Ecological modelling builds explicit representations of organisms, populations and ecosystems so that assumptions can be followed through to consequences. A useful model answers a defined ecological question, connects its hidden states to observations, and makes predictions or comparisons that can be checked. It might estimate how a pond population changes, explain why a species occupies some patches, or compare the consequences of two restoration choices.
The gold standard of ecological modelling is a model whose purpose, mechanisms, evidence and limits remain visible together. A close fit to yesterday's counts is only one test. The more demanding questions are whether the calculations represent the intended processes, whether the observations distinguish competing explanations, and whether the result remains useful in the place and period where someone plans to apply it.
This guide follows original, hypothetical pond and habitat examples. Every numerical value is invented for teaching; none is a measured property of a named species or a recommendation for managing a real reserve. Begin with Ecology for the living relationships, or Mathematical Biology for a broader tour of biological equations. Here the emphasis is the complete ecological modelling investigation, from question to defensible use.
Choose your ecological modelling route
Follow an ecological model from a clear question through observations, calculations and defensible use.
Route 1 · Chapters 1–3: Frame the ecological question
Route 2 · Chapters 4–6: Connect states, surveys and calculations
Route 3 · Chapters 7–9: Test uncertainty and predictions
Route 4 · Chapters 10–11: Use results and diagnose failure
Route 5 · Chapters 12–14: Transfer, practise and continue
Chapter index
Frame the ecological question · Chapters 1–3
1. Decide what the model must help you learn
2. Choose states, boundaries and time scales
3. Select a representation that preserves the important mechanism
Connect states, surveys and calculations · Chapters 4–6
4. Connect ecological states to what a survey records
5. Check the numerical method before interpreting its behaviour
6. Calibrate parameters without claiming more identification than the data provide
Test uncertainty and predictions · Chapters 7–9
7. Distinguish sensitivity, uncertainty and ecological variation
8. Evaluate predictions where the intended use will occur
9. Use stochastic and spatial models when averages erase the question
Use results and diagnose failure · Chapters 10–11
Transfer, practise and continue · Chapters 12–14
12. Transfer the reasoning to a changed setting
CHAPTER 1 OF 14 · Frame the ecological question
1. Decide what the model must help you learn
Imagine a group studying small aquatic animals in twelve fictional ponds. They have monthly counts, water temperatures and notes about vegetation. Three questions sound similar: why do counts differ among ponds, what will next month's count be, and would adding vegetation improve persistence? They require different evidence. Explaining an observed association, forecasting an observation and estimating an intervention's effect are separate modelling tasks.
Write a one-sentence purpose before selecting software: “Predict the distribution of survey counts one month ahead in these twelve ponds, using information available at the forecast date.” This statement identifies the output, horizon, location and information boundary. It does not promise an explanation of vegetation effects. A second investigation could address that question with an intervention design and explicit assumptions about other differences between ponds.
A conceptual model can begin as a verbal balance: animals survive, reproduce, arrive and leave; water conditions affect some of those processes; the survey detects only some animals. Mark which relationships are assumptions, which have independent evidence and which are deliberately omitted. An omitted pathway becomes especially important when a proposed action operates through it. A model with no movement cannot evaluate a corridor by quietly renaming a birth coefficient.
Make success measurable
For the count forecast, success could mean lower error than a last-count baseline and prediction intervals with useful coverage on genuinely later surveys. For an explanation, success might include predicting a pattern that competing mechanisms do not predict. For an intervention comparison, success requires attention to whether the assumed change corresponds to what the action would actually do. A single goodness-of-fit number cannot cover all three purposes.
The EPA's 2009 environmental modelling guidance frames model development, evaluation and application as connected activities. Its central practical implication here is to establish intended use early and evaluate suitability for that use. Keep a short purpose record alongside the model: question, target, domain, comparison, evidence and a condition under which you would stop using the result.
CHAPTER 2 OF 14 · Frame the ecological question
2. Choose states, boundaries and time scales
A state is a quantity needed to describe the system at a particular time. In the pond example, N might mean the number of animals alive inside a mapped pond boundary immediately before a monthly survey. That definition excludes animals temporarily outside the boundary and says nothing yet about how many will be seen. If the question concerns reproduction, separating juveniles and adults may preserve information that a single total loses.
Define inputs and parameters separately. Rainfall recorded during a month is an input. A coefficient connecting water conditions to survival is a parameter within a specified model. The initial abundance is a starting condition. These distinctions matter because uncertainty in a starting count is not repaired merely by estimating a response coefficient more precisely.
A balance that can be audited
Suppose a pond starts a month with 80 animals. During that month there are 24 births, 12 deaths, 9 immigrants and 5 emigrants. The end-of-month balance is 80 + 24 − 12 + 9 − 5 = 96. Net local demographic growth is 12; net movement adds 4. Calling the total increase of 16 “reproduction” would erase the mechanism that the model was meant to examine.
Now enlarge the boundary to include a neighbouring pond. A transfer between the two ponds is internal to the combined system. It must subtract from one local count and add to the other without creating an extra animal overall. This simple accounting test catches spatial models that accidentally count an arrival without recording the corresponding departure.
Time scale changes meaning too. A monthly survival probability applies over one month. It is not a daily probability and cannot be inserted unchanged into a daily update. If a constant daily survival probability is s and daily outcomes follow the required conditional assumptions, monthly survival over 30 days is s raised to the power 30. Conversely, taking a thirtieth root imposes a constant daily rate representation; it does not reveal actual within-month mortality timing.
Use Population Ecology to develop demographic interpretation and Ecosystem Ecology when the states are material or energy pools. In either case, write units beside every term. Adding animals to animals per month without multiplying by a duration is a broken balance, however plausible the resulting graph looks.
CHAPTER 3 OF 14 · Frame the ecological question
3. Select a representation that preserves the important mechanism
An empirical model relates measured predictors to an outcome, often without explicitly representing all intervening processes. A process model specifies how states change through mechanisms such as recruitment, mortality or movement. An individual-based model tracks distinct organisms or agents. These are useful distinctions, but an actual investigation can combine them: a demographic simulation may use an empirically estimated temperature-survival relationship.
For the twelve ponds, begin with the smallest representation that can answer the stated question. A monthly count forecast may benefit from recent counts and seasonal predictors. A question about whether a connection rescues an isolated pond requires movement and local losses to be represented. A question about which individuals reach that connection may require position, behaviour or life stage. Complexity earns its place by preserving a relevant distinction.
Compare two purposeful simplifications
Model A treats each pond as one abundance value. Model B separates juveniles and adults. Both predict the same total today. If adults survive at 0.9 and juveniles at 0.5 over the next interval, a population with 20 adults and 80 juveniles has expected survivors of 18 + 40 = 58. Another with 80 adults and 20 juveniles has 72 + 10 = 82. Identical totals conceal a difference relevant to the next step.
That calculation justifies stage structure for this question. It does not justify adding twenty further traits. Each added state requires initial values, transition rules and evidence. An unmeasured distinction can make a simulation appear more biological while leaving its conclusions dominated by assumptions. Keep a simpler comparison model so that the benefit of the extra detail remains testable.
Use a model ladder: a baseline prediction, a minimal ecological mechanism and a richer candidate. Ask what each adds and what it costs in data, interpretation and computation. A poor baseline is too easy to beat; an unnecessarily elaborate baseline can obscure the comparison. Choose a reference that a sensible observer could have used before seeing the test results.
The Movement Ecology guide helps decide when trajectories matter. Community Ecology helps when interactions among species define the question. The representation should follow the ecological distinction, not the availability of a fashionable algorithm.
CHAPTER 4 OF 14 · Connect states, surveys and calculations
4. Connect ecological states to what a survey records
The number alive and the number observed are different quantities. Suppose each of N animals is detected independently with the same probability p during a count, with no double counting or false positives. Under that simplified observation model, expected count is Np. An expected count of 30 is compatible with N = 60 and p = 0.5, or N = 100 and p = 0.3. Matching the count alone does not choose between these explanations.
Imagine that true abundance remains 100 while detection changes from 0.6 to 0.3 because survey conditions worsen. Expected counts fall from 60 to 30 without any population decline. A model fitted directly to counts might assign the change to mortality unless it represents observation conditions or has evidence that detection was stable. Adding more decimal places to the fitted mortality estimate would not fix that mistake.
Use repeated observations deliberately
For a different question, define occupancy as whether a pond contains the species during a short survey period. If occupancy remains fixed across three visits and detection probability is independently 0.4 on each visit when occupied, the probability of at least one detection is 1 − 0.6³ = 0.784. The remaining 0.216 probability of no detection shows why a sequence of zeros need not establish absence.
The foundational MacKenzie and colleagues occupancy paper develops inference with imperfect detection. Our small calculation illustrates the observation issue; it is not a complete occupancy analysis. Shared weather, movement during visits, false identifications and differing detectability among sites require additional consideration.
Survey effort is part of the model's meaning. A count from ten minutes cannot automatically be compared with a count from an hour. Nor does dividing by effort always repair the difference: detection can saturate, organisms can be counted repeatedly, and observers can choose locations preferentially. Preserve the sampling protocol and the unit of replication alongside the numbers.
For transitions between periods, continue to Dynamic Occupancy Models, Colonisation, Extinction and Detection. For the information needed to reconstruct observations, use Ecological Informatics. A state model and an observation model answer complementary parts of the same ecological question.
CHAPTER 5 OF 14 · Connect states, surveys and calculations
5. Check the numerical method before interpreting its behaviour
A model can be scientifically questionable, incorrectly implemented, or both. Separate those possibilities. Consider the continuous logistic equation dN/dt = rN(1 − N/K), with constant positive r and K. Here r is an instantaneous per-capita growth coefficient and K sets the equilibrium abundance under the specified model. The equation is a particular density-dependent representation, not a universal description of a pond.
Take N at time zero as 40, K as 100 and r as 0.5 per month. The exact solution after one month is 100 divided by [1 + 1.5 exp(−0.5)], approximately 52.362. A forward Euler approximation with one one-month step gives 40 + 0.5 × 40 × 0.6 = 52. Two half-month Euler steps give 46 after the first step and 52.21 after the second. The method approaches the continuous-model answer as its time resolution improves here.
A negative population can be a calculation warning
Use a deliberately excessive ten-month Euler step from N = 90 with the same parameters. The next value is 90 + 10 × 0.5 × 90 × 0.1 = 135. A second such step gives 135 + 10 × 0.5 × 135 × (−0.35) = −101.25. This negative result is not evidence that the real population becomes negative. It reveals a numerical approximation being used outside an adequate step-size regime.
A genuinely discrete ecological model can also oscillate, but that is a different claim. State whether the update rule is the biological model itself or an approximation to a continuous equation. Otherwise numerical artifacts can be mistaken for ecological cycles. Cambridge mathematical biology notes provide a deeper route through logistic dynamics and population models.
Useful implementation checks include zero input producing the expected zero contribution, movement conserving totals, probabilities staying within their permitted range, and known special cases matching hand calculations. Repeat calculations with tighter numerical tolerances or smaller steps where applicable. Record the comparison rather than merely reporting that the program completed without an error message.
CHAPTER 6 OF 14 · Connect states, surveys and calculations
6. Calibrate parameters without claiming more identification than the data provide
Calibration chooses parameter values according to a stated agreement criterion. Identification asks whether the available evidence can distinguish those values or mechanisms. A model can fit extremely well while leaving important combinations unresolved. The distinction is particularly important when the fitted components respond differently to a proposed management action.
Return to a simplified one-step pond model: N next = aN now + I, where a is a net demographic multiplier and I is net immigration over the interval. If abundance changes from 100 to 120, both a = 1.2 with I = 0 and a = 1.0 with I = 20 reproduce the transition. The single observation does not reveal which process created it. If the question concerns removing a connecting channel, that ambiguity directly affects the answer.
Suppose a second experiment starts at N = 50 and produces N next = 70 under the same a and I assumptions. Subtracting the two equations gives 50 = 50a, so a = 1 and I = 20. The extra condition separates the parameters algebraically in this idealised example. With noisy field observations, varying immigration or different habitat conditions, the calculation would require a suitable statistical model and would not identify the processes so cleanly.
Decide what additional evidence would distinguish the explanations
Direct movement observations, an experimental exclusion, or independent demographic information may be more informative than another total count collected under nearly identical conditions. Design the additional measurement around the ambiguous mechanism. Do not assume that a larger spreadsheet necessarily contains a new kind of evidence.
A prior distribution can regularise estimation and express existing knowledge, but its influence should remain visible. If the data barely distinguish two components, a narrow prior may largely determine their separation. That can be legitimate when the prior is well supported; it is misleading when the resulting precision is described as if it came from the pond counts alone.
Keep a parameter record with definition, units, source, estimation method, uncertainty and allowed range. Label borrowed values as borrowed. A survival estimate from another species, region or survey protocol is an assumption about transfer until evidence supports that use. Calibration is the beginning of a critical comparison, not the end of the investigation.
CHAPTER 7 OF 14 · Test uncertainty and predictions
7. Distinguish sensitivity, uncertainty and ecological variation
Sensitivity asks how an output changes when an input or parameter changes. Uncertainty concerns what is not known about the relevant quantities or model. Ecological variation concerns differences and fluctuations in the system itself. These can interact, but they are not interchangeable. A highly influential parameter may already be measured precisely; a poorly known parameter may have little effect on the decision being considered.
Use the invented linear pond update N next = aN + I, with N = 100, a = 0.9 and I = 30. The prediction is 120. Increasing a by 0.01 raises the prediction by one animal; increasing I by one raises it by one animal. The numerical derivatives have different units, so comparing their raw sizes would be misleading. Over an uncertainty range of a = 0.8 to 1.0, the prediction changes by 20; over I = 28 to 32, it changes by 4.
These ranges suggest which uncertainty matters more for this particular one-step calculation. They do not rank biological importance in general. After many steps, density dependence, parameter interactions and changing environmental inputs can alter the ranking. A sensitivity statement should always name the output and horizon.
Carry joint assumptions through the calculation
Suppose field conditions make low survival coincide with high immigration. Treating the two uncertain inputs as independent can generate combinations that the ecological setting rarely permits. Conversely, sampling only along their observed relationship may miss what happens under an intervention that breaks it. Describe whether you are exploring observed variation, uncertain knowledge or an intentionally changed scenario.
In a simulation study, propagate uncertain initial states, parameters and drivers through to an output distribution. Preserve dependencies where justified. Separate this distribution from Monte Carlo error, which is the extra approximation caused by using a finite number of simulation runs. Running the same uncertain model a million times reduces simulation noise; it does not automatically resolve uncertainty about its ecological structure.
EPA's model evaluation guidance distinguishes sensitivity and uncertainty analysis. For population-specific risk questions, Population Viability Analysis, Extinction Risk and Model Uncertainty develops the implications of thresholds, time horizons and demographic assumptions.
CHAPTER 8 OF 14 · Test uncertainty and predictions
8. Evaluate predictions where the intended use will occur
Testing a model against the data used to select its coefficients is useful for finding problems, but it is not an independent forecast test. For the twelve-pond study, reserve later survey periods before tuning the model. If the intended use is a new reserve, also ask how well it transfers to locations outside the fitting set. These are different tests and can produce different answers.
Imagine four held-out observed counts: 10, 20, 30 and 40. Model A predicts 12, 18, 33 and 37. Its absolute errors are 2, 2, 3 and 3, giving mean absolute error 2.5. A simple baseline predicts 15, 15, 35 and 35, giving mean absolute error 5. Model A improves this score by 50% on this tiny example. That statement is narrower and more informative than “the model is accurate.”
Test the output that users will receive
If the model supplies a predictive distribution, examine the distribution rather than only its central value. Do intervals contain observations at approximately their stated rate over enough relevant cases? Are they so wide that they offer little practical help? Does performance worsen during low abundance, high rainfall or a new season? Four observations are insufficient to establish reliable coverage, even if every one falls inside an interval.
Roberts and colleagues' 2017 analysis explains why dependence and structure matter when choosing cross-validation. Randomly dividing nearby or repeated observations can assess a different prediction task from transfer to a new place or period. Match the split to the intended application, and report the separation actually used.
Avoid leaking test information through preprocessing. A seasonal adjustment, scaling rule or variable selection step fitted on the complete dataset can carry information from the held-out period into training. For a prospective forecast, only information available at the stated forecast date belongs in its construction.
Finally, test against ecological patterns the model was not explicitly tuned to match. A model that reproduces average counts while missing the timing of recruitment or the frequency of empty ponds may be useful for one task and unsuitable for another. Diagnose that difference before adding complexity. A failure can reveal a missing mechanism, an observation problem or a mismatch between the test and the stated purpose.
CHAPTER 9 OF 14 · Test uncertainty and predictions
9. Use stochastic and spatial models when averages erase the question
Suppose a fictional founder produces either zero or two descendants with equal probability, independently of other founders. One founder has expected descendants equal to one, yet a one-half probability of leaving none after one generation. Three founders have expected total descendants three and a probability of zero descendants equal to 0.5³ = 0.125. Expected population size does not describe the probability of an empty population.
The independence assumption matters. If one shared event causes all founders to fail together with probability one-half, and otherwise each produces two descendants, the same three-founder expected total is three but the probability of zero is one-half. Two models with identical means can therefore answer a persistence question very differently. The ecological distinction is whether outcomes share a driver.
Preserve the spatial process that matters
Now place animals in two ponds containing 80 and 20 individuals. If 10% move from the first to the second and 20% move from the second to the first during an interval, the resulting counts are 76 and 24 before births and deaths. Total abundance remains 100. Movement changes local exposure and local persistence possibilities without changing the total immediately.
A model using only the total cannot recover those local consequences. A model that represents patches but assumes instantaneous mixing may still miss a barrier or a costly journey. Use Landscape Ecology for spatial structure and Movement Ecology for the relationship between paths and ecological outcomes.
Individual-based simulation becomes useful when individual location, history or behaviour affects the answer. Document the order of events: move, feed, reproduce and die is not necessarily equivalent to die, move, feed and reproduce. Record random-number seeds for reproducibility, but also run enough independent realisations to describe stochastic variation. One attractive animation is a single outcome, not a probability distribution.
The ODD protocol updated by Grimm and colleagues in 2020 provides a structured description of agent-based and other simulation models. Its overview, design concepts and details make assumptions and implementation easier to inspect. Documentation supports replication; it does not by itself establish that the model represents the ecosystem adequately.
CHAPTER 10 OF 14 · Use results and diagnose failure
10. Separate forecasts, scenarios and decisions
A forecast predicts an outcome under a stated information set and uncertainty treatment. A scenario explores what follows if specified conditions occur. A decision compares actions using ecological outcomes together with objectives, costs and values. A scenario is not automatically a probability claim, and the action with the highest predicted abundance is not automatically the preferred action under every objective.
Consider two hypothetical habitat actions with costs measured in arbitrary units. Action A costs 20 and produces an ecological benefit scored at 50 in a wet year or 10 in a dry year. Action B costs 12 and produces 30 in a wet year or 24 in a dry year. Suppose, solely for this exercise, the decision score subtracts cost from benefit and assigns wet and dry years equal probability.
Expected net score for A is 0.5 × 50 + 0.5 × 10 − 20 = 10. For B it is 0.5 × 30 + 0.5 × 24 − 12 = 15. B is preferred under this specified scoring rule. If the wet-year probability is p, A scores 40p − 10 and B scores 6p + 12. A exceeds B when 34p > 22, or p > 11/17, approximately 0.6471.
Expose what could reverse the choice
This threshold tells the decision-maker exactly which assumption matters in the exercise. It does not estimate p or show that the benefit units are appropriate. A minimum-persistence requirement, unequal consequences of failure, or a different cost representation could change the comparison. Keep the ecological predictions separate from the value judgement that turns them into a ranking.
For a real investigation, distinguish uncertainty in the action's ecological effect from uncertainty in future weather. An action might also change detection, maintenance effort or habitat use beyond the study boundary. Those pathways need their own evidence rather than being absorbed into a convenient benefit number.
Dietze and colleagues' work on iterative near-term forecasting emphasises repeated prediction, comparison with new observations and updating. Keep dated forecasts before outcomes arrive. Otherwise a revised explanation can be mistaken for a successful prediction. When new evidence changes the model, preserve both versions and explain why the decision comparison changed.
CHAPTER 11 OF 14 · Use results and diagnose failure
11. Diagnose failures at the right level
A useful diagnostic begins with the location of the failure. If a simulated transfer creates animals, inspect accounting and implementation. If counts fall only when observers change, inspect the observation process. If a good within-pond fit fails at new ponds, inspect transfer assumptions and environmental coverage. Each symptom points to a different next action; indiscriminately adding predictors can hide the original problem.
Residual patterns are questions to investigate
Suppose predictions are consistently too high after heavy rain. Possible explanations include mortality, emigration, lower detectability or a lagged habitat response. The residual pattern alone does not select one. Compare additional measurements or predictions that would distinguish the alternatives. A rainfall coefficient may improve forecasting while leaving the ecological mechanism unresolved.
Suppose only the smallest ponds are poorly predicted. Check whether volume, area or survey effort entered the model in compatible units. Examine whether a process scales with habitat size or edge length. Also check whether the small ponds differ systematically in observation conditions. A scale mismatch can look like a new biological effect.
Repair the claim as well as the model
When the data support short-term count prediction but not an intervention effect, narrow the claim. When a numerical scheme fails a known solution, repair the calculation before interpreting the ecology. When parameters remain confounded, report the combinations that are supported and identify useful additional evidence. Responsible repair may reduce apparent certainty while improving scientific usefulness.
Keep an explicit list of situations beyond the tested domain. A model fitted across mild seasonal changes may not describe a newly drained pond, an introduced predator or a changed survey method. Extrapolation is sometimes necessary, but it should be named and explored through alternative plausible structures and assumptions.
Check temporal aggregation too. A monthly mean can conceal a short period that determines survival. The specialist guide to why daily mean oxygen does not determine the nighttime minimum shows why the summary statistic must match the mechanism. A model can calculate an average correctly while answering the wrong ecological question.
CHAPTER 12 OF 14 · Transfer, practise and continue
12. Transfer the reasoning to a changed setting
Move from pond counts to seedlings planted in two hypothetical restoration plots. Each plot begins with 200 seedlings. At the follow-up, observers record 96 in Plot A and 108 in Plot B. A tempting conclusion is that survival was higher in B. For this exercise, independent survey calibration gives detection probabilities 0.6 in A and 0.9 in B, assumed known and applicable to the surviving seedlings.
Dividing counts by detection gives abundance estimates of 160 and 120. Under the additional assumptions of no recruitment, no movement and no unrecorded removal, estimated survival fractions are 160/200 = 0.8 and 120/200 = 0.6. The ranking reverses after the observation process is represented. These are point estimates under an invented model; a real analysis would carry count and detection uncertainty through to the comparison.
Change one assumption and explain what breaks
If detection in A is actually between 0.5 and 0.8, its count-based abundance estimate ranges from 192 to 120. That range alone reaches Plot B's estimate. If recruitment occurred, surviving original seedlings cannot be identified from total abundance without further information. If the plots were selected differently, their survival contrast does not by itself estimate the effect of a treatment.
Now ask a new question: which plot will contain more mature plants in two years? Today's survival estimates do not settle it. Growth, future survival, stage transitions and possibly competition become relevant. The modelling purpose has changed, so the state representation and evidence requirements may change too. Do not carry a successful calculation into a new question without checking its meaning.
A strong transfer explanation names four things: the ecological state, the observation mechanism, the assumptions connecting the calculation to the state, and the new evidence needed for the next question. The arithmetic is deliberately modest. The difficult work is preserving those distinctions when the organism and setting change.
This is the educational purpose described in The Gold Standard of Education: understanding remains useful beyond the original exercise. Repeat the transfer test with acoustic detections, vegetation cover or visits to habitat patches, and explain which parts of the observation model must be rebuilt rather than copied.
CHAPTER 13 OF 14 · Transfer, practise and continue
13. Build a practical learning and modelling workflow
Begin with a paper model of one ecological balance and a small hand-checkable dataset. Define the question and write a state dictionary. Draw the causal or process relationships you intend to represent. Then produce a baseline prediction and one mechanism-based alternative. At this stage, clarity about what the model counts is more valuable than an elaborate interface.
Next, implement a reproducible calculation with named inputs and units. Keep raw observations separate from transformed data. Record exclusions, missing values and changes to measurement protocols. Save the code version, parameter values, input data version and random seeds where relevant. A second person should be able to reconstruct a result without guessing which spreadsheet tab was final.
Use a compact model record
For the pond investigation, the record should identify the forecast target and horizon, the twelve-pond domain, survey protocol, ecological and observation equations, fitting period, held-out period, uncertainty treatment and baseline. Add the tests performed, known failures and reasons for each revision. Include a clear statement of what the model has not established, such as the causal effect of vegetation planting.
Plan fieldwork from the model's unresolved distinctions. If detection and abundance remain confounded, improve the observation design. If movement and reproduction cannot be separated, collect evidence about one of those processes. If transfer to new ponds fails, investigate environmental differences and sampling coverage. This turns modelling into a way of asking better empirical questions rather than a stage performed after fieldwork is finished.
Use the specialist Dynamic Occupancy route for changing site states, Population Viability Analysis for conditional persistence risks, and Ecological Informatics for traceable ecological data. Return to Mathematical Biology when the equations themselves need deeper study.
For broader connections, the How X Works hub offers mechanism-based reading routes, while the Gold Standard master connects subjects and practical applications. Choose the next article by the question that remains unresolved. Reading another model is most useful when you can explain what new distinction it helps you preserve.
CHAPTER 14 OF 14 · Transfer, practise and continue
14. Questions to check before trusting a result
Does a better fit mean a better ecological explanation?
Not necessarily. Additional flexibility can improve fit without identifying the mechanism. Ask which new predictions differ between explanations and whether independent observations support them. A useful forecast can remain mechanistically incomplete; label its achievement accurately.
Is a simple model always preferable?
No. A model is too simple when it removes a distinction necessary for the question, such as life stage, movement or imperfect detection. It is unnecessarily complicated when added detail cannot improve the relevant answer or be supported by evidence. Compare purposeful alternatives rather than treating simplicity or complexity as a virtue on its own.
Can a model be validated once and then reused everywhere?
Evidence of performance belongs to a task, domain, data process and period. A model tested for interpolation among sampled ponds has not thereby demonstrated transfer to a new climate or survey method. Keep evaluating it as the intended use changes.
Why can two models have the same average but different risk?
An average does not specify a distribution. In the founder example, shared failure and independent failure produced the same expected descendant count but different probabilities of zero. When the decision concerns a threshold or rare event, examine that event directly and state the horizon.
What should accompany a published ecological forecast?
Provide its issue date, target, horizon, location, input information boundary, uncertainty representation and comparison method. Preserve the forecast before the outcome is observed. Explain revisions and distinguish an updated forecast from a retrospective fit.
Which sources support further study?
The primary and official sources linked in this guide were checked on 27 October 2026: EPA's environmental model guidance (2009); MacKenzie and colleagues on imperfect-detection occupancy (2002); Roberts and colleagues on structured cross-validation (2017); Dietze and colleagues on iterative ecological forecasting (2018); Grimm and colleagues' ODD update (2020); and David Tong's Cambridge mathematical biology teaching notes. Their publication dates identify the works; the checking date does not imply that each is a newly issued standard.
All worked cases here are original teaching constructions. Their answers follow the explicitly stated assumptions, not a claim that real ponds or restoration plots behave exactly this way. The durable skill is to move between ecological meaning, calculation and evidence without losing track of which part supports the conclusion.