From c41381a961263c28d60976e68923157916c78b15 Mon Sep 17 00:00:00 2001 From: Johannes Ranke Date: Thu, 16 Sep 2021 15:31:13 +0200 Subject: Adapt and improve the dimethenamid vignette Adapt to the corrected data and unify control parameters for saemix and nlmixr with saem. Update docs --- docs/dev/articles/web_only/dimethenamid_2018.html | 245 +++++++++++----------- 1 file changed, 127 insertions(+), 118 deletions(-) (limited to 'docs/dev/articles/web_only/dimethenamid_2018.html') diff --git a/docs/dev/articles/web_only/dimethenamid_2018.html b/docs/dev/articles/web_only/dimethenamid_2018.html index 9a6d8388..b35d8210 100644 --- a/docs/dev/articles/web_only/dimethenamid_2018.html +++ b/docs/dev/articles/web_only/dimethenamid_2018.html @@ -32,7 +32,7 @@ mkin - 1.0.5 + 1.1.0 @@ -95,13 +95,13 @@ -
+

Using biexponential decline (DFOP) results in a slightly more random scatter of the residuals:

-plot(mixed(f_parent_mkin_const["DFOP", ]))
+plot(mixed(f_parent_mkin_const["DFOP", ]))

The population curve (bold line) in the above plot results from taking the mean of the individual transformed parameters, i.e. of log k1 and log k2, as well as of the logit of the g parameter of the DFOP model). Here, this procedure does not result in parameters that represent the degradation well, because in some datasets the fitted value for k2 is extremely close to zero, leading to a log k2 value that dominates the average. This is alleviated if only rate constants that pass the t-test for significant difference from zero (on the untransformed scale) are considered in the averaging:

-plot(mixed(f_parent_mkin_const["DFOP", ]), test_log_parms = TRUE)
+plot(mixed(f_parent_mkin_const["DFOP", ]), test_log_parms = TRUE)

While this is visually much more satisfactory, such an average procedure could introduce a bias, as not all results from the individual fits enter the population curve with the same weight. This is where nonlinear mixed-effects models can help out by treating all datasets with equally by fitting a parameter distribution model together with the degradation model and the error model (see below).

The remaining trend of the residuals to be higher for higher predicted residues is reduced by using the two-component error model:

-plot(mixed(f_parent_mkin_tc["DFOP", ]), test_log_parms = TRUE)
+plot(mixed(f_parent_mkin_tc["DFOP", ]), test_log_parms = TRUE)

@@ -177,175 +174,184 @@

nlme

-

The nlme package was the first R extension providing facilities to fit nonlinear mixed-effects models. We use would like to do model selection from all four combinations of degradation models and error models based on the AIC. However, fitting the DFOP model with constant variance and using default control parameters results in an error, signalling that the maximum number of 50 iterations was reached, potentially indicating overparameterisation. However, the algorithm converges when the two-component error model is used in combination with the DFOP model. This can be explained by the fact that the smaller residues observed at later sampling times get more weight when using the two-component error model which will counteract the tendency of the algorithm to try parameter combinations unsuitable for fitting these data.

+

The nlme package was the first R extension providing facilities to fit nonlinear mixed-effects models. We would like to do model selection from all four combinations of degradation models and error models based on the AIC. However, fitting the DFOP model with constant variance and using default control parameters results in an error, signalling that the maximum number of 50 iterations was reached, potentially indicating overparameterisation. However, the algorithm converges when the two-component error model is used in combination with the DFOP model. This can be explained by the fact that the smaller residues observed at later sampling times get more weight when using the two-component error model which will counteract the tendency of the algorithm to try parameter combinations unsuitable for fitting these data.

 library(nlme)
 f_parent_nlme_sfo_const <- nlme(f_parent_mkin_const["SFO", ])
-#f_parent_nlme_dfop_const <- nlme(f_parent_mkin_const["DFOP", ])
-# maxIter = 50 reached
+# f_parent_nlme_dfop_const <- nlme(f_parent_mkin_const["DFOP", ])
 f_parent_nlme_sfo_tc <- nlme(f_parent_mkin_tc["SFO", ])
 f_parent_nlme_dfop_tc <- nlme(f_parent_mkin_tc["DFOP", ])
-

Note that overparameterisation is also indicated by warnings obtained when fitting SFO or DFOP with the two-component error model (‘false convergence’ in the ‘LME step’ in some iterations). In addition to these fits, attempts were also made to include correlations between random effects by using the log Cholesky parameterisation of the matrix specifying them. The code used for these attempts can be made visible below.

+

Note that a certain degree of overparameterisation is also indicated by a warning obtained when fitting DFOP with the two-component error model (‘false convergence’ in the ‘LME step’ in iteration 3). However, as this warning does not occur in later iterations, and specifically not in the last of the 6 iterations, we can ignore this warning.

+

The model comparison function of the nlme package can directly be applied to these fits showing a much lower AIC for the DFOP model fitted with the two-component error model. Also, the likelihood ratio test indicates that this difference is significant. as the p-value is below 0.0001.

-f_parent_nlme_sfo_const_logchol <- nlme(f_parent_mkin_const["SFO", ],
-  random = pdLogChol(list(DMTA_0 ~ 1, log_k_DMTA ~ 1)))
-anova(f_parent_nlme_sfo_const, f_parent_nlme_sfo_const_logchol) # not better
-#f_parent_nlme_dfop_tc_logchol <- update(f_parent_nlme_dfop_tc,
-#  random = pdLogChol(list(DMTA_0 ~ 1, log_k1 ~ 1, log_k2 ~ 1, g_qlogis ~ 1)))
-# using log Cholesky parameterisation for random effects (nlme default) does
-# not converge here and gives lots of warnings about the LME step not converging
-

The model comparison function of the nlme package can directly be applied to these fits showing a similar goodness-of-fit of the SFO model, but a much lower AIC for the DFOP model fitted with the two-component error model. Also, the likelihood ratio test indicates that this difference is significant. as the p-value is below 0.0001.

-
 anova(
   f_parent_nlme_sfo_const, f_parent_nlme_sfo_tc, f_parent_nlme_dfop_tc
 )
                        Model df    AIC    BIC  logLik   Test L.Ratio p-value
-f_parent_nlme_sfo_const     1  5 818.63 834.00 -404.31                       
-f_parent_nlme_sfo_tc        2  6 820.61 839.06 -404.31 1 vs 2   0.014  0.9049
-f_parent_nlme_dfop_tc       3 10 687.84 718.59 -333.92 2 vs 3 140.771  <.0001
+f_parent_nlme_sfo_const 1 5 796.60 811.82 -393.30 +f_parent_nlme_sfo_tc 2 6 798.60 816.86 -393.30 1 vs 2 0.00 0.998 +f_parent_nlme_dfop_tc 3 10 671.91 702.34 -325.96 2 vs 3 134.69 <.0001 +

In addition to these fits, attempts were also made to include correlations between random effects by using the log Cholesky parameterisation of the matrix specifying them. The code used for these attempts can be made visible below.

+
+f_parent_nlme_sfo_const_logchol <- nlme(f_parent_mkin_const["SFO", ],
+  random = nlme::pdLogChol(list(DMTA_0 ~ 1, log_k_DMTA ~ 1)))
+anova(f_parent_nlme_sfo_const, f_parent_nlme_sfo_const_logchol)
+f_parent_nlme_sfo_tc_logchol <- nlme(f_parent_mkin_tc["SFO", ],
+  random = nlme::pdLogChol(list(DMTA_0 ~ 1, log_k_DMTA ~ 1)))
+anova(f_parent_nlme_sfo_tc, f_parent_nlme_sfo_tc_logchol)
+f_parent_nlme_dfop_tc_logchol <- nlme(f_parent_mkin_const["DFOP", ],
+  random = nlme::pdLogChol(list(DMTA_0 ~ 1, log_k1 ~ 1, log_k2 ~ 1, g_qlogis ~ 1)))
+anova(f_parent_nlme_dfop_tc, f_parent_nlme_dfop_tc_logchol)
+

While the SFO variants converge fast, the additional parameters introduced by this lead to convergence warnings for the DFOP model. The model comparison clearly show that adding correlations between random effects does not improve the fits.

The selected model (DFOP with two-component error) fitted to the data assuming no correlations between random effects is shown below.

-plot(f_parent_nlme_dfop_tc)
+plot(f_parent_nlme_dfop_tc)

saemix

-

The saemix package provided the first Open Source implementation of the Stochastic Approximation to the Expectation Maximisation (SAEM) algorithm. SAEM fits of degradation models can be performed using an interface to the saemix package available in current development versions of the mkin package.

-

The corresponding SAEM fits of the four combinations of degradation and error models are fitted below. As there is no convergence criterion implemented in the saemix package, the convergence plots need to be manually checked for every fit.

-

The convergence plot for the SFO model using constant variance is shown below.

+

The saemix package provided the first Open Source implementation of the Stochastic Approximation to the Expectation Maximisation (SAEM) algorithm. SAEM fits of degradation models can be conveniently performed using an interface to the saemix package available in current development versions of the mkin package.

+

The corresponding SAEM fits of the four combinations of degradation and error models are fitted below. As there is no convergence criterion implemented in the saemix package, the convergence plots need to be manually checked for every fit. As we will compare the SAEM implementation of saemix to the results obtained using the nlmixr package later, we define control settings that work well for all the parent data fits shown in this vignette.

 library(saemix)
-f_parent_saemix_sfo_const <- mkin::saem(f_parent_mkin_const["SFO", ], quiet = TRUE,
-  transformations = "saemix")
+saemix_control <- saemixControl(nbiter.saemix = c(800, 300), nb.chains = 15,
+    print = FALSE, save = FALSE, save.graphs = FALSE, displayProgress = FALSE)
+

The convergence plot for the SFO model using constant variance is shown below.

+
+f_parent_saemix_sfo_const <- mkin::saem(f_parent_mkin_const["SFO", ], quiet = TRUE,
+  control = saemix_control, transformations = "saemix")
 plot(f_parent_saemix_sfo_const$so, plot.type = "convergence")

Obviously the default number of iterations is sufficient to reach convergence. This can also be said for the SFO fit using the two-component error model.

-
+
 f_parent_saemix_sfo_tc <- mkin::saem(f_parent_mkin_tc["SFO", ], quiet = TRUE,
-  transformations = "saemix")
+  control = saemix_control, transformations = "saemix")
 plot(f_parent_saemix_sfo_tc$so, plot.type = "convergence")

When fitting the DFOP model with constant variance, parameter convergence is not as unambiguous (see the failure of nlme with the default number of iterations above). Therefore, the number of iterations in the first phase of the algorithm was increased, leading to visually satisfying convergence.

-
+
 f_parent_saemix_dfop_const <- mkin::saem(f_parent_mkin_const["DFOP", ], quiet = TRUE,
-  control = saemixControl(nbiter.saemix = c(800, 200), print = FALSE,
-    save = FALSE, save.graphs = FALSE, displayProgress = FALSE),
-  transformations = "saemix")
+  control = saemix_control, transformations = "saemix")
 plot(f_parent_saemix_dfop_const$so, plot.type = "convergence")

-

The same applies to the case where the DFOP model is fitted with the two-component error model. Convergence of the variance of k2 is enhanced by using the two-component error, it remains more or less stable already after 200 iterations of the first phase.

-
-f_parent_saemix_dfop_tc_moreiter <- mkin::saem(f_parent_mkin_tc["DFOP", ], quiet = TRUE,
-  control = saemixControl(nbiter.saemix = c(800, 200), print = FALSE,
-    save = FALSE, save.graphs = FALSE, displayProgress = FALSE),
-  transformations = "saemix")
-plot(f_parent_saemix_dfop_tc_moreiter$so, plot.type = "convergence")
-

-

The four combinations can be compared using the model comparison function from the saemix package:

+

The same applies in the case where the DFOP model is fitted with the two-component error model. Convergence of the variance of k2 is enhanced by using the two-component error, it remains more or less stable already after 200 iterations of the first phase.

-compare.saemix(f_parent_saemix_sfo_const$so, f_parent_saemix_sfo_tc$so,
-  f_parent_saemix_dfop_const$so, f_parent_saemix_dfop_tc_moreiter$so)
+f_parent_saemix_dfop_tc <- mkin::saem(f_parent_mkin_tc["DFOP", ], quiet = TRUE, + control = saemix_control, transformations = "saemix") +plot(f_parent_saemix_dfop_tc$so, plot.type = "convergence")
+

The four combinations and including the variations of the DFOP/tc combination can be compared using the model comparison function from the saemix package:

+
+compare.saemix(
+  f_parent_saemix_sfo_const$so,
+  f_parent_saemix_sfo_tc$so,
+  f_parent_saemix_dfop_const$so,
+  f_parent_saemix_dfop_tc$so)
Likelihoods calculated by importance sampling
     AIC    BIC
-1 818.37 817.33
-2 820.38 819.14
-3 725.91 724.04
-4 683.64 681.55
-

As in the case of nlme fits, the DFOP model fitted with two-component error (number 4) gives the lowest AIC. The numeric values are reasonably close to the ones obtained using nlme, considering that the algorithms for fitting the model and for the likelihood calculation are quite different.

+1 796.37 795.33 +2 798.37 797.13 +3 713.16 711.28 +4 666.10 664.01
+

As in the case of nlme fits, the DFOP model fitted with two-component error (number 4) gives the lowest AIC. Using more iterations and/or more chains does not have a large influence on the final AIC (not shown).

In order to check the influence of the likelihood calculation algorithms implemented in saemix, the likelihood from Gaussian quadrature is added to the best fit, and the AIC values obtained from the three methods are compared.

-
-f_parent_saemix_dfop_tc_moreiter$so <-
-  llgq.saemix(f_parent_saemix_dfop_tc_moreiter$so)
-AIC(f_parent_saemix_dfop_tc_moreiter$so)
-
[1] 683.64
-
-AIC(f_parent_saemix_dfop_tc_moreiter$so, method = "gq")
-
[1] 683.7
-
-AIC(f_parent_saemix_dfop_tc_moreiter$so, method = "lin")
-
[1] 683.17
-

The AIC values based on importance sampling and Gaussian quadrature are quite similar. Using linearisation is less accurate, but still gives a similar value.

+
+f_parent_saemix_dfop_tc$so <-
+  llgq.saemix(f_parent_saemix_dfop_tc$so)
+AIC(f_parent_saemix_dfop_tc$so)
+
[1] 666.1
+
+AIC(f_parent_saemix_dfop_tc$so, method = "gq")
+
[1] 666.03
+
+AIC(f_parent_saemix_dfop_tc$so, method = "lin")
+
[1] 665.48
+

The AIC values based on importance sampling and Gaussian quadrature are very similar. Using linearisation is known to be less accurate, but still gives a similar value.

nlmixr

-

In the last years, a lot of effort has been put into the nlmixr package which is designed for pharmacokinetics, where nonlinear mixed-effects models are routinely used, but which can also be used for related data like chemical degradation data. A current development branch of the mkin package provides an interface between mkin and nlmixr. Here, we check if we get equivalent results when using a refined version of the First Order Conditional Estimation (FOCE) algorithm used in nlme, namely First Order Conditional Estimation with Interaction (FOCEI), and the SAEM algorithm as implemented in nlmixr.

-

First, the focei algorithm is used for the four model combinations and the goodness of fit of the results is compared.

-
+

In the last years, a lot of effort has been put into the nlmixr package which is designed for pharmacokinetics, where nonlinear mixed-effects models are routinely used, but which can also be used for related data like chemical degradation data. A current development branch of the mkin package provides an interface between mkin and nlmixr. Here, we check if we get equivalent results when using a refined version of the First Order Conditional Estimation (FOCE) algorithm used in nlme, namely the First Order Conditional Estimation with Interaction (FOCEI), and the SAEM algorithm as implemented in nlmixr.

+

First, the focei algorithm is used for the four model combinations. A number of warnings are produced with unclear significance.

+
 library(nlmixr)
 f_parent_nlmixr_focei_sfo_const <- nlmixr(f_parent_mkin_const["SFO", ], est = "focei")
 f_parent_nlmixr_focei_sfo_tc <- nlmixr(f_parent_mkin_tc["SFO", ], est = "focei")
 f_parent_nlmixr_focei_dfop_const <- nlmixr(f_parent_mkin_const["DFOP", ], est = "focei")
 f_parent_nlmixr_focei_dfop_tc<- nlmixr(f_parent_mkin_tc["DFOP", ], est = "focei")
-
-AIC(f_parent_nlmixr_focei_sfo_const$nm, f_parent_nlmixr_focei_sfo_tc$nm,
-  f_parent_nlmixr_focei_dfop_const$nm, f_parent_nlmixr_focei_dfop_tc$nm)
-
                                    df    AIC
-f_parent_nlmixr_focei_sfo_const$nm   5 818.63
-f_parent_nlmixr_focei_sfo_tc$nm      6 820.61
-f_parent_nlmixr_focei_dfop_const$nm  9 728.11
-f_parent_nlmixr_focei_dfop_tc$nm    10 687.82
+
+aic_nlmixr_focei <- sapply(
+  list(f_parent_nlmixr_focei_sfo_const$nm, f_parent_nlmixr_focei_sfo_tc$nm,
+    f_parent_nlmixr_focei_dfop_const$nm, f_parent_nlmixr_focei_dfop_tc$nm),
+  AIC)

The AIC values are very close to the ones obtained with nlme which are repeated below for convenience.

-AIC(
-  f_parent_nlme_sfo_const, f_parent_nlme_sfo_tc, f_parent_nlme_dfop_tc
+aic_nlme <- sapply(
+  list(f_parent_nlme_sfo_const, NA, f_parent_nlme_sfo_tc, f_parent_nlme_dfop_tc),
+  function(x) if (is.na(x[1])) NA else AIC(x))
+aic_nlme_nlmixr_focei <- data.frame(
+  "Degradation model" = c("SFO", "SFO", "DFOP", "DFOP"),
+  "Error model" = rep(c("constant variance", "two-component"), 2),
+  "AIC (nlme)" = aic_nlme,
+  "AIC (nlmixr with FOCEI)" = aic_nlmixr_focei,
+  check.names = FALSE
 )
-
                        df    AIC
-f_parent_nlme_sfo_const  5 818.63
-f_parent_nlme_sfo_tc     6 820.61
-f_parent_nlme_dfop_tc   10 687.84
-

Secondly, we use the SAEM estimation routine and check the convergence plots for SFO with constant variance

+

Secondly, we use the SAEM estimation routine and check the convergence plots. The control parameters also used for the saemix fits are defined beforehand.

+
+nlmixr_saem_control <- saemControl(logLik = TRUE,
+  nBurn = 1000, nEm = 300, nmc = 15)
+

The we fit SFO with constant variance

 f_parent_nlmixr_saem_sfo_const <- nlmixr(f_parent_mkin_const["SFO", ], est = "saem",
-  control = nlmixr::saemControl(logLik = TRUE))
+  control = nlmixr_saem_control)
 traceplot(f_parent_nlmixr_saem_sfo_const$nm)

-

for SFO with two-component error

+

and SFO with two-component error.

 f_parent_nlmixr_saem_sfo_tc <- nlmixr(f_parent_mkin_tc["SFO", ], est = "saem",
-  control = nlmixr::saemControl(logLik = TRUE))
+  control = nlmixr_saem_control)
 traceplot(f_parent_nlmixr_saem_sfo_tc$nm)

-

For DFOP with constant variance, the convergence plots show considerable instability of the fit, which can be alleviated by increasing the number of iterations and the number of parallel chains for the first phase of algorithm.

+

For DFOP with constant variance, the convergence plots show considerable instability of the fit, which indicates overparameterisation which was already observed earlier for this model combination.

 f_parent_nlmixr_saem_dfop_const <- nlmixr(f_parent_mkin_const["DFOP", ], est = "saem",
-  control = nlmixr::saemControl(logLik = TRUE, nBurn = 1000), nmc = 15)
+  control = nlmixr_saem_control)
 traceplot(f_parent_nlmixr_saem_dfop_const$nm)

-

For DFOP with two-component error, the same increase in iterations and parallel chains was used, but using the two-component error appears to lead to a less erratic convergence, so this may not be necessary to this degree.

+

For DFOP with two-component error, a less erratic convergence is seen.

 f_parent_nlmixr_saem_dfop_tc <- nlmixr(f_parent_mkin_tc["DFOP", ], est = "saem",
-  control = nlmixr::saemControl(logLik = TRUE, nBurn = 1000, nmc = 15))
+  control = nlmixr_saem_control)
 traceplot(f_parent_nlmixr_saem_dfop_tc$nm)

-

The AIC values are internally calculated using Gaussian quadrature. For an unknown reason, the AIC value obtained for the DFOP fit using the two-component error model is given as Infinity.

+

The AIC values are internally calculated using Gaussian quadrature. For an unknown reason, the AIC value obtained for the DFOP fit using constant error is given as Infinity.

 AIC(f_parent_nlmixr_saem_sfo_const$nm, f_parent_nlmixr_saem_sfo_tc$nm,
   f_parent_nlmixr_saem_dfop_const$nm, f_parent_nlmixr_saem_dfop_tc$nm)
                                   df    AIC
-f_parent_nlmixr_saem_sfo_const$nm   5 820.54
-f_parent_nlmixr_saem_sfo_tc$nm      6 835.26
-f_parent_nlmixr_saem_dfop_const$nm  9 842.84
-f_parent_nlmixr_saem_dfop_tc$nm    10 684.51
+f_parent_nlmixr_saem_sfo_const$nm 5 798.68 +f_parent_nlmixr_saem_sfo_tc$nm 6 808.88 +f_parent_nlmixr_saem_dfop_const$nm 9 815.95 +f_parent_nlmixr_saem_dfop_tc$nm 10 669.57

The following table gives the AIC values obtained with the three packages.

 AIC_all <- data.frame(
+  check.names = FALSE,
   "Degradation model" = c("SFO", "SFO", "DFOP", "DFOP"),
   "Error model" = c("const", "tc", "const", "tc"),
   nlme = c(AIC(f_parent_nlme_sfo_const), AIC(f_parent_nlme_sfo_tc), NA, AIC(f_parent_nlme_dfop_tc)),
   nlmixr_focei = sapply(list(f_parent_nlmixr_focei_sfo_const$nm, f_parent_nlmixr_focei_sfo_tc$nm,
   f_parent_nlmixr_focei_dfop_const$nm, f_parent_nlmixr_focei_dfop_tc$nm), AIC),
   saemix = sapply(list(f_parent_saemix_sfo_const$so, f_parent_saemix_sfo_tc$so,
-    f_parent_saemix_dfop_const$so, f_parent_saemix_dfop_tc_moreiter$so), AIC),
+    f_parent_saemix_dfop_const$so, f_parent_saemix_dfop_tc$so), AIC),
   nlmixr_saem = sapply(list(f_parent_nlmixr_saem_sfo_const$nm, f_parent_nlmixr_saem_sfo_tc$nm,
   f_parent_nlmixr_saem_dfop_const$nm, f_parent_nlmixr_saem_dfop_tc$nm), AIC)
 )
 kable(AIC_all)
- - + + @@ -355,34 +361,34 @@ f_parent_nlmixr_saem_dfop_tc$nm 10 684.51 - - - - + + + + - - - - + + + + - - - + + + - - - - + + + +
Degradation.modelError.modelDegradation modelError model nlme nlmixr_focei saemix
SFO const818.63818.63818.37820.54796.60796.62796.37798.68
SFO tc820.61820.61820.38835.26798.60798.61798.37808.88
DFOP const NA728.11725.91842.84750.91713.16815.95
DFOP tc687.84687.82683.64684.51671.91666.60666.10669.57
@@ -397,6 +403,9 @@ f_parent_nlmixr_saem_dfop_tc$nm 10 684.51

EFSA. 2018. “Peer Review of the Pesticide Risk Assessment of the Active Substance Dimethenamid-P.” EFSA Journal 16 (4): 5211.

+
+

Ranke, Johannes, Janina Wöltjen, Jana Schmidt, and Emmanuelle Comets. 2021. “Taking Kinetic Evaluations of Degradation Data to the Next Level with Nonlinear Mixed-Effects Models.” Environments 8 (8). https://doi.org/10.3390/environments8080071.

+

Rapporteur Member State Germany, Co-Rapporteur Member State Bulgaria. 2018. “Renewal Assessment Report Dimethenamid-P Volume 3 - B.8 Environmental fate and behaviour, Rev. 2 - November 2017.” https://open.efsa.europa.eu/study-inventory/EFSA-Q-2014-00716.

-- cgit v1.2.1