Effects of Early Adversity and War Trauma on Learning Under Uncertainty

Abstract

In uncertain situations, individuals rely on prior experiences of successes and failures to guide future decisions. Research has shown that children exposed to early adversity, such as abuse, can exhibit atypical behaviours in probabilistic learning tasks compared to peers without such experiences, which may have long-term behavioural consequences. Building on these findings, our study investigates whether children exposed to war-related trauma and forced displacement show similar alterations in decision-making under uncertainty. We conducted a series of experiments involving tasks that required learning and decision-making under uncertainty (e.g., multi-armed bandits and foraging) in Amman, Jordan, comparing Syrian refugee children (ages 7-12) with age-matched Jordanian non-refugee peers. Although our first experiment suggested less exploratory behaviour in refugee children, results from subsequent tasks revealed a pattern better explained by heightened sensitivity to rewards. This interpretation suggests that refugee children's decision-making is driven by a stronger response to rewards across different contexts. Such heightened reward sensitivity may influence how these children approach problem-solving and decision-making, potentially leading to detrimental outcomes in environments that benefit from greater exploration and the maintenance of a stable strategy. SUMMARY: Syrian refugee children (ages 7-12) showed heightened reward sensitivity compared to age-matched Jordanian peers across decision-making tasks. Reward sensitivity influenced children's choices under uncertainty, particularly following successful risky outcomes. Findings highlight how early adversity may shape decision-making strategies relevant to resilience and long-term cognitive development.

FIGURE 1

Valence‐dependent learning. (A) Monster learning task with distinct gain and loss contexts. Red symbols above the doors (left) indicated the gain context where children could find the ‘good monster’ (the yellow smiley face), who would give them one golden coin. Conversely, blue symbols indicated a loss context and the possibility of losing one golden coin if they chose the door with the ‘bad monster’ (the blue sad face). To make the outcome of the trial evident, whenever the child won or lost a coin, it was moved in a randomly generated smooth trajectory (generated as a Bézier curve) that connected the bottom of the screen (where all the gained coins would line up) and the monster. (B) Choice probabilities as a function of trial number, split by group (refugees and non‐refugees) and trial context. Data points indicate the average choice frequency for each trial (error bars represent binomial standard errors); lines and shaded bands are the predictions of the reinforcement learning model with standard errors across children. (C) Estimated parameters of the reinforcement learning model, dots indicate parameters of individual children, whereas the columns with error bars indicate posterior means of the fixed‐effects with the associated standard errors (i.e., the standard deviation of their marginal posterior distribution). The estimated values indicate that whilst refugee and non‐refugee children did not differ in their learning rates, refugee children exhibited less variable and more exploitative choices. This can also be seen in panel D, which shows the inferred choice functions (giving choice probability as a function of the difference in value between the two options). The shaded bands are the standard error of the predicted probability and the dots above and below the functions represent the value difference in each trial (as estimated by the reinforcement learning model).

FIGURE 2

Virtual foraging task (Study 2). (A) In each trial, children decide whether to continue harvesting apples from the current tree by clicking on it or to switch to a new tree by clicking on a red arrow. Each decision (stay or move) advances the sun downward in the background; when the sun sets behind the mountain, the game ends. (B) The lower panels show each child's total score (vertical axis) plotted against their exit threshold (horizontal axis) — the minimum number of apples at which they switched trees. Small empty dots represent individual children; filled dots represent the group mean. The dark grey line indicates the expected total score for each exit threshold if it were followed consistently throughout the game, and the light grey band shows the 95% prediction interval. The dashed vertical line is the optimal exit threshold (7 apples). Both Jordanian (left) and Syrian (right) groups displayed average thresholds near the optimal value. The upper panels show the distribution of exit thresholds.

FIGURE 3

Weighting certain and uncertain rewards (Study 3). The task was designed as a multi‐armed bandit problem, in which children had to decide on every trial which door to open to maximise their rewards. (A) Example of display with the central ‘sure’ (S) door open. The monster (shown in the open door) rewards children with two coins; however, the monster has a relatively low probability of appearance (see Methods for details), and thus, if children choose the R doors, they risk not obtaining any coins. (B) Data (left) and model predictions (right). Each plot illustrates the probability of choosing the S doors as a function of the reward obtained in the previous trial. Outcomes of zero or two coins result from choosing an R door, and outcomes of one coin result from choosing the S door. In all panels, horizontal dashed lines indicate the proportion of choices that would be obtained if children were randomly choosing each door with the same frequency. Note that the data shows a tendency to alternate R and S choices: after choosing an S door, children are more likely to choose one of the R doors, and vice versa. Additionally, after choosing an R door, the next choice is influenced by the reward obtained only for refugee children (green line and dots): differently from non‐refugees, refugee children are more likely to persist with an R door after winning a reward of two coins. Instead, after choosing an R door and not winning anything, both groups are more likely to choose the S door than by chance. This pattern is fully accounted for only by the best‐fitting model (see Supporting Material for details).