Last updated: 2022-11-04
Checks: 6 1
Knit directory: spanish_data/
This reproducible R Markdown analysis was created with workflowr (version 1.6.2). The Checks tab describes the reproducibility checks that were applied when the results were created. The Past versions tab lists the development history.
The R Markdown file has unstaged changes. To know which version of
the R Markdown file created these results, you’ll want to first commit
it to the Git repo. If you’re still working on the analysis, you can
ignore this warning. When you’re finished, you can run
wflow_publish to commit the R Markdown file and build the
HTML.
Great job! The global environment was empty. Objects defined in the global environment can affect the analysis in your R Markdown file in unknown ways. For reproduciblity it’s best to always run the code in an empty environment.
The command set.seed(20210329) was run prior to running
the code in the R Markdown file. Setting a seed ensures that any results
that rely on randomness, e.g. subsampling or permutations, are
reproducible.
Great job! Recording the operating system, R version, and package versions is critical for reproducibility.
Nice! There were no cached chunks for this analysis, so you can be confident that you successfully produced the results during this run.
Great job! Using relative paths to the files within your workflowr project makes it easier to run your code on other machines.
Great! You are using Git for version control. Tracking code development and connecting the code version to the results is critical for reproducibility.
The results in this page were generated with repository version 4e764db. See the Past versions tab to see a history of the changes made to the R Markdown and HTML files.
Note that you need to be careful to ensure that all relevant files for
the analysis have been committed to Git prior to generating the results
(you can use wflow_publish or
wflow_git_commit). workflowr only checks the R Markdown
file, but you know if there are other scripts or data files that it
depends on. Below is the status of the Git repository when the results
were generated:
Ignored files:
Ignored: .Rproj.user/
Ignored: manuscript/
Ignored: manuscript2/
Ignored: output/ap_iml.RDS
Ignored: output/ap_models.RDS
Ignored: output/ap_result.RDS
Ignored: output/dis_iml.RDS
Ignored: output/dis_models.RDS
Ignored: output/dis_result.RDS
Ignored: output/np_iml.RDS
Ignored: output/np_models.RDS
Ignored: output/np_result.RDS
Ignored: output/report_files/
Ignored: output/resample_perf.RDS
Ignored: output/resampling_models.RDS
Ignored: output/stepwise_performance.RDS
Ignored: output/stepwise_performance2.RDS
Untracked files:
Untracked: analysis/7-iml-_2.Rmd
Untracked: analysis/8-descriptive.Rmd
Untracked: code/add_models.R
Unstaged changes:
Modified: analysis/7-iml _2.Rmd
Modified: analysis/index.Rmd
Note that any generated files, e.g. HTML, png, CSS, etc., are not included in this status report because it is ok for generated content to have uncommitted changes.
These are the previous versions of the repository in which changes were
made to the R Markdown (analysis/7-iml _2.Rmd) and HTML
(docs/7-iml _2.html) files. If you’ve configured a remote
Git repository (see ?wflow_git_remote), click on the
hyperlinks in the table below to view the files as they were in that
past version.
| File | Version | Author | Date | Message |
|---|---|---|---|---|
| Rmd | 4e764db | Bernard | 2022-09-16 | added interpretable codes |
# Helper
library(tidyverse)
library(data.table)
library (cowplot)
library (officer)
library (flextable)
library (dataPreparation)
# ML
library (MASS)# Step AIC
library (ncvreg)
library (abess)# Best subset regression
library (glmnet) # Lasso
library (mboost)
library (earth) # mars
library (SignifReg)
# Parallel
library (future)
# P value custom
p_extract <- function (x) {
p <- x[2, "Pr(>Chi)"]
return (p)
}
performance <- function (y_pred, y_true) {
c("Accuracy" = MLmetrics::Accuracy(y_pred, y_true),
"AUC" = MLmetrics::AUC(y_pred, y_true),
"Precision" = MLmetrics::Precision(y_true, y_pred, positive = NULL),
"Sensitivity" = MLmetrics::Sensitivity(y_true, y_pred, positive = NULL),
"Specificity" = MLmetrics::Specificity(y_true, y_pred, positive = NULL))
}# bmr <- readRDS("output/resampling_models.RDS")
#
# bmr2 <- as.data.table(bmr) %>%
# mutate (Model = mlr3misc::map (learner, "model"))
dat <- readRDS("output/df.RDS")
np_train <- dat$df_list$np$train_imp[,-1] %>% dplyr::select (-matches ("clinic"))
np_test <- dat$df_list$np$test_imp[,-1]%>% dplyr::select (-matches ("clinic"))
np_dat <- bind_rows(np_train, np_test)
ap_train <- dat$df_list$ap$train_imp[,-1] %>% dplyr::select (-matches ("clinic"))
ap_test <- dat$df_list$ap$test_imp[,-1]%>% dplyr::select (-matches ("clinic"))
ap_dat <- bind_rows(ap_train, ap_test)
dis_train <- dat$df_list$dis$train_imp[,-1]%>% dplyr::select (-matches ("clinic"))
dis_test <- dat$df_list$dis$test_imp[,-1]%>% dplyr::select (-matches ("clinic"))
dis_dat <- bind_rows(dis_train, dis_test)scales <- build_scales(np_train)
np_train <- fast_scale(np_train, scales = scales)
np_test <- fast_scale(np_test, scales = scales)o <- np_test$outcome
# P value
df <- np_train %>%
mutate (outcome = as.numeric(outcome) - 1)
uni_names <- df %>%
dplyr::select(-outcome) %>%
map(~glm(df$outcome ~ .x, data = df, family = binomial())) %>%
map(anova, test = "Chisq") %>%
map_dbl(p_extract) %>%
"<" (0.1) %>%
which() %>%
names()
form <- paste0("outcome~", paste0(uni_names, collapse = "+"))
fullmodel <- glm(form, data = df, family = binomial)
nullmodel <- glm(outcome ~1, data = df, family = binomial)
scope = list(lower=formula(nullmodel ),upper=formula(fullmodel))
np_m1 = SignifReg(fullmodel,
scope=scope,
alpha = 0.05,
direction = "both",
criterion = "p-value",
adjust.method = "none",
trace=FALSE)
p_m1 <- predict (np_m1, type = "response", newdata = np_test %>% dplyr::select (-outcome))
p_m1 <- ifelse (p_m1 > 0.5, 1, 0)
res_m1 <- performance (y_pred = p_m1,
y_true = o)
# Step AIC
full_model <- glm (outcome ~ ., data = df, family = binomial())
np_m2 <- stepAIC (full_model,
direction = "both")
summary (np_m2 )
p_m2 <- predict (np_m2, type = "response", newdata = np_test %>% dplyr::select (-outcome))
p_m2 <- ifelse (p_m2 > 0.5, 1, 0)
res_m2 <- performance (y_pred = p_m2,
y_true = o)
# Best subset regression
X_train <- model.matrix(outcome ~., data = df)[,-1]
Y_train <- as.numeric (df$outcome)
X_test <- model.matrix(outcome ~., data = np_test)[,-1]
Y_test <- as.numeric (np_test$outcome) -1
np_m3 <- abess(x = X_train,
y = Y_train,
family = "binomial",
tune.type = "cv"
)
best_np_m3 <- np_m3 [["best.size"]]
print(best_np_m3 )
coef(np_m3 , support.size = best_np_m3 , sparse = FALSE)
p_m3 <- predict (np_m3, type = "response", newx = X_test, support.size = best_np_m3)
p_m3<- as.numeric (ifelse (p_m3 > 0.5, 1, 0))
res_m3 <- performance (y_pred = p_m3,
y_true = Y_test)
# Lasso
np_m4 <- cv.ncvreg(X_train,
Y_train,
penalty = "lasso",
family = "binomial")
plot(np_m4)
coef_np_m4 <- coef(np_m4, s = "lambda.min")[coef(np_m4, s = "lambda.min") != 0]
p_m4 <- predict (np_m4, type = "response", X = X_test, lambda = np_m4$lambda.min)
p_m4<- ifelse (p_m4 > 0.5, 1, 0)
res_m4 <- performance (y_pred = p_m4,
y_true = Y_test)
## Refit lasso
X_train_refit <- model.matrix(outcome ~., data = df)[,names(coef_np_m4)]
np_m4_refit <- glm.fit(x = X_train_refit,
y = Y_train,
family = binomial())
# NCVreg
np_m5 <- cv.ncvreg(X_train,
Y_train,
type = "MCP",
family = "binomial")
plot(np_m5 )
coef_np_m5 <- coef(np_m5, s = "lambda.min")[coef(np_m5, s = "lambda.min") != 0]
p_m5 <- predict (np_m5, type = "response", X = X_test, lambda = np_m5$lambda.min)
p_m5<- ifelse (p_m5 > 0.5, 1, 0)
res_m5 <- performance (y_pred = p_m5,
y_true = Y_test)
## Refit NCVreg
X_train_refit <- model.matrix(outcome ~., data = df)[,names(coef_np_m5)]
np_m5_refit <- glm.fit(x = X_train_refit,
y = Y_train,
family = binomial())
coef(np_m5_refit)
# mboost
df1 <- df %>%
mutate (outcome = factor (outcome, levels=0:1))
np_m6 <- glmboost(outcome ~.,
data = df1,
control = boost_control(mstop = 1000, nu = 0.1),
family = Binomial(type = c("glm"))) # coefficients from Binomial(link = "logit") are 1/2
# cv10f <- cv(model.weights(np_m6), type = "kfold")
cvm <- cvrisk(np_m6) # , folds = cv10f)
plot (cvm)
np_m6 [mstop(cvm)]
summary (np_m6 [mstop(cvm)])
p_m6 <- predict (np_m6, type = "class",
newdata = np_test %>% dplyr::select (-outcome))
#p_m6<- ifelse (p_m6 > 0.5, 1, 0)
res_m6 <- performance (y_pred = p_m6,
y_true = Y_test)
## Refit mboost
coef_np_m6 <- coef(np_m6)
X_train_refit <- model.matrix(outcome ~., data = df)[,names(coef_np_m6)]
np_m6_refit <- glm.fit(x = X_train_refit,
y = Y_train,
family = binomial())
coef(np_m6_refit)
# MARS - linear
np_m7 <- earth (outcome ~.,
data = df,
linpreds=TRUE,
glm=list(family=binomial))
summary (np_m7)
p_m7 <- predict (np_m7, type = "response", newdata = np_test %>% dplyr::select (-outcome))
p_m7<- ifelse (p_m7 > 0.5, 1, 0)
res_m7 <- performance (y_pred = p_m7,
y_true = Y_test)coef_np1 <- data.frame (variables = names (coef(np_m1)),
pval = coef(np_m1))
coef_np2 <- data.frame (variables = names (coef(np_m2)),
stepaic = coef(np_m2))
coef_np3 <- data.frame (variables = rownames (coef(np_m3 , support.size = best_np_m3 , sparse = FALSE)),
bestsubset = coef(np_m3 , support.size = best_np_m3 , sparse = FALSE)[,1])
coef_np4 <- data.frame (variables = names (coef(np_m4, s = "lambda.min")[coef(np_m4, s = "lambda.min") != 0]),
lasso = coef(np_m4, s = "lambda.min")[coef(np_m4, s = "lambda.min") != 0])
coef_np4_refit <- data.frame (variables = names (coef(np_m4_refit)),
lasso_refit = coef(np_m4_refit))
coef_np5 <- data.frame (variables = names (coef(np_m5, s = "lambda.min")[coef(np_m5, s = "lambda.min") != 0]),
mcp = coef(np_m5, s = "lambda.min")[coef(np_m5, s = "lambda.min") != 0])
coef_np5_refit <- data.frame (variables = names (coef(np_m5_refit)),
mcp_refit = coef(np_m5_refit))
coef_np6 <- data.frame (variables = names (coef(np_m6)),
mboost = coef(np_m6))
coef_np6_refit <- data.frame (variables = names (coef(np_m6_refit)),
mboost_refit = coef(np_m6_refit))
coef_np7 <- data.frame (variables = names (coef(np_m7)),
mars = coef(np_m7))
np_predictors <- data.frame(variables = colnames (X_train)) %>%
left_join(coef_np1, by = "variables") %>%
left_join(coef_np2, by = "variables") %>%
left_join(coef_np3, by = "variables") %>%
left_join(coef_np4, by = "variables")%>%
left_join(coef_np4_refit, by = "variables")%>%
left_join(coef_np5, by = "variables")%>%
left_join(coef_np5_refit, by = "variables")%>%
left_join(coef_np6, by = "variables") %>%
left_join(coef_np6_refit, by = "variables")%>%
left_join(coef_np7, by = "variables") %>%
mutate_all(~ifelse(.x == 0, NA, .x)) %>%
mutate_if(is.numeric, round, 3) %>%
mutate (remove = rowSums(is.na(.[,-c(1, 6, 8, 10)])),
keep = ifelse (remove == 0, "Y", "N"))
np_predictors_best <- np_predictors %>%
filter (keep == "Y")np_res <-
cbind(Model = c("pval", "stepaic", "bestsubset", "lasso", "ncvreg", "mboost", "mars"),
as.data.frame(do.call("rbind", list(res_m1, res_m2, res_m3, res_m4, res_m5, res_m6, res_m7)))
)np_all_models <- list ("pval" = np_m1,
"stepaic" = np_m2,
"bestsubset" = np_m3,
"lasso"= np_m4,
"lasso_refit"= np_m4_refit,
"ncvreg"= np_m5,
"ncvreg_refit"= np_m5_refit,
"mboost"= np_m6,
"mboost_refit"= np_m6_refit,
"mars"= np_m7)
np <- list (performance = np_res,
predictors = np_predictors,
models = np_all_models)
saveRDS(np, "output/np_iml.RDS")scales <- build_scales(ap_train)
ap_train <- fast_scale(ap_train, scales = scales)
ap_test <- fast_scale(ap_test, scales = scales)o <- ap_test$outcome
# P value
df <- ap_train %>%
mutate (outcome = as.numeric(outcome) - 1)
uni_names <- df %>%
dplyr::select(-outcome) %>%
map(~glm(df$outcome ~ .x, data = df, family = binomial())) %>%
map(anova, test = "Chisq") %>%
map_dbl(p_extract) %>%
"<" (0.1) %>%
which() %>%
names()
form <- paste0("outcome~", paste0(uni_names, collapse = "+"))
fullmodel <- glm(formula(form), data = df, family = binomial)
nullmodel <- glm(outcome ~1, data = df, family = binomial)
scope = list(lower=formula(nullmodel ),upper=formula(fullmodel))
ap_m1 = SignifReg(fullmodel,
scope=scope,
alpha = 0.05,
direction = "both",
criterion = "p-value",
adjust.method = "none",
trace=FALSE)
p_m1 <- predict (ap_m1, type = "response", newdata = np_test %>% dplyr::select (-outcome))
p_m1 <- ifelse (p_m1 > 0.5, 1, 0)
res_m1 <- performance (y_pred = p_m1,
y_true = o)
# Step AIC
full_model <- glm (outcome ~ ., data = df, family = binomial())
ap_m2 <- stepAIC (full_model,
direction = "both")
summary (ap_m2 )
p_m2 <- predict (ap_m2, type = "response", newdata = ap_test %>% dplyr::select (-outcome))
p_m2 <- ifelse (p_m2 > 0.5, 1, 0)
res_m2 <- performance (y_pred = p_m2,
y_true = o)
# Best subset regression
X_train <- model.matrix(outcome ~., data = df)[,-1]
Y_train <- as.numeric (df$outcome)
X_test <- model.matrix(outcome ~., data = ap_test)[,-1]
Y_test <- as.numeric (ap_test$outcome) -1
ap_m3 <- abess(x = X_train,
y = Y_train,
family = "binomial",
tune.type = "cv"
)
best_ap_m3 <- ap_m3 [["best.size"]]
print(best_ap_m3 )
coef(ap_m3 , support.size = best_ap_m3 , sparse = FALSE)
p_m3 <- predict (ap_m3, type = "response", newx = X_test, support.size = best_ap_m3)
p_m3<- as.numeric (ifelse (p_m3 > 0.5, 1, 0))
res_m3 <- performance (y_pred = p_m3,
y_true = Y_test)
# Lasso
ap_m4 <- cv.ncvreg(X_train,
Y_train,
penalty = "lasso",
family = "binomial")
plot(ap_m4)
coef_ap_m4 <- coef(ap_m4, s = "lambda.min")[coef(ap_m4, s = "lambda.min") != 0]
p_m4 <- predict (ap_m4, type = "response", X = X_test, lambda = ap_m4$lambda.min)
p_m4<- ifelse (p_m4 > 0.5, 1, 0)
res_m4 <- performance (y_pred = p_m4,
y_true = Y_test)
## Refit lasso
X_train_refit <- model.matrix(outcome ~., data = df)[,names(coef_ap_m4)]
ap_m4_refit <- glm.fit(x = X_train_refit,
y = Y_train,
family = binomial())
# NCVreg
ap_m5 <- cv.ncvreg(X_train,
Y_train,
type = "MCP",
family = "binomial")
plot(ap_m5 )
coef_ap_m5 <- coef(ap_m5, s = "lambda.min")[coef(ap_m5, s = "lambda.min") != 0]
p_m5 <- predict (ap_m5, type = "response", X = X_test, lambda = ap_m5$lambda.min)
p_m5<- ifelse (p_m5 > 0.5, 1, 0)
res_m5 <- performance (y_pred = p_m5,
y_true = Y_test)
## Refit NCVreg
X_train_refit <- model.matrix(outcome ~., data = df)[,names(coef_ap_m5)]
ap_m5_refit <- glm.fit(x = X_train_refit,
y = Y_train,
family = binomial())
coef(ap_m5_refit)
# mboost
df1 <- df %>%
mutate (outcome = factor (outcome, levels=0:1))
ap_m6 <- glmboost(outcome ~.,
data = df1,
control = boost_control(mstop = 1000, nu = 0.1),
family = Binomial(type = c("glm"))) # coefficients from Binomial(link = "logit") are 1/2
# cv10f <- cv(model.weights(ap_m6), type = "kfold")
cvm <- cvrisk(ap_m6) # , folds = cv10f)
plot (cvm)
ap_m6 [mstop(cvm)]
summary (ap_m6 [mstop(cvm)])
p_m6 <- predict (ap_m6, type = "class",
newdata = ap_test %>% dplyr::select (-outcome))
#p_m6<- ifelse (p_m6 > 0.5, 1, 0)
res_m6 <- performance (y_pred = p_m6,
y_true = Y_test)
## Refit mboost
coef_ap_m6 <- coef(ap_m6)
X_train_refit <- model.matrix(outcome ~., data = df)[,names(coef_ap_m6)]
ap_m6_refit <- glm.fit(x = X_train_refit,
y = Y_train,
family = binomial())
coef(ap_m6_refit)
# MARS - linear
ap_m7 <- earth (outcome ~.,
data = df,
linpreds=TRUE,
glm=list(family=binomial))
summary (ap_m7)
p_m7 <- predict (ap_m7, type = "response", newdata = ap_test %>% dplyr::select (-outcome))
p_m7<- ifelse (p_m7 > 0.5, 1, 0)
res_m7 <- performance (y_pred = p_m7,
y_true = Y_test)coef_ap1 <- data.frame (variables = names (coef(ap_m1)),
pval = coef(ap_m1))
coef_ap2 <- data.frame (variables = names (coef(ap_m2)),
stepaic = coef(ap_m2))
coef_ap3 <- data.frame (variables = rownames (coef(ap_m3 , support.size = best_ap_m3 , sparse = FALSE)),
bestsubset = coef(ap_m3 , support.size = best_ap_m3 , sparse = FALSE)[,1])
coef_ap4 <- data.frame (variables = names (coef(ap_m4, s = "lambda.min")[coef(ap_m4, s = "lambda.min") != 0]),
lasso = coef(ap_m4, s = "lambda.min")[coef(ap_m4, s = "lambda.min") != 0])
coef_ap4_refit <- data.frame (variables = names (coef(ap_m4_refit)),
lasso_refit = coef(ap_m4_refit))
coef_ap5 <- data.frame (variables = names (coef(ap_m5, s = "lambda.min")[coef(ap_m5, s = "lambda.min") != 0]),
mcp = coef(ap_m5, s = "lambda.min")[coef(ap_m5, s = "lambda.min") != 0])
coef_ap5_refit <- data.frame (variables = names (coef(ap_m5_refit)),
mcp_refit = coef(ap_m5_refit))
coef_ap6 <- data.frame (variables = names (coef(ap_m6)),
mboost = coef(ap_m6))
coef_ap6_refit <- data.frame (variables = names (coef(ap_m6_refit)),
mboost_refit = coef(ap_m6_refit))
coef_ap7 <- data.frame (variables = names (coef(ap_m7)),
mars = coef(ap_m7))
ap_predictors <- data.frame(variables = colnames (X_train)) %>%
left_join(coef_ap1, by = "variables") %>%
left_join(coef_ap2, by = "variables") %>%
left_join(coef_ap3, by = "variables") %>%
left_join(coef_ap4, by = "variables")%>%
left_join(coef_ap4_refit, by = "variables")%>%
left_join(coef_ap5, by = "variables")%>%
left_join(coef_ap5_refit, by = "variables")%>%
left_join(coef_ap6, by = "variables") %>%
left_join(coef_ap6_refit, by = "variables")%>%
left_join(coef_ap7, by = "variables") %>%
mutate_all(~ifelse(.x == 0, NA, .x)) %>%
mutate_if(is.numeric, round, 3) %>%
mutate (remove = rowSums(is.na(.[,-c(1, 6, 8, 10)])),
keep = ifelse (remove == 0, "Y", "N"))
ap_predictors_best <- ap_predictors %>%
filter (keep == "Y")ap_res <-
cbind(Model = c("pval", "stepaic", "bestsubset", "lasso", "ncvreg", "mboost", "mars"),
as.data.frame(do.call("rbind", list(res_m1, res_m2, res_m3, res_m4, res_m5, res_m6, res_m7)))
)ap_all_models <- list ("pval" = ap_m1,
"stepaic" = ap_m2,
"bestsubset" = ap_m3,
"lasso"= ap_m4,
"lasso_refit"= ap_m4_refit,
"ncvreg"= ap_m5,
"ncvreg_refit"= ap_m5_refit,
"mboost"= ap_m6,
"mboost_refit"= ap_m6_refit,
"mars"= ap_m7)
ap <- list (performance = ap_res,
predictors = ap_predictors,
models = ap_all_models)
saveRDS(ap, "output/ap_iml.RDS")scales <- build_scales(dis_train)
dis_train <- fast_scale(dis_train, scales = scales)
dis_test <- fast_scale(dis_test, scales = scales)o <- dis_test$outcome
# P value
df <- dis_train %>%
mutate (outcome = as.numeric(outcome) - 1)
uni_names <- df %>%
dplyr::select(-outcome) %>%
map(~glm(df$outcome ~ .x, data = df, family = binomial())) %>%
map(anova, test = "Chisq") %>%
map_dbl(p_extract) %>%
"<" (0.1) %>%
which() %>%
names()
form <- paste0("outcome~", paste0(uni_names, collapse = "+"))
fullmodel <- glm(form, data = df, family = binomial)
nullmodel <- glm(outcome ~1, data = df, family = binomial)
scope = list(lower=formula(nullmodel ),upper=formula(fullmodel))
dis_m1 = SignifReg(fullmodel,
scope=scope,
alpha = 0.5,
direction = "both",
criterion = "p-value",
adjust.method = "none",
trace=FALSE)
p_m1 <- predict (dis_m1, type = "response", newdata = np_test %>% dplyr::select (-outcome))
p_m1 <- ifelse (p_m1 > 0.5, 1, 0)
res_m1 <- performance (y_pred = p_m1,
y_true = o)
# Step AIC
full_model <- glm (outcome ~ ., data = df, family = binomial())
dis_m2 <- stepAIC (full_model,
direction = "both")
summary (dis_m2 )
p_m2 <- predict (dis_m2, type = "response", newdata = dis_test %>% dplyr::select (-outcome))
p_m2 <- ifelse (p_m2 > 0.5, 1, 0)
res_m2 <- performance (y_pred = p_m2,
y_true = o)
# Best subset regression
X_train <- model.matrix(outcome ~., data = df)[,-1]
Y_train <- as.numeric (df$outcome)
X_test <- model.matrix(outcome ~., data = dis_test)[,-1]
Y_test <- as.numeric (dis_test$outcome) -1
dis_m3 <- abess(x = X_train,
y = Y_train,
family = "binomial",
tune.type = "cv"
)
best_dis_m3 <- dis_m3 [["best.size"]]
print(best_dis_m3 )
coef(dis_m3 , support.size = best_dis_m3 , sparse = FALSE)
p_m3 <- predict (dis_m3, type = "response", newx = X_test, support.size = best_dis_m3)
p_m3<- as.numeric (ifelse (p_m3 > 0.5, 1, 0))
res_m3 <- performance (y_pred = p_m3,
y_true = Y_test)
# Lasso
dis_m4 <- cv.ncvreg(X_train,
Y_train,
penalty = "lasso",
family = "binomial")
plot(dis_m4)
coef_dis_m4 <- coef(dis_m4, s = "lambda.min")[coef(dis_m4, s = "lambda.min") != 0]
p_m4 <- predict (dis_m4, type = "response", X = X_test, lambda = dis_m4$lambda.min)
p_m4<- ifelse (p_m4 > 0.5, 1, 0)
res_m4 <- performance (y_pred = p_m4,
y_true = Y_test)
## Refit lasso
X_train_refit <- model.matrix(outcome ~., data = df)[,names(coef_dis_m4)]
dis_m4_refit <- glm.fit(x = X_train_refit,
y = Y_train,
family = binomial())
# NCVreg
dis_m5 <- cv.ncvreg(X_train,
Y_train,
type = "MCP",
family = "binomial")
plot(dis_m5 )
coef_dis_m5 <- coef(dis_m5, s = "lambda.min")[coef(dis_m5, s = "lambda.min") != 0]
p_m5 <- predict (dis_m5, type = "response", X = X_test, lambda = dis_m5$lambda.min)
p_m5<- ifelse (p_m5 > 0.5, 1, 0)
res_m5 <- performance (y_pred = p_m5,
y_true = Y_test)
## Refit NCVreg
X_train_refit <- model.matrix(outcome ~., data = df)[,names(coef_dis_m5)]
dis_m5_refit <- glm.fit(x = X_train_refit,
y = Y_train,
family = binomial())
coef(dis_m5_refit)
# mboost
df1 <- df %>%
mutate (outcome = factor (outcome, levels=0:1))
dis_m6 <- glmboost(outcome ~.,
data = df1,
control = boost_control(mstop = 1000, nu = 0.1),
family = Binomial(type = c("glm"))) # coefficients from Binomial(link = "logit") are 1/2
# cv10f <- cv(model.weights(dis_m6), type = "kfold")
cvm <- cvrisk(dis_m6) # , folds = cv10f)
plot (cvm)
dis_m6 [mstop(cvm)]
summary (dis_m6 [mstop(cvm)])
p_m6 <- predict (dis_m6, type = "class",
newdata = dis_test %>% dplyr::select (-outcome))
#p_m6<- ifelse (p_m6 > 0.5, 1, 0)
res_m6 <- performance (y_pred = p_m6,
y_true = Y_test)
## Refit mboost
coef_dis_m6 <- coef(dis_m6)
X_train_refit <- model.matrix(outcome ~., data = df)[,names(coef_dis_m6)]
dis_m6_refit <- glm.fit(x = X_train_refit,
y = Y_train,
family = binomial())
coef(dis_m6_refit)
# MARS - linear
dis_m7 <- earth (outcome ~.,
data = df,
linpreds=TRUE,
glm=list(family=binomial))
summary (dis_m7)
p_m7 <- predict (dis_m7, type = "response", newdata = dis_test %>% dplyr::select (-outcome))
p_m7<- ifelse (p_m7 > 0.5, 1, 0)
res_m7 <- performance (y_pred = p_m7,
y_true = Y_test)coef_dis1 <- data.frame (variables = names (coef(dis_m1)),
pval = coef(dis_m1))
coef_dis2 <- data.frame (variables = names (coef(dis_m2)),
stepaic = coef(dis_m2))
coef_dis3 <- data.frame (variables = rownames (coef(dis_m3 , support.size = best_dis_m3 , sparse = FALSE)),
bestsubset = coef(dis_m3 , support.size = best_dis_m3 , sparse = FALSE)[,1])
coef_dis4 <- data.frame (variables = names (coef(dis_m4, s = "lambda.min")[coef(dis_m4, s = "lambda.min") != 0]),
lasso = coef(dis_m4, s = "lambda.min")[coef(dis_m4, s = "lambda.min") != 0])
coef_dis4_refit <- data.frame (variables = names (coef(dis_m4_refit)),
lasso_refit = coef(dis_m4_refit))
coef_dis5 <- data.frame (variables = names (coef(dis_m5, s = "lambda.min")[coef(dis_m5, s = "lambda.min") != 0]),
mcp = coef(dis_m5, s = "lambda.min")[coef(dis_m5, s = "lambda.min") != 0])
coef_dis5_refit <- data.frame (variables = names (coef(dis_m5_refit)),
mcp_refit = coef(dis_m5_refit))
coef_dis6 <- data.frame (variables = names (coef(dis_m6)),
mboost = coef(dis_m6))
coef_dis6_refit <- data.frame (variables = names (coef(dis_m6_refit)),
mboost_refit = coef(dis_m6_refit))
coef_dis7 <- data.frame (variables = names (coef(dis_m7)),
mars = coef(dis_m7))
dis_predictors <- data.frame(variables = colnames (X_train)) %>%
left_join(coef_dis1, by = "variables") %>%
left_join(coef_dis2, by = "variables") %>%
left_join(coef_dis3, by = "variables") %>%
left_join(coef_dis4, by = "variables")%>%
left_join(coef_dis4_refit, by = "variables")%>%
left_join(coef_dis5, by = "variables")%>%
left_join(coef_dis5_refit, by = "variables")%>%
left_join(coef_dis6, by = "variables") %>%
left_join(coef_dis6_refit, by = "variables")%>%
left_join(coef_dis7, by = "variables") %>%
mutate_all(~ifelse(.x == 0, NA, .x)) %>%
mutate_if(is.numeric, round, 3) %>%
mutate (remove = rowSums(is.na(.[,-c(1, 6, 8, 10)])),
keep = ifelse (remove == 0, "Y", "N"))
dis_predictors_best <- dis_predictors %>%
filter (keep == "Y")dis_res <-
cbind(Model = c("pval", "stepaic", "bestsubset", "lasso", "ncvreg", "mboost", "mars"),
as.data.frame(do.call("rbind", list(res_m1, res_m2, res_m3, res_m4, res_m5, res_m6, res_m7)))
)dis_all_models <- list ("pval" = dis_m1,
"stepaic" = dis_m2,
"bestsubset" = dis_m3,
"lasso"= dis_m4,
"lasso_refit"= dis_m4_refit,
"ncvreg"= dis_m5,
"ncvreg_refit"= dis_m5_refit,
"mboost"= dis_m6,
"mboost_refit"= dis_m6_refit,
"mars"= dis_m7)
dis <- list (performance = dis_res,
predictors = dis_predictors,
models = dis_all_models)
saveRDS(dis, "output/dis_iml.RDS")res_np <- readRDS("output/np_iml.RDS")
res_ap <- readRDS("output/ap_iml.RDS")
res_dis <- readRDS("output/dis_iml.RDS")model_lvl <- c("pval", "pvalAdj", "stepaic",
"bestsubset","lasso", "ncvreg", "mboost", "mars")
model_names <- c("stepP", "stepPAdj", "stepAIC",
"Best subset", "Lasso", "MCP", "mboost", "MuARS")
outcome_lvl <- c("Accuracy", "Precision", "Sensitivity", "Specificity", "AUC")
df.plot <- res_np$performance %>%
pivot_longer(-Model,
names_to = "Outcomes",
values_to = "val") %>%
mutate (Model = factor (Model, model_lvl, model_names),
Outcomes = factor (Outcomes, outcome_lvl))
f <- ggplot (df.plot) +
geom_bar(aes (x = Outcomes, y = val, fill = Model),
stat = "identity", position = "dodge") +
scale_fill_manual(values = c("#030303", "#EE6A50", "#FF0000", "#000080", "#00EE00", "#8B3E2F", "#4A4A4A", "#A6A6A6")) +
ylim (0,1) +
ylab ("Value") +
xlab ("Prediction measures") +
theme_cowplot()
tiff ("manuscript2/fig1.tiff", width = 8, height = 3, units = "in", res = 100)
f
dev.off()pred <- res_np$predictors %>%
#dplyr::select (-c(remove, keep)) %>%
mutate_if(is.numeric, round, 3) %>%
dplyr::select(-mcp_refit) %>%
mutate (Number = rowSums(!is.na(.[, c(2, 3, 4, 5, 6, 8, 9, 11)])))
pred[nrow(pred) + 1, 1] <- "Number"
pred[nrow(pred), -c(1, ncol(pred))] <- colSums(!is.na(pred[,-c(1, ncol(pred))]))
new_names <- c("Variables", "P-value", "P-valueAdj", "Step AIC",
"Best subset", "Lasso", "Lasso refit",
"Ncvreg", "mboost", "mboost refit", "MuARS", "Number")
names (pred) <- new_names
new_variables <- c("Sex - female",
"Age (years)",
"Employment - not working",
"Employment - working",
"Duration of pain (days)",
"Time since first episode (years) - 1-5",
"Time since first episode (years) - 5–10",
"Time since first episode (years) - >10",
"Chronicity - chronic",
"Baseline intensity of neck pain",
"Baseline intensity of arm pain",
"Baseline disability",
"Diagnostic procedure: X-ray - yes",
"Diagnostic procedure: MRI - yes",
"Imaging findings: disc degeneration - yes",
"Imaging findings: facet joint degeneration - yes",
"Imaging findings: scoliosis - yes",
"Imaging findings: spinal stenosis - yes",
"Imaging findings: disc protrusion - yes",
"Imaging findings: disc herniation - yes",
"Pharmacological treatment: analgesics - yes",
"Pharmacological treatment: NSAIDs - yes",
"Pharmacological treatment: steroids - yes",
"Pharmacological treatment: muscle relaxants - yes",
"Pharmacological treatment: opioids - yes",
"Pharmacological treatment: other treatments - yes",
"Non pharmacological treatments - yes",
"NRT",
"Number")
pred$Variables <- new_variables
# Export to word
my_path <- paste0("manuscript2/Table 2",
".docx")
ft <- flextable(pred) %>%
width (1, 6, unit = "cm") %>%
fontsize(size = 10) %>%
set_caption(caption = "Table 2. Beta coefficients of selected variables for the outcome of neck pain",
style = "Table Caption")
my_doc <- read_docx() %>%
body_add_flextable(ft) %>%
body_end_section_landscape()
print (my_doc, target = my_path)probe <- df.plot %>%
mutate(val = round (val, 3)) %>%
pivot_wider(names_from = Outcomes,
values_from = val)
apply (probe[,-1], 2, range)
apply (probe[,-1], 2, range) %>%
apply(2, diff)
(mean((pred$Lasso[-nrow(pred)] - pred$`Lasso refit`[-nrow(pred)])/pred$`Lasso refit`[-nrow(pred)], na.rm = TRUE)) *100
(mean((pred$mboost[-nrow(pred)] - pred$`mboost refit`[-nrow(pred)])/pred$`mboost refit`[-nrow(pred)], na.rm = TRUE)) *100
probe <- pred %>%
dplyr::select (- c(lasso, mboost, Number)) %>%
mutate (Number = rowSums(!is.na(.[, -c (1)]))) %>%
filter (Number == 8) %>%
dplyr::select (-Number) %>%
mutate (ave = rowMeans (.[,c(4:9)], na.rm = TRUE)) %>%
mutate (diffP = ((pval - ave)/ave) * 100,
diffPA = ((pvalAdj - ave)/ave) * 100) %>%
summarize (meanP = mean (diffP),
meanPA = mean (diffPA))
probe <- pred %>%
dplyr::select (- c(lasso_refit, mboost_refit, Number)) %>%
mutate (Number = rowSums(!is.na(.[, -c (1:3)]))) %>%
mutate (Impt = ifelse ((is.na(pval) | is.na (pvalAdj)) & Number == 6, 1, 0 )) %>%
filter (Impt == 1) df.plot <- res_ap$performance %>%
pivot_longer(-Model,
names_to = "Outcomes",
values_to = "val") %>%
mutate (Model = factor (Model, model_lvl, model_names),
Outcomes = factor (Outcomes, outcome_lvl))
f <- ggplot (df.plot) +
geom_bar(aes (x = Outcomes, y = val, fill = Model),
stat = "identity", position = "dodge") +
scale_fill_manual(values = c("#030303", "#EE6A50", "#FF0000", "#000080", "#00EE00", "#8B3E2F", "#4A4A4A", "#A6A6A6")) +
ylim (0,1) +
ylab ("Value") +
xlab ("Prediction measures") +
theme_cowplot()
tiff ("manuscript2/fig2.tiff", width = 8, height = 3, units = "in", res = 100)
f
dev.off()pred <- res_ap$predictors %>%
#dplyr::select (-c(remove, keep)) %>%
mutate_if(is.numeric, round, 3) %>%
dplyr::select(-mcp_refit) %>%
mutate (Number = rowSums(!is.na(.[, c(2, 3, 4, 5, 6, 8, 9, 11)])))
pred[nrow(pred) + 1, 1] <- "Number"
pred[nrow(pred), -c(1, ncol(pred))] <- as.integer(colSums(!is.na(pred[,-c(1, ncol(pred))])))
names (pred) <- new_names
pred$Variables <- new_variables
# Export to word
my_path <- paste0("manuscript2/Table 3",
".docx")
ft <- flextable(pred) %>%
width (1, 6, unit = "cm") %>%
fontsize(size = 10) %>%
set_caption(caption = "Table 3. Beta coefficients of selected variables for the outcome of arm pain",
style = "Table Caption")
my_doc <- read_docx() %>%
body_add_flextable(ft) %>%
body_end_section_landscape()
print (my_doc, target = my_path)probe <- df.plot %>%
mutate(val = round (val, 3)) %>%
pivot_wider(names_from = Outcomes,
values_from = val)
apply (probe[,-1], 2, range)
apply (probe[,-1], 2, range) %>%
apply(2, diff)
(mean((pred$Lasso[-nrow(pred)] - pred$`Lasso refit`[-nrow(pred)])/pred$`Lasso refit`[-nrow(pred)], na.rm = TRUE)) *100
(mean((pred$mboost[-nrow(pred)] - pred$`mboost refit`[-nrow(pred)])/pred$`mboost refit`[-nrow(pred)], na.rm = TRUE)) *100
probe <- pred %>%
dplyr::select (- c(lasso, mboost, Number)) %>%
mutate (Number = rowSums(!is.na(.[, -c (1)]))) %>%
filter (Number == 8) %>%
dplyr::select (-Number) %>%
mutate (ave = rowMeans (.[,c(4:9)], na.rm = TRUE)) %>%
mutate (diffP = ((pval - ave)/ave) * 100,
diffPA = ((pvalAdj - ave)/ave) * 100) %>%
summarize (meanP = mean (diffP),
meanPA = mean (diffPA))
probe <- pred %>%
dplyr::select (- c(lasso_refit, mboost_refit, Number)) %>%
mutate (Number = rowSums(!is.na(.[, -c (1:3)]))) %>%
mutate (Impt = ifelse ((is.na(pval) | is.na (pvalAdj)) & Number == 6, 1, 0 )) %>%
filter (Impt == 1) df.plot <- res_dis$performance %>%
pivot_longer(-Model,
names_to = "Outcomes",
values_to = "val") %>%
mutate (Model = factor (Model, model_lvl, model_names),
Outcomes = factor (Outcomes, outcome_lvl))
f <- ggplot (df.plot) +
geom_bar(aes (x = Outcomes, y = val, fill = Model),
stat = "identity", position = "dodge") +
scale_fill_manual(values = c("#030303", "#EE6A50", "#FF0000", "#000080", "#00EE00", "#8B3E2F", "#4A4A4A", "#A6A6A6")) +
ylim (0,1) +
ylab ("Value") +
xlab ("Prediction measures") +
theme_cowplot()
tiff ("manuscript2/fig3.tiff", width = 8, height = 3, units = "in", res = 100)
f
dev.off()pred <- res_dis$predictors %>%
#dplyr::select (-c(remove, keep)) %>%
mutate_if(is.numeric, round, 3) %>%
dplyr::select(-mcp_refit) %>%
mutate (Number = rowSums(!is.na(.[, c(2, 3, 4, 5, 6, 8, 9, 11)])))
pred[nrow(pred) + 1, 1] <- "Number"
pred[nrow(pred), -c(1, ncol(pred))] <- as.integer(colSums(!is.na(pred[,-c(1, ncol(pred))])))
names (pred) <- new_names
pred$Variables <- new_variables
# Export to word
my_path <- paste0("manuscript2/Table 4",
".docx")
ft <- flextable(pred) %>%
width (1, 6, unit = "cm") %>%
fontsize(size = 10) %>%
set_caption(caption = "Table 4. Beta coefficients of selected variables for the outcome of disability",
style = "Table Caption")
my_doc <- read_docx() %>%
body_add_flextable(ft) %>%
body_end_section_landscape()
print (my_doc, target = my_path)probe <- df.plot %>%
mutate(val = round (val, 3)) %>%
pivot_wider(names_from = Outcomes,
values_from = val)
apply (probe[,-1], 2, range)
apply (probe[,-1], 2, range) %>%
apply(2, diff)
(mean((pred$Lasso[-nrow(pred)] - pred$`Lasso refit`[-nrow(pred)])/pred$`Lasso refit`[-nrow(pred)], na.rm = TRUE)) *100
(mean((pred$mboost[-nrow(pred)] - pred$`mboost refit`[-nrow(pred)])/pred$`mboost refit`[-nrow(pred)], na.rm = TRUE)) *100
probe <- pred %>%
dplyr::select (- c(lasso, mboost, Number)) %>%
mutate (Number = rowSums(!is.na(.[, -c (1)]))) %>%
filter (Number == 8) %>%
dplyr::select (-Number) %>%
mutate (ave = rowMeans (.[,c(4:9)], na.rm = TRUE)) %>%
mutate (diffP = ((pval - ave)/ave) * 100,
diffPA = ((pvalAdj - ave)/ave) * 100) %>%
summarize (meanP = mean (diffP),
meanPA = mean (diffPA))
probe <- pred %>%
dplyr::select (- c(lasso_refit, mboost_refit, Number)) %>%
mutate (Number = rowSums(!is.na(.[, -c (1:3)]))) %>%
mutate (Impt = ifelse ((is.na(pval) | is.na (pvalAdj)) & Number == 6, 1, 0 )) %>%
filter (Impt == 1)
sessionInfo()