cif_vs_naive_km.png

def fig_cif_vs_naive_km(df) -> None:
    with warnings.catch_warnings():
        warnings.simplefilter("ignore")
        # Seeded so the tie-jitter is reproducible; clip the last jittered row
        # (it can sit just past day 30 with a tiny risk set and spike the curve).
        ajf = AalenJohansenFitter(seed=SEED)
        ajf.fit(df["event_time"], df["event_state"], event_of_interest=1)
    cif = ajf.cumulative_density_.loc[:ADMIN_HORIZON]

    kmf = KaplanMeierFitter()
    kmf.fit(df["event_time"], event_observed=(df["event_state"] == 1))
    naive_cif = (1 - kmf.survival_function_).loc[:ADMIN_HORIZON]

    fig, ax = plt.subplots(figsize=(7.5, 4.5))
    ax.step(naive_cif.index, naive_cif.values.ravel(), where="post", color=EVENT, lw=2,
            label="1 − Kaplan-Meier (ignoriert Entlassung als Konkurrenzrisiko)")
    ax.step(cif.index, cif.values.ravel(),
            where="post", color=PRIMARY, lw=2.4,
            label="Cumulative Incidence Function (korrekt, Aalen-Johansen)")
    ax.set_xlabel("Tage seit Aufnahme")
    ax.set_ylabel("Kumulative Inzidenz Tod")
    ax.set_ylim(0, 0.40)
    ax.set_title("1 − Kaplan-Meier überschätzt das Sterberisiko bei konkurrierenden Risiken")
    ax.legend(loc="upper left", fontsize=9.5)
    save(fig, ASSETS / "cif_vs_naive_km.png")