrasies and dataset size. Interestingly, we found that individuals with more Instagram followers exhibited non-linearly diminishing subjective value (utility) of likes, or in other words, derived less subjective value from each like (see Supplementary Note 4). This suggests that individuals with many followers might habituate to likes.

According to our theoretical framework, responses should be faster when the subjective reward rate is higher. Similar to how we derived model predictions (c.f., Fig. 1d), we used the model-based estimate of $$ \overline{\mathrm{R}}$$ (at t−1) dichotomized into “Low vs High” to predict the empirical τ_Post (at t), using log-linear mixed models (see Methods; the same conclusions are reached using continuous measures and regression models with cluster-corrected standard errors, see Supplementary Note 8 & Supplementary Table 7). In support of the hypothesis that people learn to maximize social rewards, the latency between posts, τ_Post, was lower when $$ \overline{\mathrm{R}}$$ was relatively high (Instagram (N_Obs = 851,946, N_Users = 2,039): β = −0.18, SE = 0.003, t = −54.59, p < 0001, see Fig. 2b. Expressed in model comparison terms, the AIC_W for the regression model including $$ \overline{\mathrm{R}}$$ was 1). Thus, increasing the average subjective reward rate from low to high reduced average latency between posts by ~18%, corresponding to ~8 h. Based on analysis with a continuous $$ \overline{\mathrm{R}}$$ term, this corresponds to a reduction of 0.34% (~5 min) in average posting latencies for each 1% increase in the subjective reward rate. Providing additional support for the logic of the $$ \overline{\mathrm{R}}\mathrm{L}$$ model, the effect of $$ \overline{\mathrm{R}}$$ on posting latencies was stronger for individuals for whom the $$ \overline{\mathrm{R}}\mathrm{L}$$ model provided a better fit (interaction Low vs High $$ \overline{\mathrm{R}}$$ * AIC_W [centered at 0.5]: β = −0.04, SE = 0.008, t = −5.5, p < 0001). To ascertain the unique effects of our variables of interest, we adjusted for R(t−1) (the number of likes received at post t−1), the specific post number, and the weekday of the preceding post in these analyses. We illustrate the relationship between τ_Post, $$ \overline{\mathrm{R}}$$, and the model policy in Fig. 2c, d, which displays the posting behavior of an example individual over a period of two years. Together, these findings strongly indicate that online behavior conforms to principles of reward learning.

Study 1: Alternative models

To test the specificity of the $$ \overline{\mathrm{R}}\mathrm{L}$$ model, we compared it with a set of plausible alternative models (see Supplementary Note 6). Specifically, we examined models (i) without effort cost (C) or net reward rate ($$ \overline{R}$$) parameters, (ii) where the effort cost was fixed, or increased (rather than decreased) with post latencies, (iii) without an instrumental response policy, and (iv) based on foraging theory. Each of these provided worse accounts of the data than the $$ \overline{\mathrm{R}}\mathrm{L}$$ model (see Supplementary Note 6, and Supplementary Tables 3–6).

Study 1: model simulations

In addition to demonstrating a model’s superior fit to the data, strong support for a model depends on its ability to reproduce the effects of interest³⁷. To confirm the model fitting results, we therefore generatively simulated the $$ \overline{\mathrm{R}}\mathrm{L}$$ model (based on the median best fitting parameters, but independent of the empirical data) and used mixed-effects models to summarize the simulations³⁸. Notably, the simulation makes very limited assumptions of how likes were generated (i.e., as random draws from a Poisson distribution, with identical parameters for all individuals, see Supplementary Methods for additional details). Nonetheless, we found that the simple reward learning $$ \left(\overline{\mathrm{R}}\mathrm{L}\right)$$ model reproduced the observed difference in response latency between high and low $$ \overline{\mathrm{R}}$$ (Fig. 2b).

Study 2

To replicate and extend the results of Study 1, we collected data from three distinct social media sites (see Methods) which, in contrast to Instagram, focus on special interest topics (Men’s fashion, Women’s fashion, Gardening, respectively). Much activity on these social media sites is focused on textual exchange rather than images, but all three contain prolific “threads”—collections of posts focused on a specific topic—with predominantly image-based content (e.g., “What are you wearing today?”, “Post pictures of your garden”), with many thousands of posts each. We limited our analyses to posts with user-generated images from such threads (see Methods and Supplementary Methods), but verify in the Supplementary Note 5 that the results are qualitatively identical when including text-based posts.

We again tested the hypothesis that social media behavior reflects social reward learning by estimating the same $$ \overline{\mathrm{R}}\mathrm{L}$$ model as in Study 1. In all three datasets (190,721 data points from 2,127 individuals), regardless of the specific topic, model comparison favored the $$ \overline{\mathrm{R}}\mathrm{L}$$ model over the model without learning (pooled mean AIC_W = 0.77, 99% CI [0.76, 0.79], t-test against equal model likelihoods: t(2126) = 38.84, p < 0001, xp = 1, see Fig. 3a–c). As in Study 1, we performed several robustness checks to verify this conclusion (see Supplementary Note 3). These findings converge with and generalize those of Study 1, providing platform-independent evidence that reward learning theory can help explain social media behavior.

Fig. 3

Signatures of reward learning on three social media sites (Study 2).

a–c Model comparison shows that the $$ \overline{\mathrm{R}}\mathrm{L}$$ model explained behavior on the three social media sites (total N = 2,127 independent individuals. a N = 543, b N = 773, c N = 813) better than a model without learning. The AIC_W expresses the relative likelihood for each model, and are presented as means +/− 99% CI. The horizontal line at 0.5 represents the chance level of no difference between models. The exceedance probability for the $$ \overline{\mathrm{R}}\mathrm{L}$$ model was 1 in all three datasets. The distribution of AIC_W is displayed in Supplementary Fig. 3. Source data are provided as a Source Data file. d–f The model derived estimate of $$ \overline{\mathrm{R}}$$, the average reward rate, predicted the latency between posts on each social media platform (d N = 543, e N = 773, f N = 813 independent individuals). In line with reward learning theory, the latency between posts was shorter with high compared to low $$ \overline{\mathrm{R}}$$. The colored points indicate the corresponding estimates from simulated data, based on ten generative simulation runs of the $$ \overline{\mathrm{R}}\mathrm{L}$$ model (see text for details). The colored lines show the average effect in the simulated data. Results are presented as means (fixed effects regression estimates) ± 99% CI from mixed-effects regressions.

As in Study 1, we used the model-based estimate of $$ \overline{\mathrm{R}}$$ to predict the empirical τ_Post, using mixed effects regression models (adjusting for the same covariates as in Study 1). As expected, a higher $$ \overline{\mathrm{R}}$$ predicted faster responding in all three datasets (see Fig. 3d–f. Men’s fashion (N_Obs = 36,139, N_Users = 541): β = −0.08, SE = 0.016, t = −5.1, p < 0001. Women’s fashion (N_Obs = 36,434, N_Users = 773): β = −0.16, SE = 0.02, t = −7.1, p < 0001. Gardening: (N_Obs = 118,148, N_Users = 813): β = −0.18, SE = 0.02, t = −12.09, p < 0001). Thus, in these respective datasets, latencies between posts were 8%, 16%, and 18% shorter when the average reward rate was high rather than low. This corresponds, based on analysis with a continuous $$ \overline{\mathrm{R}}$$ term, to a reduction of 0.18%, 0.41%, and 0.38% in average posting latencies for each 1% increase in the subjective reward rate. As in Study 1, the estimated effect of $$ \overline{\mathrm{R}}$$ on posting latencies was stronger for individuals for whom the $$ \overline{\mathrm{R}}\mathrm{L}$$ model provided a better fit (see Supplementary Note 7). Thus, regardless of platform topic, social media behavior conforms to model-based principles of reward maximization through RL.

We again conducted generative model simulations of the $$ \overline{\mathrm{R}}\mathrm{L}$$ model, based on the median estimated parameters, to test whether the model reproduced the effect of average reward rates on posting latencies. Corroborating Study 1, these simulations showed that the $$ \overline{\mathrm{R}}\mathrm{L}$$ model accurately reproduced the effect of $$ \overline{\mathrm{R}}$$ observed in the data (Fig. 3d–f) (but note that the absolute level of τ_Post is slightly overestimated in the Gardening dataset). Together with Study 1, these results confirm that basic reward learning theory provides a powerful tool for predicting and explaining the dynamics of social media use, independent of topic.

Study 2: Social comparison in social reward learning

The preceding analyses showed that people dynamically adjust their social media behavior in response to their own social rewards, as predicted by reward learning theory—a theory originally developed to test the effects of nonsocial rewards (e.g., food) in solitary contexts. However, given the intrinsically social context of social media use, we speculated that reward learning online could be modulated by the rewards others receive. In the Supplementary Note 9 we provide preliminary support for the hypothesis that reward learning, at least on the social media platforms we analyzed in Study 2, may be modulated by social comparison.

Individual differences in reward learning on social media

Having established that reward learning can help explain social media behavior, we next asked whether individuals differ in the ways they learn from rewards on social media. To address this issue, we used the parameter estimates of the basic $$ \overline{\mathrm{R}}\mathrm{L}$$ model as a compact but rich description of the mechanisms underlying behavior—a kind of computational phenotype (which is behavioral in nature and makes no direct reference to the underlying genotype)³⁹. Individual differences in these parameters can thus be viewed as differences in computational mechanisms³⁹ that are interpretable across domains. For example, individual differences in learning rates have previously been linked to both genetic⁴⁰ and developmental differences⁴¹ between individuals, while individual differences in effort cost sensitivity have been related to the dopaminergic system⁴².

More specifically, we used the three parameters of the original $$ \overline{\mathrm{R}}\mathrm{L}$$ model estimated for each individual from Study 1 & 2 (total N_Users = 4,168), as input for k-means clustering, an unsupervised, data-driven method for finding sub-groups in multidimensional data. Quantitative assessment, using multiple standard criteria, showed that four clusters gave the best sub-group solution (see Fig. 4a and Methods). In Supplementary Note 10, we report additional robustness analyses, which show that this cluster solution was stable. These clusters comprised between 41% (1739 individuals) to 7% (299 individuals) of the total dataset. Importantly, although the four datasets varied in mean τ_Post (as reflected in the P parameter), the cluster assignment was not strongly explained by dataset (Cramér’s V = 0.3; Cramér’s V is a measure of the association between two nominal variables, where 1 denotes perfect association). This indicates that clusters captured individual differences in computational learning mechanisms, rather than idiosyncrasies of social media sites.

Fig. 4

Computational phenotypes in reward learning on social media.

a Cluster analysis (n = 4168 independent individuals from Study 1 & 2) of the estimated $$ \overline{\mathrm{R}}\mathrm{L}$$ model parameters indicated that there were four distinct individual difference profiles in reward learning on social media. For illustration, the cluster assignments are plotted on the two first principal components (PC). Source data are provided as a Source Data file. b Profile of median $$ \overline{\mathrm{R}}\mathrm{L}$$ model parameter values (standardized) for each cluster; ɑ = learning rate, P = initial response policy, C = effort cost sensitivity. Source data are provided as a Source Data file.

Figure 4b illustrates the four putative computational phenotypes. For example, individuals in cluster 1 are characterized by a relatively low learning rate (ɑ). Such individuals are especially insensitive to social rewards in their behavior (and naturally, the $$ \overline{\mathrm{R}}\mathrm{L}$$ learning model provided the worst fit to these individuals relative to the model without learning: mean AIC_W = 0.11, vs AIC_W ~ 0.77 in the other three clusters). By comparison, individuals in cluster 2 are characterized by low effort cost and average learning rate, whereas cluster 4 exhibits the opposite relationship between learning rate and effort cost (and cluster 3 an intermediate profile). Individuals in both clusters 2 and 4 therefore readily post in response to social rewards, although the underlying mechanisms differ. In summary, the computational phenotyping indicates that there are important individual differences in the mechanisms underlying social media behavior.

Study 3

Finally, to provide direct evidence that the social reward rate affects posting latencies, we conducted an online experiment in which we experimentally manipulated social rewards and observed posting response latencies. The experiment was designed to capture key aspects of social media, such as Instagram. Participants (n = 176) could post “memes”—typically, an amusing image paired with a phrase that is popular on real social media—as often and whenever they wanted during a 25 min. online session (total number of posts = 2,206, see Supplementary Methods for details). Participants received feedback on their posts in the form of likes (0–19) from other ostensible online participants (“users”, see Supplementary Fig. 5 for an overview of the experiment). Participants themselves could also indicate “likes” for memes posted by other users. To test whether a higher social reward rate causes shorter response latencies in posting, we increased or decreased the average number of likes participants received between the first and second halves of the session (low reward: 0–9 likes/post, high reward: 10–19 likes/post, drawn from a uniform distribution, with direction of change counter-balanced across subjects). As expected, mixed effects regression (see Methods) showed that the average post latency was longer when the social reward rate was lower (0–9 likes/post) relative to higher (10–19 likes/post): β = 0.109, SE = 0.044, z = 2.47, p = .013 (see Fig. 5), corresponding to a 10.9% difference. Notably, participants who reported more followers on Instagram exhibited weaker effects of likes on their behavior (see Supplementary Note 11). This finding parallel how individuals with more Instagram followers in Study 1 exhibited more diminished marginal utility of likes (see “Study 1” above and Supplementary Note 4). These results further support the validity of our experiment in assessing the psychology of real-world social media use. We report additional analyses and robustness checks in Supplementary Note 11.

Fig. 5

Experimental manipulation of social reward rates (Study 3).

The estimated effect of social reward rate condition on posting latencies in the online experiment (n = 176 independent individuals, see Supplementary Fig. 5 for a design overview). Results are presented as means (fixed effects regression estimates) ±1 SE of the regression estimates from from mixed-effects regression.

To directly relate the experimental results of Study 3 to our model-based analyses of online behavior in Studies 1 and 2, we used the $$ \overline{\mathrm{R}}\mathrm{L}$$ model to generate subjective $$ \overline{\mathrm{R}}$$ time series for the subset of participants with a sufficient number of responses (see Supplementary Note 11 for details), and used these (instead of reward condition) to predict response latencies. In accordance with the model fits to the real social media data (Study 1-2), the average response latency was longer when the subjective reward rate was low, relative to high (mixed effects regression, n = 156: β = 0.28, SE = 0.045, z = 6.24, p < 0001). These experimental results demonstrate that social rewards causally influence response latencies, in support of our conclusion that social reward rates shape real social media behavior.

Social media datasets

Study 1 was based on data from a previously published study (see³⁰ for further information), in which data collection was based on a random sample of individuals who partook in a specific photography contest on Instagram in 2014. We find no evidence that contest participation was related to posting behavior (see Supplementary Note 12). The dataset was fully anonymized. To allow for analyses of learning, we excluded all individuals with less than 10 posts⁵⁰. For Study 1, the final dataset consisted of 851,946 posts from 2,039 individuals. For Study 2, we obtained three datasets from three different topic-focused (Men’s fashion: styleforum.net, Women’s fashion: forum.purseblog.com, Gardening: garden.org, see Supplementary Methods for details) social media discussion forums using web scraping techniques⁵¹ on publicly accessible data. The datasets were fully anonymized, and only included the time stamps and likes associated with posts in prolific threads focused on user-generated images (e.g., pictures of one’s clothing, see Supplementary Methods). For our analyses, we focused on posts with user-generated images, and, in analogy to Study 1, removed all users with fewer than ten image posts. The Study 2 dataset consisted of 190,721 posts from 2,127 individuals (Men’s fashion: N = 543, Women’s fashion: N = 773, Gardening: N = 813). This research was conducted in compliance with the US Office for Human Research Protections regulations (45 CFR 46.101 (b)).

Experiment

The experiment was conducted on Amazon Mechanical Turk, and approved by the ethical review board of the University of Amsterdam, The Netherlands. See the Supplementary Methods for additional information.

Description of the $$ \overline{\mathrm{R}}\mathrm{L}$$ model

The $$ \overline{\mathrm{R}}\mathrm{L}$$ model is a policy gradient version of R-learning⁵². Rather than storing action values for options and using these as a basis for decision-making, the $$ \overline{\mathrm{R}}\mathrm{L}$$ model directly updates a parametrized response policy (the mean parameter of an exponential distribution). This is beneficial for learning problems with continuous action spaces (e.g., the latency between responses)⁵³. In close similarity to standard RL models for discrete action spaces, the $$ \overline{\mathrm{R}}\mathrm{L}$$ model incrementally learns to adjust its actions (i.e., τ_Post) from prediction errors. In contrast to standard RL approaches in psychology, the prediction errors are used to directly adjust the response policy to maximize the undiscounted net rate of reward rather than to update action values.

For each post, the $$ \overline{\mathrm{R}}\mathrm{L}$$ model selects τ_Post as a draw from an exponential distribution, where the mean (i.e., the response policy or threshold) is dynamic:

The initial response policy (i.e., for t = 1) was estimated as a free parameter (0 ≤ P ≤ ∞). The subtraction term in Eq. (1) implements the momentous effect of the average reward rate (e.g., changes in motivational state), $$ {\overline{R}}^t$$, on the response rate, which is independent of learning²⁹. This term, which can be thought of as “Pavlovian” (since it is independent of instrumental behavior), slightly amplifies the effect of the average reward rate on τ_Post. Model comparison and simulation showed that this term has a small but significant contribution to model fit, but does not affect qualitative predictions.

The response policy, which determines τ_Post, is adjusted based on the prediction error, δ, the difference between the experienced reward (R^t) and the reference level:

The reference level explicitly takes into account both the effort cost of fast responding and the opportunity cost of slow responding (Fig. 1b, c), which is determined by the subjective estimate of the average net reward $$ \overline{R}$$^29,34. In our application of the model, C (0 < C ≤ ∞) is a user-specific parameter that determines the subjective effort cost of posting (e.g., taking pictures, uploading). This parameter penalizes posting in quick succession (e.g., posting three times in one day is more costly than posting three times in three days; we evaluate different effort cost formulations in the Supplementary Note 5). If actions have no intrinsic cost, the optimal policy would be to respond as quickly as possible^29,54. The last term in Eq. (2) characterizes the opportunity cost of slow responding, which increases with $$ \overline{R}$$ (Fig. 1c).

The $$ \overline{\mathrm{R}}\mathrm{L}$$ model seeks to maximize reward by dynamically updating the response policy by the “gradient”, or slope, of reward. In short, the gradient indicates the direction and distance to the hypothetical maximum net reward that could have been accrued at time t. To compute the gradient, the model tracks the sequential difference between responses (i.e., draws from the policy, which can be thought of as exploration), and combines this quantity with the net reward prediction error (Eq. 2).

Thereby, the model can learn the instrumental value of slower vs. faster responses. For example, if $$ \mathrm{\Delta }{\tau }_{Pos{t}^t}$$ is positive, which would be the case if the response latency at t was longer than at t−1, but $$ {\delta }^t$$ negative, which indicates that the net reward was lower than on average, the gradient (Eq. 4) is negative. This results in a reduced response policy (the degree of adjustment depends on the magnitude of both $$ \mathrm{\Delta }{\tau }_{Pos{t}^t}$$ and $$ {\delta }^t$$), and in turn shorter future response latencies. In contrast, if both $$ \mathrm{\Delta }{\tau }_{Pos{t}^t}$$ and $$ {\delta }^t$$ are either positive or negative, the response policy will increase, which leads to longer response latencies. The average reward rate is updated using the same reward prediction error as the policy, as it directly reflects net reward value⁵²:

However, regardless of whether slower or faster responses are rewarded, an increase in the average reward rate (Eq. 5) results in a higher opportunity cost of time (Eq. 2), which in turn results in shorter response latencies (Eq. 2, and Fig. 1d).

If either $$ \mathrm{\Delta }{\tau }_{Pos{t}^t}$$ or $$ {\delta }^t$$ = 0, the model can theoretically be trapped in local reward minima. However, the stochastic policy (Eq. 1) ensures that $$ \mathrm{\Delta }{\tau }_{Pos{t}^t}$$ is different from 0 with exceedingly high probability, which promotes continuous search by sacrificing convergence if the step size parameter is fixed (as the model policy will continuously change). For simplicity, the same step size parameter ɑ (0 < ɑ ≤ 1) was used for all update terms (Eq. 1 & 4–5). Inclusion of separate step size parameters for the different update equations did not reliably improve (parameter number penalized) model fit. Additional model information and estimation methods are detailed in the Supplementary Methods.

Statistical analysis

All model estimation, simulations, and statistical analyses were conducted using R. All reported p-values are two-tailed. Granger causality analysis was applied to first differenced data using the plm package for panel-data analysis (see Supplementary Methods for details)⁵⁵. Mixed effects modeling was conducted with the lme4⁵⁶ and glmmTMB⁵⁷ packages. All log-linear mixed effects models included a random intercept for each user. In the statistical analyses, the dependent variable τ_Post was log transformed (as the time between events follows an exponential distribution) to improve linearity. All predictors were standardized within individual (i.e., centering within cluster) to produce individual-level estimates⁵⁸. Degrees of freedom, test statistics, and p-values were derived from Satterthwaite approximations in the lmerTest package⁵⁹. The key statistical analyses were in addition repeated using log-linear regression models with cluster-corrected standard errors to ensure robustness (see Supplementary Table 7). Prior to k-means clustering, the $$ \overline{\mathrm{R}}\mathrm{L}$$ model parameter estimates were log-transformed (to improve linearity) and standardized. The optimal number of clusters was determined using the NbClst package⁶⁰.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.