The Gold Standard of Ecology | Integrated Population Models, Data Fusion and Shared-Demography Assumptions

1. Begin with the population process and estimand

Back to contents

State the population boundary, time step, life stages and demographic quantities before combining data. Decide whether abundance means breeders, all individuals or a sampled index, and whether survival is apparent or true. Write a process model that links abundance through survival, recruitment, immigration and emigration. The same word can hide different targets across datasets: a July count of adults, a winter capture history and fledglings per successful nest do not automatically describe one annual cycle. Integration begins with a coherent biological clock, not with placing likelihoods beside each other.

2. Separate the latent process from observation models

Back to contents

The population process describes unobserved demographic change. Each dataset then has its own observation model: counts may have sampling error and detection bias; capture histories have encounter probabilities; productivity surveys can miss failed nests. Keeping these layers separate prevents observation error from being mistaken for process variation. Draw the dependencies and identify which parameter each dataset informs. A precise count does not make detection equal to one, and a survival study at marked sites does not necessarily represent the full population.

3. Map every dataset onto space, time and life stage

Back to contents

Create a compatibility table before modelling. Record sampling areas, seasons, age or stage definitions, effort, missing years and inclusion rules. Some data may overlap at the same sites and individuals, creating dependence; others may sample distinct subpopulations with different demography. Align annual intervals carefully: breeding productivity in year t may contribute to recruits observed in t+1. If age classes are collapsed in counts but separated in mark–recapture, specify the aggregation. Apparent completeness in a long time series cannot repair a mismatched biological unit.

4. Use shared parameters only when sharing is biological

Back to contents

Integration often assumes that a survival or productivity parameter in one component is the same quantity driving the population projection. That equality may be justified by design, or it may be a strong exchangeability assumption. Marked adults near research stations could survive differently from unmarked adults across the range. Nest monitoring may favour accessible, successful territories. Consider spatial, temporal or individual random effects when the same mean is plausible but exact equality is not. The model should make borrowing visible and let data challenge it where information permits.

5. Understand what data fusion identifies

Back to contents

Combining components can estimate parameters that no single dataset identifies well, such as net immigration derived from population change after accounting for survival and recruitment. Such estimates are model-based residuals, not direct observations. Precision can improve because shared parameters constrain several equations, but apparent precision may be overconfident if component biases align. Inspect prior-to-posterior change, correlations and contributions from each likelihood. Fit components separately as a diagnostic, while recognising that the integrated estimand can differ from separate estimates.

6. Worked example: counts, survival and chicks

Back to contents

A coastal bird programme has annual adult counts, capture–recapture histories from two colonies and chicks per monitored nest. The first model shares one adult survival rate across the coast and uses monitored productivity directly. It implies unusually high immigration after a storm. A compatibility audit shows that counts include non-breeders, marked colonies are sheltered and failed nests are under-recorded. A revised model separates breeder and non-breeder states, allows colony-specific survival around a shared mean and models nest detection. Immigration remains positive but much more uncertain. The model adds knowledge by exposing the missing flow, not by pretending it was observed.

7. Diagnose conflict and weak identifiability

Back to contents

Posterior predictive checks can show whether the model reproduces counts, encounter histories and productivity summaries simultaneously. Leave-one-dataset-out fits reveal which component drives an estimate. Compare direct estimates with values implied by population balance, and investigate systematic disagreement rather than averaging it away. Weak identification appears as broad or highly correlated parameters, sensitivity to priors and multiple demographic combinations fitting equally well. More datasets do not guarantee more information if they repeat the same bias or introduce extra unknowns.

8. Changed-context transfer check: a threatened plant

Back to contents

Move from birds to a perennial plant. Plot censuses track stage transitions, seed traps index fecundity and genetic assignments inform recruitment sources. The transferable framework separates process and observation models, aligns life stages and tests shared parameters. The changed context adds seed dormancy and a persistent seed bank, so observed seedlings do not equal current-year production. If that hidden state is omitted, the model may assign delayed emergence to immigration or survival. Integration must follow the organism’s life cycle rather than reuse a familiar animal model.

9. Know the limits and common misconceptions

Back to contents

An IPM is not automatically unbiased or causal. It can propagate a biased component through the entire model. Conditional independence between datasets may fail when the same observers, sites or individuals contribute to several sources. Shared demography can conceal real heterogeneity. A narrow credible interval reflects the specified model and priors, not all uncertainty about population definition or measurement. Immigration estimates derived as a balance term can absorb model error. Model fit cannot prove that an unobserved mechanism is correctly named.

10. Audit the full evidence graph

Back to contents

Publish the population equation, component likelihoods, priors, code, data dictionaries, spatial and temporal coverage, overlap map and missingness rules. Report results from each component alone, the integrated model and alternative sharing structures. Test prior sensitivity, overdispersion, detection assumptions and leave-one-source-out stability. Simulate from known demographic scenarios to see whether the design can recover the quantities claimed. Archive derived data without discarding raw counts, encounter histories or nest records. A reader should be able to trace every posterior update to a biological and observational assumption.

11. Frequently asked questions

Back to contents

Is an integrated population model just a meta-analysis? No; component data are linked through one demographic process. Must datasets cover the same individuals? No, but they must support compatible population parameters or an explicit hierarchy. Can IPMs estimate immigration? Often as a balance term, though that estimate can absorb model error. Does integration always improve precision? No; incompatibility and added parameters can widen uncertainty. Should separate models still be fitted? Yes, as diagnostics for conflict and information contribution.

12. The gold standard makes borrowing testable

Back to contents

A strong IPM shows why each source belongs, how its observation process works and which demographic quantities are shared. It treats data fusion as a biological claim about common processes, not as a software convenience. Conflict, residual flows and prior sensitivity remain visible. The result is valuable when it explains population change while identifying what is directly measured, what is borrowed and what is inferred only because the demographic accounting requires it.