Long-horizon GDP models hold slopes fixed for decades. Either the level or the slopes can break in a crisis. This thesis examines the effect of the 2008 crisis on GDP growth. Mean growth fell from 2.5 percent a year before the crisis to 0.6 percent after. No break test rejects slope stability at any year. Read one way, the results imply the level fell but the structure held. Projections can re-base the starting point and keep the model, as the coefficients did not change.
The growth machine
Determinants of growth include catch-up gap, investment share, schooling, and population growth. Each country has its own intercept, while slopes come from within-country movement. Across the sample, averages are growth 1.6 percent, investment 25.6 percent of GDP, and population growth 0.7 percent. A crisis may move intercepts while keeping slopes unchanged, which the tests target.
Twenty-four economies that belonged to the OECD before 1995 make up the sample, observed annually from 1996 through 2019 in the Penn World Table, version 11.0. It closes before the COVID-19 contraction, so that a second large shock cannot confound the reading of the first.
What happened to growth
Figure 1
Growth fell after 2008
Mean per-capita growth across 24 OECD countries, 1996–2019 (shaded band: p25–p75). The panel's own post-2008 flag splits the average from 2.46% (1996–2007) to 0.64% (2008–2019).
Before 2008, growth averaged 2.5 percent a year. After 2008, growth averaged 0.6 percent a year. Every one of the 24 countries was slower after 2008. Dispersion across countries increased from 1.9 to 2.9 percentage points.
Figure 2
Every country slower
Pre- vs. post-2008 mean growth, one point per country. Every one of the 24 countries sits below the y = x line — no country matched its pre-crisis pace afterward, and the country average fell from 2.5% pre-2008 to 0.6% post-2008. Labelled points are the three lowest post-2008 averages — Greece, Italy, Finland — and the two highest — Ireland, Türkiye.
Testing for a break
Within countries, two of four regressors matter. Convergence is captured by a catch-up coefficient of minus 0.074, with a p-value of 0.009. Investment is associated with a coefficient of 0.163 and a p-value of 0.014, meaning 10 percentage points more investment is associated with about 1.6 percentage points more growth. Schooling has a p-value of 0.77 and population growth has a p-value of 0.30. At 2008, each regressor is interacted with a post-crisis dummy, but none are significant alone, with a joint F-statistic of 1.30 and a p-value of 0.351.
Figure 3
No candidate year rejects stability
Wald statistic (4 break-interaction restrictions) for a break at each candidate year. The maximum, at 2001, is 11.1 — below the Andrews (1993) 10% critical value of 14.3. (Per-year p-values aren’t valid for a max-selected statistic; compare to the critical values shown, not a chi-squared reference.)
Scanning every candidate year from 2001 to 2015, with 15 percent trimming, yields a largest statistic at 2001, with an F-statistic of 2.77 and a Wald statistic of 11.08. Clustering by country with a CR2 correction and Satterthwaite degrees of freedom is used, despite having only 24 clusters.
Figure 4
Only investment moves, and not far enough
Post-2008 shift in each coefficient, in standard-error units (two-way FE panel, CR2 cluster-robust SEs, 24 clusters). The shaded band (±2.07 SE) marks shifts not distinguishable from zero. Investment share shifts most (0.108, 1.96 of 2.07) — nearest the edge. Log GDP per capita, lagged (−1.42 of 2.07), population growth (1.37 of 2.07), and log human capital (1.11 of 2.07) stay further inside.
Alone among the regressors, investment moves, with a slope of 0.116 before the crisis and 0.224 after, and an interaction term of 0.108 with a p-value of 0.075. One slope is affected by the crisis, with investment's slope increasing from 0.116 to 0.224.
What it means for projections
Projected growth uses full-sample slopes and 2015-2019 means, starting in 2019.
Figure 5
Same slopes, nearly the same path
Illustrative projection from the full-sample slopes, 2019–2045, indexed to 2019 = 100, with other regressors frozen at their 2015–2019 country means. The pre- and post-crisis common components produce almost the same path — not a forecast.
Pre-crisis and post-crisis common components almost coincide. The crisis years have a level effect of roughly 7 percent of common output, which is permanent. A forecaster should account for a lower launch point after a financial crisis, while keeping the propagation mechanism. This projection is an illustration only.
What it cannot say
With joint tests resting on seven or eight degrees of freedom, statistical power is limited. Power calculations indicate that a shift of 0.169 can be detected with 80 percent power, given an interaction standard error of 0.055, whereas the observed shift is 0.108. Pooled averages may mask differences between core and periphery, as breaks in these groups could offset each other. Failure to reject a null hypothesis does not prove the null, as surviving a test differs from being proven. Schooling and population growth exhibit little within-country variation from year to year, so null findings are expected. Investment enters contemporaneously, allowing for the possibility of reverse causation, although lagged investment has a coefficient of 0.057 with a p-value of 0.34. Difference GMM diagnostics show a Sargan p-value of 0.93 and an AR(2) p-value of 0.53, with catch-up having a coefficient of 0.68 and a p-value of 0.71, which is imprecise and has the wrong sign.
Part of the post-2008 shortfall also predates the crisis, since productivity growth had been decelerating across the major advanced economies since the early 2000s, so the lower post-crisis level conflates the shock with a trend already underway. A longer post-crisis window, once the COVID-19 shock can be cleanly separated, would test whether the stability found here extends through a second major disruption.
Read the full paper (PDF, opens in new tab)
Get new essays and research by email. Subscribe on Substack (opens in new tab)