---
title: "Hierarchical kinetic modelling of degradation data"
author:
date:
output: mkin::hierarchical_kinetics
geometry: margin=2cm
---

\clearpage

# Setup

```{r packages, cache = FALSE, message = FALSE}
library(mkin)
library(knitr)
library(saemix)
library(parallel)
library(readxl)
```

```{r n_cores, cache = FALSE}
n_cores <- detectCores()

if (Sys.info()["sysname"] == "Windows") {
  cl <- makePSOCKcluster(n_cores)
} else {
  cl <- makeForkCluster(n_cores)
}
```

\clearpage

# Introduction

This report shows hierarchical kinetic modelling for ...
The data were obtained from ...

```{r ds}
data_path <- system.file(
  "testdata", "lambda-cyhalothrin_soil_efsa_2014.xlsx",
  package = "mkin")
ds <- read_spreadsheet(data_path, valid_datasets = c(1:4, 7:13))
covariates <- attr(ds, "covariates")
```

The covariate data are shown below.

```{r results = "asis", dependson = "ds", echo = FALSE}
kable(covariates, caption = "Covariate data for all datasets")
```

\clearpage

The datasets with the residue time series are shown in the tables below. Please
refer to the spreadsheet for details like data sources, treatment of values
below reporting limits and time step normalisation factors.

```{r results = "asis", dependson = "ds", echo = FALSE}
for (ds_name in names(ds)) {
  print(
    kable(mkin_long_to_wide(ds[[ds_name]]),
      caption = paste("Dataset", ds_name),
      booktabs = TRUE, row.names = FALSE))
  cat("\n\\clearpage\n")
}
```

# Parent only evaluations

The following code performs separate fits of the candidate degradation models
to all datasets using constant variance and the two-component error model.

```{r parent-sep, dependson = "ds"}
parent_deg_mods <- c("SFO", "FOMC", "DFOP", "SFORB")
errmods <- c(const = "constant variance", tc = "two-component error")
parent_sep_const <- mmkin(
  parent_deg_mods, ds,
  error_model = "const",
  cluster = cl, quiet = TRUE)
parent_sep_tc <- update(parent_sep_const, error_model = "tc")
```

To select the parent model, the corresponding hierarchical fits are performed below.

```{r parent-mhmkin, dependson = "parent-sep"}
parent_mhmkin <- mhmkin(list(parent_sep_const, parent_sep_tc), cluster = cl)
status(parent_mhmkin) |> kable()
```

All fits terminate without errors (status OK). The check for ill-defined
parameters shows that not all random effect parameters can be robustly
quantified.

```{r dependson = "parent_mhmkin"}
illparms(parent_mhmkin) |> kable()
```

Therefore, the fits are updated, excluding random effects that were
ill-defined according to the `illparms` function. The status of the fits
is checked.

```{r parent-mhmkin-refined}
parent_mhmkin_refined <- update(parent_mhmkin,
  no_random_effect = illparms(parent_mhmkin))
status(parent_mhmkin_refined) |> kable()
```

Also, it is checked if the AIC values of the refined fits are actually smaller
than the AIC values of the original fits.

```{r dependson = "parent-mhmkin-refined"}
(AIC(parent_mhmkin_refined) < AIC(parent_mhmkin)) |> kable()
```

From the refined fits, the most suitable model is selected using the AIC.

```{r parent-best, dependson = "parent-mhmkin"}
aic_parent <- AIC(parent_mhmkin_refined)
min_aic <- which(aic_parent == min(aic_parent), arr.ind = TRUE)
best_degmod_parent <- rownames(aic_parent)[min_aic[1]]
best_errmod_parent <- colnames(aic_parent)[min_aic[2]]
anova(parent_mhmkin_refined) |> kable(digits = 1)
parent_best <- parent_mhmkin_refined[[best_degmod_parent, best_errmod_parent]]
```

Based on the AIC, the combination of the `r best_degmod_parent` degradation
model with the error model `r errmods[best_errmod_parent]` is identified to
be most suitable for the degradation of the parent. The check below
confirms that no ill-defined parameters remain for this combined model.

```{r dependson = "parent-best"}
illparms(parent_best)
```

The corresponding fit is plotted below.

```{r dependson = "parent-best"}
plot(parent_best)
```
The fitted parameters, together with approximate confidence
intervals are listed below.

```{r dependson = "parent-best"}
parms(parent_best, ci = TRUE) |> kable(digits = 3)
```

To investigate a potential covariate influence on degradation parameters, a
covariate model is added to the hierarchical model for each of the degradation
parameters with well-defined random effects. Also, a version with covariate
models for both of them is fitted.

```{r parent-best-pH}
parent_best_pH_1 <- update(parent_best, covariates = covariates,
  covariate_models = list(log_k_lambda_free ~ pH))
parent_best_pH_2 <- update(parent_best, covariates = covariates,
  covariate_models = list(log_k_lambda_bound_free ~ pH))
parent_best_pH_3 <- update(parent_best, covariates = covariates,
  covariate_models = list(log_k_lambda_free ~ pH, log_k_lambda_bound_free ~ pH))
```

The resulting models are compared.

```{r dependson = "parent-best-pH"}
anova(parent_best, parent_best_pH_1, parent_best_pH_2, parent_best_pH_3) |>
  kable(digits = 1)
```

The model fit with the lowest AIC is the one with a pH correlation of the
desorption rate constant `k_lambda_bound_free`. Plot and parameter listing
of this fit are shown below. Also, it is confirmed that no ill-defined
variance parameters are found.

```{r dependson = "parent-best-pH"}
plot(parent_best_pH_2)
```

```{r dependson = "parent-best-pH"}
illparms(parent_best_pH_2)
parms(parent_best_pH_2, ci = TRUE) |> kable(digits = 3)
```

\clearpage

# Pathway fits

As an example of a pathway fit, a model with SFORB for the parent compound and
parallel formation of two metabolites is set up.

```{r path-1-degmod}
if (!dir.exists("dlls")) dir.create("dlls")

m_sforb_sfo2 = mkinmod(
  lambda = mkinsub("SFORB", to = c("c_V", "c_XV")),
  c_V = mkinsub("SFO"),
  c_XV = mkinsub("SFO"),
  name = "sforb_sfo2",
  dll_dir = "dlls",
  overwrite = TRUE, quiet = TRUE
)
```

Separate evaluations of all datasets are performed with constant variance
and using two-component error.

```{r path-1-sep, dependson = c("path-1-degmod", "ds")}
sforb_sep_const <- mmkin(list(sforb_path = m_sforb_sfo2), ds,
  cluster = cl, quiet = TRUE)
sforb_sep_tc <- update(sforb_sep_const, error_model = "tc")
```

The separate fits with constant variance are plotted.

```{r dependson = "path-1-sep", fig.height = 9}
plot(mixed(sforb_sep_const))
```

The two corresponding hierarchical fits, with the random effects for the parent
degradation parameters excluded as discussed above, and including the covariate
model that was identified for the parent degradation, are attempted below.

```{r path-1, dependson = "path-1-sep"}
path_1 <- mhmkin(list(sforb_sep_const, sforb_sep_tc),
  no_random_effect = c("lambda_free_0", "log_k_lambda_free_bound"),
  covariates = covariates, covariate_models = list(log_k_lambda_bound_free ~ pH),
  cluster = cl)
```

```{r dependson = "path-1"}
status(path_1) |> kable()
```

The status information shows that both fits were successfully completed.

```{r dependson = "path-1"}
anova(path_1) |> kable(digits = 1)
```
Model comparison shows that the two-component error model provides a much
better fit.

```{r dependson = "path-1"}
illparms(path_1[["sforb_path", "tc"]])
```

Two ill-defined variance components are found. Therefore, the fit is
repeated with the corresponding random effects removed.

```{r path-1-refined, dependson = "path-1"}
path_1_refined <- update(path_1[["sforb_path", "tc"]],
  no_random_effect = c("lambda_free_0", "log_k_lambda_free_bound",
    "log_k_c_XV", "f_lambda_ilr_2"))
```

The empty output of the illparms function indicates that there are no
ill-defined parameters remaining in the refined fit.

```{r dependson = "path-1-refined"}
illparms(path_1_refined)
```

Below, the refined fit is plotted and the fitted parameters are shown together
with their 95% confidence intervals.

```{r dependson = "path-1-refined", fig.height = 9}
plot(path_1_refined)
```

```{r dependson = "path-1-refined", fig.height = 9}
parms(path_1_refined, ci = TRUE) |> kable(digits = 3)
```

\clearpage

# Appendix

## Listings of initial parent fits

```{r listings-parent, results = "asis", echo = FALSE, dependson = "parent_mhmkin"}
for (deg_mod in parent_deg_mods) {
  for (err_mod in c("const", "tc")) {
    caption <- paste("Hierarchical", deg_mod, "fit with", errmods[err_mod])
    tex_listing(parent_mhmkin[[deg_mod, err_mod]], caption)
  }
}
```

## Listings of refined parent fits

```{r listings-parent-refined, results = "asis", echo = FALSE, dependson = "parent_mhmkin_refined"}
for (deg_mod in parent_deg_mods) {
  for (err_mod in c("const", "tc")) {
    caption <- paste("Refined hierarchical", deg_mod, "fit with", errmods[err_mod])
    tex_listing(parent_mhmkin_refined[[deg_mod, err_mod]], caption)
  }
}
```

## Listings of pathway fits

```{r listings-path-1, results = "asis", echo = FALSE, dependson = "path-1-refined"}
tex_listing(path_1[["sforb_path", "const"]],
  caption = "Hierarchical fit of SFORB-SFO2 with constant variance")
tex_listing(path_1[["sforb_path", "tc"]],
  caption = "Hierarchical fit of SFORB-SFO2 with two-component error")
tex_listing(path_1_refined,
  caption = "Refined hierarchical fit of SFORB-SFO2 with two-component error")
```

## Session info

```{r echo = FALSE, cache = FALSE}
parallel::stopCluster(cl)
sessionInfo()
```