The Gold Standard of Ecology | Integrated Distribution Models, Presence-Only Bias and Data Fusion

1. Define the ecological quantity to be shared

Back to contents

Decide whether the model targets occurrence probability, site use, relative intensity, abundance or environmental response. These quantities are not interchangeable across scales. Presence-only point patterns naturally inform relative intensity unless additional information anchors prevalence, while occupancy data describe a defined site and period. A shared latent surface must have a meaning compatible with every observation model.

2. Give each dataset its own observation process

Back to contents

A museum record, citizen-science checklist, atlas point and repeat survey arise through different sampling. Write a separate likelihood or observation submodel for each source, including effort, detection and selection where information exists. Do not concatenate rows and add a dataset label as though that alone repairs bias. Integration means joint estimation through explicit connections, not simply one larger spreadsheet.

3. Understand presence-only and background data

Back to contents

Presence-only records show reported locations without confirmed absences. Background points describe available environmental space or approximate a point-process integral; they are not surveyed absences. Road access, population density, observer preference and platform use can shape the record. Model sampling bias with effort covariates, target-group backgrounds or shared bias processes, and state which assumptions make the ecological and observation effects separable.

4. Use structured surveys to anchor detection and prevalence

Back to contents

Designed surveys provide known sites and non-detections, while repeat visits can estimate detection conditional on occurrence. They may anchor absolute occupancy or environmental relationships that presence-only data alone cannot identify. Yet structured surveys can still be geographically narrow, preferentially placed or inconsistently sampled. Known design is evidence to model, not a guarantee of representativeness.

5. Share only the ecological components the data can support

Back to contents

Datasets can share environmental coefficients, a spatial field or a latent intensity surface while retaining source-specific intercepts and bias terms. Full sharing gains precision but risks importing bias; partial sharing is safer when quality differs. Use simulation and identifiability diagnostics to test whether ecological response and sampling bias can be distinguished. More records do not solve two processes that have the same spatial pattern.

6. Worked example: a rare orchid near accessible trails

Back to contents

A rare orchid has 1,200 app records clustered near trails and 90 structured plots with repeat visits. A presence-only model predicts high suitability around visitor centres. The occupancy model is imprecise but shows higher occurrence on shaded limestone. An integrated model includes trail distance and observer activity in the reporting process, detection by flowering status, and a shared habitat response. Predictions improve in withheld plots, but uncertainty remains away from both trails and surveyed reserves. Fusion corrects some bias; it does not create information in unsampled landscapes.

7. Validate by source, space and prediction task

Back to contents

Hold out spatial blocks rather than random points when nearby observations share environment and sampling. Check predictions separately against structured detections, repeated surveys and independent records. Use posterior predictive checks or residual maps for each data source. A model can fit abundant presence-only points while failing to estimate occupancy, or fit surveys while missing rare range edges. Match validation to the ecological quantity claimed.

8. Changed-context transfer check: citizen reports of an invasive bird

Back to contents

Move from a rare plant to a conspicuous invasive bird. Opportunistic reports may track public concern, while structured counts capture abundance along routes. The transferable method separates ecological state, reporting intensity and survey detection. The changed context adds rapid spread and feedback: publicity can increase reports exactly where colonisation is suspected. Include time-varying effort and avoid interpreting a reporting campaign as biological acceleration.

9. Know the limits and common misconceptions

Back to contents

Integrated does not mean unbiased. A large biased dataset can dominate if its observation process is misspecified. Shared parameters can force incompatible scales together, and source-specific intercepts may be weakly identified. Presence-only data do not reveal nondetection without auxiliary information. A more precise map can still be wrong. Data fusion should not hide differences in taxonomic accuracy, timing, grain or target population.

10. Audit provenance, linkage and contribution

Back to contents

Publish data licences, deduplication, taxonomic decisions, coordinates, timing, effort fields, site definitions, background sampling, covariates, source-specific likelihoods, shared parameters, priors and code. Map every dataset separately before integration. Fit source-only and integrated models, then show how each source changes coefficients, uncertainty and predictions. Test sensitivity to bias surfaces and spatial thinning.

11. Frequently asked questions

Back to contents

Is presence-only the same as presence–absence? No. Are background points absences? No. Does integration always improve prediction? No; it helps when models and data are complementary and compatible. Can structured surveys be biased? Yes, especially when site placement is preferential. Why keep source-specific likelihoods? Because each dataset records a different observation process. Does more data remove extrapolation? No.

12. The gold standard fuses evidence without erasing provenance

Back to contents

A strong integrated model names the ecological quantity, preserves how each source was sampled and shows exactly what is shared. It tests whether bias and habitat can be separated, validates against the intended prediction task and reports where all datasets remain silent. The goal is not one seamless-looking map. It is a coherent inference in which every observation retains its history.