Exposure to and recall of violence reduce short-term memory and cognitive control

Experiment I: Exposure to Urban Violence.

We ran our first experiment among residents of the Colombian capital, Bogotá. With a yearly average of 21.6 homicides per 100,000 inhabitants over the period 2008–2011, Bogotá is one of the most violent capitals among developing countries. There exists, nonetheless, large variety in homicide rates among its districts. These rates can range from levels comparable to those of the most violent areas in the world to levels that are eight times lower. We measured individual ETV through the district-level homicide rate in the subject’s district of residence (SI Appendix, section 1.1.1, Fig. S1, and Table S1). For the ensuing analyses, we divided our sample into two groups, with one comprising subjects living in districts with relatively very high ETV—namely, those whose ETV lies above the 75th percentile. Our sample comprised 308 residents of Bogotá between 18 and 24 y of age, who came from all but two of Bogotá’s 19 districts (see SI Appendix, Tables S1 and S2 and section 1.3.1 for descriptive statistics). We focused on this age group because we aimed to minimize the risk of self-selection of subjects into districts, as people in this age group can be largely expected to live with their families.

Our dependent variable in this experiment is subjects’ responses to a test of short-term memory (37). The task consisted of memorizing a set of colored abstract geometric figures, to be later identified in a sequence of 10 figures presented in succession. The task was monetarily incentivized. We analyze the accuracy rate (i.e., the fraction of correct answers relative to the total number of questions). Crucial to our design is the VR manipulation. At the beginning of the session, a randomly determined subset of the subjects was asked to recall an event that occurred in the last year that caused anxiety or fear, explicitly hinting at violence as a possible cause of such an emotional state (9, 25). We refer to this group as VR. The remaining subjects were asked to either recall a joyful experience or a generic experience devoid of emotional content. Because results never differed when neutral or joyful experiences were recalled (Mann–Whitney–Wilcoxon test, P = 0.95, n = 200; all tests are two-tailed), we merged the two conditions in the analysis, referring to them as neutral/joy recall (abbreviated NR). In all experiments, the hypothesis that the treatment was exogenous was not rejected (SI Appendix, Tables S3–S5 and S39). For simplicity, we present results in the form of a 2 × 2 factorial design, in which the ETV component—high ETV (abbreviated HV for high violence) and low ETV (abbreviated LV for low violence)—is crossed with the recall component—VR or NR.

As shown in Fig. 1A, accuracy on the memory task was on average lowest in the HV group recalling violence (A¯HVVR=60.3% where A¯ji denotes average accuracy, the upper index i={VR,NR} denotes the recall condition, and the lower index j={HV,LV} denotes ETV level), whereas it was approximately constant in the other three groups (A¯HVNR=A¯LVNR=66.0% for both HV and LV groups in NR and, A¯LVVR=68% for the LV group in VR). According to an econometric Tobit model that controls for subjects’ demographic characteristics (Table 1 and SI Appendix, Tables S6 and S8 and Fig. S2), accuracy by the HV group recalling violence was statistically significantly lower than accuracy by any of the other three groups—albeit weakly in one case (P = 0.026 with respect to the LV group recalling violence; P = 0.075 with respect to the HV group not recalling violence; P = 0.047 with respect the LV group not recalling violence; n = 281 for each test). Conversely, accuracy by the HV group not recalling violence was not statistically different from any other LV group (SI Appendix, Table S9). These results show the combined effect of ETV and VR. People who had been highly exposed to violence—and recalled violence—performed significantly worse than any other group. Conversely, people who had been highly exposed to violence in real life, but did not recall violence, performed no differently from either of the groups exposed to LV.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 1.

Impact of ETV on short-term memory. Accuracy levels in the memory task are broken down by recall condition (VR vs. NR) and levels of ETV—high or low ETV. High ETV identifies subjects lying above the distribution’s 75th-percentile split of the district-level homicide rate in the urban sample (A) and of a self-reported ETV measure for the war-displaced sample (B). Error bars reflect ±95% confidence intervals. Top horizontal bars show the statistical significance of tests on the null hypothesis that accuracy is the same in pairs of groups, as per Tobit regression reported in SI Appendix, Table S8 (A) and SI Appendix, Table S14 (B). ^†P < 0.1; *P < 0.05; **P < 0.01; ***P < 0.001.

District-level homicide rates, our measure of ETV, may be confounded with other variables operating at the district level, such as poverty, education, and quality of public services. In SI Appendix, Tables S10 and S11, we show that when district-level measures of these variables replaced ETV as a possible causal factor in our econometric analysis, none of them yielded significant effects on accuracy. We also replicated this analysis using a self-reported measure of ETV, constructed on three items derived from a postexperiment questionnaire. Such items inquired as to whether the subject had been a victim of violence in the recent past. All subjects had some minimal degree of direct ETV. Subjects belonging to the HV group who recalled violence were, again, the worst performers among the four groups formed using this self-reported ETV measure, whereas no significant difference emerged between the HV group not recalling violence and the two LV groups (SI Appendix, Tables S12 and S13 and Fig. S3). Finally, we found no interaction effects between ETV and either joy or neutral recall (SI Appendix, Table S10). We believe that this evidence is concordant in suggesting that the observed effects were driven by ETV rather than other district-level variables.

Experiment II: Exposure to Warfare and Forced Displacement 14 y Earlier.

Colombia is one of the countries with the highest percentage of war-related internally displaced individuals (38). Between 4.9 million and 5.5 million people are believed to be internally displaced. All subjects in experiment II had been displaced in the year 2000 from their homes by attacks and death threats carried out by irregular militia groups. Most people relocated to a settlement ∼10 km from their original residence (SI Appendix, section 1.1.2). All of our subjects were officially recognized as victims of the conflict by the Colombian government, upon the conduction of an official inquiry and the provision of factual evidence (see SI Appendix, section 1.1.2 and Table S2 for descriptive statistics). As in our urban sample, a randomly determined fraction of subjects was asked to recall episodes of violence; others were asked to recall joyful or neutral experiences. We removed the constraint that the experience of violence had occurred in the last year. The wide majority (87%) of subjects involved in the VR condition talked about their experience of relocation that occurred 14 y earlier.

In this rural sample, we could not rely on an objective measure of violence as in Bogotá. We then constructed an index of ETV based on subjects’ self-reported exposure, adapted from previous studies (8). The questionnaire from which this index was derived was administered after the experimental tasks and consisted of 20 items inquiring about the number of times a subject, or a member of his or her family, was the object of violence, such as ambushes, physical or armed attack, kidnapping, forced labor, torture, or robbery, or whether they had to pay a ransom (see SI Appendix, sections 1.3.2 and 1.6.2 for questionnaire and details on the index). The prior recall of experiences of violence for subjects in the VR group did not cause overreporting of episodes of violence in the postexperiment questionnaire, in comparison with subjects in the NR group (P = 0.94, Mann–Whitney–Wilcoxon test, n = 221). Using this measure, we divided the sample in two groups: the HV group that includes the subjects whose ETV score lies above the 75th percentile and the LV group including all others.

We run a version of the memory task adapted from the one conducted in Bogotá (SI Appendix, section 1.6.2). For comparability, we restricted our analysis to the same age group used in Bogotá (i.e., subjects between 18 and 24 y of age; n = 30). Accuracy followed the same pattern observed in the urban sample (Fig. 1B and SI Appendix, Fig. S2). Mean accuracy was again lowest in the HV group recalling violence (A¯HVVR= 32.5%). Accuracy in this group was strongly significantly less than accuracy in the LV group recalling violence (A¯LVVR= 62.7%; P < 0.002, n = 30), in the HV group not recalling violence (A¯HVNR= 65.8%; P = 0.002, n = 30), and in the LV group not recalling violence (A¯LVNR= 62.7%; P = 0.005, n = 30). On the contrary, the HV group’s accuracy when not recalling violence was no different from the LV groups’ accuracy (Table 1 and SI Appendix, Tables S7, S14, and S15). The variation in accuracy appeared much larger in the war-displaced sample than in the urban sample. Presumably, this result may be due to the fact that ETV exerted much larger effects in case of exposure to traumatic and life-threating experience than urban violence. Nonetheless, we cannot rule out that different educational levels and reduced sample size in the war-displaced sample may have also affected the magnitude of the effect.

In this fieldwork, we also ran two additional tasks to measure two aspects of cognitive ability. We measured inhibitory control through the Stroop task (13, 39). This task requires selective attention (focusing on what is relevant and trying to ignore what is not) and vigilantly keeping in mind the rule to successfully execute the task. We used a numeric version of this task (26, 39), suited for a sample with high illiteracy rates. Subjects were given up to 45 numerical sequences. In each, a digit from 1 to 4 appeared one to four times, and subjects were asked to state the number of times this digit appeared. The impulsive answer is the digit that appears in the sequence. The correct answer—which supposedly requires the reflective system to overrule the automatic system—is nevertheless the number of times that such a digit appears. (For instance, the impulsive answer to the sequence “1 1 1” is 1, whereas the correct answer is 3.)

We report results from the sample including all age groups (n = 141). Accuracy levels across the groups followed the same pattern observed in the memory tasks. Average accuracy was lowest in the HV group recalling violence (A¯HVVR= 57.53%) (Fig. 2A and SI Appendix, Fig. S4). According to econometric analysis (Table 1 and SI Appendix, Table S16), accuracy in this group was statistically significantly lower, albeit weakly in one case, than accuracy by the LV group recalling violence (A¯LVVR=77.39%, P = 0.094, n = 141), the HV group not recalling violence (A¯HVNR=88.3%, P = 0.005, n = 141), and the LV group not recalling violence (A¯LVNR=80.9%, P = 0.020, n = 141). Conversely, there was no significant difference between the HV group not recalling violence and the LV groups (SI Appendix, Table S17).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 2.

Impact of ETV on cognitive control. Accuracy levels in the Stroop task and Raven sequential matrices task are broken down by recall condition (VR vs. NR) and levels of ETV—high or low ETV. High ETV identifies subjects lying above the 75th-percentile split of the distribution of the self-reported ETV measure. Error bars reflect ±95% confidence intervals. Top horizontal bars show the statistical significance of tests on the null hypothesis that accuracy is the same in pairs of groups, as per Tobit regression reported in SI Appendix, Table S16 (A) and SI Appendix, Table S18 (B). ^†P < 0.1; *P < 0.05; **P < 0.01.

We also ran a simplified version of Raven’s progressive matrices. Raven’s task is used to measure “fluid intelligence” [i.e., the capacity to reason, solve problems in novel situations independent of acquired knowledge, and see patterns or abstract relations among items (40)]. Fluid intelligence includes both inductive and deductive logical reasoning and is strongly correlated with the reasoning and problem-solving subcomponents of cognitive control. It can be considered a higher-level form of cognitive control (13). We found an overall pattern that mirrored the one found for the previous measures of cognitive abilities. Accuracy was lowest in the HV group recalling violence (A¯HVVR= 10%; Fig. 2B and SI Appendix, Fig. S4). Accuracy in this group was less than both accuracy in the HV group not recalling violence (A¯HVNR= 19%; P = 0.26, n = 67) and accuracy in the LV group recalling violence (A¯LVVR= 14.83%, P = 0.39, n = 67). However, in this case, the econometric analysis (Table 1 and SI Appendix, Tables S18 and S19) failed to find statistical significance in these patterns. This result may be partly due to the fact that accuracy levels were significantly lower for the Raven task than for other tasks, which may have reduced the power of statistical tests.

Experiments I and II: Overall Assessment.

We fitted a Tobit model on the pooled sample of observations from the three measures of accuracy reviewed above. Overall, accuracy in the HV group recalling violence was statistically significantly lower than in any other group (SI Appendix, Fig. S5; in particular, P = 0.002 with respect to the LV group recalling violence; P < 0.001 with respect to the HV group not recalling violence; and P = 0.001 with respect to the LV group not recalling violence; n = 516 for each test). Conversely, accuracy in the HV group not recalling violence was statistically no different from accuracy in the LV group either recalling or not recalling violence (Table 1 and SI Appendix, Tables S20 and S21). Shifting from the LV to the HV group in the VR condition reduced overall accuracy on our memory and cognitive tests by 38% of a SD. When not recalling violence, this same shift led to an increase of 6% of a SD. Moreover, recalling violence induced a decrease of 47.2% of a SD in accuracy in the HV group. It is also worth noting that the differential effect of ETV between the VR group and the other group was positive and significant in the pooled sample (Table 1; P = 0.004, n = 516), as well as for each individual measure, except the Raven task (Table 1 and SI Appendix, Fig. S6). The overall effect of ETV on accuracy, without conditioning on the recall group, was negative and at the border of statistical significance in the pooled sample (Table 1; P = 0.090; n = 516), whereas it was never significant for individual measures (Table 1). Likewise, the overall effects of VR, unconditional on the ETV group, was weak in the pooled sample (P = 0.073; n = 540) and not significant for the individual measures, except for the Stroop task (P = 0.013, n = 141; Table 1).

A scatterplot of our data shows that ETV had substantially similar effects in the urban sample (r = −0.29, with r being the coefficient of the interpolation line using a continuous measure for ETV) and the war-displaced sample (r = −0.21), when violence was recalled (Fig. 3A). In the group recalling neutral or joyful experiences, the coefficient was close to zero in the urban environment (r = 0.02), whereas it appeared higher (r = 0.18) in the war-displaced sample (Fig. 3B). Removing an outlier from the latter sample (SI Appendix, Fig. S7) reduced the gap between urban and war-displaced samples in the neutral/joy condition.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 3.

Intersubject variability of cognitive control measurements (experiments I and II). The scatterplot of ETV (measured on a continuous scale) and accuracy on the three cognitive tests is reported. Linear predictions for experiment I (urban sample) and experiment II (rural environment, war-displaced sample) are also plotted, and their slopes are reported. A and B report observations in the VR and NR groups, respectively.

We replicated the analyses above, defining the HV group as subjects lying above the median split of the respective sample, rather than as those lying above the 75% percentile (SI Appendix, Tables S22–S33). We also analyzed ETV measured on a continuous scale (SI Appendix, Tables S34–S37). Although the qualitative nature of the results remained generally unchanged, the statistical significance of the tests tends to be generally lower. These analyses suggest that the effects of ETV on accuracy were mainly driven by people highly exposed to violence (i.e., those lying in the upper quartile of the distribution of ETV).

Experiment III: Evaluating the Impact of Emotions and Recall.

The above results may be due to the recall of experiences of violence, to the emotions associated with the recall, or to a combination of the two factors. The lack of suitable equipment in the fieldwork in Colombia prevented us from disentangling these two factors. For this reason, we ran a replication of the study in controlled laboratory conditions in Germany. Subjects were randomly assigned to four conditions. In the baseline condition, subjects were asked to recall joyful experiences, similarly to the baseline condition in experiments I and II. In the recalled fear (RF) condition, subjects were asked to recall experiences of violence, such as the November 2015 Paris attacks. Personal ETV is extremely rare in Germany; thus, we chose to have people recalling a violent event that created widespread apprehension among Europeans. In the current fear (CF) condition, we experimentally induced emotions of fear and anxiety. This induction was achieved by exposing subjects to the prospect of receiving minor electric shocks to their fingers in the course of the experiment (41, 42). Such electric shocks were calibrated to pose some physical discomfort, but not a painful experience. The anticipation of unpleasant electric shocks leads to an increase in emotional arousal during the anticipation phase (41, 43). In CF, the emotions of fear and anxiety were elicited without any recall of experiences of violence. In a fourth condition (RF&CF), subjects were administered both RF and CF stimuli. Our experiment was designed to compare the impact on cognitive abilities of an emotional state of fear or anxiety—when CF was administered—or by the recall of experiences of violence—when RF was administered. We could thus study the impact of emotions of fear and anxiety in isolation from the recall of violence experiences and compare this effect with that of recall. The full factorial design also allowed us to study a possible interaction between such two factors. Our dependent variable was subjects’ absolute accuracy level on a computerized version of the Stroop task administered in experiment II. Our sample comprised students of a German university (SI Appendix, Table S38).

Compared with the baseline, accuracy dropped by 5.34% and 6.45%, respectively, when either CF or RF were administered, and by 10.68% when both CF and RF were administered (SI Appendix, Table S40 and Fig. S8). The joint impact of the two stimuli was thus approximately equal to the sum of each individual stimulus. An econometric model that controlled for demographic characteristics, college achievement, degree, and fieldwork wave confirmed the statistical significance of such effects (SI Appendix, Tables S41 and S42). Accuracy in both CF and RF was statistically significantly lower than the baseline (P = 0.024 for CF; P = 0.010 for RF; n = 164) and was similar in size (t test on the coefficient different being equal to zero: β = 0.83, P = 0.83, n = 164). The interaction between the two effects was not significant (β = 0.46, P = 0.93, n = 164), which confirms that the combined effect of the two stimuli was approximately the same as the sum of the two individual effects. This evidence suggests that the two effects appeared to be independent.

Subjects’ physiological reactions to the stimuli were examined by recording skin conductance responses during the experiment. Raw electrodermal signal was separated via continuous decomposition analysis (44) into a phasic stimulus-driven part and a tonic baseline level. The tonic baseline level is commonly interpreted as an indicator of general arousal, whereas the phasic component captures spikes and deviations from such a baseline level (44, 45). According to our econometric analysis (SI Appendix, Table S41), the tonic component was significantly higher for CF compared with the baseline (P = 0.027, n = 164), but this result did not hold for RF (P = 0.19, n = 164). Conversely, we found weak evidence that RF induced phasic components (P = 0.072, n = 164), whereas this result did not hold for CF (P = 0.40, n = 164). Although the differences between CF and RF were not statistically significant (t test: P = 0.47 for tonic; P = 0.49 for phasic response, n = 164), we note the tendency of physiological reactions to follow different patterns in the two conditions. This pattern appeared to be mirrored in the emotional footprints, because subjects’ self-reported emotional states tended to differ across treatments, particularly for the emotion of sadness (SI Appendix, Table S43).