Skip to contents
library(brms)
library(bicountbrms)
library(posterior)  # summaries of the posterior draws
library(dplyr)      # data manipulation
library(ggplot2)    # the figure

In a partially paired design, one source has missing values: source 2 records a count y_2 on every unit, and source 1 records a count y_1 on only some of the units. Interest usually lies in estimating the congruence and the source bias (see Getting started and The congruence–bias parameterisation), and in imputing the missing values of y_1 from all the information available, including the recorded values of y_2. Here, we simulate a partially paired design with two treatments and show how to obtain these estimates and imputations from a model fitted with binegbin_partialobs().

Structuring the data

The counts of the source that records every unit are the column y2, and the counts of the other source are the column y1. The observation indicator y1_obs is 1 on a paired row (where y_1 was recorded) and 0 on an unpaired row (where y_1 is missing), and is passed to the family with y_2 as vint(y2, y1_obs), in that order. An indicator other than 0 or 1 makes the sampler fail at initialisation, with a message in the chain output. The same check catches the two arguments of vint() written in reverse order. On an unpaired row, y1 can be any integer, because the likelihood term for that row does not depend on y1. We suggest -1, because it cannot be confused with a recorded count. y1 must not be NA on an unpaired row, because brms drops any row with a missing response before the likelihood is evaluated.

Simulated design

The design has 10 groups, such as sites, and two treatments, A and B, with 30 units for each combination of group and treatment. The midpoint M varies among groups. The congruence and the source bias differ between treatments, with f = 0.6 and \beta = 0 under treatment A, and f = 0.85 and \beta = 0.3 under treatment B, so that under treatment B source 1 records more exclusive events than source 2. Source 1 records y_1 on 18 of the 30 units (60%) in each combination of group and treatment, and every simulated y_1 is kept in y1_true, against which the imputation is checked.

set.seed(4)

# The midpoint M varies among the 10 groups: log M is normal with mean log(5)
# and SD 0.4.
groups <- data.frame(
  group = factor(1:10),
  logM  = rnorm(10, log(5), 0.4)
)

# The congruence f and the source bias beta differ between the two treatments.
treatments <- data.frame(
  treatment = factor(c("A", "B")),
  f         = c(0.6, 0.85),
  beta      = c(0, 0.3)
)

# 30 units for each combination of group and treatment, 600 units in all.
design <- merge(
  merge(groups, treatments),
  data.frame(unit = 1:30)
)

# The native parameters on each unit. Every overdispersion is 0.5.
dpars <- with(
  design,
  mfd_to_dpars(
    M         = exp(logM),
    f         = f,
    delta     = atanh(beta),
    kappas    = 0.5,
    kappaxone = 0.5,
    kappaxtwo = 0.5
  )
)

# rbicount() draws y1, y2, and the latent counts Ns, N1, and N2 on each unit.
# Source 1 records y1 on 18 of the 30 units, chosen at random, in each
# combination of group and treatment. On the other units, y1 is missing, as in
# real data, and the simulated value is kept in y1_true.
dat <- cbind(design, do.call(rbicount, c(list(n = nrow(design)), dpars))) |>
  group_by(group, treatment) |>
  mutate(
    y1_true = y1,
    y1      = if_else(
      sample(rep(c(TRUE, FALSE), times = c(18, 12))),
      y1,
      NA_integer_
    )
  ) |>
  ungroup() |>
  as.data.frame()

head(dat, 10)
#>    group     logM treatment   f beta unit y1 y2 Ns N1 N2 y1_true
#> 1      1 1.696140         A 0.6    0    1 NA  5  4  0  1       4
#> 2      2 1.392441         A 0.6    0    1  2  5  2  0  3       2
#> 3      3 1.965896         A 0.6    0    1 11  8  6  5  2      11
#> 4      4 1.847830         A 0.6    0    1  9  7  6  3  1       9
#> 5      5 2.263685         A 0.6    0    1 14 31 11  3 20      14
#> 6      6 1.885148         A 0.6    0    1  3  1  0  3  1       3
#> 7      7 1.096939         A 0.6    0    1 NA  3  3  1  0       4
#> 8      8 1.524180         A 0.6    0    1  9  3  3  6  0       9
#> 9      9 2.368054         A 0.6    0    1 13 11 10  3  1      13
#> 10    10 2.320183         A 0.6    0    1 NA 10  6  9  4      15

The model

Before the fit, the observation indicator is added and the missing values of y_1 are set to -1, as described in Structuring the data:

# Add the indicator y1_obs, 1 where source 1 recorded y1, and set the missing
# values of y1 to -1, so that brms keeps those rows
dat <- dat |>
  mutate(
    y1_obs = as.integer(!is.na(y1)),
    y1     = if_else(is.na(y1), -1L, y1)
  )

# The number of paired (y1_obs = 1) and unpaired (y1_obs = 0) units in each
# treatment
count(dat, treatment, y1_obs)
#>   treatment y1_obs   n
#> 1         A      0 120
#> 2         A      1 180
#> 3         B      0 120
#> 4         B      1 180

The model is written with bf_mfd() as in Getting started. The response is y1 | vint(y2, y1_obs), and the constructor and its Stan code are binegbin_partialobs() and binegbin_partialobs_stanvars(). The formula for logM includes a group-level effect of group, and the formulae for logitf and delta include treatment, with one coefficient for each treatment.

fit <- brm(
  bf_mfd(
    y1 | vint(y2, y1_obs),           # the indicator y1_obs follows y2
    logM      ~ 1 + (1 | group),     # M varies among groups
    logitf    ~ 0 + treatment,       # one congruence for each treatment
    delta     ~ 0 + treatment,       # one source bias for each treatment
    kappas    ~ 1,
    kappaxone ~ 1,
    kappaxtwo ~ 1
  ),
  family   = binegbin_partialobs(),
  stanvars = binegbin_partialobs_stanvars(),
  data     = dat,
  # The priors of Getting started, with a prior on the SD of the group-level
  # effect on logM
  prior    = c(
    prior(normal(1.5, 1), class = "b",  nlpar = "logM"),
    prior(normal(0, 1),   class = "sd", nlpar = "logM"),
    prior(normal(0, 1.5), class = "b",  nlpar = "logitf"),
    prior(normal(0, 0.5), class = "b",  nlpar = "delta"),
    prior(normal(0, 1),   class = "b",  nlpar = "kappas",    lb = 0),
    prior(normal(0, 1),   class = "b",  nlpar = "kappaxone", lb = 0),
    prior(normal(0, 1),   class = "b",  nlpar = "kappaxtwo", lb = 0)
  ),
  chains  = 4,
  cores   = 4,
  iter    = 3000,
  refresh = 0,
  backend = "cmdstanr",  # or "rstan"
  seed    = 1
)

An unpaired row contributes the probability of y_2 alone, which does not depend on the exclusive rate \lambda_1 or the overdispersion \kappa_{x1} of source 1. \lambda_1 and \kappa_{x1} are therefore identified by the paired rows alone. The congruence and the source bias are functions of \lambda_1, so neither can be estimated without paired rows. Where paired rows are few, the priors on logitf, delta, and kappaxone have the most influence on the posterior.

Including the unpaired rows is valid if y_1 is missing at random given the model: whether source 1 recorded a unit may depend on y_2 and on the predictors, but not otherwise on y_1.

Where there is no reason to expect the two exclusive components to differ in overdispersion, replacing kappaxone ~ 1 and kappaxtwo ~ 1 with the single formula kappax ~ 1 gives the two exclusive components one overdispersion, which is informed by the unpaired rows as well as the paired rows.

Recovery of the generating values

The table gives the posterior median and 95% credible interval of the congruence and the source bias under each treatment, the overdispersion of the exclusive component of source 1, and the SD of the group-level effect on \log M, with the values that generated the counts.

# The values that generated the counts
truth <- c(
  f_A       = 0.6,
  f_B       = 0.85,
  beta_A    = 0,
  beta_B    = 0.3,
  kappaxone = 0.5,
  sd_logM   = 0.4
)

# Each parameter on its natural scale, summarised and compared with its true
# value
recovery <- as_draws_df(fit) |>
  mutate_variables(
    f_A       = plogis(b_logitf_treatmentA),
    f_B       = plogis(b_logitf_treatmentB),
    beta_A    = tanh(b_delta_treatmentA),
    beta_B    = tanh(b_delta_treatmentB),
    kappaxone = b_kappaxone_Intercept,
    sd_logM   = sd_group__logM_Intercept
  ) |>
  subset_draws(variable = names(truth)) |>
  summarise_draws(
    est = median,
    ~ quantile2(.x, probs = c(0.025, 0.975))
  ) |>
  mutate(
    truth   = truth[variable],
    covered = truth >= q2.5 & truth <= q97.5
  )

knitr::kable(recovery, digits = 2)
variable est q2.5 q97.5 truth covered
f_A 0.58 0.49 0.66 0.60 TRUE
f_B 0.84 0.80 0.87 0.85 TRUE
beta_A -0.02 -0.10 0.06 0.00 TRUE
beta_B 0.31 0.20 0.43 0.30 TRUE
kappaxone 0.49 0.30 0.67 0.50 TRUE
sd_logM 0.45 0.29 0.82 0.40 TRUE

All six 95% intervals contain the true value.

Imputation of the missing values

posterior_predict_y1_given_y2() draws y_1 on each row from its distribution given the recorded y_2, once for each posterior draw. posterior_predict() draws y_1 from its marginal distribution, which depends on the predictors alone. The two predictions are compared below with the values of y_1 withheld from the model on the unpaired units.

# Draws of y1 given y2, and marginal draws of y1: one row per posterior draw and
# one column per unit
y1_given_y2 <- posterior_predict_y1_given_y2(fit)
y1_marginal <- posterior_predict(fit)

# The unpaired units, on which y1 was withheld from the model
unpaired <- which(dat$y1_obs == 0)

# The posterior median and 90% interval of a prediction on each unpaired unit
summarise_prediction <- function(draws, prediction) {
  data.frame(
    dat[unpaired, c("treatment", "y1_true")],
    prediction = prediction,
    est        = apply(draws[, unpaired], 2, median),
    q5         = apply(draws[, unpaired], 2, quantile, 0.05),
    q95        = apply(draws[, unpaired], 2, quantile, 0.95)
  )
}

imputed <- rbind(
  summarise_prediction(y1_given_y2, "given y2"),
  summarise_prediction(y1_marginal, "marginal")
)

# The root mean square error of the posterior median, the mean width of the 90%
# interval, and the proportion of withheld values inside the interval
imputation_summary <- imputed |>
  summarise(
    rmse     = sqrt(mean((est - y1_true)^2)),
    width    = mean(q95 - q5),
    coverage = mean(y1_true >= q5 & y1_true <= q95),
    .by      = c(treatment, prediction)
  ) |>
  arrange(treatment, prediction)

knitr::kable(imputation_summary, digits = 2)
treatment prediction rmse width coverage
A given y2 3.00 8.86 0.90
A marginal 3.79 11.90 0.92
B given y2 1.57 5.28 0.98
B marginal 4.16 13.35 0.94
imputed |>
  ggplot() +
  aes(
    x      = y1_true,
    y      = est,
    ymin   = q5,
    ymax   = q95,
    colour = treatment
  ) +
  geom_abline(
    slope     = 1,
    intercept = 0,
    linetype  = 2,
    colour    = "grey50"
  ) +
  geom_linerange(
    position = position_jitter(width = 0.25, height = 0, seed = 1),
    alpha    = 0.35
  ) +
  geom_point(
    position = position_jitter(width = 0.25, height = 0, seed = 1),
    size     = 1.2
  ) +
  facet_wrap(
    ~ prediction,
    labeller = as_labeller(
      c(
        "given y2" = "Predicted y1 conditional on y2",
        "marginal" = "Predicted y1 marginal"
      )
    )
  ) +
  scale_colour_manual(values = c(A = "#2D6A7F", B = "#C4622D")) +
  labs(
    x      = "Withheld y1",
    y      = NULL,
    colour = "Treatment"
  ) +
  theme_bw()
Two scatterplots of predicted against withheld y1 for the unpaired units, coloured by treatment. Given y2, the points lie along the line of equality. The marginal predictions lie in horizontal bands at the mean of y1 for each group and treatment.

Predictions of y_1 on the 240 unpaired units against the values of y_1 withheld from the model, given the recorded y_2 (left) and from the marginal distribution of y_1 (right). Each point is the posterior median of the prediction for one unit, and each vertical line its 90% interval. The points are jittered horizontally, because the withheld values are integers. The dashed line marks equality of the prediction and the withheld value.

The root mean square error of the posterior median of y_1 was smaller for the prediction given y_2 than for the marginal prediction: 3.0 against 3.8 under treatment A, and 1.6 against 4.2 under treatment B. The 90% intervals of the prediction given y_2 were also narrower than the 90% intervals of the marginal prediction, with a mean width of 8.9 against 11.9 under treatment A, and 5.3 against 13.3 under treatment B. The advantage of the prediction given y_2, in root mean square error and in interval width, is larger under treatment B, where the congruence is higher. The predictions given y_2 are pulled towards the mean of y_1 for each group and treatment, because the exclusive count of source 1 is independent of y_2.

Checks on the paired rows

On an unpaired row, y1 is not a recorded count. pp_check(), bayes_R2(), and loo_R2() compare predictions with y1, so for a partially paired fit, these three functions should be run with newdata restricted to the paired rows, as in pp_check(fit, newdata = subset(dat, y1_obs == 1)). loo() needs no such restriction, because the likelihood term for an unpaired row does not depend on y1. Prediction and imputation describes the prediction functions and the brms functions for post-processing.