4 — Results

Manuscript · APA section

Manuscript · Section 4 of 7 · Paper landing

4.1 Descriptive statistics

Across N = 7,327 valid district-month observations, the raw typhoid case distribution is strongly right-skewed (mean 462.5, median 325.0, max 2,771.8 cases per district-month). The log(Cases + 1) transformation substantially reduces distributional asymmetry. Precipitation exhibits wide variability (range 0–1,100 mm) reflecting Nepal’s diverse topography and monsoon concentration; temperature spans nearly 50 °C (−18 to +31.5 °C) across districts and seasons. Flood events are rare in most district-months (89 % of observations record zero events).

4.2 Temporal trends — National Trend

Figure 1: **Figure 4.1.** National annual typhoid case counts (red) and recorded flood events (blue, secondary axis), 2015–2023. Generated by `manuscript/main.py`.

The time series displays a strikingly consistent monsoon-driven seasonal cycle: every year exhibits a pronounced peak during or immediately following the June–September monsoon, with incidence rising sharply from May–June and declining through October–November. National monthly case totals ranged from ~15,000 in driest winter months to ~86,000 during the 2020 monsoon peak — a 6:1 peak-to-trough ratio. A consistent declining trend in peak incidence from 2021 onwards is observed, with 2022–2023 monsoon peaks 20–30 % below the 2015–2019 average.

4.3 Bivariate correlation — Figure 4.3

Figure 2: **Figure 4.3.** Top-10 predictors of typhoid incidence: XGBoost normalized feature-importance scores.

Among climate variables, temperature exhibits the strongest direct association with typhoid incidence: contemporaneous mean temperature r = 0.50 with raw cases (r = 0.63 with log(Cases)). One-month lagged temperature shows r = 0.48 (raw) and r = 0.60 (log). The near-equivalence reflects high month-to-month autocorrelation of temperature (r = 0.93 between lag-0 and lag-1).

Precipitation showed a weaker positive association (r = 0.37 raw; r = 0.33 log). The three-month rolling precipitation average showed even weaker correlations (r = 0.22; r = 0.21) — the more immediate monthly precipitation effect is more relevant than cumulative three-month exposure. Humidity (r = 0.22 raw) operates primarily as a co-factor. Flood events demonstrated modest but significant positive correlations (r = 0.19 raw), reflecting their episodic and spatially concentrated character.

4.4 Lag correlation ranking

Rank	Feature	Pearson r with log(Cases)
1	Cases (lag-1)	0.731
2	Mean Temperature (lag-1)	0.599
3	Temperature 3-month rolling avg	0.559
4	Precipitation (lag-1)	0.313
5	Monsoon Indicator	0.245
6	Precipitation 3-month rolling avg	0.208
7	Humidity (lag-1)	0.180
8	Flood Events (lag-1)	0.122

The autoregressive cases_lag1 (r = 0.731) confirms the strong autoregressive structure of typhoid transmission at the district level. Lagged temperature variables (r = 0.559–0.599) emerge as the strongest climate drivers — substantially stronger than precipitation (r = 0.313), reflecting Nepal’s extraordinary altitudinal range and temperature’s influence through multiple simultaneous biological pathways.

4.5 Three-model performance — Figure 4.4

Figure 3: **Figure 4.4.** RMSE, MAE, and R² across Random Forest, XGBoost, and MLP on the held-out 12-month test window (January 2022 – December 2023).

Table 4.1 — Three-model performance on the held-out 2022–2023 test set

Model	RMSE	MAE	R²	MAPE
Random Forest	131.43	72.10	0.8563	41.48 %
XGBoost ★	129.07	69.50	0.8614	44.64 %
MLP	128.07	71.45	0.8635	43.50 %

All three architectures converge within 0.01 R² of each other (0.856–0.864) on the held-out test set, indicating that the climate × autoregressive signal — not the model family — sets the predictive ceiling.

XGBoost is selected as the operational model under a single-model constraint, based on:

Lowest MAE (69.50 cases per district-month) — operationally the most relevant absolute-error metric.
Interpretable gain-based feature importance accessible to non-technical stakeholders.
Graceful missing-value handling — real-world DHIS2 ingestion is routinely incomplete.
Reproducibility under a fixed random seed — tree ensembles are deterministic; MLP trajectories vary with batch order.

4.6 Ecological-zone distribution — Figure 5

Figure 4: **Figure 4.5.** Boxplot of district-monthly typhoid case counts by ecological zone (Terai / Hill / Mountain), 2015–2023. Symmetric log-scaled y-axis.

Terai districts show the highest medians and the widest interquartile ranges — consistent with greater flood exposure, denser population, and structural WASH deficits. Mountain districts cluster at the low end, reflecting smaller, dispersed populations where density-dependent transmission chains are difficult to sustain. The symmetric log-scaled y-axis preserves visibility of both the long right tail and the bulk of the distribution.

4.7 Pearson correlation matrix — Figure 4.2

Figure 5: **Figure 4.2.** Pearson correlation matrix among typhoid cases, flood events, and three climate variables. Generated by `manuscript/main.py`.

(Pearson correlations: temperature lag-1 r = 0.60 with log(Cases); precipitation lag-1 r = 0.31; monsoon indicator r = 0.25; humidity lag-1 r = 0.18; flood lag-1 r = 0.12. All p < 0.001.)

← Back: 3 — Methodology · → Next: 5 — Discussion