How many sites do I need?
Rehearse your study on a computer before the field season: simulation-based power analysis for sites, traps, animals, years or loci.
The idea: rehearse the study on a computer
Before you spend a field season, you can ask: if the truth were what I expect, would my design detect it, and would the estimate be accurate? CascadiaTools answers this by simulation. It creates hundreds of fake data sets that look like the study you are planning, with an effect of the size you care about built in. It analyses each one with the same module you will use for the real data, and counts how often the effect is detected (that share is the power). It also checks that the estimates are unbiased and the intervals honest at each design, because power from a biased method is meaningless.
This works for any module: sites and visits for occupancy, traps and occasions for spatial capture-recapture, animals and fixes for movement, years for trend detection, loci and samples for genetics.
What the agent does (the eco-power workflow)
Pin down the effect that matters
It asks what change would matter for management, for example “a 20% drop in occupancy per standard deviation of road density”, and converts it to the model's scale, showing you the conversion.
Set realistic values for everything else
Detection probability, baseline occupancy or density, movement scale and so on, from your pilot data or the literature. Values you are unsure about become extra scenarios.
Choose what to vary
Usually one “knob” (number of sites, surveys, traps, years, animals or loci) over 4 to 6 values that span what you could afford.
Simulate and fit, many times
Typically 200 simulated data sets per design value, which pins power down to about ±0.03 near 80%. It runs on your computer, or on a computing cluster if you have one.
Read power only where the method works
At each design it checks bias and interval coverage as well as power. It also runs an effect of zero to confirm that false positives stay near 5%.
Report
A power curve, the smallest design reaching 80% power, the assumptions behind it, and how sensitive the answer is to the uncertain values. Trade-offs (more sites or more visits?) get a comparison at the same total effort.
A real result: how many sites for occupancy?
From the validation of the single-species occupancy module: 4 surveys per site, occupancy 0.5, detection 0.4 per survey, 100 simulated data sets per design. The habitat effect is a slope of 1 on the logit scale per standard deviation of the covariate.
| Sites | Power for the habitat effect | Bias of the effect | Coverage |
|---|---|---|---|
| 50 | 0.63 | +23.3% | 0.99 |
| 100 | 0.93 | +9.6% | 0.97 |
| 200 | 1.00 | +4.0% | 0.96 |
| 400 | 1.00 | +3.2% | 0.92 |
Read it like this: about 70 to 100 sites give 80% power, but below about 150 sites the slope is overestimated (+23% at 50 sites). Part of the “power” at small designs comes from that overestimation. For an accurate effect size, not just its direction, plan for 150 to 200 sites. This is exactly the kind of trade-off a power analysis should show you.
How to ask
Describe the study you are planning, what you expect, and what you can afford. Replace the parts in [square brackets].
I'm planning a camera-trap occupancy study for [species] next summer. From a pilot, detection is about [0.3] per week and about [40%] of sites are occupied. I can deploy cameras for [6] weeks, and I can afford between [60] and [150] stations. I want to detect whether occupancy is lower near roads, and an effect that matters to managers would be [a 20% drop in occupancy per km closer to a road]. How many stations do I need? Please check bias and coverage as well as power, and give me a report with a power curve.
Other examples: “How many hair snares and sessions do I need to estimate marten density within ±20%?”, “How many years of counts to detect a 5% annual decline?”, “How many loci do I need to estimate Ne?”.
Power analyses fit hundreds of models, so the Bayesian modules can take hours. The agent times one fit first and tells you how long the full run will take.
