Case Study — Detecting a Real Decoupling: GDP per Capita vs Median Household Income

The decoupling and ground-truth tools (decoupling_monitor, ground_truth_auditor) run on real, published US economic data instead of a constructed series, doing on genuine data exactly what they do in the synthetic demos. Runnable: real_data_gdp_vs_income.py.

---

Context and question

"The economy grew" is usually stated as one number: real GDP per capita. But GDP per capita is a proxy for something people actually care about — how the typical household is doing — and the closer truth for that is real median household income. The proxy-vs-truth question at the center of the whole toolkit, asked of real data: do the two move together, or does the headline number drift from the lived reality it is taken to represent?

The data (real, cited)

• Proxy — US real GDP per capita, chained 2012 dollars, annual 2000–2019 (Bureau of Economic Analysis).
• Truth — US real median household income, constant dollars, annual 2000–2019 (US Census Bureau / FRED series MEHOINUSA672N).

Both are indexed to 2000 = 100 inside the tools so the different base years drop out and the series are directly comparable. One honest caveat carried in the code: the 2019 median-income jump partly reflects the 2019 CPS ASEC processing change, not only real gains — but the core 2000–2013 decoupling does not depend on that year.

Method

Feed the two aligned series to decoupling_monitor (which flags when the proxy rises while the gap to an independent truth opens) and then to ground_truth_auditor (which asks the prior question: is the "truth" actually independent of the proxy, or merely its shadow?). No parameter was tuned to force a verdict; the co-movement window is disclosed in the code.

Findings

A sustained, real decoupling. decoupling_monitor classifies the relationship as DRIFTING from 2002 onward: the proxy kept rising while the gap to the truth widened. Indexed to 2000, real GDP per capita reached ~115 by 2014 while real median household income sat near ~96 — a peak gap of about 19 index points (19.2 in 2014). Put plainly: GDP per capita had grown ~15% while the median household was about 4% below where it started. Under a tighter co-movement window the monitor trips a strict DECOUPLED alert at the early-2010s trough (2012). The drift is robust across windows; only the sharper "DECOUPLED" label is window-sensitive, and that sensitivity is disclosed rather than hidden.

The independence check refuses to overclaim. ground_truth_auditor asks whether median income is just a restatement of GDP. Verdict: UNVERIFIED, with 51% of median income's variance not explained by GDP. Read honestly, that is the correct two-sided answer: median income is not a shadow of GDP (it carries real, independent information — which is exactly why the two can diverge), and full independence still cannot be confirmed without a labeled reference. The tool declines to certify in either direction.

Interpretation

Nobody "gamed" GDP here — this is not manipulation, and that is the point. It is the quieter, more general failure the toolkit targets: a single reported number drifting from the reality it is taken to represent, so that "the economy is growing" and "the typical household is falling behind" were both true at once for over a decade. The same shape recurs wherever a proxy is trusted without an independent check — a peg drifting from its backing, "same price, worse product" inflation, a gamed benchmark score — and it is caught the same way: watch the proxy against an independent truth, and first verify that the truth is genuinely independent.

Honest limits

• The tools describe a divergence; they do not explain it. Why median income lagged (labor-share decline, inequality, healthcare/housing costs, measurement choices) is an economics question the tools do not answer — they detect and refuse to overclaim, nothing more.
• The "DECOUPLED" label is window-sensitive; the robust finding is the drift, which holds across windows. This is stated, not smoothed over.
• Independence is UNVERIFIED, not CONFIRMED — by design. A labeled reference would be needed to confirm it, and the tool says so rather than pretending.
• Single dataset, single country, annual resolution. The result generalizes the method, not a universal economic claim.

Lessons

Run on real, cited data, the tools do exactly what they promise on synthetic inputs — detect a genuine proxy/truth divergence and refuse to certify the ground truth they can't confirm. The case shows the discipline's value precisely where it is least dramatic: no villain, no gaming, just a headline number and a lived reality that quietly came apart — the kind of decoupling that stays invisible until something independent is measured against it.

Gergely Vámossy (2026). Case Study — Detecting a Real Decoupling: GDP per Capita vs Median Household Income. https://vamossy.com/research/case-study-gdp-income-decoupling