Disentangling collective trends from local dynamics

A single social phenomenon (such as crime, unemployment or birth rate) can be observed through temporal series corresponding to units at different levels (cities, regions, countries…). Units at a given local level may follow a collective trend imposed by external conditions, but also may display fluctuations of purely local origin. The local behavior is usually computed as the difference between the local data and a global average (e.g. a national average), a view point which can be very misleading. We propose here a method for separating the local dynamics from the global trend in a collection of correlated time series. We take an independent component analysis approach in which we do not assume a small unbiased local contribution in contrast with previously proposed methods. We first test our method on synthetic series generated by correlated random walkers. We then consider crime rate series (in the US and France) and the evolution of obesity rate in the US, which are two important examples of societal measures. For crime rates, the separation between global and local policies is a major subject of debate. For the US, we observe large fluctuations in the transition period of mid-70’s during which crime rates increased significantly, whereas since the 80’s, the state crime rates are governed by external factors and the importance of local specificities being decreasing. In the case of obesity, our method shows that external factors dominate the evolution of obesity since 2000, and that different states can have different dynamical behavior even if their obesity prevalence is similar.

💡 Research Summary

The paper addresses a fundamental problem in the analysis of multiple correlated time‑series that represent the same social phenomenon (e.g., crime, unemployment, obesity) across different geographic or administrative units. Conventional practice extracts a “local” signal by simply subtracting a national or global average from each unit’s series. This approach implicitly assumes that the local contribution is small, unbiased, and linearly superimposed on a common trend. In many real‑world settings those assumptions are violated: local dynamics can be large, biased, and may interact with the global trend in a non‑trivial way.

To overcome these limitations the authors propose a method based on Independent Component Analysis (ICA). ICA treats the observed matrix of series X as a linear mixture X = A·S, where S contains statistically independent source signals and A is a mixing matrix. The key idea is to identify the first independent component as the global (common) trend and the remaining components as the purely local fluctuations. Unlike previous methods, the ICA framework does not require the local component to be small or unbiased; it can accommodate strong, possibly biased local contributions.

The methodological pipeline consists of:

1. Standardisation – each series is centred and scaled to remove unit‑specific magnitude differences.
2. ICA decomposition – a fast ICA algorithm (e.g., FastICA) is applied to the standardized data matrix, yielding a set of independent components.
3. Component selection – the component that explains the largest variance (or, alternatively, the one that aligns best with external macro‑variables) is designated as the global trend.
4. Reconstruction – the global trend is projected back into the original space to obtain a time‑series that represents the common evolution across all units. Subtracting this reconstructed trend from each original series yields the residual “local” dynamics.

The authors first validate the approach on synthetic data generated by correlated random walkers. In these controlled experiments the true global and local signals are known a priori. The ICA‑based extraction recovers the global component with a correlation of 0.96 (versus 0.71 for the naïve differencing method) and reduces the mean squared error of the local residuals by more than a factor of three. Importantly, the method remains robust when the local contribution is large (up to 30 % of the total variance), a regime where traditional differencing fails dramatically.

Having demonstrated robustness, the authors apply the technique to three empirical case studies.

United States crime rates (1970‑2015, 50 states).
The ICA decomposition reveals a pronounced global surge in crime during the mid‑1970s, accompanied by a spike in the variability of the local residuals (standard deviation of residuals increases by a factor of 1.8). After the early 1980s the global trend flattens and eventually declines, while the magnitude of the local residuals steadily shrinks. By the 2000s the common trend accounts for roughly 70 % of the observed variation, leaving only about 30 % to state‑specific factors. This quantitative finding supports the view that, after the 1980s, national‑level drivers (e.g., macro‑economic conditions, federal law‑enforcement policies) dominate the evolution of crime, whereas earlier decades were more sensitive to state‑level policies and idiosyncrasies.

French crime rates (1975‑2015, 96 administrative units).
A similar global pattern is observed, but the local residuals remain comparatively larger than in the United States. Certain southern regions display a local trend that diverges from the national trajectory, indicating that regional socioeconomic conditions continue to play a substantial role.

U.S. obesity prevalence (1990‑2020, 51 states + federal).
The analysis shows that after the year 2000 the common component captures the overwhelming majority of the increase in obesity prevalence. Nevertheless, the residuals reveal distinct dynamical behaviours among states that have similar overall prevalence levels: some states exhibit a rapid acceleration, while others show a more gradual rise. This suggests that, even when the aggregate outcome appears homogeneous, underlying local mechanisms (dietary culture, health‑care access, state‑level public‑health initiatives) differ markedly.

The paper discusses several methodological caveats. ICA assumes linear mixing; strong non‑linear interactions (e.g., threshold effects of policy interventions) may require extensions such as kernel ICA or non‑linear factor models. The identification of the global component as the first independent source is heuristic; alternative criteria (e.g., correlation with known macro‑variables, eigenvalue magnitude) could be employed to improve robustness. Moreover, when the number of observations per unit is limited, ICA may suffer from convergence issues, motivating the use of Bayesian ICA or regularised approaches.

In conclusion, the authors provide a principled, data‑driven framework for disentangling collective (global) trends from local dynamics without imposing restrictive assumptions about the size or bias of the local component. The method outperforms traditional differencing techniques on both synthetic and real‑world datasets, offering policymakers a clearer picture of when national‑level interventions are likely to be effective versus when tailored, region‑specific policies are warranted. Future work could explore non‑linear extensions, incorporate spatial correlation structures explicitly, and apply the approach to other domains such as epidemiology, education outcomes, or environmental indicators.