--- title: "Here be dragons" output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{Here be dragons} %\VignetteEncoding{UTF-8} %\VignetteEngine{knitr::rmarkdown} editor_options: chunk_output_type: console bibliography: vignette.bib --- ```{r, include = FALSE} knitr::opts_chunk$set( collapse = TRUE, comment = "#>", eval = TRUE ) # job::job({ # knitr::knit("vignettes/here-be-dragons.Rmd.orig", "vignettes/here-be-dragons.Rmd") # }) options(rmarkdown.html_vignette.check_title = FALSE) ``` ## Load dependencies ```{r} #remotes::install_github("traitecoevo/hmde") library(hmde) library(dplyr) library(ggplot2) library(deSolve) library(mixtools) library(MASS) ``` This vignette demonstrates an interaction between errors from numerical integration methods and MCMC sampling that produces a bimodal posterior distribution as a result of numerical error. We stumbled across the problem and are documenting it here, but account for it in the package through either using analytic solutions, or numerical methods with adaptive step sizes where analytic solutions are not available. The underlying issue is that if, given errors in the chosen numerical integration method, two sets of parameters for a differential equation $f$ give "the same" (more in a second) output for $$Y(t_{j+1}) = Y(t_j) + \int_{t_j}^{t_{j+1}} f(Y(t), \boldsymbol{\theta})\,dt\qquad (1)$$ the MCMC sampler will be unable to meaningfully distinguish the parameter combinations. What you see in practice is chains converging to extremely different parameter combinations, one of which is the `true' combination, the other of which is wrong but produces the same $\hat{Y}(t_j)$ values due to the numerical error. Thus, there is a form of non-identifiability that arises from numerical errors in a longitudinal model based on Equation (1) that are separate to other issues of non-identifiability, and currently under-explored in the literature. In this demonstration we use a Runge-Kutta 4th order numerical method (\cite{butcher2016numerical}) with different step sizes to show that even 'small' step sizes can give problems. It is important to state that "the same" $Y(t)$ values in this context means within statistical error of each other. We assume that our data consists of observations of the form $y_{j}$ at time $t_j$ that look like $$y_j = Y(t_j) + \text{error},$$ and have some finite level of precision. The numerical method may produce estimated values $\hat{Y}(t_j)$ that differ by some amount that is much smaller than the level of precision or observation error for the different parameter combinations, but due to the imprecision of the measurement process the MCMC sampler cannot meaningfully distinguish the estimates. For this demonstration we will simulate data as though it is measured in centimetres. We use rounding to produce simulated data with measurement precision of 0.1cm, and error of $\mathcal{N}(0, 0.1)$, analogous to the 1mm measurement precision and approximate standard deviation of the real-world source data used in @obrien2024allindividuals. ## The model We are implementing a longitudinal model of the form in Equation (1) within a hierarchical Bayesian longitudinal model where $$f(Y(t), \beta_0, \beta_1) = \beta_0 - \beta_1 Y(t)\qquad (2)$$ Equation (2) is known to produce pathological behaviour from numerical methods [@butcher2016numerical], so serves as an ideal simple example of the interaction between those pathologies and the MCMC sampling process. For the purpose of estimation we shift Equation (2) by the mean observed size which gives $$f_{fit}(Y(t), \beta_c, \beta_1) = \beta_c - \beta_1(Y(t) - \bar{y}),$$ then the back-transformation $$\beta_0 = \beta_c + \beta_1 \bar{y}.$$ We are attempting to estimate the parameters $\beta_0$ and $\beta_1$ from observations $y_j$, which is based on estimating $\hat{Y}_j$ given the prior distribution $$y_j \sim \mathcal{N}(\hat{Y}_j, 0.1).$$ As we are looking at a single individual, we have prior distributions for the default parameters which are $$\beta_1, \beta_c \sim \log\mathcal{N}(0,2),$$ and enforce $\beta_k > 0$. ## Simulating data For this demo code we use $\beta_0 = 10$ and $\beta_1 = 1$, which gives an asymptotic size of 10. If you wish to experiment with other values, input them in the next block and the rest will run based on that. We use the analytic solution to simulate true sizes over time, then add measurement error and round to the chosen measurement precision of 0.1cm to give a sequence of observations over time that become the `y_obs` data for the model fit. Notice that `y_obs` can produces values that are bigger than the theoretical asymptotic size $\beta_0/\beta_1$ due to error. ```{r} #Change these values to change the model parameters. Must be positive values. beta_0 <- 10 beta_1 <- 1 #True initial condition true_y_0 <- 1 max_time <- 9 time <- 0:max_time #Analytic solution analytic_solution <- function(x = NULL, pars = NULL){ #Pars is list of beta_0, beta_1, y_0 return( (pars[[1]]/pars[[2]]) + (pars[[3]] - (pars[[1]]/pars[[2]])) * exp(-pars[[2]] * x) ) } true_pars <- list( beta_0 = beta_0, beta_1 = beta_1, true_y_0 = true_y_0 ) true_args_list <- list(pars = c(beta_0, beta_1, true_y_0)) y_true <- analytic_solution(time, true_pars) ``` From the analytic solution we produce observations by adding measurement error and rounding to a precision of 0.1. ```{r} #Produce observations with error and limited precision y_obs <- round(y_true + rnorm(length(y_true), 0, 0.1), digits = 1) #Unrounded data if needed #y_obs <- y_true + rnorm(length(y_true), 0, 0.1) #Observed data frame obs_data_frame <- tibble( time = time, y_obs = y_obs, obs_index = 1:length(y_obs) ) #Have a look at the true and 'observed' data plot_data <- tibble( x = c(time, time), y = c(y_true,y_obs), Data = c(rep("True sizes", times = length(y_true)), rep("Obs. sizes", times = length(y_obs))) ) sizes_over_time <- ggplot(plot_data, aes(x = x, y = y, group = Data)) + geom_point(aes(colour = Data, shape = Data), size = 2) + geom_line(aes(colour = Data, linetype = Data), linewidth = 1) + scale_linetype_manual(values = c("dashed", "dotted")) + ylim(0, 10.5) + labs(x = "Time", y = "Y") + theme_classic() + theme(legend.position = "inside", legend.position.inside = c(0.7, 0.3)) sizes_over_time #Have a look at the observations against the analytic solution analytic_observed <- ggplot(obs_data_frame, aes(x = time, y = y_obs)) + geom_function(fun=analytic_solution, args = true_args_list, linewidth = 1, colour = "black") + geom_point(colour = "darkorchid", size = 3) + geom_line(colour = "darkorchid", linewidth = 1, linetype = "dashed") + labs(x = "Time", y = "Y", title = "Data simulation") + ylim(0, 10.5) + theme_classic() analytic_observed ``` A note on error and precision: the same bad behaviour occurs even if you use unrounded data with all the precision R offers and much smaller error. You can test this by uncommenting the line that rounds the data and/or changing the measurement error standard deviation. The chosen values, and rounding, are intended to demonstrate that this is a problem that may occur with realistic precision and error. ## Implementing models and collecting posterior estimates In this section we are going to run a sequence of sets of models. The first batch is about checking for the existence of bimodal posterior distributions across different step sizes for the same numerical method. We then use a different numerical method to see if the problem persists. Lastly, we use different means in the parameter priors to see if the posterior can be constrained by such methods. ### Step size data We're going to run 100 fits with a step size of 1 using the custom RK4 solver and a single chain each. Each chain is expected to converge to a parameter combination that gives estimated $\hat{Y}(t_j)$ close to the analytic solution, but which combination is converged to is random. We do single chains because each can fall into the numerical error trap, and the easiest way to identify that trap is to extract the estimates afterwards. Each fit takes a few seconds to run, so allow several minutes for the following block. There are likely to be diagnostic problems but that is part of what we are here to explore so we will be ignoring them. Divergent transitions in particular are to be expected when we have the very different parameter estimates we see. Results are hidden for this block. ```{r, results='hide'} runs <- 100 step_size = 0.5 par_est_tibble <- tibble(run = c(), step_size = c(), beta_0 = c(), beta_1 = c()) for(i in 1:runs){ #Run the model suppressWarnings( fit <- hmde_model("affine_single_ind") |> hmde_assign_data(n_obs = nrow(obs_data_frame), y_obs = obs_data_frame$y_obs, obs_index = obs_data_frame$obs_index, time = obs_data_frame$time, y_bar = mean(obs_data_frame$y_obs), step_size = step_size, int_method = 1) |> #RK4 hmde_run(chains = 1, cores = 1, iter = 2000) ) #Extract parameter estimates ests <- hmde_extract_estimates(fit = fit, input_measurement_data = obs_data_frame) temp <- tibble( run = i, step_size = step_size, beta_0 = ests$individual_data$ind_beta_0_mean, beta_1 = ests$individual_data$ind_beta_1_mean ) par_est_tibble <- rbind(par_est_tibble, temp) } ``` ## Analysis We are going to fit a finite mixture model to tell us about the clustering in the posterior distributions. We assume that there are two clusters (you can check with scatter plots), that one is close to the true values and the other some distance away with much larger estimates for both, and use the mean for rows where $\hat{\beta_}_0 > mean(\hat{\beta_}_0)$ as our starting value for the iterative process to avoid singularities. The overall mean of the estimates works as a threshold for extreme values because of the bimodality and distance between clusters. ```{r} step_size_mix_models <- list() step_size_mix_model_plots <- list() step_size_mix_models_par_ests <- tibble( good_beta_0 = c(), good_beta_1 = c(), error_beta_0 = c(), error_beta_1 = c(), step_size = c(), error_fraction = c(), dist = c() ) for(i in 1:length(unique(par_est_tibble$step_size))){ #Get data for single step size step_size_selected <- unique(par_est_tibble$step_size)[i] analysis_data <- par_est_tibble %>% filter(step_size == step_size_selected) #Get some extreme estimates possible_error <- analysis_data %>% filter(beta_0 > mean(analysis_data$beta_0)) #To speed up the iterative algorithm we provide some initial conditions mu <- list( #Means from true parameters and extreme estimates true = c(beta_0, beta_1), error = c(mean(possible_error$beta_0), mean(possible_error$beta_1)) ) #Fit multivariate normal finite mixture model to the estimates step_size_mix_models[[i]] <- mvnormalmixEM(x = analysis_data[,c(3,4)], mu = mu) print(paste0("Summary of mixture model for step size ", step_size_selected)) print(summary(step_size_mix_models[[i]])) step_size_mix_model_plots[[i]] <- plot(step_size_mix_models[[i]], whichplots = 2, xlab2 = "Beta 0", ylab2 = "Beta 1") dist_table <- tibble( #Data to calculate distance b_0 = c(step_size_mix_models[[i]]$mu[[2]][1], step_size_mix_models[[i]]$mu[[1]][1]), b_1 = c(step_size_mix_models[[i]]$mu[[2]][2], step_size_mix_models[[i]]$mu[[1]][2]) ) #Extract values step_size_mix_models_par_ests_temp <- tibble( good_beta_0 = step_size_mix_models[[i]]$mu[[1]][1], good_beta_1 = step_size_mix_models[[i]]$mu[[1]][2], error_beta_0 = step_size_mix_models[[i]]$mu[[2]][1], error_beta_1 = step_size_mix_models[[i]]$mu[[2]][2], step_size = step_size_selected, error_prob = step_size_mix_models[[i]]$lambda[2], dist = dist(dist_table) ) step_size_mix_models_par_ests <- rbind(step_size_mix_models_par_ests, step_size_mix_models_par_ests_temp) } #Have a look at the estimates step_size_mix_models_par_ests ``` We get bimodality in the posterior distributions. If you look at multiple step sizes you will see that smaller step sizes push the second cluster further away. Bias in the estimates closest to the true values is due to the same measurement error in the 'observed' data for all the fits. The second mode in the estimates arises from the numerical integration error as we will verify shortly. The extreme bimodality of the posterior is consistent behaviour, even though the point of the second mode shifts based on the step size. Some aesthetics before plots. ```{r} legend_spec <- tibble( step_size_name = c("0.5", "0.25", "0.125"), step_size = c(0.5, 0.25, 0.125), x = c(-10, -10, -10), y = c(-10, -10, -10), colours = c("#f8766d", "#00ba38", "#609cff"), linetypes = c("longdash", "dashed", "dotted"), shapes = c(19, 17, 15) ) legend_spec_with_true <- tibble( step_size_name = c("0.5", "0.25", "0.125", "True pars"), step_size = c(0.5, 0.25, 0.125, NA), x = c(-10, -10, -10, -10), y = c(-10, -10, -10, -10), colours = c("#f8766d", "#00ba38", "#609cff", "black"), linetypes = c("longdash", "dashed", "dotted", "solid"), shapes = c(19, 17, 15, 3) ) for(i in 1:nrow(legend_spec)){ fancy_name_no_step_size <- paste0("Beta_0 = ", signif(step_size_mix_models_par_ests$error_beta_0[i], digits = 3), ",\n Beta_1 = ", signif(step_size_mix_models_par_ests$error_beta_1[i], digits = 3)) legend_spec$fancy_name_no_step_size[i] <- fancy_name_no_step_size legend_spec_with_true$fancy_name_no_step_size[i] <- fancy_name_no_step_size fancy_name <- paste0("Step size ", step_size_mix_models_par_ests$step_size[i], "\n", fancy_name_no_step_size) legend_spec$fancy_name[i] <- fancy_name legend_spec_with_true$fancy_name[i] <- fancy_name } legend_spec_with_true$fancy_name_no_step_size[4] <- paste0("Beta_0 = ", beta_0, ",\n Beta_1 = ", beta_1) legend_spec_with_true$fancy_name[4] <- paste0("True values\n Beta_0 = ", beta_0, ",\n Beta_1 = ", beta_1) ``` We use scatter plots of the clusters for qualitative analysis. As the clusters are so distant from each other, and so tight around their means, we separate them out for plots. The mixture models gives a classification of each point in the data that we use to filter observations. As the clusters are so distant we can use other heuristics such as $\hat{\beta_}_0 > 2\beta_0$ which agree perfectly with the cluster analysis. The contours over the scatter plot come from data simulated from the cluster's multivariate normal distribution identified by the finite mixture model. ```{r} scatterplot_errors_only <- list() scatterplot_good_only <- list() for(i in 1:length(unique(par_est_tibble$step_size))){ step_size_select <- unique(par_est_tibble$step_size)[i] plot_data <- par_est_tibble %>% filter(step_size == step_size_select) #Get classification from mixture model plot_data$good_est <- step_size_mix_models[[i]][["posterior"]][,1] error_ests_scatter <- plot_data %>% filter(!as.logical(good_est)) good_ests_scatter <- plot_data %>% filter(as.logical(good_est)) #Scatter plot of erroneous parameters xpos <- (min(error_ests_scatter$beta_0) + 0.2*(max(error_ests_scatter$beta_0) - min(error_ests_scatter$beta_0))) ypos <- (max(error_ests_scatter$beta_1) - 0.1*(max(error_ests_scatter$beta_1) - min(error_ests_scatter$beta_1))) norm_data <- as.data.frame(mvrnorm(n = 10000, mu = step_size_mix_models[[i]][["mu"]][[2]], Sigma = step_size_mix_models[[i]][["sigma"]][[2]])) names(norm_data) <- c("beta_0", "beta_1") scatterplot_errors_only[[i]] <- ggplot(data = error_ests_scatter, aes(x = beta_0, y = beta_1)) + geom_point(colour = legend_spec$colours[i], shape = legend_spec$shapes[i], alpha = 0.5, size = 2) + geom_density_2d(data = norm_data, colour = "black") + labs(x = "beta_0 est.", y = "beta_1 est.", title = "Second cluster") + annotate("text", x = xpos, y = ypos, label = paste0("Probability: \n", step_size_mix_models_par_ests$error_prob[i])) + theme_classic() #Scatter plot of good parameter estimates xpos <- (min(good_ests_scatter$beta_0) + 0.2*(max(good_ests_scatter$beta_0) - min(good_ests_scatter$beta_0))) ypos <- (max(good_ests_scatter$beta_1) - 0.1*(max(good_ests_scatter$beta_1) - min(good_ests_scatter$beta_1))) norm_data <- as.data.frame(mvrnorm(n = 10000, mu = step_size_mix_models[[i]][["mu"]][[1]], Sigma = step_size_mix_models[[i]][["sigma"]][[1]])) names(norm_data) <- c("beta_0", "beta_1") scatterplot_good_only[[i]] <- ggplot(data = good_ests_scatter, aes(x = beta_0, y = beta_1)) + geom_point(colour = legend_spec$colours[i], shape = legend_spec$shapes[i], alpha = 0.5, size = 2) + geom_density_2d(data = norm_data, colour = "black") + labs(x = "beta_0 est.", y = "beta_1 est.", title = "First cluster") + annotate("text", x = xpos, y = ypos, label = paste0("Probability: \n", (1-step_size_mix_models_par_ests$error_prob[i]))) + theme_classic() } ``` We can double-check the numerical error by using an independent solver with the same step size. We use the `deSolve` package which has an implementation of RK4 and allows us to choose the step sizes using the time parameter. ```{r} #install.packages("deSolve") library(deSolve) #Create DE function DE <- function(Time, State, Pars) { #Implementation of DE with(as.list(c(State, Pars)), { dY <- beta_0 - beta_1 * Y return(list(c(dY))) }) } ``` ### Second cluster analysis We want to look at the behaviour of the numerical method for the bad estimate clusters. To do so we project forward from the initial condition using the chosen step size, bad parameter combination, and see what happens. We can compare the numerical solution to both the true sizes over time, and to the analytic solution with those same bad parameter estimates. First we generate the numerical and analytic solution data. ```{r} yini <- c(Y = true_y_0) #Initial condition y_over_time <- tibble(model="True Sizes", y_analytic = y_true, y_numeric = y_true, time = 0:max_time, beta_0_par = beta_0, beta_1_par = beta_1 ) #Generate Y(t) with RK4 for(i in 1:nrow(step_size_mix_models_par_ests)){ pars_combo <- c(beta_0 = step_size_mix_models_par_ests$error_beta_0[i], beta_1 = step_size_mix_models_par_ests$error_beta_1[i]) times <- seq(0, max_time, by = step_size_mix_models_par_ests$step_size[i]) solution_pars <- c(pars_combo, true_y_0) y_true_temp <- analytic_solution(times, solution_pars) numerical_output <- ode(yini, times, DE, pars_combo, method = "rk4") y_over_time_temp <- tibble( model = step_size_mix_models_par_ests$step_size[i], y_analytic = y_true_temp, y_numeric = numerical_output[,2], time = times, beta_0_par = step_size_mix_models_par_ests$error_beta_0[i], beta_1_par = step_size_mix_models_par_ests$error_beta_1[i] ) y_over_time <- rbind(y_over_time, y_over_time_temp) } ``` Here is a figure that shows all of the estimated sizes over time for the bad parameter combinations across different step sizes, compared to the true values. ```{r} y_over_time_filtered <- y_over_time %>% filter(time %in% 0:max_time) #Plot sizes over time for all models compare_sizes_over_time <- ggplot(y_over_time_filtered, aes(x=time, y=y_numeric, group_by = as.factor(model))) + geom_point(aes(colour = as.factor(model), shape = as.factor(model)), alpha=0.5, size = 2, stroke = 1.5) + geom_line(aes(colour = as.factor(model)), alpha=0.5, linewidth = 1) + scale_colour_manual(values = legend_spec_with_true$colours) + scale_shape_manual(values = legend_spec_with_true$shapes) + labs(x = "Time", y = "Y(t)", title = "Estimated Y(t) with bad parameters", colour = "Step size", shape = "Step size") + theme_classic() + theme(legend.position = "inside", legend.position.inside = c(0.7, 0.3)) compare_sizes_over_time ``` These are indistinguishable values. If you look extremely closely you get some deviation due to parameter bias, but that is not statistically relevant to the process. To demonstrate that a smaller step size to test the method is enough to identify bad estimates, we show that RK4 with step size 0.001 diverges from the true sizes over time. The lines in this plot are based on the small step size numerical estimates, while the points come from the $Y(t_j)$ values for the discrete observation times. ```{r} #Generate y(t) with RK4 given the parameter estimates y_over_time_smallstep <- tibble(model=legend_spec_with_true$fancy_name[4], y_hat = y_true, time = 0:max_time ) for(i in 1:nrow(step_size_mix_models_par_ests)){ pars_combo <- c(beta_0 = step_size_mix_models_par_ests$error_beta_0[i], beta_1 = step_size_mix_models_par_ests$error_beta_1[i]) times <- seq(0, max_time, by = 0.001) numerical_output <- ode(yini, times, DE, pars_combo, method = "rk4") y_over_time_temp <- tibble( model = legend_spec_with_true$fancy_name_no_step_size[i], y_hat = numerical_output[,2], time = times ) y_over_time_smallstep <- rbind(y_over_time_smallstep, y_over_time_temp) } point_data <- y_over_time_smallstep %>% filter(time %in% 0:max_time) #Plot sizes over time compare_sizes_over_time_smallstep <- ggplot(y_over_time_smallstep, aes(x=time, y=y_hat, grouping = as.factor(model))) + geom_line(aes(colour = as.factor(model), linetype = as.factor(model)), alpha=0.5, linewidth = 1) + geom_point(data = point_data, aes(colour = as.factor(model), shape = as.factor(model)), alpha=0.5, size = 2, stroke = 1.5) + geom_function(fun=analytic_solution, args=true_args_list, colour="black", linetype = "solid", linewidth=1) + scale_shape_manual(values = legend_spec_with_true$shapes) + scale_colour_manual(values = legend_spec_with_true$colours) + scale_linetype_manual(values = c(legend_spec$linetypes, NA)) + labs(x = "Time", y = "Y(t)", title = "Small step size", colour = "Parameters", shape = "Parameters", linetype = "Parameters") + theme_classic() + theme(legend.position = "inside", legend.position.inside = c(0.7, 0.4), legend.key.spacing.y = unit(2, 'mm')) + guides(colour = guide_legend(byrow = TRUE)) compare_sizes_over_time_smallstep ``` Here are two plots showing the ODEs and analytic solutions with the bad estimates compared to the true parameter values. To plot the ODEs we exploit the fact that they are straight lines rather than plotting the functions properly. ```{r} #Get asymptotic size step_size_mix_models_par_ests$Y_max <- step_size_mix_models_par_ests$error_beta_0/step_size_mix_models_par_ests$error_beta_1 #Build points for start and end of lines plot_data <- tibble( x = c(0, (beta_0/beta_1), rep(0, times = nrow(step_size_mix_models_par_ests)), step_size_mix_models_par_ests$Y_max), y = c(beta_0, 0, step_size_mix_models_par_ests$error_beta_0, rep(0, times = nrow(step_size_mix_models_par_ests))), step_size = c("True pars", "True pars", step_size_mix_models_par_ests$step_size, step_size_mix_models_par_ests$step_size) ) #Plot DEs error_de_plot <- ggplot(data = plot_data, aes(x,y)) + geom_line(aes(colour = as.factor(step_size), linetype = as.factor(step_size)), linewidth = 1) + scale_colour_manual(values = c(legend_spec$colours[3:1], "black")) + scale_linetype_manual(values = c(legend_spec$linetypes[3:1], "solid")) + labs(title = "ODEs", x = "Y(t)", y = "f", colour = "Step size", linetype = "Step size") + theme_classic() + theme(legend.position = "inside", legend.position.inside = c(0.7, 0.7)) error_de_plot #Plot analytic solutions error_solution_plot <- ggplot() + geom_function(fun=analytic_solution, args=true_args_list, colour="black", linetype = "solid", linewidth=1) for(i in 1:nrow(step_size_mix_models_par_ests)){ #Add the analytic solutions args_list <- list(pars=c(step_size_mix_models_par_ests$error_beta_0[i], step_size_mix_models_par_ests$error_beta_1[i], true_y_0)) error_solution_plot <- error_solution_plot + geom_function(fun=analytic_solution, args=args_list, colour=legend_spec$colours[i], linetype = legend_spec$linetypes[i], linewidth=1) } error_solution_plot <- error_solution_plot + geom_line(data = legend_spec_with_true, linewidth=1, aes(colour = fancy_name_no_step_size, linetype = fancy_name_no_step_size, x = x, y = y)) + scale_colour_manual(values = c(legend_spec_with_true$colours[4], legend_spec$colours[c(2,3,1)])) + scale_linetype_manual(values = c(legend_spec_with_true$linetypes[4], legend_spec$linetypes[c(2,3,1)])) + xlim(0, max_time) + ylim(true_y_0, (beta_0/beta_1+0.5)) + labs(x = "Time", y = "Y(t)", title = "Analytic solutions", colour = "Parameters", linetype = "Parameters") + theme_classic() + theme(legend.position = "inside", legend.position.inside = c(0.7, 0.4), legend.key.spacing.y = unit(2, 'mm')) + guides(colour = guide_legend(byrow = TRUE)) error_solution_plot ``` The limiting behaviour at $\beta_0/\beta_1$ is consistent, what changes is how fast line approaches the asymptote. # Where to from here? For the purpose of the hmde package that this vignette is a part of, we account for the pathologies in this particular model by using the analytic solution for the von Bertalanffy equation. More work needs to be done to understand the interaction between numerical methods and MCMC sampling. What we have demonstrated is that the problem exists, and it is not enough to have numerical stability at the true parameter values because MCMC estimation moves around, you need numerical stability in a potentially quite large part of the parameter space. The good news is that simulated data and posterior plots with more accurate numerical methods can at least identify that something is going wrong.