Testing, Stress, and Performance: How Students Respond Physiologically to High-Stakes Testing

The results of high-stakes standardized tests determine course placement, graduation, and college admission for students, result in sanctions or rewards for schools, and inform education policy. There is substantial resistance to testing regimes, often predicated on the notion that students are “stressed” by tests.¹ Yet, to our knowledge, no evidence exists on test-induced physiological stress among K–12 students in a real-world setting.² Understanding variation in test-induced stress responses and implications for performance is important for determining whether scores on high-stakes tests are reliable measures of ability and knowledge, or if they are biased by “stress disparities” between children (see review in Heissel, Levy, and Adam 2017).

This study raises important questions about the use of high-stakes testing. We document how high-stakes testing affects low-income children's stress biology in one charter school network, and we show how changes in children's physiological responses to high-stakes tests relate to performance on the standardized test. Our goal in this study is to identify these patterns in one setting, and to call for more research into this area in the field. Expanding our understanding of the relationship between stress and test performance will affect our understanding of how high-stakes test results should be used and interpreted. Throughout this paper, our footnotes include suggestions for future research in this area.

We use saliva-based measures of cortisol—a primary stress hormone that indicates how the biological stress system is functioning—among low-income students in New Orleans to document how cortisol levels change in response to a high-stakes standardized test administered to students in grades 3–8, relative to a regular baseline school week. We call this change “cortisol reactivity.” We find that students have 18 percent higher cortisol levels in the homeroom period just before taking the high-stakes test, relative to that same timeframe during weeks without testing. These differences are driven by boys, whose homeroom cortisol is 35 percent higher during testing weeks than regular weeks.³

Everyone has a natural cortisol rhythm over the course of the day (described in more detail in section 2). Acute stressors are associated with increases in cortisol above these natural rhythms. An increase in cortisol is not necessarily bad—in the best case, it can provide the energetic boost one needs to respond to a challenge with attention and focus. How an individual responds to a given stressor is based on what is adaptive in that person's particular context (Del Giudice, Ellis, and Shirtcliff 2011; Shirtcliff et al. 2014). What is adaptive to a given individual may differ by various background characteristics, and it is not necessarily adaptive in the context of an academic test. Although our entire sample can be considered economically disadvantaged, we find suggestive evidence of differences in cortisol changes by level of disadvantage, with the largest cortisol effects for those living in high-poverty and high-crime neighborhoods.

We next examine whether differences in cortisol reactivity are associated with test performance on a subsample of students for whom we have test score data. Large increases in cortisol can make concentration difficult, while reduced cortisol may be a sign of disengagement with a task. We show that both increases and decreases in cortisol from the baseline week to the high-stakes testing week are associated with lower test scores on the high-stakes test, relative to how we would expect students to perform based on other in-school academic performance (i.e., grades).

Descriptive studies show that children from low socioeconomic status and racial/ethnic minority groups have lower average scores on standardized academic tests relative to high socioeconomic status and white families (Reardon 2011; Bradbury et al. 2015). Low socioeconomic status and racial/ethnic minority individuals are also more likely to be exposed to stressful life events relative to higher income or white individuals (see review in Hatch and Dohrenwend 2007). These patterns are correlated, but the physiological stress response may provide a link between them. In particular, students who experience chronic stress may respond differently to new stressors, such as high-stakes tests. Persistent socioeconomic gaps in academic performance could be due in part to different responses to the stress of testing disparately affecting test performance. This disparate effect, in turn, has implications for whether standardized tests are a fair means of evaluating student ability and school quality.

This study makes several contributions. For one, we document cortisol patterns for a low-income 7-to-15-year-old student population about which there is limited evidence. This is the first study to take cortisol samples from such young students during the timeframe surrounding high-stakes testing, and our experience provides guidance for researchers interested in measuring cortisol levels in similar populations. Second, we document how cortisol patterns change for this population in response to a stressful event. This is relevant to understanding how students respond to tests in their actual school settings. Third, we provide the first evidence on how differences in cortisol responses are related to performance on real-world standardized tests. This is crucial for understanding the validity of those tests themselves and the interpretation of individual differences in test results, which can have important real-world consequences. Our analysis of test performance is necessarily correlational—who has larger cortisol responses is not randomly assigned. Changes to cortisol could be attributed to other outside shocks (e.g., changes to family income) that are also correlated with test scores. Still, after conditioning on in-school performance (grades), demographics, and neighborhood characteristics, we find that large changes to cortisol are associated with worse outcomes on the test. We use this evidence as a call for more research into the relationship between stress and test performance, across a variety of settings.

This paper proceeds as follows: In section 2 we provide more background on the science of biological stress responses and the cortisol hormone. Section 3 describes our data. Section 4 describes our analytic strategy. Section 5 presents our results. Section 6 discusses the results and conclusions.

Biological stress responses include multiple systems, but this paper focuses on the hypothalamic-pituitary-adrenal (HPA) axis and its primary hormonal product, cortisol. Cortisol levels show a strong circadian rhythm across the day, known as the diurnal cortisol rhythm, with the highest cortisol levels occurring shortly after waking and the lowest levels occurring about thirty minutes after sleep begins (see Gunnar and Quevedo 2007 for more details). Two key measures in cortisol research are the waking cortisol level and the daily cortisol slope (i.e., the rate at which cortisol levels drop from wake to bedtime). The cortisol awakening response (CAR), a sharp increase in cortisol thirty to forty minutes after waking, is an additional measure. The CAR provides an energetic boost to help individuals meet the expected demands of the upcoming day (see review in Clow et al. 2010).

Real or perceived stressors can increase cortisol above typical diurnal levels.⁴ For routine stressors (e.g., momentary loneliness [Doane and Adam 2010]), cortisol levels return to their usual daily pattern approximately an hour after the stressor has passed. According to the adaptive calibration model, the stress response is generally adaptive; for instance, the HPA axis may mobilize psychological and physiological responses when presented with a stressor (Del Giudice, Ellis, and Shirtcliff 2011; Shirtcliff et al. 2014). One at-home study had twenty-four participants (aged 21–42 years) recruited from a university community provide hourly cortisol samples over a 48-hour period. Rising cortisol was associated with subsequent-hour increases in positive emotions such as activeness, alertness, and relaxation, and marginally significant decreases in nervousness (Hoyt et al. 2016).

Broadly, high or rising cortisol occurs when individuals are in personally relevant situations, are engaged with their environment, and are facing a difficult (but not impossible) task. Low or diminishing cortisol occurs if an individual is disengaged from the environment, a task is impossible, or a task is no longer novel.⁵ The HPA axis can also be anticipatory, with rising cortisol levels before an expected stressful event or changes to the CAR if the prior day was particularly stressful.⁶ In the context of high-stakes testing, we may expect moderately increased cortisol before the test, particularly if the student expects the test to be difficult but manageable, with stakes that matter for them. Limited (or lowered) cortisol responses to stressors may be related to disengagement or “shutting down” in the face of the test; large increases in cortisol may reflect feeling threatened or overwhelmed in a way that is likely to prevent productive focus.

Stress patterns also differ by sex. Females’ CARs tend to peak later in the day than males’ CARs (Stalder et al. 2016). Moreover, males may be more responsive to achievement-related stressors, whereas females may be more responsive to social rejection (Stroud, Salovey, and Epel 2002). A meta-analysis of twenty-eight studies similarly found larger cortisol responses to stressors in males than females (Sauro, Jorgensen, and Pedlow 2003). In the context of high-stakes testing, we may then expect larger cortisol responses to high-stakes testing from male students.

Of particular concern in this context, long-term stress exposure can lead to changes in the HPA axis that can be maladaptive in some contexts, including school. For instance, hypocortisolism is a condition that can follow a period of chronic stress, wherein the HPA axis shows low levels of cortisol and no longer responds to stressors (see summaries in McEwen 1998; McEwen and Gianaros 2010). This is one reason we might expect that children with high-stress backgrounds respond less-optimally (physiologically) to a high-stakes test. However, our results are more consistent with a story that chronic stress is associated with high cortisol reactivity in this population.

HPA axis activity may affect cognitive performance during test-taking by affecting memory recall. Associations between cortisol and memory recall generally displays an inverse-U pattern in laboratory-based studies.⁷ In particular, inducing large increases or decreases in cortisol results in worse memory recall. If cortisol and memory recall are related, then differences in stress response may lead to different test outcomes even among students with equal ability who have learned the same amount of material. If the students most likely to be “stressed testers” come from already-disadvantaged backgrounds, this pattern may exacerbate the observed achievement gaps on high-stakes tests.

Two previous studies compared a baseline week of normal activity against a stressful testing week. Weekes et al. (2006) found that male undergraduate students had an increase in examination-week cortisol levels, while female undergraduates did not. The authors found no link between psychological (self-reported) stress and physiological stress as measured by cortisol. In contrast, Malarkey et al. (1995) collected cortisol and other measures on medical students one month before, during, and two weeks after examinations. They found increases in cortisol during the test week but only for those students who perceived the test as stressful. Neither set of authors examined performance on the tests and its relationship to cortisol.

Other research has not included baseline stress levels but instead examined same-day changes in cortisol in response to stressors. Perceiving a researcher-administered test during the school day as stressful was correlated with higher same-day cortisol and lower test performance in Swedish adolescents (Lindahl, Theorell, and Lindblad 2005). Conversely, among young, low-income children in a Head Start program, having a larger same-day cortisol response to a stressor was correlated with better cognition and behavioral outcomes than those without a cortisol response (Blair, Granger, and Razza 2005). Adults with higher anxiety had larger increases in cortisol in response to performance tasks than those who did not (Malarkey et al. 1995; Schlotz et al. 2006). Whether cortisol improves or detracts from performance may depend on anxiety about the task at hand (Mattarella-Micke et al. 2011).

Overall, the relationships between perceived stress, stress hormones, and performance on a task are complicated and related to a wide variety of background characteristics. These relationships highlight the importance of accounting for baseline differences in cortisol patterns for individual students: Do students perform poorly because of elevated cortisol, or do the students who perform poorly in general also tend to have high cortisol levels in regular, non-tested weeks? In addition, it is not obvious that a real-world high-stakes test will lead to a physiological reaction in a group of young, low-income students. If reactions do occur, it is not obvious who would be most affected, or how such reactions might correspond to performance on the test. This study contributes to our understanding of these dynamics by measuring how cortisol changes in response to a high-stakes test for grade-school students from disadvantaged backgrounds.

Our data consist of cortisol measures, student diaries, and administrative data on student demographics and academic performance, for students from a charter school network in New Orleans. Descriptive statistics are in table 1. The participants were almost all black (95 percent), economically disadvantaged (97 percent),⁸ and from high-poverty neighborhoods (with 40 percent of block group households in poverty, mean block group income of $27,000, and mean block group unemployment of 13 percent). The households were also in neighborhoods with a great deal of police activity, with a mean of 416 high-priority emergency (911) calls within a quarter-mile of their home in the prior year. However, these averages mask heterogeneity: The fraction of neighborhood households in poverty ranged from 14 to 91 percent, mean neighborhood incomes ranged from $9,000 to $58,000, neighborhood unemployment rates ranged from 0 to 74 percent, and the number of nearby high-priority 911 calls ranged from 0 to 1,380 in the prior year. On average, the participants are disadvantaged relative to the overall population, but there is significant variation within the sample.⁹

We collected salivary cortisol samples from ninety-three pre-adolescent and adolescent volunteers in grades 3–8, across three schools from the charter school network. We recruited participants through flyers distributed by their school, obtained parental consent and participant assent, and briefed participants on the protocol during homeroom on their first day of collection. Some participants joined the study late (N = 13 joined in week 2) and were briefed on the protocol individually. These students were mainly from school 2.¹⁰

To provide the samples, participants let saliva collect in their mouth, then used a small straw to drain the saliva into a small vial; this is called the passive drool technique. Participants watched a saliva sample demonstration at the first collection, had a video demonstration available, and received reminder texts from the research team during the data collection to ensure that they followed protocol. Participants were instructed to avoid eating, drinking, and brushing their teeth 30 minutes prior to each sample collection. A kitchen timer preset to 30 minutes was provided to aid in the timing of sample 2. Participants were instructed to refrigerate their home samples as soon as possible after collection and return their home samples to the research team in homeroom every day.

We dropped week 1 samples for two students who had extremely high cortisol levels (mean = 15.43 µg/dL and 3.50 µg/dL, relative to the overall mean of 0.15 µg/dL), likely indicating that they were taking cortisol-containing medication. Per typical protocol for cortisol, we top-code cortisol levels to 1.80 µg/dL.¹⁴ Online appendix table A.1 contains descriptive statistics for the homeroom cortisol samples by school.

Participants filled out diaries at each saliva collection. The sample 1 diary included questions about the prior night's sleep and that morning's wake time. We coded wake time for each day as the minimum reported timing across the sample 1 cortisol sample, the sample 1 diary entry, and diary-reported daily waking time. Students took sample 1 on days 2 and 3; by design day 1 did not have a reported wake time. If students were missing the wake time measure, we imputed it using the mean wake time by individual by week, then (if still missing) the mean wake time by individual, then (if still missing) the mean wake time by school by week.

We calculated time since wake for each sample as the length of time between that day's wake time and the reported timing of the cortisol sample. If missing sample timing, we imputed it using that sample's diary time, then (if still missing) the mean of the sample timing by individual by sample number, then (if still missing) the mean of the sampling timing by school by sample number.

The charter network provided administrative data including participants’ scores on low-stakes math, science, English Language Arts (ELA), and social studies tests and high-stakes math, science, and ELA tests. The administrative data also included in-school grades (on a 0–100 scale) for each academic quarter in math, science, ELA, and social studies. In the test score analysis, we dropped students’ missing test score data or missing cortisol data in the baseline week or high-stakes testing week, leaving us with N = 68 students in the test score subsample. We lose fifteen students who did not have baseline week cortisol (thirteen who joined in week 2 and two who did not have homeroom-specific data), eight students who appear to have moved out of the charter network,¹⁵ and two with missing test score data (but with cortisol data, meaning they were in school at least part of the week). We used a linear probability model to regress an indicator for being in the final test score sample on the demographic characteristics and in-school grades for the ninety-three participants. Those with 10-point higher science grades had a 13.1-percentage point higher probability of being in the final sample; all other demographic characteristics were not statistically related to being in the final test score sample.

We converted each test score into standardized Z-score units by grade; the resulting scores should thus be interpreted as the distance from the average score in standard deviations.¹⁶

The results of the high-stakes test in our study mattered for the school, as they contributed to the letter grade (A–F) rating given to the school by Louisiana's Department of Education. However, the test had no direct repercussions for individual students. In addition, the students took a variety of other tests in their school system throughout the year, including a series of tests that were only used for internal assessment but mimicked the structure of the year-end high-stakes test.¹⁷ Given how often these students were tested, we might expect them to be so accustomed to the process that even high-stakes tests would not be perceived as stressful. This will reduce the likelihood of finding any effect of testing on cortisol responses.

Supplementary analyses test for variation based on proxies for chronic stress—specifically, poverty rates and crime rates in students’ neighborhoods. We might expect students’ responsiveness to the stress of the test to differ if they are chronically stressed. We also tested for differences by gender.

The primary variable of interest is Responsivity_i, which is a vector of indicator variables representing 20 percentage-point bins for the change in homeroom cortisol levels from baseline to the high-stakes testing week. Bins are grouped as follows: −30 percent or lower, −10 to −30 percent, −10 to 10 percent (the reference bin), 10 percent to 30 percent, 30 percent to 50 percent, 50 percent to 70 percent, and 70 percent or higher. We will show that alternative bin cutoffs lead to qualitatively similar conclusions. Statistically significant coefficients for CurrentCortisol_i would indicate that the same-day level of cortisol is related to test performance; statistically significant coefficients for Responsivity_i would indicate that the change in cortisol level from baseline to the test week is related to performance.

The vector of in-school grades accounts for regular, non-high-stakes student performance, and any observed effects of responsivity or current cortisol would indicate underperformance or overperformance on the test beyond what is predicted by those test scores and demographic factors. To the extent that some demographics (e.g., homelessness) themselves cause chronic stress that in turn could affect cortisol responses, our main estimates could underestimate the effect of cortisol responsivity.¹⁹ Still, other unobserved shocks to individuals (e.g., parental job loss) could conceivably be related to both changes in cortisol and test performance.²⁰ Moreover, some participants were missing either requisite cortisol or testing data, and the N in the final analysis is 68.²¹ Given potential omitted variable bias and this smaller sample, we interpret the academic performance results as suggestive and do not conduct subgroup analyses.

Figure 2 displays the cortisol patterns from wake to eight hours post-wake for baseline, low-stakes, and high-stakes weeks using locally weighted scatter plot smoothing, which does not impose parameters on the pattern. Cortisol followed the expected diurnal pattern in the baseline week. We see the sharp rise in the cortisol awakening response (15–60 minutes after waking), following by falling cortisol as time passes.

The pattern visibly differs in the high-stakes testing week, with a less-pronounced CAR and much higher levels of cortisol during the homeroom period. In the baseline week, cortisol levels were not elevated above the expected slope during homeroom, which provides an important test on our hypothesis: We would not expect elevated cortisol during homeroom in a regular school week. Cortisol elevations during the low-stakes test week were in between the baseline and high-stakes test weeks in the homeroom period.

The estimates in table 2 show whether, within individuals, homeroom cortisol levels differed from baseline to the testing weeks. All columns include individual fixed effects and a quadratic of time relative to waking for a given day. The coefficient on low-stakes (high-stakes) testing approximates whether the level of cortisol differs from the baseline week to the low-stakes (high-stakes) testing week. Column 1 does not include wake time or controls for whether the sample was taken during the CAR period (15-60 minutes post-wake), as these were necessarily imputed on day 1 of each week. However, later wake times were associated with higher waking cortisol, so we added controls for wake time and CAR in column 2.²² The homeroom estimates were similar whether controlling for these variables or not, and going forward we prefer the more conservative estimate that controls for wake time and CAR. On average, student levels of cortisol were 18 percent higher in homeroom in the high-stakes week relative to the same students’ homeroom cortisol at baseline. There was no statistical difference in cortisol in the low-stakes week relative to the baseline week, though as expected the coefficients are positive. The high-stakes and low-stakes weeks do not statistically differ from one another; future analyses should explore this relationship further.²³

Columns 3 through 6 of table 2 examine subgroups. All estimates within a column come from the same regression, and the bottom rows of the table test whether the sum of the main effect and the interaction for a given test differs from zero. Male students had large average increases in homeroom cortisol in the low-stakes testing week (31 percent) and high-stakes testing week (33 percent), relative to the baseline week. The female effect sizes statistically differed from the male estimates; the difference relative to baseline was −4 percent in the low-stakes week (calculated as 0.310–0.353) and +7 percent in the high-stakes week. Neither of the female estimates statistically differed from zero, as indicated by the p-values at the bottom of the table. Turning to neighborhood characteristics, we first divide individuals by the median neighborhood poverty rate observed in our sample (40 percent); lower-poverty neighborhood ranged from 14 percent to 40 percent (with a mean of 28 percent) and higher-poverty neighborhoods ranged from 41 percent to 91 percent (with a mean of 53 percent). Those from higher-poverty neighborhoods had larger average increases in homeroom cortisol than those from lower-poverty neighborhoods in the high-stakes week (30 percent versus 6 percent), relative to baseline, though the difference between groups was not statistically significant. Similarly, those from neighborhoods with an above-median number of high-priority 911 calls within 0.25 miles (median = about 340 calls) had larger average increases in homeroom cortisol levels than those from below-median neighborhoods in the high-stakes week (32 percent versus 5 percent).²⁴ The difference between groups was again not statistically significant. While the stress responses of those facing chronic stress do differ from their less-stressed peers, we do not find evidence of pervasive hypocortisolism (the lack of ability to respond to a stressor), per se; indeed, those participants have moderately larger increases in cortisol. Note that our sample size is a bit smaller in the neighborhood analysis due to missing or difficult-to-geocode addresses.

Our final subgroup analysis examines changes by student ability, based on median first-observed-quarter grades expressed in Z-scores (median = −0.15 standard deviation).²⁵ Increases in cortisol were driven by lower-achieving students; there is no average change for above-median students.

There was higher cortisol before the test, on average, relative to the baseline week, but there was also substantial variation in reactivity. Figure 3 displays the density of the change (“responsivity”) from baseline to the low-stakes testing week and from baseline to the high-stakes testing week. Although, on average, cortisol was higher in testing weeks, some individuals had little change and others actually had lower cortisol in the testing weeks—either due to the noisiness of the cortisol sampling or perhaps due to disengagement from the stressful situation. We next test whether these different responses were associated with different performances on the test.

Figure 4 examines how cortisol reactivity was related to test outcomes. It is unclear how best to measure this relationship, and we include several approaches for transparency. First, panel A breaks the subgroup of participants with the requisite data into quintiles based on the percentage change in their homeroom cortisol from baseline to the high-stakes testing week. Quintile 1, the reference group, includes those whose cortisol fell 22 to 78 percent relative to baseline during high-stakes testing. Quintile 2 includes those with little change, ranging from −21 percent to +12 percent. Quintile 3 participants had moderate increases, from 13 percent to 52 percent. The final two quintiles cover those with large increases, from 53 to 119 percent in quintile 4 and over 119 percent increases in quintile 5.

Quintile 2 differs significantly from quintile 1, with test scores 0.38 standard deviation higher (conditional on demographic controls, concurrent cortisol, and in-school grades; p-value = 0.027). The final three quintiles do not differ significantly from quintile 1. We reject that the five quintiles are the same at the 10 percent level with an F-test (F(4, 42) = 2.15; p-value = 0.091), but we cannot reject that quintiles 1, 3, 4, and 5 are statistically the same (F(3, 42) = 0.87; p-value = 0.465). In other words, it appears that participants in quintile 2, who have the least amount of change from baseline to the high-stakes test, outperform the other quintiles, conditional on the other control variables (although it is only marginally statistically significant).

An alternative, parametric approach to the estimate is displayed in panel B. Prior research has found an inverse-U shape in the relationship between cortisol and outcomes (Het, Ramlow, and Wolf 2005; Schilling et al. 2013). Here, conditional on demographic controls and in-school grades, we model a quadratic estimate of the relationship between test score (the outcome) and the raw level of change in cortisol in micrograms per deciliter.²⁶ Panel B plots this estimate and its 95 percent confidence interval, as well as binned scatter plots for test scores and change in cortisol (with five to six observations per bin).²⁷ The pattern appears as an inverse-U, but contrary to prior work, we find no evidence of an improvement in outcomes for moderate increases in cortisol. Moreover, the quadratic term is not statistically significant, at least at our level of power.²⁸

Finally, our preferred model groups the estimates into 20-percentage point bins. The estimates are somewhat noisy, but relative to those in the low reactivity group (from −10 percent to +10 percent homeroom cortisol change from baseline to the high-stakes week), those with either large increases or decreases in cortisol from the baseline week performed worse on the standardized test. In other words, decreases and increases in cortisol were associated with underperformance on the high-stakes test. Grouping the “change” bins together, an increase of more than 10 percent or a decrease of more than 10 percent was associated with a 0.443 standard deviation decrease in the test score (p-value = 0.009), relative to those with little cortisol responsivity (−10 percent to +10 percent), holding school-year academic grades, demographic characteristics, and concurrent cortisol constant. Table 3 contains these results. The estimates are fairly similar when broken up by those who increase more than 10 percent (0.437 standard deviation lower scores relative to those with −10 percent to +10 percent change, p-value = 0.015) and those who decrease more than 10 percent (0.458 standard deviation lower scores, p-value = 0.021).

A potential concern is that grades do not adequately account for student testing ability. Thus, column 2 of table 3 uses low-stakes test scores instead of quarters 1–3 grades to control for baseline ability. Results were similar (−0.558 standard deviation lower for large reactivity participants, relative to the ±10 percent group, p-value = 0.005; N = 62), though the N was also slightly lower because some students were missing low-stakes test results. Results were also similar without controlling for concurrent cortisol (−0.439 standard deviation, p-value = 0.007; N = 67) and when adding the low-stakes test as an additional outcome (−0.261 standard deviations, p-value = 0.019; N = 136 tests for 73 participants).²⁹ The estimates are based on the average score across the math, ELA, and science high-stakes tests to decrease variability in scores; post hoc analyses demonstrated that the effects were negative for all three individual tests, with the largest estimate in science.³⁰ There was no relationship between a quadratic of concurrent level of homeroom cortisol during the testing week itself and outcomes on the test, with or without including reactivity from baseline.

One concern with the analysis is that cortisol variability, rather than responsivity to the high-stakes test, is associated with worse outcomes on the test. We test this with three placebo measures in columns 5–7 of table 3. These all test whether changes between days unrelated to a given test predict performance on the test. Column 5 uses whether changes from day 1 to day 2 (or day 2 to day 3 for three individuals who joined on day 2) in the baseline week predict performance on the high-stakes test. It does not. As a second placebo in column 6, we test whether cortisol responsivity to the low-stakes test predicted performance months later on the high-stakes test. It did not, nor did responsivity to the high-stakes test predict performance on the low-stakes test in column 7. Thus, it does not appear that cortisol variability in general predicts performance on the high-stakes test. Instead, it is specifically changes from baseline to the high-stakes test week that are associated with performance on the high-stakes test itself.

We do not conduct the full binning exercise by subgroups due to small sample size. However, when we compared lower-reactivity participants (±10 percent cortisol change) to higher-reactivity participants (greater than ±10 percent change), we found no statistically significant differences in the patterns by gender, neighborhood poverty, neighborhood crime, or prior grades.³¹ So, although some groups are more likely to be high-reactivity than others, the relationship with test scores is similar among all high-reactivity participants, at least to the extent that we can test it in this setting.

Figures 4 and 5 show that the estimates are noisy, with considerable unexplained fluctuation in test scores, and that the best outcomes appear around where there is little cortisol change. Overall, we take this as suggestive evidence that large changes in cortisol in response to high-stakes tests are associated with worse performance on the test, but there is much more to be done in this area.

One hypothesis is that a cortisol spike could be associated with “acting out” and misbehavior during the test, which could inhibit performance. We can assess this hypothesis because the charter network tracked behavior using a daily points-based system.³² Throughout the year, the average student got into at least some trouble on 35 percent of school days.

Relative to a regular day, there were no differences in the probability of getting in trouble on a low-stakes test day. However, for the most important week of high-stakes testing, students were 26 percentage points less likely to get in trouble than on regular school days (p-value = 0.000).³³ We do not take these estimates as a measure of acting out, necessarily, given the discretion that teachers have in assigning points to students.³⁴ Perhaps teachers were more lenient in general on test days, or perhaps students had fewer opportunities to get in trouble. However, we did test whether those with large increases in cortisol had different drops in infractions than those who had decreased cortisol or those who did not have a strong cortisol response.³⁵ We found no evidence of a difference in the probability of getting in trouble by those with very large increases (or decreases) in cortisol level. As best as we can measure, then, we find no evidence that misbehavior is driving the results on the tests. Instead, we hypothesize that the ability to focus and recall information relevant to the test is affected.

This study examined whether children responded physiologically to high-stakes testing in a naturalistic setting, and how any responses were associated with performance on a high-stakes test. Children in one charter school network displayed a statistically significant increase in cortisol level in anticipation of high-stakes testing; this pattern was driven by male students. We also find some evidence that, among a sample of disadvantaged students, the most-disadvantaged students had the largest increase in cortisol in anticipation of the high-stakes test. These changes were driven by the occurrence of a test that mattered for schools but had limited consequence for individual students.

Moderate decreases and increases in cortisol were associated with underperformance on the high-stakes test, relative to what we would have expected from students given their in-school academic performance and other characteristics. Even the average increase in cortisol shown in table 2 (18 percent) was associated with lower test scores, relative to those with little change in cortisol. An increase of more than 10 percent or a decrease of more than 10 percent was associated with a 0.4 standard deviation decrease in test scores, relative to those with little change. This is equivalent to approximately 80 points on the 1,600-point SAT scale. Concurrent cortisol measured as linear, quadratic, or bins during the test was not a statistically significant predictor of performance; it was cortisol change relative to baseline that predicted outcomes.

Of course, one study on a small, nonrandom sample of students may not give us the true population-level effect of high-stakes tests on cortisol or how cortisol relates to test performance. We view this study as a call for additional work in this area. Specifically, we identify four non-exclusive themes in need of future research. First, future analyses should replicate that students do indeed have an increase in cortisol and that it differs by various attributes. Such research should examine a more diverse population of students, rather than the largely low-income, mostly black population we examined here. A larger sample size would permit a greater degree of heterogeneity analysis than is possible in the present study in order to more robustly test whether different groups respond differentially to high-stakes testing. One possibility is collecting cortisol samples around the time of SAT or ACT testing, as these tests have real-world implications for students.

Second, research must confirm whether moderate changes in cortisol are associated with worse performance in other settings. Researchers rely on high-stakes tests as a measure of academic performance to evaluate various education and social policies. Such research may accept that high-stakes tests are noisy measures of ability or knowledge, but it generally assumes that the noise is evenly distributed across the socioeconomic spectrum. If, however, certain groups are systematically “stressed testers”—that is, they have large physiological reactions to the high-stakes testing setting—the policies recommended by such research may be suboptimal. As an extreme example, consider a world where all children learn the same amount of material during the year but group A has a bigger physiological reaction, and subsequently lower scores, than group B. Examining test scores would lead researchers to conclude there is an achievement gap between these groups and that group A needs intervention. But in reality, both groups learn the same amount of material and can perhaps even apply that material similarly in the real world. The policy solution in this case would be much different than if learning differed between groups. Such test-day stress deficits are not the only cause of achievement gaps, but they may explain part of existing disparities. Future research should examine how large a role they play. A key consideration in any such research is causality. Researchers cannot assign stress responses to students in real-world tests, but real-world tests may be more stressful to students than lab-based tests. Carefully designed research, which includes measures of performance outside of the high-stakes tests, will be necessary to move understanding forward.

Third, researchers should consider how school policies that use test scores may exacerbate or alleviate disparities among groups. If certain groups are more likely to be stressed testers, then, holding baseline knowledge constant, those stressed testers will be disadvantaged by admission or graduation policies based on high-stakes tests. Researchers should carefully consider how policy decisions interact with biological responses to testing.

Finally, if new work confirms that testing causes stress for students in ways that impact their performance, a logical question is what schools can do to mitigate the stress response—or at least the effects of the stress response on performance. Potential options include mindfulness programs (Zenner, Herrnleben-Kurz, and Walach 2014), integrated mental health interventions (Fazel et al. 2014), or yoga (Ehud, An, and Avshalom 2010), among other interventions. Though such programs may have benefits well beyond test scores, researchers could investigate whether they are associated with changes in biological stress reactions to high-stakes testing.

Given the prevalence of high-stakes testing in U.S. education policy, much more work is needed in this area. If the patterns of test-induced stress that we find in this study continue to hold up, it might suggest that high-stakes testing results should be used and interpreted differently than the way they are currently implemented in education policy and practice.

We thank the anonymous school district and its staff for their invaluable cooperation, as well as Kaho Arakawa, Chernjen Lee, Royette Tavernier, members of the COAST Lab at Northwestern University, and seminar participants at Northwestern University and the AEFP, APPAM, and Western Economic Association meetings. Laura Scaramella at the University of New Orleans provided access to laboratory space. We are grateful for funding from the Spencer Foundation (grant no. 2015000117) and the Institute for Policy Research at Northwestern University.