The ultimate test of a retrospective trigger: would it have flagged the outbreaks that actually drew emergency money, and early enough to act? Northeast Nigeria received two CERF Rapid Response cholera allocations in the study period.
c = cerf.copy()
c["date"] = c["date"].dt.date
c.columns = ["Allocation date", "Label", "Note"]
c.style.hide(axis="index")| Allocation date | Label | Note |
|---|---|---|
| 2018-10-31 | CERF RR 2018 | Response to 2018 BAY-states cholera outbreak |
| 2021-10-22 | CERF RR 2021 | Response to 2021 BAY-states cholera outbreak |
fig, ax = plt.subplots(figsize=(11, 4.5))
y_of = {s: i for i, s in enumerate(U.STATES)}
for st in U.STATES:
d = act[act["state"] == st]
ax.scatter(d["week"], [y_of[st]] * len(d), s=70, color=U.STATE_COLORS[st], zorder=3, label=st)
for _, r in cerf.iterrows():
ax.axvline(r["date"], color=U.HDX["gray_dark"], ls="--", lw=1.3, zorder=2)
ax.text(r["date"], 2.55, r["label"], rotation=90, va="bottom", ha="right",
color=U.HDX["gray_dark"], fontsize=9)
ax.set_yticks(list(y_of.values()))
ax.set_yticklabels(U.STATES)
ax.set_ylim(-0.6, 2.9)
ax.xaxis.set_major_locator(mdates.YearLocator())
ax.xaxis.set_major_formatter(mdates.DateFormatter("%Y"))
ax.set_title("Trigger activations vs CERF Rapid Response dates")
ax.grid(axis="y", visible=False)
plt.tight_layout()
plt.show()
rows = []
for _, r in cerf.iterrows():
yr = r["date"].year
prior = act[(act["week"] <= r["date"]) & (act["week"].dt.year == yr)]
if len(prior):
first = prior.sort_values("week").iloc[0]
lead_days = (r["date"] - first["week"]).days
rows.append([
r["label"], r["date"].date(), first["ADM2_EN"], first["state"],
first["week"].date(), len(prior), f"{lead_days} days (~{lead_days // 7} wks)",
])
else:
rows.append([r["label"], r["date"].date(), "—", "—", "—", 0, "no prior activation"])
lt = pd.DataFrame(rows, columns=[
"CERF event", "CERF date", "First-firing LGA", "State",
"First activation", "Activations before CERF", "Lead time",
])
lt.style.hide(axis="index")| CERF event | CERF date | First-firing LGA | State | First activation | Activations before CERF | Lead time |
|---|---|---|---|---|---|---|
| CERF RR 2018 | 2018-10-31 | Mubi North | Adamawa | 2018-05-21 | 9 | 163 days (~23 wks) |
| CERF RR 2021 | 2021-10-22 | Damaturu | Yobe | 2021-08-16 | 5 | 67 days (~9 wks) |
In both CERF years the trigger activates weeks-to-months ahead of the allocation date, in multiple LGAs. That is the encouraging headline: on this retrospective data the design would have surfaced the 2018 and 2021 outbreaks with genuine lead time, consistent with the earlier exploration’s finding that the trigger anticipates the CERF events by roughly 3–9 weeks.
The lead times above are measured against onset-dated cases. In production the clock that matters is the reporting clock — when the LGA’s cases actually reach an analyst. Nothing in this file measures that lag, and it is the single biggest open question for a BAY cholera framework. A trigger that fires 6 weeks before CERF on onset dates delivers zero lead time if the data arrives 6 weeks late.
Other caveats:
On the 2018–2023 line-list, a percentile-plus-growth trigger behaves the way an AA trigger should: dormant between outbreaks, lit up during them, and firing ahead of both historical CERF allocations. The epidemiology (sharp seasonal waves concentrated in a handful of LGAs) supports an LGA-level observational trigger. The blockers are not epidemiological but operational and evidentiary: data timeliness is unconfirmed, the 2020 gap limits calibration, and the threshold and priority-LGA choices need explicit, documented justification before any of this becomes an endorsed framework.