Conservation decisions in R
Most conservation analysis ends one step short of the thing it was commissioned for. A model says how many species a site holds, or how likely a founded population is to persist, or how far a restored patch has come. None of those is a decision. The decision is which sites to buy, how to spend a fixed number of animals, whether to survey again before acting, and it involves objectives that pull against each other, a budget, and uncertainty that will not be resolved in time.
These tutorials build that last step by hand in base R. They share one working assumption: the arithmetic is the easy part, and almost every recommendation that falls apart falls apart on a choice made before the arithmetic started. Which currency the plan is measured in, how the objectives were scaled, which alternatives were on the table, what the model was allowed to assume about the future. Each tutorial simulates a system whose truth is known, so the plan a method produces can be scored against the plan that was actually best, rather than against itself.
Choosing where to act
Reserve selection is the oldest formal decision problem in conservation and the clearest illustration of why a ranking is not a plan.
- Complementarity and reserve selection asks what a site adds rather than what it holds, which is the difference between protecting a set of species and protecting the same rich place several times over.
- Conservation costs and return on investment puts a price on land and shows what changes when the budget is money rather than hectares, including the case where costing the land buys nothing at all.
- Compactness in reserve design writes the boundary length penalty and the annealing search out in full, then prices tidiness: what a compact reserve costs in the currency the plan was meant to maximise.
- Implementing a plan while land is lost buys a finished plan out slowly while unprotected units are cleared, and measures the annual loss rate above which following the list stops paying and re-solving every year takes over.
- Checking a conservation prioritisation scores greedy selection against the exact optimum, replaces the single map with a selection frequency, and shows that the choice of target moves the answer further than the algorithm does.
Putting populations back, and measuring recovery
Once a site is secured the questions become demographic, and the projections run far past the data that supports them.
- Reintroduction release strategies treats a fixed number of available animals as a budget to allocate across years and life stages, and finds where the marginal animal stops being worth taking from the donor population.
- SLOSS and reserve configuration answers the single large or several small question twice, once for species richness and once for persistence, and shows that the two answers turn on different properties of the system and can point in opposite directions.
- Restoration trajectories and recovery measures how far a restored site has come, and takes apart the three decisions hiding inside that percentage: the metric, the reference, and whether recovery was watched through time or inferred across sites of different ages.
- Checking a reintroduction analysis breaks the viability projection on purpose: the horizon it cannot support, the uncertainty source that standard software does not propagate, correlated bad years, and a density dependence assumption no young population can test.
Making the choice explicit
The formal decision layer sits above all of the above, and its job is to separate what the ecology says from what the values say.
- Structured decision making in R builds a consequence table, removes the alternatives that need no value judgement at all, elicits weights the way that respects how much is actually at stake, and shows the scaling choice that quietly changes the winner.
- The expected value of information prices the survey before it is run, and demonstrates the result that surprises most ecologists: uncertainty that does not change the best action is worth nothing to resolve, however large it is.
- Adaptive management and learning manages a system while still deciding how it works, and finds the conflict at the centre of the idea, which is that the action paying best today is usually the one that teaches least.
- Eliciting probabilities from experts fills the cell of the table that has no data behind it, and treats the expert as an instrument in need of calibration rather than an authority to be averaged: stated intervals come out too narrow, and the pooling rule decides how much of that narrowness survives into the combined number.
- When decision makers disagree starts where a consequence table nobody disputes still fails to produce agreement, applies the aggregation rules that all look reasonable written down, and shows that a recommendation is partly a statement about the rule that produced it.
- Checking a decision analysis asks how firm a recommendation really is, whether the uncertainty everyone worries about is the uncertainty that matters, and what a single new alternative does to the whole table.
Deciding again next year
A consequence table settles one choice at one moment. When the action taken now changes the state the next choice starts from, the object being solved for stops being a decision and becomes a rule.
- Markov decision processes for management sets out the three pieces a dynamic problem needs, solves for the rule rather than the action, and measures how much the short-sighted answer costs when the state is allowed to move.
- Stochastic dynamic programming for harvest derives the harvest rule instead of comparing rules somebody proposed, and separates the assumptions that bend the rule from the ones that only move its level.
- When to stop monitoring treats the decision to survey again as the decision it is, and shows why more data can be worth less than the delay it buys.
- Checking a sequential decision model attacks the machinery: the discount rate, the grid the state was carved into, a transition matrix that was assumed rather than measured, and the end of the horizon.
What these tutorials assume
Base R and one plotting package throughout. The optimisers, the annealing search, the Bayesian updates and the value of information calculations are written out rather than called from a decision package, because the failures worth understanding live inside those steps. Every system is simulated, so the best available plan is known while the method runs. Costs, weights and utilities are treated as inputs supplied by people rather than quantities estimated from data, and the tutorials try to be explicit about where one stops and the other starts.
Where this connects
A conservation decision borrows its inputs from most of the rest of this site, and reads better once those inputs are understood as models rather than facts.
- Population models in R supplies the projection machinery that a viability analysis and a release schedule both run on, including what a matrix model can and cannot say about the future.
- Landscape connectivity in R provides the between-site distances that spatial reserve design and reintroduction siting depend on, and is candid about how much of those distances is assumption.
- Species coexistence in R is where the question of what a restored community will settle into gets its theory, including the possibility that it settles somewhere other than the reference.
- Choosing a decision threshold from costs is the same logic applied to a single yes or no call from a model, which is the smallest decision problem there is.
- Population viability analysis and extinction risk is the projection that most conservation decisions quietly rest on, with an honest account of what its numbers are entitled to claim.
- Tail dependence and joint extremes attacks the assumption underneath every spread-the-risk argument on this page, that a correlation coefficient settles how often the sites fail together, and shows two networks agreeing on every correlation you can measure while disagreeing wildly about the year they all go under.
- Copulas for dependent ecological data is the machinery that makes that possible: dependence chosen separately from the marginal distributions, which is what a viability model needs once the species it simulates stop being symmetric.