Beta and Sigma Convergence — Interactive Lab

A pedagogical companion to Beta and Sigma Convergence Across Countries: A Stata Tutorial ↗ Back to the post

Are poorer countries catching up?

For decades the answer was no. From 1960 to 2000 richer countries grew faster, and the world's income gaps widened. After about the year 2000 the sign flipped — poorer countries finally grew faster on average. This is the "new era of unconditional convergence" documented by Patel, Sandefur, and Subramanian (2021). But convergence is real, slow, and counter-intuitive: even after the 2000 sign flip, the actual spread of world incomes kept growing until 2008.

This app lets you turn the dials yourself. Across four tabs you will simulate a Solow-style cross-section and watch β-convergence emerge from the speed parameter; explore the post's actual rolling-β estimates as the start year slides from 1960 to 2010; and step through the year-by-year income variance to see the 8-year lag between β-convergence and σ-convergence in the data.

β vs σ — two views of the same race

Both views ask "is the gap closing?" but measure it differently. β-convergence watches whether the country that started behind is now running faster. σ-convergence watches the width of the pack. The animation below contrasts the two: the blue runner (rich) and the orange runner (poor) close on each other (β-convergence), while the dashed bar in the middle tracks the variance of log income across a small set of countries — the variance can stay flat or even rise while individual runners are converging, because random shocks redistribute incomes even as the catch-up force operates.

Tab 2

Solow Simulator

Set the convergence speed β and the noise level. Watch a fresh cross-section of 84 simulated countries grow over 19 years. See the OLS scatter and the implied half-life update live.

Tab 3

Rolling β

The post's actual rolling-β estimates — slide the start year from 1960 to 2010 and watch the coefficient cross zero around 1990 and stabilise near +0.0036 for the recent windows.

Tab 4

σ Evolution

Sixty years of cross-country income variance from the post's balanced 84-country panel. The peak is in 2008, eight years after β-convergence begins — the lag is visible at a glance.

Glossary (open a card if a term is unfamiliar)

β-convergence
Poorer countries grow faster than richer ones. Tested by regressing annualised growth on log initial income; a negative slope λ (or a positive structural β) is convergence.
σ-convergence
The variance of log income across countries shrinks over time. The actual gap is narrowing.
Slope λ vs. speed β
λ is the raw OLS slope (depends on period length s). β is the structural speed: β = −ln(1 + λs)/s. Different sign convention — negative λ corresponds to positive β.
Half-life τ
Years to close half the existing income gap: τ = ln(2) / β. For 2000–2019, β = 0.00365 gives τ = 190 years.
Structural break
A point in time where the convergence coefficient changes sign or magnitude. In this data the break sits around 2000.
Rolling window
Re-estimate the regression for every possible start year (end year fixed at 2019). The sequence of estimates traces out how convergence has evolved.
NLS
Nonlinear Least Squares. Estimates β directly without going through λ. Gives identical point estimates to OLS conversion in this setting.
Why β alone is not enough
Young, Higgins, and Levy (2008) showed β-convergence is necessary but not sufficient for σ-convergence. Random growth shocks can widen the distribution even as poor countries grow faster on average.

Solow Simulator — generate a cross-section and watch β-convergence emerge

Each "country" is drawn with a random initial log-income and a true steady-state. Growth over the next T years is governed by the Barro–Sala-i-Martin convergence equation:

gi = α − [(1 − e−βT)/T] · ln(yi,0) + εi

Drag β (the true convergence speed) and σ (the noise) and watch the OLS fit on the simulated data update live. Compare the recovered λ̂ and half-life to the values you set.

Default 84 matches the post's balanced panel.
Default 19 matches the 2000–2019 window from the post.
0.0036 ≈ post's 2000–2019 estimate. 0.020 ≈ classic conditional benchmark. Negative ⇒ divergence.
Standard deviation of the annualised growth shock. Larger ⇒ noisier scatter.
Recovered OLS λ̂
slope of growth on log initial
Implied β̂
−ln(1 + λ̂T)/T
True β (you set)
0.0040
target speed
Implied half-life
τ = ln(2)/β̂ (years)

What to look for

  • Pull β positive and watch the fitted line tilt downward in the scatter: poorer countries grow faster. Pull β negative and the line tilts upward — the divergence regime of 1960–2000.
  • Notice the recovered β̂ is noisy. Even when the true β is the 2%/year benchmark, a single 19-year cross-section recovers it imprecisely. This is the §6 lesson: unconditional convergence is real but statistical power is limited.
  • Shrink the noise σ and the OLS line snaps to its true slope. Real cross-country growth has σ ≈ 0.015 — the deck is stacked against detecting convergence from one window alone.
  • Press the reseed button to draw a fresh sample with the same parameters. The variation across draws is the standard error you see in the post's regression tables.

The post's rolling-β plot — interactively

These 51 estimates come straight from convergence_rolling_beta_ols.csv in the post's folder — the same data behind Figure 6. Each dot is one OLS regression of annualised growth on log initial income, with start year on the x-axis and end year fixed at 2019. Use the highlight slider to read out a single window; toggle the convergence-benchmark line to compare to the 2% conditional-convergence baseline.

Slide to read out one rolling window. The window covers start → 2019.
The classic Barro–Sala-i-Martin (1992) conditional-convergence rate.
The width of the CI shrinks as the sample grows but stays wide at every start year.
β at selected start year
structural speed (per year)
Speed (% per year)
100 × β
Half-life
years (ln 2 / β)
95% CI
[lower, upper]

What to look for

  • Slide the highlight to 1960: β = −0.00056. The post's headline "no convergence over 1960–2019" sits right there. The estimate is statistically indistinguishable from zero.
  • Slide to 1990: β crosses zero. This is the moment the new era of unconditional convergence becomes visible in the rolling plot.
  • Slide to 2000: β = +0.00365, half-life = 190 years. Statistically significant (CI excludes zero) — the post's headline result.
  • Slide to 2008: β peaks at +0.00525, the strongest convergence in the data. After this the post-financial-crisis estimates moderate.
  • Compare every window to the 2% benchmark: the dashed teal line shows where Barro–Sala-i-Martin's 1992 conditional-convergence rate sits. Unconditional convergence is real but five-to-six times slower.

σ evolution — when does the actual gap start closing?

The year-by-year variance of log GDP per capita across the post's 84-country balanced panel (convergence_sigma_evolution.csv). β-convergence began around 2000; σ-convergence did not start until 2008. This 8-year lag is the §13 message of the post made visible at a glance.

Slide to read out one year's variance and 95% CI.
Vertical line marking the year β-convergence begins.
Vertical line marking the variance peak.
Chi-squared CI for a variance under N = 84.
Variance at selected year
var(ln y) across 84 countries
Standard deviation
√variance
% above 1960
vs. base 0.924
95% CI for variance
chi-squared, N = 84

What to look for

  • From 1960 to 2008 the variance rises almost monotonically: 0.924 → 1.918, a 108% widening. The world income distribution was getting wider, not narrower.
  • The variance peaks in 2008 at 1.918, then declines. This is when σ-convergence finally begins — eight years after β-convergence.
  • By 2019 the variance is 1.764 — still 91% above the 1960 level. Sigma-convergence since 2008 has only undone a small fraction of the prior divergence.
  • Toggle the 2000 and 2008 markers off to see the smooth shape of the curve. Toggle them on to spot the lag visually.
  • Connect back to Tab 3: in the rolling-β plot the coefficient turns positive around 1990; in the σ plot the variance does not turn down until 2008. Both pictures together show why §13 calls β necessary but not sufficient.