The better performance on admission tests could be explained by the high proportion of Honors students

in our sample (22% compared to 18% in the ASU population). The last four columns of Table 1 show how

Honors students compare with ASU students and the average college student at a top-10 university. We see

that they perform better than the average ASU student (which is expected) and just slightly worse than the

average college student at a top-10 university. The share of white Honors students in our sample (60%) is

higher than the proportion in the ASU population and much higher than the proportion of white students

in the top-10 universities.

Overall, we believe our sample of ASU students is a reasonable representation of students at other large

public schools, while the Honors students may provide insight into the experiences of students at more elite

institutions.

3 Analytic Framework

We next outline a simple analytic framework that guides the empirical analysis . Let Oi (COV ID-19) be

the potential outcome of individual i associated with COVID-19 treatment. We are interested in the causal

impact of COVID-19 on student outcomes:

∆i (O) = Oi (COV ID-19 = 1) − Oi (COV ID-19 = 0), (1)

where the first term on the right-hand side is student i’s outcome in the state of the world with COVID-19,

and the second term being student i’s outcome in the state of the world without COVID-19. Recovering

the treatment effect at the individual level entails comparison of the individual’s outcomes in two alternate

states of the world. With standard data on realizations, a given individual is observed in only one state of the

world (in our case, COV ID-19 = 1). The alternate outcomes are counterfactual and unobserved. A large

econometric and statistics literature studies how to identify these counterfactual outcomes and moments of

the counterfactual outcomes (such as average treatment effects) from realized choice data (e.g., Heckman

and Vytlacil, 2005; Angrist and Pischke, 2009; Imbens and Rubin, 2015). Instead, the approach we use in

this paper is to directly ask individuals for their expected outcomes in both states of the world. From the

collected data, we can then directly calculate the individual-level subjective treatment effect. As an example,

consider beliefs about end-of-semester GPA. The survey asked students “What semester-level GPA do you

expect to get at the end of this semester?” This is first-term on the right-hand side of equation (1). The

counterfactual is elicited as follows “Were it not for the COVID-19 pandemic, what semester-level GPA

would you have expected to get at the end of the semester?”. The difference in the responses to these two

6
questions gives us the subjective expected treatment effect of COVID-19 on the student’s GPA. For certain

binary outcomes in the survey, we directly ask students for the ∆i . For example, regarding graduation plans,

we simply ask a student if the ∆i is positive, negative, or zero: “How has the COVID-19 pandemic affected

your graduation plan? [graduate later; graduation plan unaffected; graduate earlier].”

The approach we use in this paper follows a small and growing literature that uses subjective expectations

to understand decision-making under uncertainty. Specifically, Arcidiacono et al. (2020) and Wiswall and

Zafar (2018) ask college students about their beliefs for several outcomes associated with counterfactual

choices of college majors, and estimate the ex-ante treatment effects of college majors on career and family

outcomes. Shapiro and Giustinelli (2019) use a similar approach to estimate the subjective ex-ante treatment

effects of health on labor supply. There is one minor distinction from these papers: while these papers

elicit ex-ante treatment effects, in our case, we look at outcomes that have been observed (for example,

withdrawing from a course during the semester) as well as those that will be observed in the future (such

as age 35 earnings). Thus, some of our subjective treatment effects are ex-post in nature while others are

ex-ante.

The soundness of our approach depends on a key assumption that students have well-formed expectations

for outcomes in both the realized state and the counterfactual state. Since the outcomes we ask about are

absolutely relevant and germane to students, they should have well-formed expectations for the realized state.

In addition, given that the counterfactual state is the one that had been the status quo in prior semesters

(and so students have had prior experiences in that state of the world), their ability to have expectations for

outcomes in the counterfactual state should not be a controversial assumption.6

4 Empirical Analysis

4.1 Treatment Effects

We start with the analysis of the aggregate-level treatment effects, which are presented in Table 2. The

outcomes are organized in two groups, academic and labor market (see Appendix Table A1 for a complete

list of outcomes). The first two columns of the table show the average beliefs for those outcomes where the

survey elicited beliefs in both states of the world. The average treatment effects shown in column (3) are of

particular interest. Since we can compute the individual-level treatment effects, columns (4)-(7) of the table

show the cross-sectional heterogeneity in the treatment effects.

campus activities (COVID-19) since most students would not have had any experience with this counterfactual prior to March
this year.

7
 We see that the average treatment effects are statistically and economically significant for all outcomes.

The average impacts on academic outcomes, shown in Panel A, are mostly negative. For example, the average

subjective treatment effect of COVID-19 on semester-level GPA is a decline of 0.17 points. More than 50%

of the students in our sample expect a decrease in their GPA due to the treatment (versus only 7% expecting

an increase). Additionally, on average, 13% of the participants delayed their graduation, 11% withdrew from

a class during the spring semester, and 12% stated that their major choice was impacted by COVID-19. An

interesting and perhaps unanticipated result is that, on average, students are 4 percentage points less likely

to enroll in an online class given their experience with online instruction due to the pandemic.7,8 However,

there is a substantial amount of variation in terms of the direction of the effect: 31% (47%) of the participants

are now more (less) likely to enroll in online classes. We explore this heterogeneity in more detail in the

next section, but it seems that prior experience with having taken online classes somewhat ameliorates the

negative experience: the average treatment effect for students with prior experience with online classes is a

2.4 percentage points decrease in their likelihood of enrolling in online classes, versus a 9.5 percentage points

decline for their counterparts (difference statistically significant at the 0.1% level).

This large variation in the treatment effects of COVID-19 is apparent in several of the other outcomes,

such as study hours, where the average treatment effect of COVID-19 on weekly study hours is -0.9 (that is,

students spend 0.9 less hours studying per week due to COVID-19). The interquartile range of the across-

subject treatment effect demonstrates substantial variation, with the pandemic decreasing study time by 5

hours at the 25th percentile and increasing study time by 4 hours at the 75th.

Overall, these results suggest that COVID-19 represents a substantial disruption to students’ academic

experiences, and is likely to have lasting impacts through changes in major/career and delayed gradua-

tion timelines. Students’ negative experiences with online teaching, perhaps due to the abruptness of the

transition, also has implications for the willingness of students to take online classes in the future.

Turning to Panel B in Table 2, we see that students’ current and expected labor market outcomes were

substantially disrupted by COVID-19. As for the extensive margin of current employment, on average, 29%

of the students lost the jobs they were working at prior to the pandemic (67% of the students were working

prior to the pandemic), 13% of students had their internships or job offers rescinded, and 61% of the students

reported that a close family member had lost their job or experienced an income reduction. The last statistic

is in line with findings from other surveys of widespread economic disruption across the US.9 Respondents
7 The questions that were asked to elicit this were: “ Suppose you are given the choice to take a course online/remote or

in-person. [Had you NOT had experience with online/remote classes this semester], what is the percent chance that you would
opt for the online/remote option?”
8 This result is in line with a survey about eLearning experiences across different universities in Washington and New York

that concludes that 75% of the students are unhappy with the quality of their classes after moving to online learning due to
COVID-19.
9 According to the US Census Bureau Household Pulse Survey Week 3, 48% of the surveyed households have experienced a

8
experienced an average decrease of 11.5 hours of work per week and a 21% decrease in weekly earnings,

although there was no change in weekly earnings for 52% of the sample, which again reflects substantial

variation in the effects of COVID-19 across students.

In terms of labor market expectations, on average, students foresee a 13 percentage points decrease in

the probability of finding a job by graduation, a reduction of 2 percent in their reservation wages, and a

2.3 percent decrease in their expected earnings at age 35. The significant changes in reservation wages and

expected earnings at age 35 due to COVID-19 demonstrate that students expect the treatment effects of

COVID-19 to be long-lasting. This is consistent with Oreopoulos et al. (2012) which finds that graduating

during a recession implies an initial loss in earnings of 9% that decreases to 4.5% within 5 years and disappears

after 10 years. Rothstein (2020) estimates that the Great Recession may have even longer-lasting effects on

graduates’ employment.

4.2 Heterogeneous Effects

We next explore demographic heterogeneity in the treatment effects of COVID-19. Figure 1 plots the

average treatment effects across several relevant demographic divisions including gender, race, parental

education, and parental income. Honors college status and expected graduation cohort are also included as

interesting dimensions of heterogeneity in the COVID-19 context. The figure shows the impacts for six of

the more economically meaningful outcomes from Table 2 (additional outcomes can be found in Figure A1).

At least four patterns of note emerge from Figure 1. First, compared to their classmates, students from

disadvantaged backgrounds (lower-income students defined as those with below-median parental income,

racial minorities, and first-generation students) experienced larger negative impacts for the academic out-

comes, as shown in the first three panels of the figure.10 The trends are most striking for lower-income

students, who are 55% more likely to delay graduation due to COVID-19 than their more affluent class-

mates (0.16 increase in the proportion of those expecting to delay graduation versus 0.10), expect 30%

larger negative effects on their semester GPA due to COVID-19, and are 41% more likely to report that

COVID-19 impacted their major choice (these differences are statistically significant at the 5% level). For

some academic outcomes, COVID-19 had similarly disproportionate effects on nonwhite and first-generation

students, with nonwhite students being 70% more likely to report changing their major choice compared

to their white peers, and first-generation students being 50% more likely to delay their graduation than

students with college-educated parents. Thus, while on average COVID-19 negatively impacted several

measures of academic achievement for all subgroups, the effects are significantly more pronounced for socioe-
loss in employment income since March 13 2020.
10 The cutoff for median parental income in our sample is $80,000

9
conomic groups which were predisposed towards worse academic outcomes pre-COVID.11 The pandemic’s

widening of existing achievement gaps can be seen directly in students’ expected Semester GPA. Without

COVID-19, lower-income students expected a 0.052 lower semester GPA than their higher-income peers.

With COVID-19, this gap nearly doubles to 0.098.12

Second, Panel (d) of Figure 1 shows that the switch to online learning was substantially harder for some

demographic groups; for example, men are 7 percentage points less likely to opt for an online version of a

course as a result of the COVID-19 treatment, while women do not have a statistically significant change in

their online preferences. We also see that Honors students revise their preferences by more than 2.5 times

the amount of non-Honors students. As we show later (in Table 4), these gaps persist after controlling

for household income, major, and cohort, suggesting that the switch to online learning mid-semester may

have been substantially more disruptive for males and Honors students. While the effect of COVID-19 on

preferences for online learning looks similar for males and Honors students, our survey evidence indicates

that different mechanisms underpin these shifts. Based on qualitative evidence, it appears that Honors

students had a negative reaction to the transition to online learning because they felt less challenged, while

males were more likely to struggle with the learning methods available through the online platform.13 One

speculative explanation for the gender difference is that consumption value of college amenities is higher for

men (however, Jacob et al. (2018), find little gender difference in willingness to pay for the amenities they

consider).

The third trend worth highlighting from Figure 1 is that Honors students were better able to mitigate

the negative effect of COVID-19 on their academic outcomes (panels a, b, and c), despite appearing to be

more disrupted by the move to online learning (panel d). Honors students report being less than half as

likely as non-Honors students to delay graduation and change their major due to COVID-19. Extrapolating

from these patterns provides suggestive evidence that academic impacts for students attending elite schools–

the group more comparable to these Honors students– are likely to have been small relative to the impacts

for the average student at large public schools.

Finally, the last two panels of Figure 1 present the COVID effect on two labor market expectations and

show much less meaningful heterogeneity across demographic groups compared to the academic outcomes

in previous panels. This suggests that, while students believe COVID-19 will impact both their academic
11 Based on analysis of ASU administrative data including transcripts, we find that, relative to their counterparts, first-

generation, lower-income, and non-white students drop out at higher rates, take longer to graduate, have lower GPAs at
graduation, and are more likely to switch majors when in college (see Appendix Table A3)
12 The difference is significant at 1% in both cases.
13 Honors students were as likely as non-Honors students to say that classes got easier after they went online but, conditional on

saying classes got easier, were 47% more likely to say “homework/test questions got easier.” Conversely, males were marginally
more likely to say classes got harder after they went online (10% more likely, p=0.055) and, conditional on this, were 14% more
likely to say that “online material is not clear”.

10