![Lifecycle: experimental](https://camo.githubusercontent.com/e0545638bb59b45a142a632a941ffa63e6816e0966382f7492030ec90a76313b/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f6c6966656379636c652d6578706572696d656e74616c2d6f72616e67652e737667)
The goal of rmot is to implement hierarchical Bayesian longitudinal models to solve the Bayesian inverse problem of estimating differential equation parameters based on repeat measurement surveys. Estimation is done using Markov Chain Monte Carlo, implemented through
Stan via RStan, built under R 4.3.3. The inbuilt models are based on case studies in ecology.
The general use case is to estimate a vector of parameters $\boldsymbol{\theta}$ for a chosen differential equation
$$f\left( Y \left( t \right), \boldsymbol{\theta} \right) = \frac{dY}{dt}$$
based on the longitudinal structure
$$Y \left( t_{j+1} \right) = Y\left( t_j \right) + \int_{t_j}^{t_{j+1}}f\left( Y \left( t \right), \boldsymbol{\theta} \right),dt. $$
The input data are observations of the form $y_{ij}$ for individual $i$ at time $t_j$, with repeated observations coming from the same individual. We parameterise $f$ at the individual level by estimating $\boldsymbol{\theta}_i$ as the vector of parameters. We have hyper-parameters that determine the distribution of $\boldsymbol{\theta}_i$ with typical prior distribution
$$\boldsymbol{\theta}_i \sim \log \mathcal{N}\left(\boldsymbol{\mu}_{\log\left(\boldsymbol{\theta}\right)}, \boldsymbol{\sigma}_{\log \left( \boldsymbol{\theta} \right)}\right), $$
where $\boldsymbol{\mu}_{\log\left(\boldsymbol{\theta}\right)}$ and $\boldsymbol{\sigma}_{\log\left(\boldsymbol{\theta}\right)}$ are vectors of means and standard deviations. In the case of a single individual, these are chosen prior values. In the case of a multi-individual model $\boldsymbol{\mu}_{\log\left(\boldsymbol{\theta}\right)}$ and $\boldsymbol{\sigma}_{\log\left(\boldsymbol{\theta}\right)}$ have their own prior distributions and are fit to data.
Rmot comes with four DEs built and ready to go, each of which has a version for a single individual and multiple individuals.
The constant model is given by
$$f \left( Y \left( t \right), \beta \right) = \frac{dY}{dt} = \beta,$$
and is best understood as describing the average rate of change over time.
The power law model is given by
$$f \left( Y \left( t \right), \beta_0, \beta_1, \bar{Y} \right) = \frac{dY}{dt} = \beta_0 \bigg( \frac{Y \left( t \right)}{\bar{Y}} \bigg)^{\beta_1},$$
where $\beta_0>0$ is the coefficient, $\beta_1$ is the power, and $\bar{Y}$ is a user-provided parameter that centres the model in order to avoid correlation between the $\beta$ parameters.
The von Bertalanffy mode is given by
$$f \left( Y \left( t \right), \beta, Y_{max} \right) = \frac{dY}{dt} = \beta \left( Y_{max} - Y \left( t \right) \right),$$
where $\beta$ is the growth rate parameter and $Y_{max}$ is the maximum value that $Y$ takes.
The Canham (Canham et
al. 2004)
model is a hump-shaped function given by
$$f \left( Y \left( t \right), f_{max}, Y_{max}, k \right) = \frac{dY}{dt} = f_{max} \exp \Bigg( -\frac{1}{2} \bigg( \frac{ \ln \left( Y \left( t \right) / Y_{max} \right) }{k} \bigg)^2 \Bigg), $$
where $f_{max}$ is the maximum growth rate, $Y_{max}$ is the $Y$-value at which that maximum occurs, and $k$ controls how narrow or wide the peak is.
‘rmot’ is under active development. You can install the current
developmental version of ‘rmot’ from GitHub with:
# install.packages("remotes")
remotes::install_github("traitecoevo/rmot")
Create constant growth data with measurement error:
y_obs <- seq(from=2, to=15, length.out=10) + rnorm(10, 0, 0.1)
Measurement error is necessary as otherwise the normal likelihood
$$s_{ij} \sim \mathcal{N}\left( 0, \sigma_e \right)$$
blows up as $\sigma_e$ approaches 0.
Fit the model:
constant_fit <- rmot_model("constant_single_ind") |>
rmot_assign_data(n_obs = 10, #Integer
y_obs = y_obs, #vector length n_obs
obs_index = 1:10, #vector length n_obs
time = 0:9, #Vector length n_obs
y_0_obs = y_obs[1] #Real
) |>
rmot_run(chains = 1, iter = 1000, verbose = FALSE, show_messages = FALSE)
Please submit a GitHub issue with details of the bug. A reprex would be particularly helpful with the bug-proofing process!