The Gold Standard of Ecology | Joint Species Distribution Models, Residual Associations and Latent Factors

1. Begin with the community question

Back to contents

Decide whether the aim is prediction, environmental response, ordination, community turnover, rare-species borrowing or residual association. These aims favour different likelihoods, hierarchies and validation targets. A model chosen because it returns an association matrix may answer the wrong ecological question. Define the sampling unit, response scale and prediction domain before choosing a framework.

2. Respect the response distribution

Back to contents

Presence–absence, counts, biomass and percentage cover have different observation processes. Choose Bernoulli, Poisson, negative-binomial, hurdle, ordinal or compositional structures that match how data were generated. Overdispersion, excess zeros and detection error do not vanish when species are modelled jointly. A poorly fitted marginal model can manufacture cross-species residual structure.

3. Separate measured environment from residual structure

Back to contents

The fixed part of a JSDM estimates how each species responds to recorded covariates. Residual association describes covariance left after those predictors and other model components. Positive residual association means species co-vary beyond the fitted explanation; negative association means they separate. Neither label supplies a mechanism. Omitted habitat, shared dispersal, observer effects and direct interaction can produce similar signs.

4. Use latent factors as a low-rank summary

Back to contents

A small number of latent variables can represent patterns shared by many species without estimating every covariance freely. Species loadings describe how each taxon relates to those latent dimensions. The factors may act like unmeasured gradients or a regularised ordination device, but their axes can rotate and their biological meaning is not guaranteed. Choose factor number through predictive checks and sensitivity, not a visually pleasing biplot.

5. Put space, time and hierarchy where they belong

Back to contents

Residual association can arise because nearby sites share unmeasured conditions or because surveys recur through time. Spatially structured latent factors, site random effects and temporal components can reduce confounding, but they introduce scale and prior choices. Include sampling hierarchy—plots within sites, visits within seasons—rather than letting one latent layer carry every source of dependence.

6. Worked example: woodland birds across 80 plots

Back to contents

A presence–absence JSDM includes canopy cover, elevation and distance to water for 22 bird species. Two insectivores retain a positive residual association. Adding a spatial latent factor weakens it, and adding observer identity weakens it further; prediction improves at held-out plots. The study reports shared unmeasured spatial habitat or observation as the leading explanation, not facilitation. A targeted food survey and repeated detection design are proposed before any interaction claim.

7. Validate the task the model is meant to perform

Back to contents

Randomly splitting observations can leak spatial or temporal dependence into both training and test sets. Use site-blocked, spatially separated or forward-in-time validation that matches the intended transfer. Assess calibration and discrimination for species, but also community summaries and joint predictions. A model can predict common species well while failing rare taxa or the covariance pattern that motivated joint modelling.

8. Changed-context transfer check: reef recovery after bleaching

Back to contents

A reef JSDM models coral and fish assemblages before and after bleaching. The transferable rule is to separate measured environmental response from residual association. The context adds compositional cover, imperfect fish detection and a strong shared disturbance. A new negative association after bleaching could reflect constrained space, differential survival or survey conditions; causal competition requires process evidence beyond the fitted covariance.

9. Know the common misconceptions

Back to contents

JSDMs do not solve causal inference merely by modelling species together. Residual correlations are conditional on the chosen covariates, scales, likelihood and random effects. Latent variables are not directly measured environmental gradients. More factors can fit noise, while too few can leave structure. Borrowing strength can help rare species but may pull genuinely unusual responses toward a community mean. Predictions beyond sampled environments remain extrapolations.

10. Audit data, model and ecological interpretation

Back to contents

Publish the community matrix, effort variables, taxonomic decisions, covariates, transformations, spatial coordinates, hierarchy, priors, factor choices and code. Inspect posterior predictive checks or residual diagnostics by species and site. Repeat the analysis with alternative likelihoods, factor counts, spatial structures and covariate sets. Report which associations persist and which disappear when plausible missing structure is represented.

11. Frequently asked questions

Back to contents

Is a positive residual association facilitation? No; it is compatible with several mechanisms. How many latent factors are needed? Enough for the intended predictive and residual structure, judged with diagnostics and sensitivity. Can rare species benefit? Yes, through hierarchical sharing, provided shrinkage is checked. Must the model be Bayesian? No; several likelihood-based and Bayesian JSDM frameworks exist.

12. The gold standard keeps association conditional

Back to contents

Joint models become valuable when they clarify what many species share and where their responses differ. The gold standard fits the observation process, represents space and hierarchy, validates the intended task and treats latent factors as conditional summaries. Residual association is evidence that the fitted explanation is incomplete—not automatic proof of the ecological mechanism that completes it.