EpiPvr bridge

Cropmix keeps the user-facing workflow in Python. R and EpiPvr remain behind a subprocess bridge.

External installation

Install R, then once in R:

install.packages("EpiPvr")

Check from the shell:

cropmix doctor

SPT access-period experiment

from cropmix.epipvr import (
    AccessPeriodAssay,
    AccessPeriodExperiment,
    EpiPvrBackend,
)

aap = AccessPeriodAssay(
    duration=(2, 3, 4, 5),
    tested=(30, 30, 30, 30),
    infected=(10, 15, 18, 20),
)

iap = AccessPeriodAssay(
    duration=(0.25, 0.5, 0.75, 1.0),
    tested=(30, 30, 30, 30),
    infected=(5, 10, 15, 20),
)

experiment = AccessPeriodExperiment.spt(
    acquisition=aap,
    inoculation=iap,
    fixed_inoculation_for_acquisition=6,
    fixed_acquisition_for_inoculation=4,
    vectors_per_plant=20,
)

fit = EpiPvrBackend().fit(experiment)

The bridge maps the Python experiment to EpiPvr's d_AAP, d_IAP, d_durations, d_vectorspp, and d_virusType structure.

For PT, it additionally supplies d_LAP and a 3x3 duration matrix.

Output

EpiPvrFit exposes:

fit.posterior(unit="per_day")
fit.parameter_summary(unit="per_day")
fit.convergence_report()
fit.median_host_transmission()
fit.median_pathogen_parameters()

Posterior variables are kept jointly by draw. Cropmix does not independently resample marginal alpha, beta, and mu distributions.

The bridge also exposes EpiPvr's branching-process epidemic-probability function through EpiPvrBackend.epidemic_probability().

Diagnostics

EpiPvr itself emphasizes assessing model fit before reporting or propagating parameter estimates. Cropmix exports the EpiPvr summary table, Bayesian R2 values, divergent-transition count, and tree-depth diagnostic.

Use fit.require_usable() to enforce a conservative diagnostic gate before parameter propagation.

Important model boundary

PT parameters can be inferred through EpiPvr and used in Cropmix. The current spatial engine does not explicitly represent the latent period of the virus within the vector. However, PT-like dynamics can be approximated within the current engine by using the inferred acquisition and inoculation parameters together with a vector recovery/clearance rate equal to zero. This provides a reduced approximation of PT transmission rather than a mechanistic PT model with an explicit exposed-vector compartment.