Romeb Package: An Introduction

Introduction

Romeb implements robust median‑based Bayesian growth curve modeling that accommodate the three classical missing‑data mechanisms—MCAR, MAR and MNAR-and complete data, particularly beneficial when data are nonnormally distributed or include outliers. A detailed tutorial can be found in Tang & Tong (2025).

Function usage

The main interface is

Romeb(
  Missing_Type,      # "MNAR", "MAR", "MCAR", or "no missing"
  data,              # matrix / data frame
  time,              # Numeric vector of measurement times (e.g., c(0,1,2,3)).
  seed,              # reproducibility seed
  K      = 0,        # number of auxiliary variables
  chain  = 1,        # number of MCMC chains
  Niter  = 6000,     # iterations per chain
  burnIn = 3000      # burn‑in iterations
)

Arguments

Argument	Description
`Missing_Type`	Character string specifying the assumed missing‑data mechanism. One of `"MNAR"`, `"MAR"`, `"MCAR"`, `"no missing"`.
`data`	Matrix or data frame. If `K = 0`, all columns are treated as outcomes y; otherwise the first K columns are auxiliary variables and the next `Time` columns are outcomes.
`time`	Numeric vector of measurement times (e.g., c(0,1,2,3)).
`seed`	Integer seed ensuring reproducibility.
`K`	Non‑negative integer (default 0) giving the number of auxiliary variables.
`chain`	Number of parallel MCMC chains (default 1).
`Niter`	Total iterations per chain (default 6000).
`burnIn`	Iterations discarded as burn‑in (default 3000).

Output object

Running

Res <- Romeb(...)
print(Res)             # or simply type Res

returns a compact table with the posterior median, Geweke z‑scores, the 95% equal‑tail credible interval, and the 95% highest‑posterior‑density (HPD) interval for each monitored parameter.

Further elements can be accessed directly:

Element	Content
`Res$quantiles`	Posterior mean, SD, naïve and time‑series SEs, plus selected quantiles for every parameter after burn‑in.
`Res$geweke`	Vector of Geweke diagnostic z‑scores; values within ±1.96 indicate no evidence against lack of convergence.
`Res$credible_intervals`	95% equal‑tail credible intervals (2.5% & 97.5% quantiles).
`Res$hpd_intervals`	95% HPD intervals (shortest 95% region).
`Res$samps_full`	Complete `coda::mcmc.list` (including burn‑in). Inspect with `coda::traceplot(Res$samps_full[,'par[i]'])` for par[i] .

Quick examples

Below we illustrate a workflow.

set.seed(123)
Y <- matrix(rnorm(300*5), nrow = 300, ncol = 5)  # tiny complete data set
result_full <- Romeb("no missing", data = Y, time = c(0, 1, 2, 3, 4), seed = 123)

## Compiling model graph
##    Resolving undeclared variables
##    Allocating nodes
## Graph information:
##    Observed stochastic nodes: 1500
##    Unobserved stochastic nodes: 1804
##    Total graph size: 14432
## 
## Initializing model

print(result_full)

## Romeb GCM summary
## ==================
## 
## Posterior medians (50% quantiles):
##       par[1]       par[2]       par[3]       par[4]       par[5] 
##  0.010448884  0.008549176  0.171543659 -0.056132922  0.046658630 
## 
## Geweke test:
##     par[1]     par[2]     par[3]     par[4]     par[5] 
## -0.4295213  0.8113677  0.1677203 -0.2686674  0.5478983 
## 
## 95% credible intervals:
##               2.5%       97.5%
## par[1] -0.08372728  0.09968858
## par[2] -0.03105851  0.04984433
## par[3]  0.09413207  0.27461789
## par[4] -0.09527199 -0.02609691
## par[5]  0.03244535  0.06536003
## 
## 95% hpd intervals:
##            par[1]      par[2]     par[3]      par[4]     par[5]
## lower -0.08410641 -0.03164083 0.09030855 -0.09251707 0.03161084
## upper  0.09900967  0.04894551 0.26590422 -0.02422712 0.06381646
## 
## Use x$samps_full to access full MCMC samples, and coda::traceplot(x$samps_full[,'par[i]']) for the trace plot of par[i].

Note: par [1]: latent intercept, par [2]: latent slope: par [3]: variance of the latent intercept, par [4]: covariance between intercept and slope, par [5]: variance of the latent slope.

MCAR example

set.seed(456)
Y <- matrix(rnorm(300 * 5), nrow = 300)
miss <- runif(length(Y)) < 0.1       # 10% missing completely at random
Y[miss] <- NA
result_mcar <- Romeb("MCAR", data = Y, time = c(0, 1, 2, 3, 4), seed = 456)

MNAR with auxiliary variables

set.seed(789)
X  <- matrix(rnorm(300 * 2), 300, 2)     # two auxiliaries
Y  <- matrix(rnorm(300 * 5), 300, 5)
Data <- cbind(X, Y)
result_mnar <- Romeb("MNAR", data = Data, time = c(0, 1, 2, 3, 4), K = 2, seed = 789)

Inspecting convergence visually

# Uses the tiny example result_full from above
coda::traceplot(result_full$samps_full[,'par[1]'])

Trace plot for the first chain (par[1])

How to cite

Please cite the package as:

Tang,D.and Tong,X.(2025). Romeb: An R Package for Robust Median-Based Bayesian Growth Curve Modeling with Missing Data.

Bibliographic metadata can also be obtained via citation("Romeb").