Panel Time-Series & Cross-Sectional Dependence Services
In panels with a long time dimension—countries or firms tracked over many years—units are rarely independent, and effects rarely identical across them. Second-generation panel estimators (CCEMG, AMG, DCCE, CS-ARDL, PMG) handle the cross-sectional dependence and slope heterogeneity that standard panel methods assume away.
Panel time-series methods analyse panels with a large time dimension where units are cross-sectionally dependent (linked by common shocks or factors) and effects may differ across units. Second-generation estimators—CCEMG, AMG, DCCE, and CS-ARDL—correct for cross-sectional dependence, while pooled mean group (PMG) and mean group estimators handle heterogeneous short- or long-run slopes across units.
The problems these methods solve
Standard (“first-generation”) panel estimators make two assumptions that often fail in macro-panels—countries, industries, or firms observed over long spans. The first is cross-sectional independence: that units are unrelated to one another. In reality, common shocks—global cycles, oil prices, financial crises, technology waves—hit many units at once, making their outcomes cross-sectionally dependent. Ignoring this dependence produces biased estimates and badly overstated significance. The second is slope homogeneity: that a predictor’s effect is identical across all units. Across diverse countries or firms, that is frequently untrue, and imposing a common slope misrepresents the relationship.
Panel time-series methods (the “second generation”) are built for exactly these conditions—panels where the time dimension is large enough to treat each unit as its own time series while still pooling information across units. They test for and correct cross-sectional dependence, and they allow effects to vary across units, so the estimates reflect a genuinely heterogeneous, interconnected panel rather than an idealised one.
The estimators, and when each fits
Several complementary estimators address these issues. The common correlated effects (CCE) approach—including CCEMG (mean group) and its dynamic extension DCCE—controls for cross-sectional dependence by augmenting each unit’s regression with cross-sectional averages that proxy the unobserved common factors. The augmented mean group (AMG) estimator handles dependence through a common dynamic process and also allows heterogeneous slopes. For long-run relationships in dependent panels, CS-ARDL combines the ARDL approach with cross-sectional augmentation.
On the heterogeneity side, the pooled mean group (PMG) estimator is widely used for panel error-correction models: it constrains the long-run relationship to be common across units while letting the short-run dynamics differ—a middle ground between fully pooled and fully heterogeneous (mean group) estimation, appropriate when theory implies a shared long-run equilibrium. Which estimator fits depends on the nature of the dependence, whether slopes are homogeneous, and whether the interest is short-run, long-run, or both—choices we make on the basis of the relevant tests, not habit.
Second-generation panel estimators
| Estimator | Handles | Best for |
|---|---|---|
| CCEMG | Cross-sectional dependence + heterogeneous slopes | Static relationships in dependent panels |
| DCCE | Dependence in dynamic panels | Dynamic models with a lagged outcome |
| AMG | Dependence via a common dynamic process | Heterogeneous panels with a common trend |
| CS-ARDL | Long-run relationships + dependence | Cointegration in dependent panels |
| PMG / MG | Slope heterogeneity | Common long-run, heterogeneous short-run (PMG) |
Testing first, then choosing the estimator
The defining discipline of panel time-series work is that the estimator is chosen on the basis of tests, not assumed. The first step is a cross-sectional dependence test (such as the Pesaran CD test): if dependence is present, first-generation estimators and first-generation unit-root tests are invalid, and second-generation methods are required. Where dependence matters, second-generation panel unit-root tests (for example, CIPS) establish the integration properties, and slope-homogeneity tests indicate whether a pooled or a mean-group approach is appropriate.
Only once these properties are established is the estimator selected—CCEMG, DCCE, AMG, CS-ARDL, or PMG/MG—to match the dependence structure, the integration orders, and the homogeneity finding. Because these methods rely on the time dimension to estimate unit-specific dynamics, an adequate time span matters, and we are explicit about the demands and limits of the data. Where a long-run relationship is claimed, we report the appropriate panel cointegration evidence and the error-correction adjustment. Reported honestly, this sequence is what distinguishes a credible dependent-panel analysis from a first-generation model applied where its assumptions do not hold.
Cross-sectional dependence invalidates standard panel methods—so test for it first. If units share common shocks, first-generation estimators and unit-root tests are biased and overstate significance. The CD test, second-generation unit-root tests, and a slope-homogeneity test decide which estimator is valid—before any model is fitted.
Software
We deliver second-generation panel estimation in established, reproducible tools—Stata (xtcce/xtdcce2, xtmg, xtpmg) and R—with cross-sectional dependence testing, second-generation unit-root and cointegration tests, slope-homogeneity tests, and the chosen estimator with its diagnostics, all with versioned code.
How we deliver a panel time-series study
This work sits within our wider econometrics practice—so dependence and heterogeneity are tested first, and the estimator is matched to what the data actually show.
We start by testing for cross-sectional dependence and, where present, using second-generation unit-root and cointegration tests to establish the integration properties, along with a slope-homogeneity test. We then select the estimator—CCEMG, DCCE, AMG, CS-ARDL, or PMG/MG—that matches the dependence, integration, and homogeneity findings and the short-run/long-run focus of the question.
Reporting sets out the dependence and unit-root/cointegration tests, the homogeneity finding, the estimator and why it was chosen, and the long-run and short-run (error-correction) results—so the analysis can be judged against its assumptions.
You receive the estimates from the chosen second-generation estimator with appropriate inference, the cross-sectional dependence, unit-root, cointegration, and slope-homogeneity test results, the long-run coefficients and error-correction dynamics where relevant, robustness across estimators, and reproducible analytical code and analysis-ready files (where appropriate and permitted).
Panel time-series across Management & Allied Studies
Long macro-panels of countries, industries, and firms are typically both interconnected and heterogeneous—precisely the conditions these estimators are built for—so they are widely used across the quantitative disciplines we serve.
Economics & Public Policy
Cross-country panels for growth, trade, and policy where common global shocks and heterogeneous effects are the norm. A core setting.
Energy & Environmental Economics
Country panels on energy, emissions, and growth—a very common application of CCEMG, AMG, and CS-ARDL.
Finance & Financial Markets
Cross-country or cross-market panels where common financial shocks link units and effects differ across them.
Development & International Economics
Heterogeneous country panels with shared shocks, where pooled first-generation estimates would mislead.
Banking & Monetary Economics
Cross-country banking and monetary panels with common cycles and country-specific dynamics.
Industry & Regional Studies
Industry- or region-level panels over long spans where common trends and heterogeneous responses coexist.
Panel time-series: common questions
A long macro-panel with common shocks?
When countries or firms in your panel share global shocks and respond differently, second-generation estimators give valid results where standard panel methods fail—with dependence and heterogeneity tested first, and the estimator matched to the data.