rctme: Latent-outcome models for RCTs with differential measurement error

Description

Tools for fitting latent continuous-outcome models in randomized trials with a binary response process and a surrogate measurement subject to response-dependent error. The package provides EM estimators implemented with compiled C code for fast E-steps and observed-data log-likelihood evaluation, plus Bayesian Stan wrappers for corresponding models.

Author(s)

Maintainer: Bin Wang bwang831@gmail.com

Internal input validation

Description

Internal input validation

Usage

.validate_inputs(W, T, G)

Unblinding bias correction

Description

Unblinding bias correction

Usage

blindingMLE(
  W,
  T,
  G,
  Bayes = FALSE,
  ActiveOnly = TRUE,
  max_iter = 200,
  tol = 1e-06,
  grid_size = 81,
  grid_sd = 6,
  verbose = FALSE,
  compute_se = TRUE,
  hess_step = 1e-04,
  init = NULL,
  chains = 4,
  cores = chains,
  iter_warmup = 1000,
  iter_sampling = 1000,
  seed = 123,
  control = list(adapt_delta = 0.95, max_treedepth = 12),
  refresh = 0
)

Arguments

Value

An object of class 'rctme_fit'.

Examples

## Not run: 
if (requireNamespace("rstan", quietly = TRUE)) {
sigma = 1.2
sig.theta = 1.0
beta0 = 0
beta1 = 1
beta2 = 2
ntreat = nsham = 100
Tind  = c(rep(1, ntreat), rep(0,nsham))  #treatment group indicator
u1v   = rep(u1,ntreat)
u2v   = rep(u2,nsham)
uv    = c(u1v,u2v)
tauv  = uv - rep(u2, ntreat+nsham)
r = rnorm(ntreat + nsham, mean = 0, sd = sigma)
x = uv + r #actual endpoint outcome
q = 1/(1 + exp(-(beta0 + beta1*Tind + beta2*(tauv+r))))
bernGen = function(qq){rbinom(1,1,qq)}
I = sapply(q,bernGen)
rsham = rnorm(ntreat + nsham, mean = 0, sd = sig.theta)
w = x + (theta + rsham)*I
lm0 = lm(w~Tind)
tau1 = lm0$coef[2]; tau1
u12 = tapply(w,Tind,FUN=mean,na.rm=TRUE)
lm1 = lm(w~Tind+I)
tau2 = lm1$coef[2]; tau2
mydata <- data.frame(y=w,group=Tind,guess=I)
out5 <- blindingMLE(W = mydata$y, T = mydata$group, G = mydata$guess,
                    Bayes=FALSE,ActiveOnly=TRUE)
out5
}

## End(Not run)

EM fit with free 'delta0' and 'delta1'

Description

Fits the latent-outcome model

X|T \sim N(\mu_0 + \tau T, \sigma_X^2),

G|X,T \sim \text{Bernoulli}(\text{logit}^{-1}(\alpha_0 + \alpha_T T + \alpha_X (X-\mu_0))),

W|X,G=g \sim N(X + \delta_g, \sigma_{Wg}^2).

Usage

fit_rctme_em(
  W,
  T,
  G,
  max_iter = 200,
  tol = 1e-06,
  grid_size = 81,
  grid_sd = 6,
  verbose = FALSE,
  compute_se = TRUE,
  hess_step = 1e-04,
  init = NULL
)

Arguments

Details

The E-step and observed-data log-likelihood are evaluated in compiled C code for speed.

Value

An object of class 'rctme_fit'.

EM fit with 'delta0' fixed at 0

Description

Same model as [fit_rctme_em()] but with the measurement-error mean shift in the 'G = 0' group fixed at zero.

Usage

fit_rctme_em_delta0_fixed(
  W,
  T,
  G,
  max_iter = 200,
  tol = 1e-06,
  grid_size = 81,
  grid_sd = 6,
  verbose = FALSE,
  compute_se = TRUE,
  hess_step = 1e-04,
  init = NULL
)

Arguments

Value

An object of class 'rctme_fit'.

Bayesian Stan fit with free 'delta0' and 'delta1'

Description

Fits the unrestricted Bayesian latent-outcome model using 'rstan' and the Stan program bundled with the package.

Usage

fit_rctme_stan(
  W,
  T,
  G,
  chains = 4,
  cores = chains,
  iter_warmup = 1000,
  iter_sampling = 1000,
  seed = 123,
  control = list(adapt_delta = 0.95, max_treedepth = 12),
  refresh = 200
)

Arguments

Value

A list with components 'fit', 'summary', and 'stan_file'.

Bayesian Stan fit with 'delta0' fixed at 0

Description

Fits the restricted Bayesian latent-outcome model using 'rstan' and the Stan program bundled with the package.

Usage

fit_rctme_stan_delta0_fixed(
  W,
  T,
  G,
  chains = 4,
  cores = chains,
  iter_warmup = 1000,
  iter_sampling = 1000,
  seed = 123,
  control = list(adapt_delta = 0.95, max_treedepth = 12),
  refresh = 200
)

Arguments

Value

A list with components 'fit', 'summary', and 'stan_file'.

Locate a bundled Stan file

Description

Locate a bundled Stan file

Usage

rctme_stan_file(model = c("delta0_fixed", "delta_free"))

Arguments

Value

Path to the '.stan' file shipped with the package.

Simulate a small RCT measurement-error example dataset

Description

Generates a toy dataset for examples and testing under the latent outcome model implemented in this package.

Usage

simulate_rctme_data(
  n = 200,
  mu0 = 0.5,
  tau = 1,
  sigmaX = 1,
  alpha0 = -0.5,
  alphaT = 0.4,
  alphaX = 1,
  delta0 = 0,
  delta1 = -0.4,
  sigmaW0 = 0.8,
  sigmaW1 = 1,
  seed = NULL
)

Arguments

Value

A data frame with columns W, T, G, and latent X.