The Gold Standard of Ecology | Integrated Population Models, Shared Parameters and Data Conflict

1. Define the population process and target parameters

Back to contents

Specify the population boundary, life stages, annual cycle and transitions that connect abundance through time. Decide whether the target is total abundance, growth rate, survival, fecundity, immigration or a management contrast. Draw the state process before attaching datasets. If immigration is only a residual needed to balance the population equation, call it that rather than presenting it as directly observed movement.

2. Give every dataset its own observation model

Back to contents

Counts may include detection error, productivity surveys sample nests, and mark–recapture data condition on encounter histories. Model each sampling process on its own scale while linking it to the shared biological state. Treating a count as true abundance or a monitored nest sample as population-wide fecundity can transmit bias into every connected parameter.

3. Use shared parameters deliberately

Back to contents

A survival parameter may appear in both the population projection and encounter-history likelihood; fecundity may connect brood data to recruitment. Sharing is the source of integration, but the definitions must match in age, interval, geography and eligibility. Annual apparent survival from marked adults is not automatically the same quantity as true survival in a population balance equation.

4. Check dependence among data sources

Back to contents

Standard formulations often factor component likelihoods as if datasets were conditionally independent. The same animals, sites or observers may contribute to several sources, creating residual dependence and double use of information. Map overlap explicitly. Model shared latent processes or covariance where possible, or run sensitivity analyses that remove overlapping components. More data streams do not equal more independent evidence.

5. Separate process variation from measurement error

Back to contents

Population size changes because demographic events vary, while observations differ because sampling is imperfect. State-space structure separates those layers. If all variation is assigned to observation, biological volatility is understated; if all is process noise, surveys appear more precise than they are. Use replicated or calibration information where available and check whether the inferred variation matches field knowledge.

6. Worked example: seabird counts, nests and marked adults

Back to contents

A colony has annual total counts, chick production from monitored nests and capture–recapture histories for adults. The integrated model links adult survival and fecundity to a stage-structured population process. Counts rise while marked-adult survival is stable; the model attributes the residual mostly to recruitment and immigration. When detection in early counts is allowed to be lower, the immigration estimate shrinks sharply. The study reports immigration as weakly identified and shows both specifications rather than declaring a newly measured movement rate.

7. Diagnose conflict between components

Back to contents

Fit or predict each component separately and compare its parameter support with the joint result. Posterior predictive checks can reveal that the model matches counts but misses recovery ages or brood distributions. Conflict may indicate scale mismatch, unmodelled heterogeneity or genuine population change. The integrated estimate is not automatically a compromise truth; a dominant, precise but biased component can overwhelm the others.

8. Changed-context transfer check: an urban butterfly population

Back to contents

Transect counts, citizen-science sightings and larval surveys are integrated. The transferable principle is separate observation models linked to one biological process. The new context adds opportunistic effort and spatial mismatch: sightings concentrate near accessible parks, while larval plots cover selected host plants. Shared phenology and detection terms are needed before the sources can support city-wide abundance or recruitment.

9. Know the common misconceptions

Back to contents

An IPM is not just several datasets placed in one software file. Shared parameters do not guarantee compatible definitions, and narrower intervals do not prove less bias. Data integration cannot identify every unobserved rate; residual immigration is especially model-dependent. Good chain convergence does not show ecological adequacy. Adding a weak or conflicting component can worsen inference when its observation model is wrong.

10. Audit component fit and influence

Back to contents

Publish the state equations, component likelihoods, data overlap map, priors, code and preprocessing. Report fits to each dataset as well as the joint posterior. Refit after removing one component at a time, vary detection and process-error assumptions and examine prior-to-posterior learning for weak parameters. Show which dataset changes each conclusion and where apparent precision comes from.

11. Frequently asked questions

Back to contents

Must datasets cover identical years? Not necessarily, but the process model must bridge gaps without hiding extrapolation. Can unobserved immigration be estimated? Sometimes as a residual under strong assumptions; uncertainty and interpretation should reflect that. Does integration always improve precision? It often can, but incompatibility or dependence can create false precision. Are Bayesian methods required? No, though hierarchical Bayesian computation is common.

12. The gold standard integrates evidence without hiding disagreement

Back to contents

A credible integrated population model starts with one explicit population process, gives each survey its own observation model and shares only compatible parameters. It checks overlap, component fit and data conflict before celebrating precision. The gold standard uses integration to expose how evidence connects—and also where the connections strain.