RainbowStats Research Note

When One Slope Isn't Enough

How tax-period LP regressions can sharpen our inspection of corporate profits, inequality, and labor share—sometimes greatly, and sometimes just enough to keep us honest.

Open the 31-slide live analysis →

A standard regression asks for one relationship that works everywhere. Public policy rarely returns the favor. Tax rules change, institutions react, and the economy keeps adding plot twists as if it were being paid by the episode.

Justice Potter Stewart once quipped, “I know it when I see it.” The same can be said for correlation.

Unfortunately, a convincing-looking line can stroll into a chart wearing a very respectable suit. The useful question is not merely whether two series move together, but whether their relationship survives changes in the policy environment—and whether a more flexible description earns its extra complexity.

To investigate, we compare ordinary least squares (OLS) with a constrained linear-programming regression at the median. The predictor is U.S. corporate profits after tax divided by gross national product (CP/GNP). The outcomes are the U.S. Gini index and private nonfarm business labor share. OLS supplies the clean baseline. The LP model lets the profit coefficient adjust during selected federal tax-policy periods.

OLS is the baseline, not the villain

OLS estimates the single line that minimizes squared residuals. That makes it simple, familiar, and deliberately unforgiving of large errors. Its one-slope discipline is often a feature. But if a policy regime changes how a predictor relates to an outcome, a full-sample average can hide that change in plain sight.

The LP model used here estimates a median relationship by minimizing absolute deviations. It adds a period-specific adjustment to the corporate-profits coefficient:

yₜ = α + βxₜ + Σᵣ δᵣxₜI(t ∈ r) + εₜ    and    minimize Σₜ|εₜ|

Here, β is the base slope and each δ changes that slope only inside tax period r. The Gini model also requires every effective period slope to remain nonnegative. The labor-share model uses a broad lower bound of −100; it allows negative slopes while preventing an extreme solution.

The key test: a richer model will almost always fit at least a little better. Adjusted R² makes it pay rent for the extra period coefficients. If the adjusted measure still rises, the gain is harder to dismiss as mere flexibility.

The result in one picture

Grouped bars comparing adjusted R-squared for OLS and tax-period LP models of Gini and labor share
Adjusted R² rises from 0.283 to 0.438 for Gini and from 0.888 to 0.898 for labor share. Period-adjustment parameters are included in the complexity penalty.
OutcomeModelObservationsAdjusted R²MAE
GiniOLS620.3070.2831.567
GiniTax-period LP620.5210.4380.959
Labor shareOLS380.8940.8880.00833
Labor shareTax-period LP380.9120.8980.00652

For Gini, the gain is substantial: adjusted R² improves by 0.155 and mean absolute error falls by 38.8%. For labor share, OLS already explains most of the variation, so LP produces a smaller adjusted-R² gain of 0.010 while reducing mean absolute error by 21.8%.

Grouped bars showing indexed mean absolute error for OLS and tax-period LP models
Each outcome's OLS mean absolute error is indexed to 100 so the differently scaled outcomes can share a chart. LP at the median directly minimizes absolute deviations, so this comparison should be read alongside adjusted R²—not as an independent trophy.

Gini: the average slope leaves history on the table

Adjusted R²0.283 → 0.438+0.155 after the penalty
Mean absolute error1.567 → 0.95938.8% lower

The OLS baseline estimates a positive average association between CP/GNP and Gini. The tax-period LP fit preserves a positive base relationship while allowing its strength to vary. One period—the 1976–1978 regime—lands exactly at the zero-slope boundary. Other periods range from roughly 24.9 to 77.8 Gini-index units per unit of the CP/GNP ratio.

That is more informative than declaring one eternal slope and going home early. It says the association is not equally strong in every policy window. It does not say which tax law caused what; history has more confounders than a family group chat.

Annual U.S. Gini index and fitted tax-period LP values with shaded tax-policy periods from 1963 to 2024
Annual Gini observations and tax-period LP fitted values, 1963–2024. Shading identifies the modeled policy periods; numbers correspond to the regime table below.
RegimeStartEndGini effective slope
11954-08-161968-06-2824.923
21969-12-301971-12-1030.739
31976-10-041978-11-060.000 (binding)
41982-09-031984-07-1836.318
51986-10-221990-11-0557.961
61993-08-102001-06-0777.809
72003-05-282013-01-0260.887
82017-12-222022-08-1655.650

Labor share: a smaller gain can still matter

Adjusted R²0.888 → 0.898+0.010 after the penalty
Mean absolute error0.00833 → 0.0065221.8% lower

The labor-share baseline is already strong. Its OLS slope is negative (−1.207), and the LP period-effective slopes remain negative, ranging from about −1.00 to −1.24 across the four policy windows represented in the 1987–2024 sample.

Here LP is a refinement rather than a rescue. That distinction matters. A method worth trusting should be able to say “small improvement” without reaching for a fog machine.

Annual U.S. private nonfarm business labor share and fitted tax-period LP values with shaded policy periods from 1987 to 2024
Annual private nonfarm business labor share and tax-period LP fitted values, 1987–2024. Only policy periods overlapping this shorter sample are modeled.

What improved—and what did not

The exercise supports a practical claim: LP regression can improve inspection of policy changes by exposing period-specific relationships that a single OLS coefficient averages away. The Gini result shows that the improvement can be large. The labor-share result shows that it need not be.

But this is still descriptive modeling. Tax-policy periods coincide with recessions, monetary changes, demographic shifts, trade shocks, technology, transfers, and many other forces. Annual frequency and same-date alignment simplify timing. Neither a higher adjusted R² nor a lower error establishes a causal tax effect. For causal claims, the next step would require a research design—credible counterfactuals, controls, timing assumptions, and robustness tests—not merely a more charming regression.

A note on fit measures: OLS minimizes squared error; median LP minimizes absolute error. MAE therefore speaks the LP model's native language. Adjusted R² is reported as a descriptive comparison and includes the added period-adjustment coefficients, but it does not turn a median regression into OLS or grant causal immunity.

Reproduce the analysis

The complete RainbowStats slideshow contains both OLS baselines, the LP model summaries, coefficients, fitted values, residual diagnostics, shadow prices, and annotated actual-versus-fit charts.

RainbowStats script Model comparison data Gini actual and fitted data Labor-share actual and fitted data LP coefficient data Tax-regime data Chart-generation code

Sources and definitions

The predictor uses Corporate Profits After Tax (CP) divided by Gross National Product (GNP). Outcomes are the World Bank's GINI Index for the United States and the BLS Private Nonfarm Business Sector: Labor Share, all accessed through FRED. The Stewart line refers to his concurrence in Jacobellis v. Ohio (1964). Calculations and graphics are reproduced from the linked RainbowStats analysis.