Methods

The paper uses two complementary econometric strategies: a difference-in-differences (DiD) model that estimates the average effect of the pandemic on each outcome, and a difference-in-differences-in-differences (DDD) model that estimates whether the effect differed across institutional and student subgroups.


Strategy 1: Difference-in-Differences

Intuition

The DiD approach exploits the fact that the COVID-19 pandemic hit Mexico in the 2019-2020 academic year (March 2020) and continued through 2020-2021. The identifying idea is to compare how outcomes changed from the first to the second year within the pandemic window (2019-2020 to 2020-2021), relative to how the same outcomes changed from the first to the second year within the pre-pandemic window (2017-2018 to 2018-2019).

Treatment group: higher education institutions observed during academic years 2019-2020 and 2020-2021 (the pandemic cohort).

Control group: the same higher education institutions observed during academic years 2017-2018 and 2018-2019 (the pre-pandemic cohort).

Post indicator: equals 1 for the second year in each two-year window — academic years 2018-2019 and 2020-2021 — and 0 for the first year in each window — 2017-2018 and 2019-2020.

The interaction Post × Treatment is the DiD estimator: it captures the pandemic’s effect by comparing the year-over-year change in the pandemic cohort to the year-over-year change in the pre-pandemic cohort.

Specification

Difference-in-Differences Equation (1)

\[Y_{icmy} = \alpha + \gamma\,\text{Treatment}_{imy} + \delta\,\text{Post}_{imy} + \beta\,(\text{Post} \times \text{Treatment})_{imy} + \delta X_{imy} + e_{icmy}\]

where:

Term Definition
\(Y_{icmy}\) Outcome (new entry, enrollment, or graduation) for institution \(i\), campus \(c\), area of study \(m\), year \(y\)
\(\text{Treatment}_{imy}\) = 1 for the pandemic cohort (2019-2020 and 2020-2021), = 0 for pre-pandemic (2017-2018 and 2018-2019)
\(\text{Post}_{imy}\) = 1 for the second year in each cohort (2018-2019 and 2020-2021), = 0 for the first year (2017-2018 and 2019-2020)
\(\beta\) The average treatment effect on the treated (ATT): the pandemic’s causal effect on the outcome
\(X_{imy}\) Controls: delivery format dummy, public/private dummy, top-20 dummy
\(e_{icmy}\) Residual; standard errors clustered at the municipality level

What \(\beta\) estimates. The DiD estimator \(\beta\) compares the year-over-year change during the pandemic window (2019-20 to 2020-21) with the year-over-year change during the pre-pandemic window (2017-18 to 2018-19). If both cohorts would have followed the same trend absent the pandemic, then any additional change in the pandemic cohort is attributable to COVID-19.

Strategy 2: Difference-in-Differences-in-Differences

Intuition

The DDD approach extends the DiD to test whether the pandemic’s effect differed across subgroups. For each heterogeneity dimension (public vs. private funding, top-20 vs. non-top-20 status, synchronous vs. asynchronous delivery), a third interaction term is added: Post × Treatment × Subgroup indicator.

Specification

DDD Equation (2)

\[Y_{icmy} = \alpha + \beta_1\,\text{Treatment}_{imy} + \beta_2\,\text{Het}_{icy} + \beta_3\,\text{Post}_{my} + \beta_4\,(\text{Post} \times \text{Treatment})_{imy}\] \[+ \beta_5\,(\text{Treatment} \times \text{Het})_{icmy} + \beta_6\,(\text{Het} \times \text{Post})_{icmy} + \beta_7\,(\text{Post} \times \text{Treatment} \times \text{Het})_{icmy} + e_{icmy}\]

where \(\text{Het}_{icy}\) is a dummy equal to 1 if the institution belongs to the heterogeneous subgroup of interest (e.g., public, top-20, synchronous), and 0 otherwise.

The parameter \(\beta_7\) captures the differential pandemic effect for the subgroup relative to the complement. A positive \(\beta_7\) means the subgroup fared better than the non-subgroup; a negative \(\beta_7\) means it fared worse.


Robustness Suite

The paper validates the main DiD estimates with four robustness checks:

1. Placebo test (pre-parallel trends). Shifts the treatment window one year earlier. Finding null effects supports the parallel trends assumption.

2. False Discovery Rate correction. Because three outcomes are tested simultaneously, the paper applies the FDR q-value correction of Anderson (2008). No significant result becomes insignificant after this correction.

3. Oster (2019) bounds for omitted variable bias. Simulates bounds around \(\beta\) based on an expected R². When bounds exclude zero, the estimate is robust to omitted variable bias. All main bounds exclude zero.

4. Leave-one-major-out sensitivity. Re-estimates the DiD specification ten times, each time excluding one area of study. Estimates remain stable across all exclusions, confirming the results are not driven by any single field.