Dynamic Spatial Durbin Model (DSDM)

• DSDM is a panel econometric approach that extends classical spatial models by incorporating dynamic temporal dependence and multiple channel spillovers.
• It simultaneously models contemporaneous and lagged effects of outcomes and treatments to robustly analyze spatial and network interactions.
• Estimation methodologies such as MLE, QMLE, and Bayesian MCMC provide reliable parameter identification and enable decomposition of direct and indirect impacts.

The Dynamic Spatial Durbin Model (DSDM) is a panel econometric framework that extends classical spatial econometric models by allowing for (i) dynamic temporal dependence, (ii) multiple channels of cross-sectional or network spillovers, and (iii) flexible modeling of contemporaneous and lagged effects of both outcomes and key covariates. The DSDM is particularly suited for the analysis of spatial or network interactions under temporal persistence, as illustrated in recent applications to GenAI adoption and productivity in the U.S. banking sector (Kikuchi, 2 Feb 2026), and in the broader econometric theory literature (Kuersteiner et al., 2018).

1. Formal Specification and Key Features

A canonical DSDM for $N$ cross-sectional units (e.g., banks) observed over $T$ time periods is given by

$$y_{it} = \tau\,y_{i,t-1} + \rho\;\sum_{j=1}^N w_{ij}y_{jt} + \eta\;\sum_{j=1}^N w_{ij}y_{j,t-1} + \beta\,D_{it} + \theta\;\sum_{j=1}^N w_{ij}D_{jt} + \gamma'X_{it} + \mu_i + \delta_t + \varepsilon_{it}$$

where:

• $y_{it}$: outcome for unit $i$ at time $t$ (e.g., ROA or ROE).
• $D_{it}$: focal binary or continuous treatment (e.g., GenAI adoption).
• $X_{it}$: $K$-vector of exogenous controls.
• $W = [w_{ij}]$: row-normalized $T$0 spatial or network weight matrix.
• $T$1, $T$2: unit and time fixed effects.
• $T$3: idiosyncratic error.

The model coefficients are:

• $T$4: temporal autoregressive parameter (lag persistence).
• $T$5: spatial lag coefficient (contemporaneous dependence).
• $T$6: spatial-temporal lag (past-neighbor effect).
• $T$7: direct within-unit effect of $T$8.
• $T$9: contemporaneous network spillover from $$y_{it} = \tau\,y_{i,t-1} + \rho\;\sum_{j=1}^N w_{ij}y_{jt} + \eta\;\sum_{j=1}^N w_{ij}y_{j,t-1} + \beta\,D_{it} + \theta\;\sum_{j=1}^N w_{ij}D_{jt} + \gamma'X_{it} + \mu_i + \delta_t + \varepsilon_{it}$$0.

This structure generalizes to richer panel settings including interactive effects, endogenous weights, and higher-order lags (Kuersteiner et al., 2018).

2. Spatial Weight Matrix Construction

The weight matrix $$y_{it} = \tau\,y_{i,t-1} + \rho\;\sum_{j=1}^N w_{ij}y_{jt} + \eta\;\sum_{j=1}^N w_{ij}y_{j,t-1} + \beta\,D_{it} + \theta\;\sum_{j=1}^N w_{ij}D_{jt} + \gamma'X_{it} + \mu_i + \delta_t + \varepsilon_{it}$$1 encodes the structure of economic, competitive, or geographical relationships. In (Kikuchi, 2 Feb 2026), two main specifications are used:

• Network-based (asset similarity kernel):

$$y_{it} = \tau\,y_{i,t-1} + \rho\;\sum_{j=1}^N w_{ij}y_{jt} + \eta\;\sum_{j=1}^N w_{ij}y_{j,t-1} + \beta\,D_{it} + \theta\;\sum_{j=1}^N w_{ij}D_{jt} + \gamma'X_{it} + \mu_i + \delta_t + \varepsilon_{it}$$2

where $$y_{it} = \tau\,y_{i,t-1} + \rho\;\sum_{j=1}^N w_{ij}y_{jt} + \eta\;\sum_{j=1}^N w_{ij}y_{j,t-1} + \beta\,D_{it} + \theta\;\sum_{j=1}^N w_{ij}D_{jt} + \gamma'X_{it} + \mu_i + \delta_t + \varepsilon_{it}$$3 is average assets of unit $$y_{it} = \tau\,y_{i,t-1} + \rho\;\sum_{j=1}^N w_{ij}y_{jt} + \eta\;\sum_{j=1}^N w_{ij}y_{j,t-1} + \beta\,D_{it} + \theta\;\sum_{j=1}^N w_{ij}D_{jt} + \gamma'X_{it} + \mu_i + \delta_t + \varepsilon_{it}$$4 and $$y_{it} = \tau\,y_{i,t-1} + \rho\;\sum_{j=1}^N w_{ij}y_{jt} + \eta\;\sum_{j=1}^N w_{ij}y_{j,t-1} + \beta\,D_{it} + \theta\;\sum_{j=1}^N w_{ij}D_{jt} + \gamma'X_{it} + \mu_i + \delta_t + \varepsilon_{it}$$5 is the cross-sectional standard deviation of $$y_{it} = \tau\,y_{i,t-1} + \rho\;\sum_{j=1}^N w_{ij}y_{jt} + \eta\;\sum_{j=1}^N w_{ij}y_{j,t-1} + \beta\,D_{it} + \theta\;\sum_{j=1}^N w_{ij}D_{jt} + \gamma'X_{it} + \mu_i + \delta_t + \varepsilon_{it}$$6.

• Alternative specifications: $$y_{it} = \tau\,y_{i,t-1} + \rho\;\sum_{j=1}^N w_{ij}y_{jt} + \eta\;\sum_{j=1}^N w_{ij}y_{j,t-1} + \beta\,D_{it} + \theta\;\sum_{j=1}^N w_{ij}D_{jt} + \gamma'X_{it} + \mu_i + \delta_t + \varepsilon_{it}$$7 can be based on geographic proximity, regulatory affiliation, or loan-portfolio similarity. All choices enforce $$y_{it} = \tau\,y_{i,t-1} + \rho\;\sum_{j=1}^N w_{ij}y_{jt} + \eta\;\sum_{j=1}^N w_{ij}y_{j,t-1} + \beta\,D_{it} + \theta\;\sum_{j=1}^N w_{ij}D_{jt} + \gamma'X_{it} + \mu_i + \delta_t + \varepsilon_{it}$$8 and $$y_{it} = \tau\,y_{i,t-1} + \rho\;\sum_{j=1}^N w_{ij}y_{jt} + \eta\;\sum_{j=1}^N w_{ij}y_{j,t-1} + \beta\,D_{it} + \theta\;\sum_{j=1}^N w_{ij}D_{jt} + \gamma'X_{it} + \mu_i + \delta_t + \varepsilon_{it}$$9.

This flexible construction allows $y_{it}$0 to reflect competition, information flows, or contagion risk pathways in empirical networks.

3. Dynamic Structure: Temporal and Spatial Lags

Two layers of persistence are modeled:

• Temporal autoregression ($y_{it}$1): $y_{it}$2 captures own history, essential for phenomena exhibiting inertia, as in profit dynamics driven by loan portfolio stickiness and management continuity.
• Spatial-temporal lags ($y_{it}$3): $y_{it}$4 allows last-period performance of neighbors to affect current outcomes, interpreting peer influence with a lag.

Omission of these terms would bias estimates for spillover strength and distort inference on network dependence.

4. Estimation Methodologies

The DSDM’s simultaneity and spatial dependence render OLS inconsistent. Three estimation strategies are prominent (Kikuchi, 2 Feb 2026):

• Maximum Likelihood (MLE): Direct maximization with Jacobian correction $y_{it}$5 accounts for simultaneity.
• Quasi-MLE (QMLE): Robust “sandwich” variance estimators accommodate heteroskedasticity and serial correlation.
• Bayesian MCMC: Weakly informative priors on $y_{it}$6 yield full posterior inference and enforce stability constraints (e.g., $y_{it}$7).

Identification of spatial parameters ($y_{it}$8, $y_{it}$9) is achieved via likelihood structure or, in the orthogonal forward differencing GMM approach (Kuersteiner et al., 2018), by exploiting linear and quadratic moment conditions for general weight matrices, including endogenous or data-driven $i$0.

5. Interpretation of Parameters and Impact Decomposition

Parameter interpretation is central due to model feedback:

• $i$1: Direct (own) effect of adoption or treatment, conditional on peer behavior.
• $i$2: Contemporaneous peer spillover; positive values indicate knowledge diffusion or network complementarity.
• $i$3: Instantaneous propagation of outcome shocks across the network (algorithmic coupling).
• $i$4: Lagged peer effect; negative estimates can signal competitive displacement.

Because spatial models transmit shocks through multiple pathways, effects are typically decomposed using: $i$5

• Direct impact: Average diagonal entry (own-unit effect).
• Indirect (spillover) impact: Off-diagonal row-mean (peer effect).
• Total impact: Sum of direct and indirect.

Table: Typical Parameter Estimates in Bank Productivity Analysis (Kikuchi, 2 Feb 2026)

Outcome	Direct Effect ($i$6)	Spillover ($i$7)	Top-Quartile Spillover ($i$8)
ROA	0.0373	0.1606	—
ROE	0.4199	0.6787	3.13

Significant and large $i$9 values, especially for large banks, indicate nontrivial system-wide synchronization.

6. Empirical Application: GenAI Adoption and Systemic Risk

In the application to U.S. banks (Kikuchi, 2 Feb 2026), the DSDM quantifies the association between GenAI adoption and productivity (ROA/ROE), while also revealing:

• Productivity Paradox: DSDM shows $t$0 for adopters but SDID uncovers a significant negative short-run causal effect (“Implementation Tax”), suggesting selection effects dominate contemporaneous impacts.
• Spillover Dynamics: Substantial positive $t$1 implies that AI adoption by competitors boosts own performance, with spillovers for large banks ($t$2) signaling tight “algorithmic coupling.”
• Systemic Risk Implications: Algorithmically coupled networks are susceptible to correlated shocks (e.g., technical failures in widely-used AI).

7. Limitations and Robustness

Identified limitations include:

• Endogeneity of Adoption: Causality between adoption and performance cannot be ascribed solely to DSDM; selection effects persist. Causal inference requires external strategies such as Synthetic Difference-in-Differences.
• Weight Matrix Misspecification: Results depend on the choice of $t$3; robustness checks under alternative specifications are essential.
• Model Misspecification: Simplistic dynamics may neglect nonlinearity or higher-order lags; robustness is enhanced by alternative treatmen specifications (e.g., adoption mention-counts), placebo testing, and bootstrapping.

Addressing these, the paired use of DSDM for associational inference and SDID for causal estimation yields a more nuanced view of technology adoption effects. The DSDM further identifies network spillovers and system-level risks not discernible in unit-level panel analyses (Kikuchi, 2 Feb 2026, Kuersteiner et al., 2018).