discrepancy in ART allocation by expected mortality

[jeffeaton]

There is a difference between Spectrum and Leapfrog implementations for ART allocation "proportional to expected mortality” allocation. This is also a likely difference in EPP-ASM, however, slightly different and not interrogated previously because ART allocation in EPP-ASM is only by coarse age groups.

The image below shows an example of the ART allocation by age for women in the first year of ART scale-up (2000) in a test file with allocation set to fully “proportional to eligibility” and “proportional to expected mortality”.

Three observations:

• In the “proportional to eligibility”, everything matches exactly (hurray!). So the difference is in the ‘expected mortality’ allocation calculation.
• Leapfrog allocates more ART to older ages (higher mortality -> higher expected ART allocation) and less ART to younger ages (lower mortality).
  • Compared to the “proportional to eligibility” allocation, the Spectrum “expected mortality” shifts ART coverage slightly to older ages, but less than the Leapfrog allocation
• The Leapfrog allocation has a sharp change in the allocation at age 25, 35, and 45 coinciding with the mortality rate changes (which makes sense), but the Spectrum allocation is smooth over these age category breaks.
  • Suspect some smoothing step in Spectrum that Leapfrog is not implementing? (This smoothing step is the component that is likely not relevant in EPP-ASM)

Contrived file: To explore further, I varied the relative mortality rates by age groups to see how that affected the ART allocation. I contrived an extreme file with near 0 mortality in 25+ age groups. I expected nearly all of the ART allocation in age group 15-24.

Features of the file:

• 80% incidence pulse in year 1999
• All new infections go into CD4 <50 (to remove any effects of disease progression and just isolate the mortality parameters)
• First year of ART in 2000
• ART allocation 100% based on ‘expected mortality’ option
• Natural history mortality is the default value for 15-24, and reduced to 0.001 for age groups 25-34, 35-44, 45+.

In the Leapfrog simulation, essentially all ART is allocated to age group 15-24 (image below; what I expected). But the Spectrum pattern for ART by age looks very different, and unsure reasons why.

[jeffeaton]

Uploading PJNZ files and R script used to create the above examples for safe keeping.

artinit-mx-elig-example_2024-01-03.zip

[jeffeaton]

This is addressed in: f74c14e73

Re-running the checks above with this commit gives the following comparison plots:

There may still be a difference with Spectrum calculating ART eligibility weights only once at the start of the year vs. Leapfrog / EPP-ASM at every time step.