The Gold Standard of Ecology | Population Viability Analysis, Extinction Risk and Model Uncertainty

1. Define viability as a measurable event

Back to contents

State whether the event is extinction, fewer than ten breeding females, loss of one subpopulation, failure to meet a recovery target or another threshold. Choose a time horizon and justify it relative to species generation time and the decision. “Ninety per cent viable” is incomplete without threshold and horizon. Quasi-extinction is often more operational than literal zero because populations can become functionally lost before the last individual dies.

2. Draw the life cycle before choosing software

Back to contents

Map stages, survival, reproduction, maturation, movement and density dependence. Decide which processes the question requires. A stage-structured matrix can be sufficient for one management comparison; an individual-based model may be needed when genetics, social structure or spatial movement matters. Complexity should follow mechanism and evidence. A detailed interface does not compensate for unsupported transitions.

3. Separate demographic, environmental and observation uncertainty

Back to contents

Demographic stochasticity arises from chance births and deaths among individuals. Environmental stochasticity changes vital rates across years or places. Observation uncertainty concerns what the data reveal about abundance and rates. Treating estimated vital rates as known can make simulation precise for the wrong population. Catastrophes and correlations among rates require explicit evidence or scenario bounds rather than a single dramatic assumption.

4. Set the initial population and spatial boundary carefully

Back to contents

Initial abundance, stage structure, sex ratio and distribution can dominate near-term risk. Use an observation model or uncertainty distribution when counts are imperfect. Define whether immigration and emigration connect the focal population to neighbours. A closed model can overstate risk for a rescued population or understate it when apparent survival hides permanent emigration. Spatial boundaries should match the management action, not merely the available spreadsheet.

5. Use sensitivity and elasticity for decisions, not slogans

Back to contents

Sensitivity asks how outputs change when a parameter changes; elasticity expresses proportional response under a particular model. A vital rate with high elasticity may be impossible to alter, while a lower-elasticity rate may be the practical lever. Explore management actions through plausible effect sizes, costs and uncertainty. Do not infer that the most influential mathematical parameter is automatically the correct conservation priority.

6. Worked example: a small island bird population

Back to contents

An invented population begins with 120 birds in three life stages. Under estimated survival and reproduction, 18 per cent of 10,000 simulations fall below twenty adults within thirty years. A predator-control scenario raises first-year survival and reduces the point estimate to 7 per cent. When sampling uncertainty in survival, uncertain starting abundance and correlated bad years are added, plausible risk spans roughly 4 to 31 per cent across model runs. The action still improves most scenarios, but the exact probability is not stable. The report therefore emphasises the direction and size of management contrast, identifies adult survival data as valuable and avoids presenting 7 per cent as a measured fact.

7. Validate structure and behaviour before trusting forecasts

Back to contents

Check that the model reproduces observed stage distributions, growth patterns and variability over periods not used for fitting where possible. Conduct extreme-value tests, conservation-of-individual checks and simulation recovery. Compare alternative density dependence, catastrophe and dispersal structures. Validation does not prove future truth; it can reveal coding errors, impossible dynamics and assumptions that dominate outputs.

8. Changed-context transfer check: a fragmented orchid metapopulation

Back to contents

Move from one bird population to orchids in habitat patches. The transferable workflow still defines a threshold, life cycle, horizon and uncertainty. The new context adds seed banks, dormancy, local extinction and colonisation. Counting flowering plants alone is an observation of one stage, not total abundance. A patch-level occupancy or metapopulation model may answer connectivity questions better than forcing the problem into an individual count model.

9. Know the limits and common misconceptions

Back to contents

A PVA does not predict the date a species will disappear. Long-horizon probabilities can be extremely sensitive to poorly known rates and structural assumptions. Small datasets cannot support every parameter separately. A low modelled risk does not mean no action is needed, and a high risk does not by itself rank feasibility, cost or cultural priorities. IUCN status criteria and a bespoke PVA can inform one another, but they are not interchangeable labels.

10. Audit scenarios, uncertainty and decisions

Back to contents

Publish model equations or algorithms, life-cycle diagram, data provenance, parameter distributions, covariance, initial states, thresholds, horizons, number of simulations, random seeds and code. Distinguish fitted parameters from expert scenarios. Show point estimates beside uncertainty and alternative structures. Register management comparisons before tuning assumptions to favour one action. Record which new data would most reduce decision uncertainty.

11. Frequently asked questions

Back to contents

Is extinction probability a species property? No; it is conditional on model, horizon, threshold and assumptions. How many simulations are enough? Enough for stable summaries at the precision claimed; Monte Carlo error should be checked. Must PVA be individual-based? No. Can expert judgment be used? Yes when labelled, elicited and stress-tested. Does sensitivity identify the best action? Not alone; feasibility and achievable parameter change matter. Can models include climate change? Yes through explicit scenarios whose uncertainty remains visible.

12. The gold standard uses PVA as a decision laboratory

Back to contents

A credible PVA defines the event, builds the simplest defensible population process and carries observation, parameter and structural uncertainty into the comparison. It tests whether management conclusions survive different plausible worlds. The goal is not a single impressive extinction percentage. It is to learn which risks are robust, which actions help across scenarios and which missing evidence could change the choice.