Examples of fitting DFA models with lots of data

Eric J. Ward, Sean C. Anderson, Mary E. Hunsicker, Mike A. Litzow, Luis A. Damiano, Mark D. Scheuerell, Elizabeth E. Holmes, Nick Tolimieri

2024-02-26

Source: vignettes/a7_bigdata.Rmd

For some applications, there may be a huge number of observations (e.g. daily stream flow measurements, bird counts) making estimation with MCMC slow. While estimation (and uncertainty) for final models should be done with MCMC, there are a few much faster alternatives that we can use for these cases. They may be generally useful for other DFA problems – both in diagnosing convergence problems, and doing preliminary model selection.

Let’s load the necessary packages:

Data simulation

The sim_dfa function normally simulates loadings \(\sim N(0,1)\), but here we will simulate time series that are more similar with loadings \(\sim N(1,0.1)\)

set.seed(1)
s = sim_dfa(num_trends = 1, num_years = 1000, num_ts = 4,
            loadings_matrix = matrix(nrow = 4, ncol = 1, rnorm(4 * 1,
    1, 0.1)), sigma=0.05)
matplot(t(s$y_sim), type="l")

Sampling argument

In the examples below, we’ll take advantage of the estimation argument. By default, this defaults to MCMC (“sampling”) but can be a few other options described below. If you want to construct a model object, but do no sampling, you can also set this to “none”.

fit <- fit_dfa(..., estimation = "sampling")

Posterior optimization

The fastest estimation approach is to do optimze the posterior (this is similar to maximum likelihood but also involves the prior distribution). We can implement this with by setting the estimation argument to “optimizing”

Note – because this model has a lot of parameters, estimation can be finicky, and can get stuck in local minima. You may have to start this from several seeds to get the model to converge successfully – or if there is a mismatch between the model and data, it may not converge at all.

For example, this model does not converge

set.seed(123)
m <- fit_dfa(y = s$y_sim, estimation = "optimizing")
## Warning in .local(object, ...): non-zero return code in optimizing

The optimizing output is saved here (value = log posterior, par = estimated parameters)

names(m$model)
## [1] "par"         "value"       "return_code" "theta_tilde"

And if convergence is successful, the optimizer code will be 0 (this model isn’t converging)

m$model$return_code
## [1] 70

But if we change the seed, the model will converge ok:

set.seed(124)
m <- fit_dfa(y = s$y_sim, estimation = "optimizing")
m$model$return_code
## [1] 0

Posterior approximation

A second approach to quickly estimating parameters is to use Variational Bayes, which is also implemented in Stan. This is implemented by changing the estimation to “vb”, as shown below. Note: this gives a helpful message that the maximum number of iterations has been reached, so these results should not be trusted.

m <- fit_dfa(y = s$y_sim, estimation = "vb", seed=123)

There are a number of other arguments that can be passed into rstan::vb(). These include iter (maximum iterations, defaults to 10000), tol_rel_obj (convergence tolerance, defaults to 0.01), and output_samples (posterior samples to save, defaults to 1000). To use these, a function call would be

m <- fit_dfa(y = s$y_sim, estimation = "vb", seed=123, iter=20000, 
             tol_rel_obj = 0.005, output_samples = 2000)