Spatial Marginal Structural Model
Spatial Marginal Structural Model with Inverse Probability Weighting · Also known as: Spatial MSM, Geospatial MSM, Spatial IPW-MSM, Space-time marginal structural model
The Spatial Marginal Structural Model (Spatial MSM) extends the classical marginal structural model to settings where units are geographically distributed and spatial dependencies — such as neighborhood spillovers, clustering, and spatial confounding — may bias causal estimates. It estimates causal effects of spatially varying exposures by constructing inverse probability weights that account for both individual covariates and spatial location, then fitting a weighted outcome model in the resulting pseudo-population.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use a Spatial MSM when you are estimating the causal effect of a spatially structured exposure (e.g., pollution levels, access to services, neighborhood interventions) on an outcome measured at geographically referenced units, and when both measured and unmeasured spatial confounding are plausible. It is particularly suited to large areal or point-level observational datasets where treatment varies across space and spillover effects are limited. Do not use it when there is no credible positivity (some areas have near-zero probability of a particular exposure level), when spillover between nearby units is substantial (which violates SUTVA), when spatial clustering of unmeasured confounders overwhelms any plausible weighting correction, or when the dataset is too small to estimate a stable spatial propensity score.
Strengths & limitations
- Explicitly accounts for spatial confounding by incorporating location into the propensity score, reducing bias that standard MSMs ignore.
- Targets marginal causal effects in the whole population, not conditional effects for a particular covariate profile, making results directly policy-relevant.
- Stabilized weights allow transparent diagnostics: weight distributions reveal where positivity fails across the map.
- Compatible with a wide range of spatial propensity-score estimators, from GAMs to spatial random forests.
- Produces doubly-robust estimates when combined with outcome-model adjustment (augmented IPW variant).
- Relies on the no-unmeasured-spatial-confounders assumption, which is untestable and often implausible when spatial processes are driven by unobserved neighborhood factors.
- Estimating a spatially-informed propensity score requires sufficiently dense geographic coverage; sparse or irregular spatial sampling degrades weight quality.
- Violation of SUTVA (interference between nearby units) — common in spatial settings — invalidates the standard causal interpretation.
- Extreme weights can inflate variance substantially; trimming reduces variance but introduces bias whose magnitude is hard to quantify.
Frequently asked
How does a Spatial MSM differ from a standard MSM?
A standard MSM uses inverse probability weights based solely on individual-level covariates. A Spatial MSM additionally incorporates geographic location — via spatial random effects, smooths, or spatially-structured propensity scores — so that the weights also balance spatial confounders that cluster geographically, not just individual-level imbalances.
Does a Spatial MSM handle interference between neighboring units?
No. Like the standard MSM, it assumes no interference (SUTVA). If treatment in one spatial unit causally affects outcomes in neighboring units — which is common for air pollution, contagion, or policy spillovers — a different framework such as interference-aware causal models is needed.
What propensity score model should I use for the spatial component?
Common choices include logistic regression with spatial random effects (e.g., via INLA or MCMC), generalized additive models with thin-plate spline smooths of coordinates, or spatial random forests. The choice depends on data density, the expected form of spatial variation in treatment, and computational resources.
How do I check if the spatial weights are reasonable?
Inspect the distribution of stabilized weights: they should be centered near 1, with most between roughly 0.1 and 10. Map the weights spatially — extreme weights concentrated in a particular region signal that positivity fails there and the causal effect is not identified for that area.
Can I use a Spatial MSM with continuous spatial treatments?
Yes. The generalized propensity score framework extends MSMs to continuous treatments by modeling the conditional density of the continuous exposure given covariates and location, then constructing weights from that density. The principles are the same but estimation is more demanding.
Sources
- Robins, J. M., Hernan, M. A., & Brumback, B. (2000). Marginal structural models and causal inference in epidemiology. Epidemiology, 11(5), 550-560. DOI: 10.1097/00001648-200009000-00011 ↗
- Schnell, P. M., & Papadogeorgou, G. (2020). Mitigating unobserved spatial confounding when estimating the effect of supermarket access on cardiovascular disease deaths. Annals of Applied Statistics, 14(2), 793-816. DOI: 10.1214/20-aoas1377 ↗
How to cite this page
ScholarGate. (2026, June 3). Spatial Marginal Structural Model with Inverse Probability Weighting. ScholarGate. https://scholargate.app/en/causal-inference/spatial-marginal-structural-model
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Inverse Probability WeightingCausal inference↔ compare
- Marginal Structural ModelCausal inference↔ compare
- Propensity Score WeightingCausal inference↔ compare
- Spatial Doubly Robust EstimationCausal inference↔ compare
- Spatial Instrumental VariablesCausal inference↔ compare
- Spatial Propensity Score MatchingCausal inference↔ compare