EVENTSTUDY_KB: Grounding Knowledge Base for Event Study Diagnostics

In EventStudy: Event Study Analysis

Grounding Knowledge Base for Event Study Diagnostics

Description

EVENTSTUDY_KB is a pure-R list of rule records that maps diagnostic conditions (read from an es_diagnostics object) to grounded methodological recommendations, each carrying a structured academic citation and a severity level.

Package-level list; use es_kb() to access.

Usage

EVENTSTUDY_KB

Details

The KB is built at package-load time from calls to .kb_rule() and never depends on a network connection or LLM provider. It is the correctness-critical layer that the offline advice functions (recommend_stat, flag_robustness) and the Phase 7 LLM layer both evaluate against.

KB-04 scope note: Phase 5 delivers the KB data structure only – exported and serializable, ready for Phase 7 system-prompt injection. The actual prompt-injection behavior is a Phase 7 deliverable and is deliberately out of scope here.

Rule record fields

Each element of EVENTSTUDY_KB is a list with:

id

Unique character identifier (e.g. "KB-NORM-PATELL").

category

Either "stat_choice" (steers which test to use) or "robustness" (data-quality / reliability warning).

condition

A function(diag) that accepts an es_diagnostics list and returns a length-1 logical. Returns FALSE when any required field is NA (NA-safe; never errors on missing data).

recommendation

Character. The methodological action to take.

citation

Named list: author (character), year (integer), key (character), venue (character).

severity

One of "info", "warning", "error".

Threshold notes

The following thresholds are literature-informed defaults and are marked [ASSUMED] – they can be tuned without breaking the KB contract:

• Shapiro-Wilk p-value threshold: 0.05

• Proportion for normality-holds rule: \>= 70% of events

• Proportion for non-normality rule: \>= 50% of events

• Durbin-Watson bounds: [1.5, 2.5]

• R-squared low-fit threshold: < 0.05

• Small-N threshold: < 10 valid events

• Degenerate-event threshold: < 80% of events fitted

• High-dispersion CAR IQR threshold: > 0.10 or SD > 0.15

EventStudy documentation built on Oct. 31, 2026, 1:07 a.m.