test.GRID.h2o.xgboost.10Kdata <- function() {
reqxgb <- requireNamespace("xgboost", quietly = TRUE)
reqh2o <- requireNamespace("h2o", quietly = TRUE)
if (!reqxgb || !reqh2o) return(TRUE)
## -----------------------------------------------------------------------
## ****************************** IMPORTANT ******************************
## -----------------------------------------------------------------------
## Make sure to always install the latest versions of these packages: "gridisl" and "stremr":
# devtools::install_github('osofr/gridisl')
# devtools::install_github('osofr/stremr')
## -----------------------------------------------------------------------
## -----------------------------------------------------------------------
## Analyses by intervention
## **** makes it easier to read the individual analyses ****
## -----------------------------------------------------------------------
library("stremr")
library("magrittr")
library("data.table")
library("testthat")
library("ggplot2")
library("tibble")
library("tidyr")
library("purrr")
library("dplyr")
library("sl3")
data.table::setDTthreads(1)
options(stremr.verbose = FALSE)
options(sl3.verbose = FALSE)
options(gridisl.verbose = FALSE)
# set_all_stremr_options(estimator = "speedglm__glm")
data(OdatDT_10K)
# OdatDT_10K[is.na(N), "N" := 0][is.na(CVD), "CVD" := 0][is.na(highA1c), "highA1c" := 0][is.na(TI), "TI" := 0]
Odat_DT <- OdatDT_10K
# select only the first 1,000 IDs
Odat_DT <- Odat_DT[ID %in% (1:100), ]
setkeyv(Odat_DT, cols = c("ID", "t"))
## -----------------------------------------------------------------------
## Define some summaries (lags C[t-1], A[t-1], N[t-1])
## -----------------------------------------------------------------------
ID <- "ID"; t <- "t"; TRT <- "TI"; I <- "highA1c"; outcome <- "Y.tplus1";
lagnodes <- c("C", "TI", "N")
newVarnames <- paste0(lagnodes, ".tminus1")
Odat_DT[, (newVarnames) := shift(.SD, n=1L, fill=0L, type="lag"), by=ID, .SDcols=(lagnodes)]
# indicator that the person has never been on treatment up to current t
Odat_DT[, ("barTIm1eq0") := as.integer(c(0, cumsum(get(TRT))[-.N]) %in% 0), by = eval(ID)]
Odat_DT[, ("lastNat1.factor") := as.factor(lastNat1)]
## ------------------------------------------------------------------
## Propensity score models for Treatment, Censoring & Monitoring
## ------------------------------------------------------------------
gform_TRT <- "TI ~ CVD + highA1c + N.tminus1"
stratify_TRT <- list(
TI=c("t == 0L", # MODEL TI AT t=0
"(t > 0L) & (N.tminus1 == 1L) & (barTIm1eq0 == 1L)", # MODEL TRT INITATION WHEN MONITORED
"(t > 0L) & (N.tminus1 == 0L) & (barTIm1eq0 == 1L)", # MODEL TRT INITATION WHEN NOT MONITORED
"(t > 0L) & (barTIm1eq0 == 0L)" # MODEL TRT CONTINUATION (BOTH MONITORED AND NOT MONITORED)
))
gform_CENS <- c("C ~ highA1c + t")
gform_MONITOR <- "N ~ 1"
## ------------------------------------------------------------
## **** As a first step define a grid of all possible parameter combinations (for all estimators)
## **** This dataset is to be saved and will be later merged in with all analysis
## ------------------------------------------------------------
trunc_IPW <- 10
# tvals <- 0:8
tvals <- 0:2
tmax <- 13
## number of folds for CV:
nfolds <- 3
tbreaks = c(1:8,11,14)-1
## This dataset defines all parameters that we like to vary in this analysis (including different interventions)
## That is, each row of this dataset corresponds with a single analysis, for one intervention of interest.
analysis <- list(intervened_TRT = c("gTI.dlow", "gTI.dhigh"),
trunc_wt = c(FALSE, TRUE),
stratifyQ_by_rule = c(TRUE, FALSE)) %>%
cross_df() %>%
arrange(stratifyQ_by_rule) %>%
mutate(nfolds = as.integer(nfolds)) %>%
mutate(trunc_MSM = map_dbl(trunc_wt, ~ ifelse(.x, trunc_IPW, Inf))) %>%
mutate(trunc_TMLE = trunc_MSM*10)
## ----------------------------------------------------------------
## IMPORT DATA
## ----------------------------------------------------------------
# library("h2o")
# h2o::h2o.init(nthreads = 2)
OData <- stremr::importData(Odat_DT, ID = "ID", t = "t", covars = c("highA1c", "CVD", "lastNat1", "lastNat1.factor"), CENS = "C", TRT = "TI", MONITOR = "N", OUTCOME = outcome)
OData <- define_CVfolds(OData, nfolds = 3, fold_column = "fold_ID", seed = 12345)
OData$dat.sVar[]
OData$fold_column <- NULL
OData$nfolds <- NULL
fold_column <- "fold_ID"
## ----------------------------------------------------------------
## Define ensemble of models for fitting propensity scores (g).
## ----------------------------------------------------------------
## This example uses a discrete SuperLearner: best model will be selected on the basis of CV-MSE.
## Use cross-validation to select best model for g (set 'fit_method_g <- "none"' to just fit a single model w/out CV)
fit_method_g <- "cv"
# fit_method_g <- "none"
## Note that 'interactions' CANNOT be used with h2o (for now).
## The only learners that allow interactions are: "glm" ,"speedglm", "xgboost".
# models_g <-
# defModel(estimator = "h2o__glm", family = "binomial",
# # lambda_search = FALSE,
# # nlambdas = 5,
# param_grid = list(
# alpha = 0
# # alpha = c(0.5)
# ))
# defModel(estimator = "xgboost__glm", family = "binomial", nthread = 1)
# nrounds = 100,
# early_stopping_rounds = 2,
# interactions = list(c("CVD", "highA1c")))
# +
# defModel(estimator = "speedglm__glm", family = "quasibinomial")
lrn_glm <- Lrnr_glm_fast$new()
lrn_glm_sm <- Lrnr_glm_fast$new(covariates = c("CVD"))
# lrn_glmnet_binom <- Lrnr_pkg_SuperLearner$new("SL.glmnet", family = binomial())
lrn_glmnet_binom <- Lrnr_glmnet$new(nlambda = 5)
# lrn_glmnet_gaus <- Lrnr_pkg_SuperLearner$new("SL.glmnet", family = "gaussian")
lrn_glmnet_gaus <- Lrnr_glmnet$new(nlambda = 5)
sl <- Lrnr_sl$new(learners = Stack$new(lrn_glm, lrn_glm_sm, lrn_glmnet_binom),
metalearner = Lrnr_nnls$new())
# models_g <- lrn_glm
models_g <- sl
## ----------------------------------------------------------------
## AN EXAMPLE OF A GIANT GRID OF MODELS.
## This will perform an extensive search of model hyper-parameters for fitting;
## Define grids and do random grid search, as shown below.
## USE THIS IN CASE OF UNLIMITED COMPUTATIONAL RESOURCES (OR VERY SMALL DATA).
## OTHERWISE LIMIT THE NUMBER OF RANDOM MODELS BEING DRAWN FROM THE GRID.
## ----------------------------------------------------------------
# h2o_GBM_hyper <- list( # max_depth = c(3:10, 15),
# max_depth = c(seq(3, 19, 4), 25),
# # ntrees = c(500),
# ntrees = c(100),
# learn_rate = c(.05, .1), # 0.01, 0.03, , 0.005,
# # sample_rate = seq(0.2, 1, 0.05),
# sample_rate = c(.5, .75, 1),
# # col_sample_rate = seq(0.1, 1, 0.05),
# col_sample_rate_per_tree = c(.4, .6, .8, 1),
# balance_classes = c(TRUE)
# # col_sample_rate_change_per_level = seq(0.9, 1.1, 0.01),
# # nbins = 2^seq(4,10,1),
# # nbins_cats = 2^seq(4,12,1),
# # min_split_improvement = c(0,1e-8,1e-6,1e-4),
# # histogram_type = c("UniformAdaptive","QuantilesGlobal","RoundRobin"),
# )
# # class_sampling_factors, ## Desired over/under-sampling ratios per class (in lexicographic order).
# # max_after_balance_size ## Maximum relative size of the training data after balancing class counts (Default is 5)
# # categorical_encoding = c("AUTO", "Enum", "OneHotInternal", "OneHotExplicit", "Binary", "Eigen")
# RF_hyper <- list(
# mtries = -1,
# # mtries = c(-1, seq(4, length(covars), by = 10)),
# max_depth = c(3,10,15,20),
# ntrees = c(100),
# # ntrees = c(500),
# # ntrees = c(50, 100, 200, 500, 1000),
# sample_rate = c(0.632, seq(0.7, 1, 0.1)),
# col_sample_rate_per_tree = seq(0.6, 1, 0.1),
# balance_classes = TRUE
# # col_sample_rate_change_per_level = seq(0.9, 1.1, 0.01)
# # min_rows = 2^seq(0,log2(nrow(inputDT))-1,1),
# # nbins = 2^seq(4,10,1), # n bins for split-finding for continuous and integer columns
# # nbins_cats = 2^seq(4,12,1), # n bins for split-finding for categorical columns
# # min_split_improvement = c(0,1e-8,1e-6,1e-4), # min req rel error improvement thres for a split to happen
# # histogram_type = c("UniformAdaptive", "Random", "QuantilesGlobal","RoundRobin")
# )
# models_g <<-
# defModel(estimator = "xgboost__gbm", family = "binomial",
# nrounds = 200, # nrounds = 500,
# early_stopping_rounds = 3,
# interactions = list(c("CVD", "highA1c")),
# learning_rate = .1,
# max_depth = 3,
# gamma = .5,
# colsample_bytree = 0.8,
# subsample = 0.8,
# lambda = 2,
# alpha = 0.5,
# max_delta_step = 2) +
# defModel(estimator = "xgboost__drf", family = "binomial",
# nrounds = 200, # nrounds = 500,
# early_stopping_rounds = 3,
# interactions = list(c("CVD", "highA1c")),
# learning_rate = .1,
# max_depth = 3,
# gamma = .5,
# # colsample_bytree = 0.8,
# # subsample = 0.8,
# lambda = 2,
# alpha = 0.5,
# max_delta_step = 2) +
# defModel(estimator = "xgboost__gbm",
# family = "binomial",
# search_criteria = list(strategy = "RandomDiscrete", max_models = 100),
# seed = 23,
# nrounds = 200, # nrounds = 500,
# early_stopping_rounds = 3,
# interactions = list(c("CVD", "highA1c")),
# param_grid = list(
# learning_rate = c(.05, .1, .3), # .05,
# max_depth = c(seq(3, 19, 4), 25),
# min_child_weight = c(1, 3, 5, 7),
# gamma = c(.0, .05, seq(.1, .9, by=.2), 1),
# # colsample_bytree = c(.4, .6, .8, 1),
# subsample = c(.5, .75, 1),
# lambda = c(.1, .5, 1, 2, 5), # lambda = c(1,2,5),
# alpha = c(0, .1, .5),
# ## Maximum delta step we allow each tree’s weight estimation to be.
# ## If the value is set to 0, it means there is no constraint.
# ## If it is set to a positive value, it can help making the update step more conservative.
# ## Might help in logistic regression when class is extremely imbalanced.
# ## Set it to value of 1-10 to help control the update
# max_delta_step = c(0, 1, 2, 5, 10)
# )
# ) +
# defModel(estimator = "h2o__glm", family = "binomial", alpha = 0, lambda = 0, lambda_search = FALSE) +
# defModel(estimator = "h2o__glm", family = "binomial",
# nlambdas = 5, lambda_search = TRUE,
# param_grid = list(
# alpha = c(0, .5, 1)
# )) +
# defModel(estimator = "h2o__randomForest",
# distribution = "bernoulli",
# seed = 23,
# search_criteria = list(
# strategy = "RandomDiscrete", max_models = 100, max_runtime_secs = 3*60*60),
# param_grid = RF_hyper,
# binomial_double_trees = TRUE,
# stopping_metric = "MSE", stopping_rounds = 3, score_tree_interval = 1) +
# defModel(estimator = "h2o__gbm",
# distribution = "bernoulli",
# seed = 23,
# search_criteria = list(
# strategy = "RandomDiscrete", max_models = 100, max_runtime_secs = 3*60*60), # stopping_rounds = 5
# param_grid = h2o_GBM_hyper,
# stopping_metric = "MSE", stopping_rounds = 3, score_tree_interval = 1)
## ------------------------------------------------------------------------
## Define models for iterative G-COMP (Q) -- PARAMETRIC LOGISTIC REGRESSION
## ------------------------------------------------------------------------
## regression formulas for Q's:
Qforms <- rep.int("Qkplus1 ~ CVD + highA1c + N + lastNat1 + TI + TI.tminus1", (max(tvals)+1))
## no cross-validation model selection, just fit a single model specified below
fit_method_Q <- "none"
## Use speedglm to fit all Q.
## NOTE: it is currently not possible to use fit_method_Q <- "cv" with speedglm or glm.
## To perform cross-validation with GLM use 'estimator="h2o__glm"' or 'estimator="xgboost__glm"'
# models_Q <- defModel(estimator = "speedglm__glm", family = "quasibinomial")
models_Q <- Lrnr_glm_fast$new()
## ------------------------------------------------------------------------
## Alternative specifications of Q models for iterative G-COMP (Q) -- NONPARAMETRIC REGRESSION (GBM) + REGULARIZED GLM
## ------------------------------------------------------------------------
# ## This example uses a discrete SuperLearner: best model will be selected on the basis of CV-MSE.
# fit_method_Q <- "cv"
# models_Q <<-
# defModel(estimator = "xgboost__gbm", family = "binomial",
# nrounds = 200,
# early_stopping_rounds = 3,
# interactions = list(c("TI", "highA1c"), c("TI", "CVD"), c("TI", "lastNat1")),
# learning_rate = .1,
# max_depth = 5,
# gamma = .5,
# colsample_bytree = 0.8,
# subsample = 0.8,
# lambda = 2,
# alpha = 0.5,
# max_delta_step = 2) +
# defModel(estimator = "xgboost__glm",
# family = "quasibinomial",
# seed = 23,
# nrounds = 500,
# early_stopping_rounds = 5,
# interactions = list(c("TI", "highA1c"), c("TI", "CVD"), c("TI", "lastNat1")),
# param_grid = list(
# alpha = c(.0, 0.5, 1.0),
# lambda = c(.01, .1, .5, .9, 1.5, 5)
# )
# )
## ----------------------------------------------------------------
## Fit propensity score models.
## We are using the same model ensemble defined in models_g for censoring, treatment and monitoring mechanisms.
## ----------------------------------------------------------------
OData <- fitPropensity(OData,
gform_CENS = gform_CENS, gform_TRT = gform_TRT,
stratify_TRT = stratify_TRT, gform_MONITOR = gform_MONITOR,
models_CENS = models_g,
models_TRT = models_g,
models_MONITOR = Lrnr_glm_fast$new(),
fit_method = fit_method_g,
fold_column = fold_column)
## ------------------------------------------------------------
## RUN IPW ANALYSES
## **** For each individual analysis do filter()/subset()/etc to create a grid of parameters specific to given estimator
## ------------------------------------------------------------
IPW <- analysis %>%
rename(trunc_weight = trunc_MSM) %>%
distinct(intervened_TRT, trunc_weight) %>%
group_by(intervened_TRT) %>%
mutate(wts_data = map(first(intervened_TRT), getIPWeights, OData = OData, tmax = tmax)) %>%
## save the tables of weights summaries (sep for each regimen)
mutate(wts_tabs = map(wts_data,
~ get_wtsummary(.x, cutoffs = c(0, 0.5, 1, 10, 20, 30, 40, 50, 100, 150), by.rule = TRUE))) %>%
## save the tables with number at risk / following each rule (sep for each regimen)
mutate(FUPtimes_tabs = map(wts_data,
~ get_FUPtimes(.x, IDnode = ID, tnode = t))) %>%
ungroup() %>%
## IPW-Adjusted KM (Non-Parametric or Saturated MSM):
mutate(NPMSM = map2(wts_data, trunc_weight,
~ survNPMSM(wts_data = .x,
trunc_weights = .y,
OData = OData))) %>%
mutate(NPMSM = map(NPMSM, "estimates")) %>%
## Crude MSM for hazard (w/out IPW):
mutate(MSM.crude = map(wts_data,
~ survMSM(wts_data = .x,
OData = OData,
tbreaks = tbreaks,
use_weights = FALSE,
glm_package = "speedglm"))) %>%
mutate(MSM.crude = map(MSM.crude, "estimates")) %>%
## IPW-MSM for hazard (smoothing over time-intervals in tbreaks):
mutate(MSM = map2(wts_data, trunc_weight,
~ survMSM(wts_data = .x,
trunc_weights = .y,
OData = OData,
tbreaks = tbreaks,
glm_package = "speedglm"))) %>%
mutate(MSM = map(MSM, "estimates")) %>%
rename(trunc_MSM = trunc_weight)
## save IPW tables (will be later merged with main results dataset)
IPWtabs <- analysis %>%
left_join(IPW) %>%
distinct(intervened_TRT, trunc_MSM, wts_tabs, FUPtimes_tabs) %>%
nest(intervened_TRT, wts_tabs, FUPtimes_tabs, .key = "IPWtabs")
IPW <- IPW %>% select(-wts_data, -wts_tabs, -FUPtimes_tabs)
## ------------------------------------------------------------
## GCOMP ANALYSIS
## ------------------------------------------------------------
GCOMP <-analysis %>%
distinct(intervened_TRT, stratifyQ_by_rule) %>%
mutate(GCOMP = map2(intervened_TRT, stratifyQ_by_rule,
~ fit_GCOMP(intervened_TRT = .x,
stratifyQ_by_rule = .y,
tvals = tvals,
OData = OData,
models = models_Q,
Qforms = Qforms,
fit_method = fit_method_Q,
fold_column = fold_column))) %>%
mutate(GCOMP = map(GCOMP, "estimates"))
## ------------------------------------------------------------
## TMLE ANALYSIS
## ------------------------------------------------------------
TMLE <- CVTMLE <- analysis %>%
rename(trunc_weight = trunc_TMLE) %>%
distinct(intervened_TRT, stratifyQ_by_rule, trunc_weight)
TMLE <- TMLE %>%
mutate(TMLE = pmap(TMLE, fit_TMLE,
tvals = tvals,
OData = OData,
models = models_Q,
Qforms = Qforms,
fit_method = fit_method_Q,
fold_column = fold_column)) %>%
mutate(TMLE = map(TMLE, "estimates")) %>%
rename(trunc_TMLE = trunc_weight)
# CVTMLE <- CVTMLE %>%
# mutate(CVTMLE = pmap(CVTMLE, fit_CVTMLE,
# tvals = tvals,
# OData = OData,
# models = models_Q,
# Qforms = Qforms,
# fit_method = fit_method_Q,
# fold_column = fold_column)) %>%
# mutate(CVTMLE = map(CVTMLE, "estimates")) %>%
# rename(trunc_TMLE = trunc_weight)
## ------------------------------------------------------------
## COMBINE ALL ANALYSES INTO A SINGLE DATASET
## ------------------------------------------------------------
results <- analysis %>%
left_join(IPW) %>%
left_join(GCOMP) %>%
left_join(TMLE)
# %>%
# left_join(CVTMLE)
## Nest each estimator by treatment regimen (we now only show the main analysis rows)
results <- results %>%
# nest(intervened_TRT, NPMSM, MSM.crude, MSM, .key = "estimates")
nest(intervened_TRT, NPMSM, MSM.crude, MSM, GCOMP, TMLE, .key = "estimates")
## Calculate RDs (contrasting all interventions, for each analysis row & estimator).
## The RDs data no longer needs the intervened_TRT column
results <- results %>%
mutate(RDs =
map(estimates,
~ select(.x, -intervened_TRT) %>%
map(~ get_RDs(.x)) %>%
as_tibble()
))
## ------------------------------------------------------------
## Uncomment and run this to remove the individual EIC estimates from MSM and TMLE estimates.
## This is useful if saving results to a file. The EIC take up a lot of space, hence removing them
## will significantly reduce the final file size.
## ------------------------------------------------------------
## Clean up by removing the subject-level IC estimates for EVERY SINGLE ESTIMATE / ANALYSIS
## WARNING: THIS IS A SIDE-EFFECT FUNCTION!
# res <- results[["estimates"]] %>%
# map(
# ~ select(.x, -intervened_TRT) %>%
# map(
# ~ map(.x,
# ~ suppressWarnings(.x[, ("IC.St") := NULL]))))
# rm(res)
## ------------------------------------------------------------
## Add models used for g and Q. Create the final analysis file.
## Add IPWtabs
## ------------------------------------------------------------
results <- results %>%
left_join(IPWtabs) %>%
mutate(fit_method_g = fit_method_g) %>%
mutate(fit_method_Q = fit_method_Q) %>%
mutate(models_g = map(fit_method_g, ~ models_g)) %>%
mutate(models_Q = map(fit_method_Q, ~ models_Q))
## ------------------------------------------------------------
## VARIOUS WAYS OF PLOTTING SURVIVAL CURVES
## ------------------------------------------------------------
ests <- "TMLE"
SURVplot <- results[1, ][["estimates"]][[1]][[ests]] %>%
ggsurv %>%
print
results[["estimates"]]
# GCOMP <-GCOMP %>%
# mutate(plotGCOMP = map(GCOMP, ~ ggsurv(estimates = .x))) %>%
## FLATTEN THE results data (long format)
## then add a new column with ggplot survival plot objects for THESE 3 estimators;
ests <- c("MSM", "GCOMP", "TMLE")
longSURV <- results %>%
select(trunc_wt, stratifyQ_by_rule, trunc_MSM, trunc_TMLE, estimates) %>%
unnest(estimates) %>%
gather(key = est, value = estimates, NPMSM, MSM.crude, MSM, GCOMP, TMLE) %>%
filter(est %in% ests) %>%
select(-intervened_TRT) %>%
nest(estimates, .key = "estimates") %>%
mutate(SURVplot = map(estimates, ~ ggsurv(.x)))
## Visualize all survival curves at once with a single interactive trelliscope panel
## (install via: devtools::install_github("hafen/trelliscopejs"))
reqtrell <- requireNamespace("trelliscopejs", quietly = TRUE)
if (reqtrell && FALSE) {
longSURV %>%
trelliscopejs::trelliscope(name = "test", panel_col = "SURVplot")
longSURV
}
## ------------------------------------------------------------
## VARIOUS WAYS OF PLOTTING RD TABLEs
## ------------------------------------------------------------
## THE TABLE OF RDs FOR TMLE:
results %>% filter(trunc_wt == TRUE, stratifyQ_by_rule == TRUE) %>% select(RDs) %>% unnest(RDs) %>% select(TMLE) %>% unnest(TMLE)
## PLOT RD FOR TMLE FOR SPECIFIC SCENARIO:
ests <- "TMLE"
RDplot <- results[["RDs"]][[1]][[ests]][[1]] %>%
ggRD(t_int_sel = 1:5) %>%
print
## GENERERATE RD PLOTS across all scenarios for these two estimators:
ests <- c("MSM", "TMLE")
longRDs <- results %>%
select(trunc_wt, stratifyQ_by_rule, trunc_MSM, trunc_TMLE, RDs) %>%
unnest(RDs) %>%
gather(key = est, value = RDs, NPMSM, MSM.crude, MSM, GCOMP, TMLE) %>%
filter(est %in% ests) %>%
mutate(RDplot = map(RDs, ~ ggRD(.x)))
reqtrell <- requireNamespace("trelliscopejs", quietly = TRUE)
if (reqtrell && FALSE) {
longRDs %>%
trelliscopejs::trelliscope(name = "test", panel_col = "RDplot")
longRDs
}
# h2o::h2o.shutdown(prompt = FALSE)
}
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.