After the largest household debt expansions, median three-year growth is 6.56%, against 10.01% in the middle of the distribution

What Mian, Sufi and Verner claimed

Their paper regressed subsequent real GDP growth on the three-year change in the household and non-financial business debt-to-GDP ratios, entered with a one-year lag and country fixed effects, on an unbalanced panel of 30 countries running largely over 1960–2012. Two of its claims are tested here: first, that a three-year rise in the household debt-to-GDP ratio predicts lower subsequent GDP growth; second, that the corresponding rise in non-financial business debt predicts much less. Their coefficients, standard errors, and the global household-debt cycle they extract are not reproduced, because this note runs no regression and holds nothing constant; only the qualitative ordering is the object of comparison.

Source: JST Macrohistory Database (jst): total household loans (thh), total business loans (tbus), nominal GDP, real GDP per capita and population. The claim under test is Mian, A., A. Sufi, and E. Verner (2017), “Household Debt and Business Cycles Worldwide,” Quarterly Journal of Economics 132(4), 1755–1817.

What we run it on

The JST panel view jst carries total loans to households (thh), total loans to business (tbus), nominal GDP in the same currency units, real GDP per capita (rgdpmad), and population, for 18 advanced economies; the usable household episodes span 1873–2017.

• The exposure is a 3-year change in a ratio. For every country-year with all four inputs, the household debt ratio is thh/gdp and the exposure is that ratio minus its value 3 calendar years earlier in the same country, in percentage points. Real output is rgdpmad×pop and the outcome is 100 times the log difference between output 3 calendar years ahead and output today, so the growth window starts where the debt window ends. An episode at year t exists only if the country reports data at exactly t−3 and t+3, so a gap in a series drops the episode rather than silently widening the window. The paper’s own regressor enters with a one-year lag; the no-gap windows here are this note’s simplification, stated as such.
• Quartiles are pooled, not per country. The 1,260 episodes with both windows observable are sorted into four near-equal groups (315/315/315/315) by the debt change over the full panel, so a “large expansion” means large by the whole panel’s standard, not the country’s own.
• The outcome is a count against the pooled median. Each quartile reports its median subsequent growth and the share of its episodes whose growth falls below the pooled median of 8.00%. Over the whole sample that share is 50.0% (630 of 1,260) by construction of the median, which is the benchmark every quartile is read against.
• One pre-declared variant, nothing else. The single variant keeps episodes whose base year t falls in 1960–2012, the closest cut to the paper’s own sample years, with quartiles re-drawn inside that set (859 episodes; debt windows may reach back to 1957 and growth windows forward to 2015). The run with tbus in place of thh (1,191 episodes) is not a variant: it is the test of the paper’s second claim, the household-versus-business asymmetry.

Source: JST Macrohistory Database (jst): total household loans (thh), total business loans (tbus), nominal GDP, real GDP per capita and population. The claim under test is Mian, A., A. Sufi, and E. Verner (2017), “Household Debt and Business Cycles Worldwide,” Quarterly Journal of Economics 132(4), 1755–1817. All numbers on this page are computed at build time from src/lib/householdDebtCycles.ts; nothing is typed in.

The household quartiles: weakest growth after the biggest expansions

The top quartile, household debt-ratio rises of 4.9 to 26.6pp over three years, is followed by median growth of 6.56%, the lowest of the four, and by below-median growth in 61.0% of episodes (192 of 315). The second quartile grows fastest at 10.01%, a 3.45pp gap to the top, and the bottom quartile, debt-ratio changes of -30.4 to -0.2pp, is also weak: 7.10% median growth and 57.1% below the pooled median (180 of 315).

Source: JST Macrohistory Database (jst): total household loans (thh), total business loans (tbus), nominal GDP, real GDP per capita and population. The claim under test is Mian, A., A. Sufi, and E. Verner (2017), “Household Debt and Business Cycles Worldwide,” Quarterly Journal of Economics 132(4), 1755–1817. "Below pooled median" counts episodes with subsequent growth under the 8.00% pooled median; the unconditional share is 50.0% (630/1,260).

The 1960–2012 window: the paper’s own era

Keeping the episodes whose base year falls in 1960–2012, the closest cut to the sample years behind the paper’s estimates (debt windows may reach back to 1957 and growth windows forward to 2015), and re-drawing the quartiles inside that set leaves 859 episodes and the same ordering: the top quartile’s median growth of 6.85% is again the lowest, 3.44pp under the second quartile’s 10.29%, and 60.7% of top-quartile episodes (130 of 214) land below that window’s own pooled median of 8.13%.

Source: JST Macrohistory Database (jst): total household loans (thh), total business loans (tbus), nominal GDP, real GDP per capita and population. The claim under test is Mian, A., A. Sufi, and E. Verner (2017), “Household Debt and Business Cycles Worldwide,” Quarterly Journal of Economics 132(4), 1755–1817. Quartiles and the pooled median are recomputed within 1960-2012; the unconditional share is 49.9% (429/859).

The asymmetry test: the same cut on business debt

Replacing household loans with business loans (tbus) and re-running the identical design yields 1,191 episodes. The ordering survives but compresses: the top quartile’s median growth of 7.09% sits 1.33pp under the second quartile’s 8.42%, against a 3.45pp gap on the household side, and the top quartile’s below-median share of 54.2% (161 of 297) stands much closer to the 50.0% benchmark (595 of 1,191) than the household top quartile’s 61.0% (192 of 315).

Source: JST Macrohistory Database (jst): total household loans (thh), total business loans (tbus), nominal GDP, real GDP per capita and population. The claim under test is Mian, A., A. Sufi, and E. Verner (2017), “Household Debt and Business Cycles Worldwide,” Quarterly Journal of Economics 132(4), 1755–1817. Same base rule, windows, pooled quartiles, and pooled-median count as the household tables, with tbus/gdp as the ratio; the pooled median here is 7.98%.

Reading the result

• The paper’s direction reproduces. Top-quartile household debt expansions are followed by the weakest median growth of any quartile in both the full sample (6.56%) and the 1960–2012 window (6.85%), and by below-median growth in 61.0% and 60.7% of episodes respectively (192/315 and 130/214), against benchmarks of 50.0% (630/1,260) and 49.9% (429/859).
• The household-business asymmetry reproduces. The fastest-to-slowest quartile gap is 3.45pp for household debt against 1.33pp for business debt, and the business top quartile’s below-median share of 54.2% (161/297) is far nearer its 50.0% benchmark (595/1,191) than the household top quartile’s 61.0% (192/315) is to its own.
• The counting view adds a hump. The bottom household quartile, debt-ratio changes of -30.4 to -0.2pp, is followed by below-median growth in 57.1% of episodes (180/315), a worse share than the third quartile’s 44.1% (139/315): on this panel, heavy deleveraging and heavy leveraging are both followed by weak growth, and the middle of the distribution grows fastest. A regression with a single linear debt-change term would average these two tails against each other.

What this cannot tell you

• The windows overlap. Consecutive episodes within a country share up to 2 of their 3 debt years and 2 of their 3 growth years, so the 1,260 episodes are serially dependent and the counts overstate the number of independent observations; no share on this page carries a standard error for that reason.
• No country fixed effects. The quartiles are pooled across all 18 countries, so countries whose debt ratios swing more, or whose trend growth differs, load unevenly onto the quartiles; the paper’s within-country estimates absorb exactly the level differences this counting design leaves in.
• Annual data. The panel is annual, so each 3-year window is a difference between two annual endpoint observations and sub-year timing between the debt change and the growth response is invisible.
• Advanced economies only. The JST panel covers 18 advanced economies against the paper’s 30-country panel, so the emerging-market part of their sample is untested here.
• Counting is not their estimator. The paper’s claim is about distributed-lag coefficients with controls; a quartile count with none is a weaker, assumption-light check, so agreement here supports the direction of their finding without reproducing its magnitude, and disagreement would not have refuted the regression.