DRM methods and results

Scope and Inputs

This report documents the DIPI workflow and outputs generated from the DIPI panel file:

• Panel input: dipi_hiv_tb_malaria_PANEL_20260527_215631.xlsx
• Report generated: 20260527_215634

Methods (Step-by-step)

1. Start from disease spending series from the production outputs.
2. Merge DALY inputs and compute country-specific DALY shares:
   • DALY_share_H, DALY_share_T, DALY_share_M = disease DALYs / all-cause DALYs.
   • DALY shares are treated as constant by country across years in this implementation.
3. Merge government health spending intensity (ghes_per_gdp_mean) and WEO macro inputs.
4. Compute total government health spending in real USD:
   • GHE_total_R24 = ghes_per_gdp_mean * NGDPD_R24.
5. Compute disease shares of GHE:
   • HIV DIPI uses total HIV spending where available: hiv_spend_total_hat_public_plus_private (fallback to hiv_spend_final_ggxpc).
   • GHE_share_H = hiv_dipi_spend_basis / GHE_total_R24
   • DRMT_HH_ggte is used as TB disease spending in DIPI calculations.
   • tb_adjusted_public_spend = DRMT_HH_ggte - (tb_spend_total_hat_public_plus_private - tb_spend_final_ggx) is retained in the panel for reference but is not used for TB DIPI targets in the current specification.
   • GHE_share_T = DRMT_HH_ggte / GHE_total_R24
   • Malaria DIPI uses total malaria spending where available: malaria_spend_total_hat_public_plus_private (fallback to malaria_spend_final_ggx).
   • GHE_share_M = malaria_dipi_spend_basis / GHE_total_R24
6. Compute DIPI by disease:
   • DIPI = (Disease spending share of GHE) / (Disease DALY share).
7. Assign income tiers using a GDP-per-capita proxy (NGDPD_R24 / LP) and split LMIC into:
   • LLMIC and ULMIC.
8. Compute historical benchmark medians using years 2017-2024:
   • Pooled across tiers (single benchmark scenario used in forecasting/targets)
   • For all countries and separately for Global Fund-eligible countries only (reported for comparison).
9. Compute target spending to reach the pooled benchmark median DIPI (dipi50_target_spend_*) and gap vs DIPI-basis spending.
10. Build DIPI catch-up scenarios for 2026-2035:
    • Gap-fill schedule in FCST years:
      • 2026: 20%
      • 2027: 40%
      • 2028: 60%
      • 2029: 80%
      • 2030: 100%
    • FAR5 (2031-2035) assumption:
      • DIPI catch-up spending follows the same year-to-year absolute increment (same slope) as the DIPI-basis forecast.
      • DIPI target path in FAR5 is extended from end-FCST (2030) using the same DIPI-basis slope.
      • Therefore, in FAR5 the DIPI catch-up series is set equal to the DIPI target path (100% of target).

Median DIPI Benchmarks (Pooled Across Income Tiers)

Distribution of Gap Coverage (pct_target_ggx)

The histogram below summarizes the distribution of pct_target_ggx for 2027-2029, where:

• pct_target_ggx = DIPI-basis spending / DIPI target spending
• lower values indicate larger gaps relative to the DIPI target

Countries Below Target in Future Years (2025-2035)

All-country Pooled Trajectory Plots (Shiny-style aggregate view)

These plots sum spending across all countries by year and show:

• DIPI-basis spending totals
• DIPI catch-up totals (pooled GF-eligible benchmark)
• DIPI target paths (pooled GF-eligible benchmark; hidden in 2017-2024 for readability)

HIV

TB

Malaria

Notes on Interpretation

• Historical-year DIPI targets are benchmark comparisons; the DIPI catch-up spending scenario is constrained to equal DIPI-basis spending through 2025.

Objective

Estimate country-year private HIV spending using IHME HIV historical relationships between public spending (ghes_total_mean) and private share, then apply to the latest HIV production public-spending panel.

This report focuses on predictive use (not causal inference). Coefficients describe associations useful for forecast scaling of private share from projected public spending levels.

Data Coverage

File names are abbreviated here for readability; full source file names are listed in Source Files below.

Model coverage in production predictions: private share estimates are available for 98.08% of panel rows (1,836/1,872). Private spending amount estimates are available for 98.02% of rows.

Model Specification

Primary predictive model (fractional logit with country effects):

logit(private_share_it) = α + β₁ log(1 + ghes_it) + β₂ year_c_t + β₃ year_c_t² + γ_i

Fallback model (for countries not in HIV training set):

logit(private_share_it) = α + β₁ log(1 + ghes_it) + β₂ year_c_t + β₃ year_c_t²

Conversion from predicted share to private amount:

private_amount̂_it = [private_sharê_it / (1 − private_sharê_it)] × public_hiv_spend_it

with public_hiv_spend_it = hiv_spend_final_ggxpc.

Why this specification:

• private_share is bounded in [0,1], so a logit-link fractional model is appropriate.
• Country fixed effects capture persistent country-level financing structure differences.
• Year and year-squared allow smooth temporal drift in shares.
• Global fallback prevents dropped predictions for countries not represented in HIV training data.

Prediction Summary (Production Panel)

Across all predicted rows, total estimated private HIV spending is 16,557,368,306 versus total public HIV spending 82,006,460,959, implying an aggregate private/public ratio of 0.202.

Model Performance Diagnostics

How to read these diagnostics:

• Lower RMSE and MAE indicate better predictive accuracy.
• Correlation close to 1 indicates better rank-order tracking across observations.
• mean_pred vs mean_actual indicates calibration bias (over/under prediction).
• Holdout results are more informative than in-sample for expected forward performance.

Coefficient Snapshot

Top non-country fixed-effect terms from saved coefficients:

Source Files

• IHME combined input: DRM2026/Output/ihme_disease_spending_country_year_20260527_215226.csv
• Production panel with private estimates: DRM2026/Output/hiv_spending_production_panel_with_private_hiv_20260527_215356.csv
• Coefficients file: DRM2026/Output/hiv_private_share_model_coefficients_20260527_215356.csv
• Model stats file: DRM2026/Output/hiv_private_share_model_stats_20260527_215356.csv

Methods

• Historical target: DRMM_pub = max(DRMMwmr, DRMMgovcat).
• Historical missing values are imputed with a BRMS mixed model on log(DRMM_pub) with country random effects and country trend.
• Predictors are selected from available covariates: year_c, income_tier, log_ngdpdpc_r24, log_lp, log_cases, log_incidence.
• First forecast year (typically 2025) is anchored by a recency-weighted historical log-linear trend fit within each country.
• Remaining forecast years use real GGX per-capita growth tracking (NGGXDPC_R24) from WEO.
• FAR5 period (post-WEO) uses each country’s average annual GGX per-capita growth over WEO forecast years.

Model Fit Diagnostics

• Diagnostics CSV: malaria_drmm_pub_mcmc_diagnostics_20260527_215617.csv
• In-sample fit CSV: malaria_drmm_pub_fit_metrics_20260527_215617.csv

Row Source Counts

• CSV: malaria_drmm_pub_row_source_counts_20260527_215617.csv

Total DRMM_pub_final_ggx by Calendar Year

• CSV: malaria_drmm_pub_total_by_year_20260527_215617.csv

Annual public spending totals and country coverage are included once in the public, private and combined totals in section 5, alongside separate public and private missing-value counts.

Historical Imputation Share by Year

• CSV: malaria_drmm_pub_imputed_share_by_year_20260527_215617.csv

Run Metadata

• Historical years: 2017-2024
• Forecast years: 2025-2030
• FAR5 years: 2031-2035
• Forecast anchor year: 2025
• Forecast anchor model: Weighted log-linear trend on DRMM_pub_hist_filled (half_life_years=3), fallback to one-step GGX growth from prior year
• BRMS formula: log_drmm_pub ~ year_c + income_tier + log_ngdpdpc_r24 + log_lp + (1 + year_c | ISO)

Objective

Estimate private malaria spending from public malaria spending using IHME historical malaria relationships, then produce total malaria spending (public + private) in the production panel.

Model

• Primary model: fractional logit with country fixed effects: private_share ~ log(1 + ghes_total_mean) + factor(iso3)
• Fallback model: global fractional logit: private_share ~ log(1 + ghes_total_mean)
• Conversion: private_amount_hat = (share_hat / (1 - share_hat)) * DRMM_pub_final_ggx

Diagnostics

Diagnostics CSV: malaria_private_model_diagnostics_20260527_215624.csv

Prediction Source Coverage

Source-count CSV: malaria_private_model_source_counts_20260527_215624.csv

Yearly Totals: Public, Private, Combined

Totals CSV: malaria_public_private_total_by_year_20260527_215624.csv

Run Stats

• Training rows (Malaria): 2226
• Training countries (Malaria): 106
• Pseudo R2 (ISO FE): 0.9596
• Pseudo R2 (Global fallback): 0.0141

Historical Period (observed values) 2017-23

Reasons for difference:

• Updated observed data for public DRM
• New WEO forecast
• Uses IHME for private HIV instead of derivation from aggregate private spending from UNAIDS
• Private markup is not a constant, but varies by year and is a function of public spending level.

Aggregated comparison: GC8

Additional reasons for GC8 differences (the historical-comparison reasons above also apply):

• Baseline 2025 instead of 2023. Baseline based on weighted historical trend in prior years (instead of old 2-step approach)
• DIPI spending gaps recalculated each year.

Notes

• Old workbook source is Results_IC8.xlsx; comparison rows were read from compare_ic8_vs_new_panel_long_20260611_072806.csv.
• New workbook source is the latest dipi_hiv_tb_malaria_PANEL_*.xlsx used when the comparison file was created.
• Comparison is restricted to Global Fund eligible country-years by disease (inH, inT, inM = 1).
• Year 2016 is excluded.

Aggregated comparison: GC7 (supplement)