Telling Coupling from Composition in Size–Volatility Scaling

Large organizations fluctuate less than small ones, but not by as much as they would if they were bags of independent parts. If a firm, a country or a city were a sum of independent pieces, the spread of its yearly growth rates would shrink with size $S$ roughly as $S^{-1/2}$. Measured exponents are instead near 0.15 for national economies and 0.25 for U.S. manufacturing firms. For two decades this gap has been read as evidence that the parts are causally coupled, with shocks spreading through internal hierarchies. There is a competing explanation. Independent processes that start at different times and have very unequal sizes can produce the same shallow exponent over the finite range of sizes we can actually observe. The size bins at the large end are also so thin that their volatility estimates are noisy and biased.

This note proposes a test that separates the two explanations. The test uses one cross-section: each entity's size, age and one-period growth rate. It combines an independence null model calibrated to the data, which predicts a specific finite-size crossover toward the $S^{-1/2}$ exponent, with age-stratified exponent estimates that the two explanations predict to behave differently. The test is validated where longitudinal ground truth exists. The same question comes up whenever someone reads a causal claim ("the parts interact") off a single-time size distribution.

Core Claim

A cross-sectional exponent below 0.5 is not, by itself, evidence of coupling. It becomes evidence only once two things have been ruled out: that heterogeneous entry times mix the size bins, and that finite size makes the independent-parts limit invisible over the observed range. Both are testable from snapshot data. Consider an entity whose growth is the sum of independent subunit shocks with common variance $\sigma_0^2$. Its growth-rate variance is

$$\sigma^2(S) = \sigma_0^2 \, \mathbb{E}[H \mid S], \qquad H = \sum_i (s_i/S)^2,$$

where $s_i$ are subunit sizes and $H$ is their concentration. Under independence, any apparent exponent comes entirely from how concentration falls with size. That depends on the entity's age, because older entities have accumulated more subunits. This gives two predictions. First, within a narrow age stratum the exponent should move toward 0.5 once entities hold many subunits, so the curve should bend at large $S$. Second, the pooled exponent should shift when the mixture of ages across size bins is changed. Genuine coupling predicts the opposite: the exponent is stable across age strata and there is no crossover. The contribution is a likelihood-based test of these predictions that accounts for thin bins at large size.

Where This Sits

The coupling reading comes from Universal Features in the Growth Dynamics of Complex Organizations. That paper fits log-log volatility scaling and a data collapse across 152 countries ($\beta = 0.15 \pm 0.03$) and argues that the exponent rules out independent subunits. Universal fluctuations in growth dynamics of economic systems extends this to Compustat firms ($\beta = 0.25 \pm 0.02$ and $0.26 \pm 0.01$). It also shows that young sectors such as biotechnology have no clean power law at first (regression standard error 0.2 in 1981–1986) and acquire one later (0.02 by 2005–2010). That age dependence is exactly what a composition account would predict, and the paper leaves it unexplained. These two papers motivate the proposal and supply its main targets.

The challenge comes from How Do Online Social Networks Grow?. It shows that apparent long-range correlations in city-level user growth can be reproduced by superposing independent processes with staggered entry. We take their superposition idea as a starting point, but they reproduce the pattern qualitatively and never test it against coupling. We turn it into a calibrated null with a decision rule.

The finite-size side constrains how the estimates are read. Scale-Free Graph with Preferential Attachment and Evolving Internal Vertex Structure shows a degree exponent sitting at 4.92 after $10^9$ vertices when its true limit is 5. That is a warning that an exponent stable over the observed range can still be a transient. MSE-optimal adjustment sets in linear Gaussian causal models with finite sample size shows that estimators that are optimal in the large-sample limit can lose to biased ones at small $n$. This matters for how we weight thin large-size bins.

Causal-discovery methods frame the ground truth but cannot answer the question directly. PLaCy (Robust Causal Discovery in Real-World Time Series with Power-Laws) orients causal edges from the evolution of power-law spectral parameters, but it needs long time series. Foundations of Testing for Finite-Sample Causal Discovery gives finite-sample-valid edge tests, but assumes a known skeleton and Gaussian data. Neither applies to snapshots.

Research Program

The first component is a generative testbed with known answers. The independence family draws entity entry times from an entry-rate distribution and adds subunits over time. Subunit sizes come from heavy-tailed distributions, and each subunit gets independent shocks. The coupling family runs the same processes but propagates shocks along a hierarchy, with a tunable coupling strength. Both families are calibrated so that their marginal size distributions and their pooled log-log exponents match the empirical 0.15–0.26 range. By construction, the naive test then cannot tell them apart.

The second component is the estimator. We fit $\sigma(S)$ bin by bin within age strata by maximum likelihood, using the Laplace-like, tent-shaped growth-rate distribution the collapse literature reports rather than a Gaussian. The model includes an explicit crossover term that allows the local exponent to bend toward 0.5 above a fitted scale. Bin-level sample sizes enter the likelihood, so thin large-size bins count as uncertain rather than as clean points.

The decision statistic compares two models: a crossover-plus-age-dependence model, and a single-exponent model with no age dependence. Its null distribution comes from a parametric bootstrap under the calibrated independence model, which avoids relying on large-sample test asymptotics that the heavy tails would undermine.

The third component is validation against longitudinal ground truth and application to the data in the file. On Compustat, we compute the test from single-year cross-sections. We then check its verdict against panel quantities: within-firm volatility trajectories as firms age, and whether sector-level shocks show directed lagged propagation. On the Gowalla, Renren and Twitter city data, the superposition account is already plausible, so the test should favor independence. Where that account fails, it should say so. On the Penn World Table, the test addresses the original claim for countries.

Baselines

The naive reference point is the published practice. We fit the pooled log-log regression of $\sigma$ on $S$ by ordinary least squares, conclude coupling whenever $\beta$ is significantly below 0.5, and check the data collapse visually. The benchmarks for this reference are the reported exponents: $\beta = 0.15 \pm 0.03$ for countries (Amaral et al.) and $0.26 \pm 0.01$ for U.S. manufacturing firms from 1993 to 2015 (Matin et al.). We rerun it on exactly the same cohorts and bins that our method uses. On the synthetic testbed, we expect it to classify every calibrated independence model as coupled.

The strongest existing competitor is the superposition-reproduction approach of Zhu et al.: fit a heterogeneous-entry independent model and ask whether it reproduces the observed statistics. We give it the same entry-rate model family and the same calibration budget as our null. We score it by the rule "coupled if the independent model fails to reproduce $\beta$ within its confidence interval." No reported result was found on Lacuna for any coupling-versus-superposition classification accuracy. Beating this competitor therefore means a higher area under the ROC curve for recovering the true generating family on the synthetic testbed, at matched false-positive rates. It also means higher agreement with the panel-derived verdict on Compustat sectors.

For longitudinal subsets, PLaCy is a reference for what directed-propagation analysis finds when time series exist. The file reports that PLaCy outperforms PCMCI, DYNOTEARS and Rhino on the Rivers data, but gives no reported F1 values. We use it only as a ground-truth proxy, not as a competitor.

Two ablations isolate the key ingredients. One removes the crossover term but keeps age stratification. The other keeps the crossover term but pools across ages. Each is compared against the full test.

Why not just…?

Why not just use longitudinal panels and panel regression or lagged autoregressions? Where long panels exist, they are the better tool, and we use them as ground truth. The case for a cross-sectional test rests on settings where panels are short or absent: nascent sectors, city-size snapshots, and historical cross-sections. Wu et al. needed five years and 20,218 person-years to settle direction for a handful of blood markers, which shows how costly the longitudinal route is. If Compustat panel results show that short panels of three to five years already classify as well as our cross-sectional test, the proposal's practical value shrinks to the snapshot-only settings. We will report that honestly.

Why not just fit the exponent more carefully, with maximum likelihood and bootstrap intervals? Tighter intervals around a misspecified single-exponent model do not help. A calibrated independence model can produce $\beta \approx 0.2$ with a narrow interval, and the graph-growth result of Matuszak et al. shows exponents that stay far from their limit over nine decades of size. The first ablation measures how much of the gain comes from modeling the crossover rather than from better fitting.

Why not just stratify by age and refit the slope? This is the real threat, because age stratification alone captures the composition half of the argument. We therefore make it an ablation arm rather than dismissing it. We expect it to fail in thin strata, where each age band has few large entities and the crossover region is poorly sampled. That is exactly where the size-aware likelihood should help. If plain stratification matches the full test on both the synthetic and the Compustat validation, we will adopt it as the recommended practice and report the crossover model as unnecessary.

Decisive Checks

The first check is synthetic. At matched pooled exponents, the full test must separate the independence and coupling families substantially better than both baselines. It must also degrade gracefully as the number of entities per bin falls to realistic Compustat and country counts; there are only 152 countries.

The second check is a prediction about biotechnology. The rise and stabilization of the biotech power law reported by Matin et al. should be accounted for by the composition model's age dynamics without invoking coupling. If it is not, that counts as evidence that coupling is real there.

The third check targets the exceptions. Utilities and finance, which deviate from the pattern in the same paper, should be where the two verdicts diverge most clearly.

A result showing that the country-level $\beta = 0.15$ survives the calibrated independence null would strengthen the original coupling claim rather than overturn it. That outcome is informative too.