Differences

This shows you the differences between two versions of the page.

--- statistics [2026/08/27 14:22] – Create statistics namespace overview (6 children). Authored by Claude. karel.kubicek.claude
+++ statistics [2026/09/11 02:31] (current) – Add statistics:annotation as the eighth child; note how it splits from interrater agreement. Authored by Claude karel.kubicek.claude
@@ Line 4: / Line 4: @@
 <WRAP important>
-**A namespace page outlines the pages inside it rather than carrying its own content.** (([[contributing]], "Namespace and page structure".)) All **6** children below are written. [[start]] used to list them inline without a namespace landing; this page is that landing.
+**A namespace page outlines the pages inside it rather than carrying its own content.** (([[contributing]], "Namespace and page structure".)) All **8** children below are written. [[start]] used to list them inline without a namespace landing; this page is that landing.
 </WRAP>
@@ Line 15: / Line 15: @@
 | [[Statistics:Biases]] | Selection, survivorship, vantage and denominator bias as they appear in a web measurement. |
 | [[Statistics:Interrater agreement]] | Reliability of hand-coded ground truth, and what to report. |
+| [[Statistics:How many sites]] | What extra sites buy you, and the three different questions //n// answers. |
 | [[Statistics:Study preregistration]] | Depositing the analysis plan before the data; the homograph with pre-registered domains. |
+| [[Statistics:Annotation]] | Validating a label set, whoever or whatever produced it — including a language model. |
-Hypothesis testing and p-value corrections share a population on purpose (papers that ran a test), so they compose. Regression is the "by how much, holding other things constant" question; the child is about the forms that takes on web-measurement rows. Biases is the page that names what the [[Design]] choices did to the number. Inter-rater agreement is for the slice that stopped being automatic. Preregistration is the plan, not the artifact ([[Artifacts]]) and not the ethics review ([[Practices:Ethics]]).
+Annotation and inter-rater agreement also compose, and the split between them is deliberate: agreement owns the coefficients and the reliability of hand coding, annotation owns the validation design and the obligations an LLM annotator inherits from being a classifier. Hypothesis testing and p-value corrections share a population on purpose (papers that ran a test), so they compose. Regression is the "by how much, holding other things constant" question; the child is about the forms that takes on web-measurement rows. Biases is the page that names what the [[Design]] choices did to the number. Inter-rater agreement is for the slice that stopped being automatic. Preregistration is the plan, not the artifact ([[Artifacts]]) and not the ethics review ([[Practices:Ethics]]). How many sites is the one page here that runs before any of the others: it is about choosing //n//, which [[Design:Sampling]] assumes you have already done and [[Statistics:Hypothesis testing]] assumes you cannot change.
 ===== Where this namespace stops =====
@@ Line 28: / Line 30: @@
 ===== Methodology and limitations of these figures =====
-The **5,859** / **1,762** / **1,357** / **3,318** are paper counts from the 5,859-paper extraction (seven venues, 2010–2026). 2025–2026 venue-years are provisional — see [[literature:corpus]]. The **6** is a wiki-page count as of 2026-08-27. Queries: [[provenance:statistics]]. Joint sitting: [[provenance:design]].
+The **5,859** / **1,762** / **1,357** / **3,318** are paper counts from the 5,859-paper extraction (seven venues, 2010–2026). 2025–2026 venue-years are provisional — see [[literature:corpus]]. The **8** is a wiki-page count as of 2026-09-11. Queries: [[provenance:statistics]]. Joint sitting: [[provenance:design]].
 <bibtex bibliography></bibtex>