Simulation with sdmTMB

class: center, middle, inverse, title-slide

.title[
# Simulation with sdmTMB
]
.subtitle[
## DFO DSAF workshop
]
.author[
### 
]
.date[
### January 12–16 2026
]

---

# Useful things to do with sdmTMB simulation

- Survey design

- Evaluating model distributional assumptions

- Evaluating the ability to estimate quantities / model bias

- To understand how a model behaves

---

# Decisions when simulating from a fitted model

Fixed effects: fixed (MLE), MVN draw, MCMC draw

Random effects:

- empirical Bayes (EB) estimates, MVN draw, MCMC draw *from the precision matrix*

- draw from a new random field based on the estimated "hyper" parameters (e.g. SD and range)

---

# Do I want to draw from the precision matrix?

.small[
Do I want my spatial "wiggles" to reflect my fitted system?
 - If yes: draw from the precision matrix
]

.small[
Do I want entirely new "wiggles" that just have the same properties (same variance and same rate of correlation decay)?
 - If yes: simulate new GMRFs
 ]
 
--

.small[
Do I want to ensure I have simulated data that can be recovered without any bias in ideal conditions?
 - If yes: simulate new GMRFs
]

---

# `simulate.sdmTMB()` vs. `sdmTMB_simulate()` !?

`simulate.sdmTMB()`: simulate from an already fitted model

`sdmTMB_simulate()`: simulate data from scratch; may be renamed `simulate_new()`.

The lines blur: you can take estimates from a fitted model and feed them to `sdmTMB_simulate()`.

---

class: center, middle, inverse

# Examples

---

# Examples

`simulate.sdmTMB()` by default only simulates new observation error and 1 draw:

``` r
simulate(fit)
simulate(fit, nsim = 100)
```

---

# Examples

Simulating with entirely new random effects based on the estimated SDs and ranges:

``` r
simulate(fit, re_form = NA)
```

---

# Examples

Simulating with a new draw of random effects for each iteration:

``` r
simulate(
  fit, 
* type = "mle-mvn",
* mle_mvn_samples = "multiple",
  nsim = 100
)
```

---

# Examples

Simulating from a fitted model and refitting to it

``` r
fit <- sdmTMB(y ~ 1, ..., data = dat)

*dat$y_sim <- simulate(fit, type = "mle-mvn")

fit2 <- sdmTMB(y_sim ~ 1, ..., data = dat)
```

---

# Examples

Simulating new random fields based on the estimated field properties

``` r
fit <- sdmTMB(...)

*dat$y_sim <- simulate(fit, re_form = NA)

fit2 <- sdmTMB(y_sim ~ 1, ..., data = dat)

# repeat in a loop and assess bias in parameters etc.
```

---

# Examples

.xsmall[
Simulate from a fitted model to assess whether your model is probabilistically consistent with the observations.
]

.small[

``` r
*s <- simulate(fit, nsim = 100)
prop_zeros <- apply(s, 2, \(x) mean(x == 0))
hist(prop_zeros)
abline(v = obs_prop_zero, col = "red", lwd = 2)
```

<img src="13-simulation_files/figure-html/unnamed-chunk-7-1.png" width="60%" style="display: block; margin: auto;" />
]

---

# Examples

Use `newdata` in `sdmTMB.simulate()` to simulate new observations. E.g., to compare survey designs.

``` r
fit <- sdmTMB(...)

# new = new survey locations
new$y_sim <- simulate(fit, newdata = new, ...)

fit_sim <- sdmTMB(y_sim ~ ..., data = new)

p <- predict(fit_sim, ...)

get_index(p)1

# compare against known truth
# (simulate on your survey grid) and repeat
```

---

# Examples

Feeding in random effects from a fitted model to `sdmTMB_simulate()`.

``` r
fit <- sdmTMB(...)
omega <- get_pars(fit)$omega_s
sdmTMB_simulate(
  ~ depth, 
  data = data, 
* fixed_re = list(omega_s = omega)
)
```

---

# Examples

.xsmall[
Check simulations conditioned on a model as a graphical "Turing test". Can you, or a colleague, or a reviewer distinguish which panel is real?
]

.center[
<img src="images/coastwide-posterior-predictive-sims-slides-spiny.png" width="280px"/>
]