- engine.js / analyze.py: one MISSING set ("Unknown", "Unsure",
"Unsure/unknown", "Multiple", "(not recorded)", "(not asked)", ...).
Factor tables compare only runs that answered each question and report how
many were left out; models are complete-case for every choice variable;
route rankings group only fully known routes; the route flow skips
missing steps.
- "I've lost count" now counts as a later attempt.
- The chase goal keeps its "not asked" stratum as an adjustment only; it is
no longer displayed as a finding.
- Page: no unknown levels in charts, tables or filters; factor explorer notes
how many runs were left out; Methods explains the rule.
JS/Python parity still exact.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UCzTd31GphpMXDhQqWrCMQ
61 lines
4.0 KiB
Python
61 lines
4.0 KiB
Python
"""Recompute route rankings in Python (crude Fisher and stratified Mantel-Haenszel) and compare with engine.js."""
|
||
import json, subprocess, sys
|
||
sys.path.insert(0, "research/python")
|
||
import pandas as pd
|
||
from stats_core import odds_ratio, mantel_haenszel, bh
|
||
d = pd.DataFrame(json.load(open("tests/out/runs_js.json")))
|
||
js = json.loads(subprocess.check_output(["node", "-e", """
|
||
const E=require('./packages/core/src/engine.js'),fs=require('fs');const raw=JSON.parse(fs.readFileSync('data/public/form_responses.json'));
|
||
const {runs}=E.codeRows(raw.rows,raw.headers);const o={};for(const t of ['see_Blue','bond_Blue','see_shadow','bond_Green']){o[t]=E.rankRoutes(runs,t,10)}
|
||
console.log(JSON.stringify(o));"""]))
|
||
MISSING = {"Unknown", "Unsure", "Unsure/unknown", "Multiple", "(not recorded)", "(not asked)", "(not asked/blank)", "Unknown/lost count", "Blank/not asked"}
|
||
miss = lambda v: not isinstance(v, str) or v == "" or v in MISSING # noqa: E731
|
||
def stage(r):
|
||
# a route is only ranked from runs whose every step is known (mirrors rankRoutes in engine.js)
|
||
if r.scream.startswith("DIED") or miss(r.scream) or miss(r.resolution): cad = None # noqa
|
||
else: cad = ("Help → " + r.resolution) if r.scream == "Investigate & help" else (r.scream + " → " + r.resolution)
|
||
hour = {"Tree": "Climb the tree", "Valley": "Valley centre"}.get(r.hour)
|
||
fs = None if r.scream == "DIED at first sighting" or miss(r.first_sighting) else r.first_sighting
|
||
post_null = not isinstance(r.post_path, str)
|
||
post = None if post_null or miss(r.post_path) else r.post_path
|
||
vals = [None if miss(r.company) else r.company, None if miss(r.direction) else r.direction, fs, cad, post, hour]
|
||
if any(v is None for i, v in enumerate(vals) if not (i == 4 and post_null)): return None
|
||
return " › ".join(v if v else "—" for v in vals)
|
||
def stratum(r):
|
||
if r.source != "Form": return "bulk"
|
||
return {"Blue": "blue", "Black": "black", "Blue or Black": "blue-or-black"}.get(r.goal, "other")
|
||
d["route"] = d.apply(stage, axis=1); d["stratum"] = d.apply(stratum, axis=1)
|
||
d["branch"] = d.branch.where(d.branch.notna(), None)
|
||
T = {"see_Blue": (lambda x: x.branch == "Passed the Red: meadow", lambda x: x.saw_blue_meadow),
|
||
"bond_Blue": (lambda x: x.hour == "Valley", lambda x: (x.outcome == "Bonded") & (x.outcome_color == "Blue")),
|
||
"see_shadow": (lambda x: (x.hour == "Tree") & x.branch.notna(), lambda x: x.saw_shadow),
|
||
"bond_Green": (lambda x: x.hour == "Tree", lambda x: (x.outcome == "Bonded") & (x.outcome_color == "Green"))}
|
||
fails = 0
|
||
for t, (pool, ev) in T.items():
|
||
P = d[pool(d)].copy(); P["e"] = ev(P).astype(bool)
|
||
K, N = int(P.e.sum()), len(P)
|
||
rows = {}
|
||
for route, g in P[P.route.notna()].groupby("route"):
|
||
if len(g) < 10: continue
|
||
k, n = int(g.e.sum()), len(g)
|
||
orr = odds_ratio(k, n - k, K - k, N - n - (K - k))
|
||
tabs = []
|
||
for s, sg in P.groupby("stratum"):
|
||
inr = sg.route == route
|
||
a = int((inr & sg.e).sum()); nin = int(inr.sum()); c = int((~inr & sg.e).sum()); nout = len(sg) - nin
|
||
tabs.append([a, nin - a, c, nout - c])
|
||
rows[route] = (k, n, orr, mantel_haenszel(tabs))
|
||
jr = {r["route"]: r for r in js[t]["rows"]}
|
||
if set(jr) != set(rows): print(t, "route sets differ", set(jr) ^ set(rows)); fails += 1
|
||
for route, (k, n, orr, mh) in rows.items():
|
||
j = jr.get(route)
|
||
if not j: continue
|
||
if (k, n) != (j["k"], j["n"]) or abs(orr[0] - j["or"]) > 1e-6 * max(1, orr[0]) or abs(orr[3] - j["fisher_p"]) > 1e-6:
|
||
print("crude diff", t, route); fails += 1
|
||
if mh is None:
|
||
if j["mh_or"] is not None: print("mh none diff", t, route); fails += 1
|
||
elif j["mh_or"] is None or abs(mh[0] - j["mh_or"]) > 1e-6 * max(1, mh[0]) or abs(mh[3] - j["mh_p"]) > 1e-5:
|
||
print("mh diff", t, route, mh, j["mh_or"], j["mh_p"]); fails += 1
|
||
print(f"route rankings for {len(T)} outcomes: {'ALL MATCH' if not fails else str(fails) + ' differences'}")
|
||
sys.exit(1 if fails else 0)
|