One question runs through everything:
Would a wider uncertainty make this claim easier to support?
If it would, the claim is built wrongly, whatever else is true of it.
Before that makes sense, one thing has to be cleared up. Data is not good or bad. A blurred photograph is a perfectly faithful record of what a blurred photograph records. Nothing about it is bad. What differs between one frame and another is not quality but how much it constrains — how wide the uncertainty is.
A wide uncertainty, honestly stated, is not a fault. It is frequently the most useful thing you can report, and nothing in this course treats a wide result as a poor one. What is a fault is a stated uncertainty that is wrong: a chart that has faded since last year, a camera that was never calibrated. Those look exactly like good numbers and misrepresent themselves, which is why you will be recording batches and versions of things that seem like they should not need recording.
With that settled, the test is a counterfactual you run on your own work. If my measurement had been less certain, would this conclusion have been easier to reach? If the answer is yes, the conclusion is not coming from the reef. It is coming from the uncertainty.
Four ways that happens, all of them the same way:
| The claim | Widening the uncertainty about... | ...increases |
|---|---|---|
| "A third of this coral is bleached" | whether a white pixel is bleached tissue or a clipped highlight | the bleached count |
| "No change was detected" | the mean, by surveying fewer stations | the chance nothing clears the spread |
| "Damage is widespread" | which of two similar things you are seeing, by merging them | the merged class |
| "The two surveys agree" | how the methods differed, by re-scoring the older one | apparent agreement |
The second one is the most common and the hardest to spot, because it looks like caution. Less effort made "nothing detected" more likely. That is not a conservative claim; it is an anti-conservative claim in modest clothing.