There are a number of reasons why phenomix models may
not converge. First, it’s likely that all models won’t converge for a
single dataset. As an example, we can create some data representing an
asymmetric distribution, with gaussian tails (no extremes) and random
variability by year (both in the mean and standard deviations).
# create 20 years of data
set.seed(123)
df <- expand.grid("doy" = 100:200, "year" = 1:20)
df$mu <- rnorm(unique(df$year), 150, 5)[df$year]
df$sig1 <- rnorm(unique(df$year), 30, 5)[df$year]
df$sig2 <- rnorm(unique(df$year), 30, 5)[df$year]
df$sig <- ifelse(df$doy < df$mu, df$sig1, df$sig2)
df$pred <- dnorm(df$doy, df$mu, sd = df$sig, log = TRUE)
df$pred <- exp(df$pred + 8)
df$number <- round(rnorm(nrow(df), df$pred, 0.1))The model with student-t errors that is fit to these data
set.seed(1)
fit_t <- fit(create_data(df, asymmetric_model = TRUE, min_number = 1,
tail_model = "student_t"),
silent = TRUE,
control = list(eval.max = 4000, iter.max = 5000, rel.tol = 1e-7)
)But the model with the generalized normal tails struggles.
fit_gnorm <- fit(create_data(df, asymmetric_model = TRUE, min_number = 1,
tail_model = "gnorm"),
silent = TRUE,limits=TRUE,
control = list(eval.max = 4000, iter.max = 5000, rel.tol = 1e-7)
)This is an example where simplifying helps greatly. The model with generalized normal tails will converge when the random effects in the mean and variance are turned off, but has a hard time estimating the shape parameters with them on.
fit_gnorm <- fit(create_data(df, asymmetric_model = TRUE, min_number = 1,
tail_model = "gnorm",
est_mu_re = FALSE,
est_sigma_re = FALSE),
silent = TRUE,
control = list(eval.max = 4000, iter.max = 5000, rel.tol = 1e-7)
)Sometimes, it helps to specify the initial values for a model. We can do that in a couple ways, but it may be easiest to first construct a model to see what initial values are needed.
Using the above example, we can specify
fit_model = FALSE.
fit_gnorm <- fit(create_data(df, asymmetric_model = TRUE, min_number = 1,
tail_model = "gnorm",
est_mu_re = FALSE,
est_sigma_re = FALSE),
silent = TRUE,
control = list(eval.max = 4000, iter.max = 5000, rel.tol = 1e-7),
fit_model = FALSE
)The $init_vals contains the initial values that would be
used to fit the model.
fit_gnorm$init_vals
#> theta theta theta theta theta
#> 2.0308397 2.2015325 3.5295172 3.1596658 2.9785872
#> theta theta theta theta theta
#> 0.9586761 1.9551192 3.1481328 2.4014047 2.4826709
#> theta theta theta theta theta
#> 4.3289806 3.4550282 2.2355953 2.8107439 2.9945364
#> theta theta theta theta theta
#> 2.5991447 2.3153863 3.2276358 4.0206346 3.5062919
#> log_beta_1 log_obs_sigma b_mu b_sig1 b_sig2
#> 0.1000000 0.0000000 150.0000000 1.0000000 1.0000000Suppose we wanted to start at a higher value for
log_beta_1. We could change that with
Now we can try re-fitting the model,
There are two ways to specify limits on estimated parameters. First,
we include hard coded limits that may be turned on with
limits = TRUE, e.g.
fit_gnorm <- fit(create_data(df, asymmetric_model = TRUE, min_number = 1,
tail_model = "gnorm",
est_mu_re = FALSE,
est_sigma_re = FALSE),
silent = TRUE,
limits = TRUE,
control = list(eval.max = 4000, iter.max = 5000, rel.tol = 1e-7),
fit_model = TRUE
)However these limits may not be reasonable for every situation. We can also change these manually, but including a list of limits based on the parameters being estimated. We do this using the same approach above, with specifying initial values. First we construct the model but don’t do estimation,
fit_gnorm <- fit(create_data(df, asymmetric_model = TRUE, min_number = 1,
tail_model = "gnorm",
est_mu_re = FALSE,
est_sigma_re = FALSE),
silent = TRUE,
control = list(eval.max = 4000, iter.max = 5000, rel.tol = 1e-7),
fit_model = FALSE
)The parameters need to be in the same order as
init_vals, so we can try something like this:
lower = c(rep(0, 20), # lower for theta
0.05, # lower log_beta_1
-3, # lower log_obs_sigma
140, # lower on mu
15,# lower on b_sig1
15)# lower on b_sig2
upper = c(rep(10, 20),# upper for theta
log(10),# upper log_beta_1
0, # upper log_obs_sigma
160,# upper on mu
40,# upper on b_sig1
40)# upper on b_sig2Then we can pass these in as a list,