Section navigation is enabled for this long article; the full text remains continuous.
BSTVC and Spatiotemporal Interpretability · Part 2
Part 1 introduced three questions: how relationships vary, how factors contribute, and how well a model predicts. Today we take the first question one step further: what do we assume when we describe an entire study area and period with one regression coefficient?
Conceptual illustration, generated for this article. The shapes, colors, and layers represent an idea; they are not observations, estimated coefficients, or a map of any actual region.
One coefficient is a modeling choice
In a regression with a shared slope for a predictor, the same change in that predictor enters the model with the same coefficient at every included place and time. Other parts of the model can still vary. For example, places may have different baseline outcome levels even when their slopes are constrained to be equal.
For a linear predictor written schematically as η(i,t) = α(i,t) + β × x(i,t), varying α(i,t) changes the baseline, while the shared β keeps the slope the same. Replacing it with β(i,t) changes the relationship itself. This is explanatory notation, not the exact specification of a particular BSTVC implementation. In a nonlinear model, the coefficient is interpreted on the selected link scale; the corresponding change on the outcome scale need not be constant.
That shared coefficient is useful when it matches the question and is supported by the data. It provides a compact summary and can make estimation more stable. However, it is a model-based pooled relationship, not necessarily the arithmetic average of independently estimated local slopes.
Consider a purely hypothetical study of an environmental exposure and a health outcome across several districts and years. A single slope asks for one common conditional association. A varying-coefficient analysis asks whether the association differs by location, period, or both, after accounting for the other terms included in the model. Neither analysis, by itself, identifies the effect of an intervention.
Three patterns worth separating
| Pattern | Question | What must not be confused with it |
|---|---|---|
| Spatial variation in a relationship | Does the exposure–outcome association differ between places? | Different outcome levels alone |
| Temporal variation in a relationship | Does the association change across periods? | A rise or fall in the outcome alone |
| Space–time interaction in a relationship | Does the temporal pattern of the association differ between places? | Merely including separate place and time terms |
The third question needs particular care. A model with separate spatial and temporal coefficient components does not automatically represent every possible place-specific time trajectory. The selected model specification determines which patterns can be estimated.
The Bayesian STVC methodological work provides a framework for investigating spatial and temporal nonstationarity through local coefficient estimates. Its relevance here is the move from a shared relationship to explicitly modeled variation—not a claim that every dataset requires the most flexible model. Song, Shi & Wang, 2020
A changing map is not enough
A map of posterior means can be visually striking while leaving important questions unanswered. Before interpreting different colors as strong evidence of different relationships, examine uncertainty and the data behind each estimate.
- Check precision. A local estimate with a broad credible interval may not establish the direction of an association.
- Compare directly. If the scientific question concerns the difference between two coefficients, examine the posterior distribution of their difference when available. Comparing colors or whether each interval crosses zero is not the same test.
- Inspect support. Sparse observations, limited predictor variation, collinearity, missingness, or changing measurement practices can affect local estimates.
- Check robustness. Consider whether the pattern persists under reasonable choices of priors, spatial structure, temporal structure, and model specification.
Structured Bayesian models can share information across observations. This can stabilize estimates, but the resulting patterns still depend on the model and its assumptions. More detail on a map does not automatically mean more reliable knowledge.
When should you investigate heterogeneity?
Start with the research question rather than the software menu. A useful sequence is:
- State why the relationship might plausibly vary across space or time.
- Inspect the consistency and coverage of the place–time dataset.
- Fit an interpretable baseline with a shared coefficient.
- Compare a scientifically motivated varying-coefficient specification.
- Examine uncertainty and use validation appropriate to the intended use: interpolation within the observed domain is different from predicting a new place or a future period.
Keep a simpler model when added flexibility is not sufficiently supported. Keep a richer model only with a clear account of what it helps explain or predict and what remains uncertain.
What this means for a BSTVC workflow
The BSTVC-R project organizes local coefficient interpretation, global factor assessment, and prediction as related but distinct analytical outputs. In practice, begin this lesson with the local coefficient results and their uncertainty. Do not turn the largest coefficient into an automatic global importance ranking, and do not infer predictive performance from a visually persuasive coefficient map.
For software and documentation, visit the BSTVC organization website and BSTVC Desktop. BSTVC is developed through the contributions of the HEOA–West China Health and Medical Geography Research Group and collaborators. Current software use is free for academic and other noncommercial purposes; consult the applicable official license for the version you use.
中文摘要
一个全局回归系数可以概括研究区域和时期内的共同条件关联，但它不一定等于局部系数的算术平均，也不能自动代表每个地区、每个时期。需要区分“结局水平不同”与“变量关系不同”，以及空间差异、时间变化和时空交互。BSTVC 的局部系数分析用于研究这些关系的异质性；解读时应同时考察不确定性、数据支持和模型设定。地图颜色不同不能直接证明关系显著不同，统计关联也不能直接写成因果效应。是否增加模型复杂度，应由研究问题、稳健性检查和合适的验证共同判断。
Next step: When reading your next coefficient map, ask: “Which relationship varies, on what scale, and with how much uncertainty?”
Further reading
- Song, C., Shi, X., & Wang, J. (2020). Spatiotemporally Varying Coefficients (STVC) model: a Bayesian local regression to detect spatial and temporal nonstationarity in variables relationships. Annals of GIS, 26(3), 277–291. Publisher / DOI
- BSTVC-R official repository: framework, documentation, and license
- Previous article: BSTVC and Spatiotemporal Interpretability
Credits and disclosure
Educational article by Chao Song and the BSTVC team, drawing on the work of the cited researchers and collaborators. This article is not a new empirical study. Sources checked 25 October 2026.