Forecasting Longitudinal Networks with lame

Cassy Dorff, Tosin Salau, Shahryar Minhas

2026-11-17

What this vignette covers

lame() with at least one dynamic component (dynamic_beta, dynamic_ab, dynamic_uv) is a state-space model, and predict(fit, h = ...) carries its posterior forward to future periods. This vignette walks through:

  1. Fit a model with dynamic coefficients.
  2. Forecast h periods ahead on the link scale and the response scale.
  3. Compute counterfactual forecasts by feeding in alternative future covariates.
  4. Visualise the forecast and the per-coefficient ribbon plot.
  5. Check the forecast: choose the transition model, score held-out periods and assess calibration.

Step 1: Fit

We simulate 30 actors over 6 periods, a single dyadic covariate trade, and a directed binary outcome cooperation. The effect of trade on cooperation is time-varying and mean-reverting in the simulation: it follows an AR(1) around 0.5 with persistence 0.7, starting high (0.9) and settling toward its long-run level. The network is moderately dense (~37%), more than 300 ties in each period. dynamic_beta estimates one coefficient per period, and the ties of a period are what inform its coefficient.

library(lame)
set.seed(2026)

n_fc <- 30
T_fc <- 6
rho_true <- 0.7
beta_bar <- 0.5
beta_t_true <- numeric(T_fc)
beta_t_true[1] <- 0.9
for (t in 2:T_fc) {
    beta_t_true[t] <- beta_bar + rho_true * (beta_t_true[t - 1] - beta_bar) +
        rnorm(1, 0, 0.15)
}

Xdyad <- lapply(seq_len(T_fc), function(t) {
    x <- matrix(rnorm(n_fc * n_fc), n_fc, n_fc)
    array(
        x,
        dim = c(n_fc, n_fc, 1),
        dimnames = list(NULL, NULL, "trade")
    )
})
a_fc <- rnorm(n_fc, 0, 0.4)
b_fc <- rnorm(n_fc, 0, 0.4)
Y <- lapply(seq_len(T_fc), function(t) {
    eta <- -0.5 + beta_t_true[t] * Xdyad[[t]][, , 1] + outer(a_fc, b_fc, "+")
    Yt <- matrix(rbinom(n_fc * n_fc, 1, pnorm(eta)), n_fc, n_fc)
    diag(Yt) <- NA
    rownames(Yt) <- colnames(Yt) <- sprintf("a%02d", seq_len(n_fc))
    Yt
})
names(Y) <- paste0("t", seq_len(T_fc))

round(beta_t_true, 3) # simulated coefficient path
#> [1] 0.900 0.858 0.589 0.583 0.545 0.432
sapply(Y, function(y) round(mean(y, na.rm = TRUE), 3)) # per-period density
#>    t1    t2    t3    t4    t5    t6 
#> 0.400 0.374 0.378 0.367 0.352 0.357

fit <- lame(
    Y,
    Xdyad = Xdyad,
    family = "binary",
    R = 0,
    dynamic_beta = "dyad", # the trade coefficient is AR(1)
    dynamic_beta_kind = "ar1", # mean-reverting; switch to "rw1" for drift
    nscan = 300,
    burn = 30,
    odens = 5,
    verbose = FALSE
)

# estimated path of the trade coefficient
coef_path <- coef(fit)
trade_row <- grep("^trade[._]dyad$|trade", rownames(coef_path), value = TRUE)[1]
round(coef_path[trade_row, ], 3)
#>    t1    t2    t3    t4    t5    t6 
#> 0.871 0.816 0.592 0.519 0.519 0.506

The estimated path follows the simulated one: it starts high and settles toward the long-run level, and the largest gap between the two is 0.07. lame() ran two independently seeded chains and compared them before returning the fit; the largest R-hat among the monitored quantities is 1.03. summary(fit) reports the posterior of the persistence \(\rho_\beta\), which Step 5 uses to choose the transition model.

Step 2: Forecast h periods ahead

# forecasts draw the AR(1) state at random, so seed them
set.seed(1)
# 3-step-ahead forecast on the link scale
fc_link <- predict(fit, h = 3, type = "link")
length(fc_link) # 3, one matrix per future period
#> [1] 3
dim(fc_link[[1]]) # n x n (30 x 30 here)
#> [1] 30 30

# response scale: the inverse link is applied to each draw
fc_resp <- predict(fit, h = 3, type = "response")
# posterior-mean predicted probabilities
range(fc_resp[[1]], na.rm = TRUE)
#> [1] 0.005357869 0.955282626

# by_draw = TRUE returns the full [n, n, h, n_draws] array
fc_full <- predict(fit, h = 3, type = "response", by_draw = TRUE)
dim(fc_full) # n x n x 3 x n_draws
#> [1]  30  30   3 352

# lower, median and upper matrices per period (95% by default)
fc_ci <- predict(fit, h = 3, type = "response", interval = "credible")
str(fc_ci[[1]]) # lower, median and upper, each n x n
#> List of 3
#>  $ lower : num [1:30, 1:30] 0.0795 0.0389 0.1298 0.1198 0.1084 ...
#>   ..- attr(*, "dimnames")=List of 2
#>   .. ..$ : chr [1:30] "a01" "a02" "a03" "a04" ...
#>   .. ..$ : chr [1:30] "a01" "a02" "a03" "a04" ...
#>  $ median: num [1:30, 1:30] 0.2204 0.0757 0.3048 0.1863 0.1787 ...
#>   ..- attr(*, "dimnames")=List of 2
#>   .. ..$ : chr [1:30] "a01" "a02" "a03" "a04" ...
#>   .. ..$ : chr [1:30] "a01" "a02" "a03" "a04" ...
#>  $ upper : num [1:30, 1:30] 0.411 0.128 0.523 0.287 0.264 ...
#>   ..- attr(*, "dimnames")=List of 2
#>   .. ..$ : chr [1:30] "a01" "a02" "a03" "a04" ...
#>   .. ..$ : chr [1:30] "a01" "a02" "a03" "a04" ...

Each element of fc_ci holds lower, median and upper matrices of tie probabilities for one future period. The 95% interval of a dyad is on average 0.237 wide on the probability scale at h = 1 and 0.293 at h = 3. Uncertainty about the coefficient path accumulates with the horizon; because the path is mean-reverting (the posterior mean of rho_beta is 0.8), it levels off at the rate given in Step 5.

Visualising the forecast

Two views make the forecast concrete. The left panel is the predicted tie-probability matrix for the first future period (fc_resp[[1]]) – the actual network the model expects next. The right panel collapses each horizon to its across-dyad average predicted probability with the average 95% credible band, so you can watch the band widen as the horizon grows.

library(ggplot2)
library(patchwork)

# left: forecast heatmap for the first future period
P1 <- fc_resp[[1]]
rn <- rownames(P1); if (is.null(rn)) rn <- sprintf("a%02d", seq_len(nrow(P1)))
cn <- colnames(P1); if (is.null(cn)) cn <- sprintf("a%02d", seq_len(ncol(P1)))
heat_df <- data.frame(
    sender = factor(rn[row(P1)], levels = rev(rn)),
    receiver = factor(cn[col(P1)], levels = cn),
    prob = as.vector(P1)
)
p_heat <- ggplot(heat_df, aes(receiver, sender, fill = prob)) +
    geom_tile() +
    scale_fill_distiller(palette = "Blues", direction = 1, limits = c(0, 1), name = "P(tie)") +
    labs(title = "Forecast for t7", x = "Receiver", y = "Sender") +
    theme_bw(base_size = 9) +
    theme(
        panel.border = element_blank(),
        axis.line = element_blank(),
        axis.text = element_blank(),
        axis.ticks = element_blank(),
        legend.position = "top",
        plot.title = element_text(hjust = 0),
        plot.subtitle = element_text(hjust = 0),
        plot.caption = element_text(hjust = 0)
    )

# right: across-dyad mean probability with the mean 95% band
horizon_df <- data.frame(
    period = factor(
        paste0("t", T_fc + seq_along(fc_ci)),
        levels = paste0("t", T_fc + seq_along(fc_ci))
    ),
    median = sapply(fc_ci, function(m) mean(m$median, na.rm = TRUE)),
    lower = sapply(fc_ci, function(m) mean(m$lower, na.rm = TRUE)),
    upper = sapply(fc_ci, function(m) mean(m$upper, na.rm = TRUE))
)
p_horizon <- ggplot(horizon_df, aes(period, median, group = 1)) +
    geom_ribbon(aes(ymin = lower, ymax = upper), fill = "grey85") +
    geom_line() + geom_point(size = 2) +
    labs(
        title = "Average forecast by horizon",
        x = "Forecast period", y = "Mean P(tie)"
    ) +
    theme_bw(base_size = 9) +
    theme(
        panel.border = element_blank(),
        axis.line = element_blank(),
        axis.ticks = element_blank(),
        legend.position = "top",
        plot.title = element_text(hjust = 0),
        plot.subtitle = element_text(hjust = 0),
        plot.caption = element_text(hjust = 0)
    )

p_heat + p_horizon

Left: heatmap of predicted tie probabilities among the 30 actors for the first forecast period, darker cells indicating higher predicted probability. Right: the across-dyad average predicted probability at horizons h = 1, 2, 3 with a shaded 95 percent credible band that widens with the horizon.

The heatmap shows the model does not forecast a uniform network: some sender/receiver pairs carry a much higher predicted tie probability than others, inherited from their estimated additive effects and the recovered trade coefficient. The horizon panel shows the average predicted probability staying roughly level (the coefficient is mean-reverting) while its credible band widens with h – the visual signature of accumulating forecast uncertainty.

The forecast propagates (β_t, a_t, b_t, U_t, V_t) forward by drawing from the joint posterior of the dynamic state-space hyperparameters (ρ_β, σ_β, ρ_ab, σ_ab, ρ_uv, σ_uv) for each posterior draw. The same forecasting machinery covers the supported outcome families; the binary example above is illustrative. For family = "normal", type = "response" returns the forecasted conditional mean. For family = "poisson", each matched parameter draw has conditional mean e * exp(η + ve / 2), where e is the future-period exposure (set with newexposure). With by_draw = FALSE the point forecast is the mean count at the draw-averaged predictor and variance, e * exp(mean(η) + mean(ve) / 2), the form of the in-sample fitted() value; by_draw = TRUE returns each draw’s own conditional mean. Use interval = "prediction" for predictive count quantiles, or interval = "credible" for quantiles of the per-draw conditional means. These are distinct summaries.

Poisson exposure offsets via period_exposure

For family = "poisson", pass period_exposure = e to lame() (a non-negative numeric vector of length T) when the per-period rate should be scaled by a known exposure (e.g. weeks observed, population size, number of edge-formation opportunities). The sampler treats log(period_exposure[t]) as a fixed offset on the linear predictor: conditional on the latent log-rate Z, the Poisson mean is period_exposure[t] * exp(Z_t). Integrating the Gaussian observation noise within a parameter draw gives period_exposure[t] * exp(η_t + ve / 2); at the posterior means of η_t and ve this is the mean count fitted() reports. Zero exposure gives a structural zero count. period_exposure belongs to the Poisson family. When you call predict(fit, h = K) later without newexposure, the response-scale forecast reuses the last observed exposure for every future period; pass newexposure = rep(e_new, K) to forecast under a different exposure path.

Step 3: Counterfactual forecasts

Pass newdata = list_of_h_future_X_arrays to combine alternative future covariates with the forecast. Each array is n x n x p and the slice order along the 3rd dimension must match the original dimnames(Xdyad[[1]])[[3]] from the fit (here a single slice, trade).

# seed the forecast draws
set.seed(1)
# baseline: hold trade at its last observed value
last_X <- Xdyad[[length(Xdyad)]]
X_future <- list(last_X, last_X, last_X)
fc_cf <- predict(fit, h = 3, type = "response", newdata = X_future)

# counterfactual: one more unit (one standard deviation) of trade for every dyad
X_future_up <- lapply(X_future, function(x) {
    x[, , 1] <- x[, , 1] + 1
    x
})
fc_up <- predict(fit, h = 3, type = "response", newdata = X_future_up)

# change in each dyad's predicted probability at h = 3
delta_h3 <- fc_up[[3]] - fc_cf[[3]]
summary(as.vector(delta_h3))
#>     Min.  1st Qu.   Median     Mean  3rd Qu.     Max. 
#> 0.009329 0.150519 0.189317 0.171677 0.208352 0.219691

The trade coefficient is positive, so one more unit of trade raises the forecast cooperation probability at h = 3: by 0.17 on average, and by between 0.01 and 0.22 across dyads. The marginal probit effect \(\phi(\eta)\,\beta\) is largest for dyads whose baseline linear predictor \(\eta = a_i + b_j + \beta x_{ij}\) is near zero (a tie probability near one half) and smallest for dyads whose probability is already close to 0 or 1. For substantive interpretation, summarise delta_h3 for the dyads that matter rather than reporting the aggregate summary().

Two properties of counterfactual covariates. (1) Because dynamic_beta draws the coefficient path forward from its posterior, the same covariate change produces different magnitudes at different horizons: the multiplier on trade moves with the horizon. (2) For a covariate centred at zero, such as trade, an additive shift (+ 1) is the meaningful counterfactual; a multiplicative one (* 2) suits strictly positive covariates.

Step 4: Visualise the coefficient path

library(ggplot2)
autoplot(fit, probs = c(0.025, 0.5, 0.975)) +
    labs(
        title = "Posterior coefficient paths",
        subtitle = "median + 95% interval",
        x = "Time period", y = "Coefficient value"
    )

In-sample posterior coefficient paths from autoplot.lame: each panel shows the posterior median line and a 95 percent credible interval ribbon across the fitted training periods (t1-t6) for one regression coefficient. This plots the estimated in-sample path, not a forecast.

The ribbon shows the posterior 95% credible interval around the posterior median per coefficient and period; faceted by coefficient.

Step 5: Forecast diagnostics

The transition model decides how forecast uncertainty grows with the horizon:

In words: an AR(1) coefficient returns to its long-run mean and its forecast settles into a stationary distribution; an RW1 coefficient keeps its last level and adds innovation variance every period.

lame checks which of the two the data support. After an AR(1) fit, summary(fit) recommends dynamic_beta_kind = "rw1" when the posterior of \(\rho_\beta\) is concentrated at or above 0.97, and predict() says so once per fit when the upper 97.5% bound of \(\rho_\beta\) reaches 0.99. For the Step 1 fit that bound is 0.979, below 0.99: the mean-reverting forecast stands.

A drifting coefficient: AR(1) or RW1

On a short panel the choice rests on the substantive question: does the effect return to a long-run level, or keep the level it has reached? Fit both transition models and compare their forecasts. Here a covariate effect drifts upward over four periods of a 20-actor network. The cells are simulated independently, so the fits set dcor = FALSE:

set.seed(2026)
n_sp <- 20
T_sp <- 4
beta_drift <- c(0.4, 0.9, 1.4, 1.9) # drifts upward without mean reversion
X_sp <- lapply(seq_len(T_sp), function(t) {
    x <- matrix(rnorm(n_sp * n_sp), n_sp, n_sp)
    array(
        x,
        dim = c(n_sp, n_sp, 1),
        dimnames = list(NULL, NULL, "sanction_risk")
    )
})
Y_sp <- lapply(seq_len(T_sp), function(t) {
    eta <- -2 + beta_drift[t] * X_sp[[t]][, , 1]
    Yt <- matrix(rbinom(n_sp * n_sp, 1, pnorm(eta)), n_sp, n_sp)
    diag(Yt) <- NA
    rownames(Yt) <- colnames(Yt) <- sprintf("s%02d", seq_len(n_sp))
    Yt
})
names(Y_sp) <- paste0("t", seq_len(T_sp))
sapply(Y_sp, function(y) round(mean(y, na.rm = TRUE), 3)) # sparse early on
#>    t1    t2    t3    t4 
#> 0.034 0.055 0.118 0.161

fit_drift <- lapply(c(ar1 = "ar1", rw1 = "rw1"), function(kind) {
    lame(
        Y_sp,
        Xdyad = X_sp,
        family = "binary",
        R = 0,
        dcor = FALSE,
        dynamic_beta = "dyad",
        dynamic_beta_kind = kind,
        nscan = 250,
        burn = 50,
        odens = 5,
        verbose = FALSE
    )
})
# estimated coefficient path under each transition model
round(t(sapply(fit_drift, function(f) coef(f)["sanction_risk_dyad", ])), 3)
#>        t1    t2    t3    t4
#> ar1 0.471 0.854 1.404 1.862
#> rw1 0.514 0.842 1.404 1.873

Both fits recover the upward drift, and they agree on the fitted periods. Four periods leave the posterior of \(\rho_\beta\) close to its prior (prior mean 0.8, posterior median 0.82 in the AR(1) fit), which is why the choice is made on substance. The two fits differ in what they carry forward. The forecast coefficient at horizon h is the link-scale difference between a forecast with sanction_risk = 1 for every dyad and one with sanction_risk = 0:

x_at <- function(value) {
    lapply(seq_len(6), function(h) {
        array(
            value,
            dim = c(n_sp, n_sp, 1),
            dimnames = list(NULL, NULL, "sanction_risk")
        )
    })
}
# the same seed gives both forecasts the same coefficient draws
beta_forecast <- function(f) {
    set.seed(1)
    fc_one <- predict(f, h = 6, type = "link", newdata = x_at(1))
    set.seed(1)
    fc_zero <- predict(f, h = 6, type = "link", newdata = x_at(0))
    mapply(function(a, b) mean(a - b, na.rm = TRUE), fc_one, fc_zero)
}
beta_fc <- sapply(fit_drift, beta_forecast)
round(t(beta_fc), 3)
#>      [,1]  [,2]  [,3]  [,4]  [,5]  [,6]
#> ar1 1.660 1.500 1.361 1.288 1.193 1.151
#> rw1 1.861 1.849 1.816 1.833 1.790 1.821

drift_df <- data.frame(
    h = rep(seq_len(6), 2),
    model = rep(c("AR(1)", "RW1"), each = 6),
    coefficient = c(beta_fc[, "ar1"], beta_fc[, "rw1"])
)
ggplot(drift_df, aes(h, coefficient, linetype = model)) +
    geom_line() + geom_point(size = 1.8) +
    labs(
        title = "Forecast coefficient by horizon",
        x = "Forecast horizon h", y = "Coefficient on sanction_risk"
    ) +
    theme_bw(base_size = 9) +
    theme(
        panel.border = element_blank(),
        axis.line = element_blank(),
        axis.ticks = element_blank(),
        legend.position = "top",
        plot.title = element_text(hjust = 0),
        plot.subtitle = element_text(hjust = 0),
        plot.caption = element_text(hjust = 0),
        legend.title = element_blank()
    )

Line plot of the forecast coefficient on sanction_risk at horizons one through six under the AR(1) and RW1 transition models fitted to the drifting panel. The AR(1) line declines toward the long-run mean; the RW1 line stays at the last fitted level.

The last fitted coefficient is 1.86 (AR(1)) and 1.87 (RW1). The AR(1) forecast draws it back toward its long-run mean, to 1.66 at h = 1 and 1.15 at h = 6; the RW1 forecast holds the last level, 1.82 at h = 6. The same two fits at the default chain length give 1.25 and 1.87 at h = 6, and their 95% forecast intervals on the link scale behave as the two variance formulas say: under AR(1) the average width goes from 2.0 at h = 1 to 2.9 at h = 5 and stays there, under RW1 it goes from 1.9 to 3.4 at h = 6 and is still growing. Choose AR(1) when the effect is expected to settle and RW1 when a shift is expected to persist; lfo() below scores either choice on held-out periods.

To score forecasts on held-out periods, use the exact rolling-origin leave-future-out CV:

# refit before each of the last two periods and score that period
# seed the forecast draws inside lfo()
set.seed(1)
T_fit <- length(fit$YPM)
lfo_periods <- tail(seq_len(T_fit), max(1L, min(2L, T_fit - 2L)))
lfo_res <- lfo(
    fit,
    periods = lfo_periods,
    refit = TRUE,
    nscan = 500,
    burn = 100,
    odens = 5,
    verbose = FALSE
)
print(lfo_res)
#> 
#> ── Exact rolling-origin LFO ──
#> 
#> Periods evaluated: 5 and 6
#> Refit per leave-out: TRUE
#> Total elpd_lfo: -955.34
#>  period      elpd n_obs
#>       5 -481.7928   870
#>       6 -473.5436   870

lfo() refits the model on the periods before each origin and scores the held-out period with that period’s own covariates. elpd is the log predictive density summed over the n_obs observed dyads of the period; higher is better. nscan, burn and odens set the chain of each refit and, left out, the refits run at the length of fit. Period 5 scores -481.8 and period 6 -473.5 here; refits at the default length (nscan = 10000) score -481.8 and -473.1.

Compare the scores across origins. A sharp decline at the last origin points to a late change in the data-generating process, and detect_change_point(fit) then names the coefficient path and the period to inspect.

Probability-integral-transform (PIT) calibration

forecast_pit() complements lfo() by asking a sharper question: given the held-out outcomes, is the h-step posterior-predictive distribution itself well calibrated? The PIT uses the posterior predictive CDF, averaging the conditional distributions over matched posterior draws. For binary / Poisson / ordinal outcomes it randomises between the CDF just below and at the observation, using the Czado-Gneiting-Held PIT (Czado, Gneiting & Held 2009). A well-calibrated forecast produces PIT values that are Uniform(0, 1).

# fit on the first five periods, hold out the sixth
fit_train <- lame(
    Y[1:5],
    Xdyad = Xdyad[1:5],
    family = "binary",
    R = 0,
    dynamic_beta = "dyad",
    nscan = 100,
    burn = 30,
    odens = 5,
    verbose = FALSE
)
# seed the randomised PIT
set.seed(1)
pit <- forecast_pit(fit_train, y_future = list(Y[[6]]))
print(pit)
#> 
#> ── Forecast PIT calibration check ──
#> 
#> • Family: "binary"
#> • Forecast horizon: 1
#> • Held-out dyads: 870
#> • Kolmogorov-Smirnov statistic against Uniform(0, 1): 0.0519
#> • Global calibration p-value: unavailable for dependent network cells
#> • Fraction in [0.025, 0.975]: 0.933 (expect 0.95)
#> These are descriptive diagnostics. Network dependence requires a calibrated
#> joint reference for global inference.

plot(pit) renders the PIT values against the Uniform(0, 1) reference that a perfectly calibrated forecast would follow:

plot(pit)

PIT calibration plot for the held-out period: the empirical distribution of probability-integral-transform values compared with the Uniform(0,1) reference. Bars or a curve tracking the reference indicate calibration; a U-shape signals an over-confident (too-narrow) forecast and a central hump signals an under-confident (too-wide) one.

Bars that hug the uniform reference indicate a well-calibrated forecast. A U-shape (mass piling up in the tails) is the signature of an over-confident forecast whose intervals are too narrow; a central hump is the opposite, an under-confident forecast with intervals that are too wide. Read the plot for its overall shape; cover_95 summarises it in one number.

pit$cover_95 is the fraction of PIT values in [0.025, 0.975]; here it is 0.933, compared with 0.95 under marginal calibration (0.929 when fit_train runs at the default chain length). The KS statistic (0.052 here) is the largest distance between the empirical PIT distribution and the uniform reference. Network cells are dependent, so it is reported as a descriptive distance and pit$ks_p is NA. To pool evidence over several held-out periods, pass one matrix per horizon in y_future.

Comparing two fits with loo_compare()

When you have two candidate specifications, say a static-beta baseline and a dynamic_beta = "dyad" variant, loo::loo_compare() ranks them on out-of-sample predictive performance. Fit both with save_log_lik = TRUE and independent residuals (dcor = FALSE), so that each observed cell is one likelihood factor:

# two candidates with independent residuals and saved log-likelihoods
fit_static <- lame(
    Y,
    Xdyad = Xdyad,
    family = "binary",
    R = 0,
    dcor = FALSE,
    nscan = 5000,
    burn = 1000,
    odens = 25,
    save_log_lik = TRUE,
    verbose = FALSE
)
fit_dyn <- lame(
    Y,
    Xdyad = Xdyad,
    family = "binary",
    R = 0,
    dcor = FALSE,
    dynamic_beta = "dyad",
    nscan = 5000,
    burn = 1000,
    odens = 25,
    save_log_lik = TRUE,
    verbose = FALSE
)

# the best model has elpd_diff = 0; a named list labels the rows
cmp <- loo::loo_compare(list(
    fit_static = loo::loo(fit_static),
    fit_dyn = loo::loo(fit_dyn)
))
cmp
best <- rownames(cmp)[1]
runnerup <- rownames(cmp)[2]
diff_val <- abs(round(cmp[2, "elpd_diff"], 1))
se_val <- round(cmp[2, "se_diff"], 1)
ratio <- round(diff_val / se_val, 1)

How to read this output. The top row is the preferred model and has elpd_diff = 0. Each subsequent row gives the difference in expected log pointwise predictive density relative to that model and the standard error of the difference. A difference that is large relative to its standard error supports the preferred model; when the two are similar, the specifications are not distinguishable on this predictive criterion. The Pareto-\(k\) diagnostics in the loo output report which pointwise estimates are reliable. Run as written on this panel, the comparison prefers the dynamic model: the static model has elpd_diff = -22.2 with standard error 6.6, and every Pareto-\(k\) is below 0.5.

loo::loo() and loo::waic() read the saved scores of normal, binary and poisson fits, and of ordinal fits with ordinal_cutpoints = "explicit", whenever the observed cells are independent given the parameters: directed fits with dcor = FALSE, symmetric fits (each unordered dyad counted once) and bipartite fits. Fits with residual dyad correlation are compared on whole held-out periods with lfo() above; capped-binary, fixed-rank-nomination and rank-likelihood ordinal fits are compared with the posterior predictive checks of gof().

Memory-conscious long runs

If the in-memory log-likelihood matrix from save_log_lik = TRUE would exceed RAM, pass save_log_lik = "chunked" instead. Each chain writes its scores to disk as it samples. loo() reads the on-disk chunks as it reads the in-memory matrix, and read_log_lik(fit) returns them as a matrix:

fit_chk <- lame(
    Y,
    Xdyad = Xdyad,
    family = "binary",
    R = 0,
    dcor = FALSE,
    dynamic_beta = "dyad",
    nscan = 2000,
    burn = 500,
    odens = 5,
    save_log_lik = "chunked", # binary files, one per chunk of columns
    log_lik_chunk_size = 5000L,
    verbose = FALSE
)
loo::loo(fit_chk) # reads the chunks transparently

See also