eohi/.history/eohi1/regressions e1 - assumptions_20251016142956.r
2025-12-23 15:47:09 -05:00

393 lines
14 KiB
R

# Regression Analysis - Assumption Checking
# IVs: demo_sex, demo_age, demo_edu
# DVs: eohiDGEN_mean, ehi_global_mean
options(scipen = 999)
library(car)
library(performance)
library(see)
library(ggplot2)
library(gridExtra)
library(dplyr)
library(lmtest) # For bptest and durbinWatsonTest
setwd("C:/Users/irina/Documents/DND/EOHI/eohi1")
data <- read.csv("ehi1.csv")
# Check for missing values
missing_summary <- data %>%
select(demo_sex, demo_age, demo_edu, eohiDGEN_mean, ehi_global_mean) %>%
summarise_all(~sum(is.na(.)))
print("Missing values check:")
print(missing_summary)
# Remove rows with missing values
data_clean <- data %>%
select(pID, demo_sex, demo_age, demo_edu, eohiDGEN_mean, ehi_global_mean) %>%
filter(complete.cases(.))
print(paste("Clean data dimensions:", paste(dim(data_clean), collapse = " x ")))
# Recode demo_sex as numeric for regression (0 = Female, 1 = Male)
data_clean$demo_sex_numeric <- ifelse(data_clean$demo_sex == "Male", 1, 0)
# Check demo_edu levels and recode if needed
print("Education levels:")
print(table(data_clean$demo_edu))
# Recode education as ordinal (assuming higher values = more education)
edu_levels <- c("High School (or equivalent)", "College Diploma/Certificate",
"University - Undergraduate", "University - Graduate")
data_clean$demo_edu_numeric <- match(data_clean$demo_edu, edu_levels)
# Verify recoding
print("Sex recoding (0=Female, 1=Male):")
print(table(data_clean$demo_sex_numeric))
print("Education recoding (1=HS, 2=College, 3=Undergrad, 4=Grad):")
print(table(data_clean$demo_edu_numeric))
# =============================================================================
# REGRESSION MODELS
# =============================================================================
# Define the 6 regression models
models <- list()
# Model 1: demo_sex → eohiDGEN_mean
models$sex_eohiDGEN <- lm(eohiDGEN_mean ~ demo_sex_numeric, data = data_clean)
# Model 2: demo_age → eohiDGEN_mean
models$age_eohiDGEN <- lm(eohiDGEN_mean ~ demo_age, data = data_clean)
# Model 3: demo_edu → eohiDGEN_mean
models$edu_eohiDGEN <- lm(eohiDGEN_mean ~ demo_edu_numeric, data = data_clean)
# Model 4: demo_sex → ehi_global_mean
models$sex_ehi_global <- lm(ehi_global_mean ~ demo_sex_numeric, data = data_clean)
# Model 5: demo_age → ehi_global_mean
models$age_ehi_global <- lm(ehi_global_mean ~ demo_age, data = data_clean)
# Model 6: demo_edu → ehi_global_mean
models$edu_ehi_global <- lm(ehi_global_mean ~ demo_edu_numeric, data = data_clean)
# =============================================================================
# ASSUMPTION CHECKING FUNCTIONS
# =============================================================================
# Function to check linearity assumption
check_linearity <- function(model, model_name) {
print(paste("=== LINEARITY CHECK:", model_name, "==="))
# Residuals vs Fitted plot
plot(model, which = 1, main = paste("Linearity:", model_name))
# Component + residual plot (partial residual plot)
crPlots(model, main = paste("Component+Residual Plot:", model_name))
return(NULL)
}
# Function to check normality of residuals
check_normality <- function(model, model_name) {
print(paste("=== NORMALITY CHECK:", model_name, "==="))
# Q-Q plot
plot(model, which = 2, main = paste("Q-Q Plot:", model_name))
# Shapiro-Wilk test
residuals <- residuals(model)
shapiro_test <- shapiro.test(residuals)
print(paste("Shapiro-Wilk test p-value:", format(shapiro_test$p.value, digits = 5)))
# Kolmogorov-Smirnov test
ks_test <- ks.test(residuals, "pnorm", mean(residuals), sd(residuals))
print(paste("Kolmogorov-Smirnov test p-value:", format(ks_test$p.value, digits = 5)))
# Histogram of residuals
hist_plot <- ggplot(data.frame(residuals = residuals), aes(x = residuals)) +
geom_histogram(bins = 30, fill = "lightblue", color = "black") +
ggtitle(paste("Residuals Histogram:", model_name)) +
theme_minimal()
print(hist_plot)
return(list(shapiro_p = shapiro_test$p.value, ks_p = ks_test$p.value))
}
# Function to check homoscedasticity (constant variance)
check_homoscedasticity <- function(model, model_name) {
print(paste("=== HOMOSCEDASTICITY CHECK:", model_name, "==="))
# Scale-Location plot
plot(model, which = 3, main = paste("Scale-Location Plot:", model_name))
# Breusch-Pagan test
bp_test <- bptest(model)
print(paste("Breusch-Pagan test p-value:", format(bp_test$p.value, digits = 5)))
# White test (if available)
tryCatch({
white_test <- bptest(model, ~ fitted(model) + I(fitted(model)^2))
print(paste("White test p-value:", format(white_test$p.value, digits = 5)))
}, error = function(e) {
print("White test not available for this model")
})
return(list(bp_p = bp_test$p.value))
}
# Function to check independence (no autocorrelation)
check_independence <- function(model, model_name) {
print(paste("=== INDEPENDENCE CHECK:", model_name, "==="))
# Durbin-Watson test
dw_test <- durbinWatsonTest(model)
print(paste("Durbin-Watson statistic:", format(dw_test$dw, digits = 5)))
print(paste("Durbin-Watson p-value:", format(dw_test$p, digits = 5)))
# Residuals vs Order plot
residuals_vs_order <- ggplot(data.frame(
residuals = residuals(model),
order = seq_along(residuals(model))
), aes(x = order, y = residuals)) +
geom_point(color = "black") +
geom_hline(yintercept = 0, linetype = "dashed", color = "red") +
ggtitle(paste("Residuals vs Order:", model_name)) +
theme_minimal()
print(residuals_vs_order)
return(list(dw_stat = dw_test$dw, dw_p = dw_test$p))
}
# Function to check for influential observations
check_influence <- function(model, model_name) {
print(paste("=== INFLUENCE CHECK:", model_name, "==="))
# Cook's Distance plot
plot(model, which = 4, main = paste("Cook's Distance:", model_name))
# Calculate influence measures
cooks_d <- cooks.distance(model)
leverage <- hatvalues(model)
dffits_val <- dffits(model)
# Identify influential observations
cooks_threshold <- 4/length(cooks_d) # Cook's D threshold
leverage_threshold <- 2 * (length(coef(model))/nobs(model)) # Leverage threshold
dffits_threshold <- 2 * sqrt(length(coef(model))/nobs(model)) # DFFITS threshold
influential_cooks <- which(cooks_d > cooks_threshold)
influential_leverage <- which(leverage > leverage_threshold)
influential_dffits <- which(abs(dffits_val) > dffits_threshold)
print(paste("Cook's Distance threshold:", format(cooks_threshold, digits = 5)))
print(paste("Influential observations (Cook's D):", length(influential_cooks)))
print(paste("Leverage threshold:", format(leverage_threshold, digits = 5)))
print(paste("High leverage observations:", length(influential_leverage)))
print(paste("DFFITS threshold:", format(dffits_threshold, digits = 5)))
print(paste("Influential observations (DFFITS):", length(influential_dffits)))
if (length(influential_cooks) > 0) {
print(paste("Cook's D influential cases:", paste(influential_cooks, collapse = ", ")))
}
if (length(influential_leverage) > 0) {
print(paste("High leverage cases:", paste(influential_leverage, collapse = ", ")))
}
if (length(influential_dffits) > 0) {
print(paste("DFFITS influential cases:", paste(influential_dffits, collapse = ", ")))
}
return(list(influential_cooks = influential_cooks,
influential_leverage = influential_leverage,
influential_dffits = influential_dffits))
}
# Function to get comprehensive model summary
get_model_summary <- function(model, model_name) {
print(paste("=== MODEL SUMMARY:", model_name, "==="))
# Basic model summary
summary_model <- summary(model)
print(summary_model)
# R-squared and adjusted R-squared
print(paste("R-squared:", format(summary_model$r.squared, digits = 5)))
print(paste("Adjusted R-squared:", format(summary_model$adj.r.squared, digits = 5)))
# AIC and BIC
aic_val <- AIC(model)
bic_val <- BIC(model)
print(paste("AIC:", format(aic_val, digits = 5)))
print(paste("BIC:", format(bic_val, digits = 5)))
return(list(summary = summary_model, r_squared = summary_model$r.squared,
adj_r_squared = summary_model$adj.r.squared, aic = aic_val, bic = bic_val))
}
# =============================================================================
# RUN ASSUMPTION CHECKS FOR ALL MODELS
# =============================================================================
# Create results storage
assumption_results <- list()
model_summaries <- list()
# Model names for reference
model_names <- c("Sex → EOHI-DGEN", "Age → EOHI-DGEN", "Education → EOHI-DGEN",
"Sex → EHI-Global", "Age → EHI-Global", "Education → EHI-Global")
cat("\n", rep("=", 80), "\n")
cat("COMPREHENSIVE REGRESSION ASSUMPTION ANALYSIS\n")
cat(rep("=", 80), "\n")
# Run assumption checks for each model
for (i in seq_along(models)) {
model_name <- model_names[i]
model <- models[[i]]
cat("\n", rep("-", 60), "\n")
cat("ANALYZING MODEL", i, ":", model_name, "\n")
cat(rep("-", 60), "\n")
# Store results
assumption_results[[i]] <- list()
assumption_results[[i]]$model_name <- model_name
# 1. Model Summary
model_summaries[[i]] <- get_model_summary(model, model_name)
assumption_results[[i]]$summary <- model_summaries[[i]]
# 2. Linearity Check
assumption_results[[i]]$linearity <- check_linearity(model, model_name)
# 3. Normality Check
assumption_results[[i]]$normality <- check_normality(model, model_name)
# 4. Homoscedasticity Check
assumption_results[[i]]$homoscedasticity <- check_homoscedasticity(model, model_name)
# 5. Independence Check
assumption_results[[i]]$independence <- check_independence(model, model_name)
# 6. Influence Check
assumption_results[[i]]$influence <- check_influence(model, model_name)
}
# =============================================================================
# SUMMARY TABLE OF ASSUMPTION VIOLATIONS
# =============================================================================
cat("\n", rep("=", 80), "\n")
cat("ASSUMPTION VIOLATION SUMMARY\n")
cat(rep("=", 80), "\n")
# Create summary table
violation_summary <- data.frame(
Model = character(),
Linearity = character(),
Normality = character(),
Homoscedasticity = character(),
Independence = character(),
Influential_Obs = character(),
stringsAsFactors = FALSE
)
# Populate summary table
for (i in seq_along(models)) {
model_name <- model_names[i]
# Check normality (Shapiro-Wilk p < 0.05 indicates violation)
normality_violation <- ifelse(assumption_results[[i]]$normality$shapiro_p < 0.05, "VIOLATED", "OK")
# Check homoscedasticity (Breusch-Pagan p < 0.05 indicates violation)
homosced_violation <- ifelse(assumption_results[[i]]$homoscedasticity$bp_p < 0.05, "VIOLATED", "OK")
# Check independence (Durbin-Watson p < 0.05 indicates violation)
independence_violation <- ifelse(assumption_results[[i]]$independence$dw_p < 0.05, "VIOLATED", "OK")
# Check for influential observations
influential_count <- length(assumption_results[[i]]$influence$influential_cooks)
influential_status <- ifelse(influential_count > 0, paste("YES (", influential_count, ")", sep = ""), "NO")
# Linearity is assessed visually, so we'll mark as "CHECK VISUALLY"
linearity_status <- "CHECK VISUALLY"
violation_summary <- rbind(violation_summary, data.frame(
Model = model_name,
Linearity = linearity_status,
Normality = normality_violation,
Homoscedasticity = homosced_violation,
Independence = independence_violation,
Influential_Obs = influential_status
))
}
print(violation_summary)
# =============================================================================
# MODEL COMPARISON TABLE
# =============================================================================
cat("\n", rep("=", 80), "\n")
cat("MODEL COMPARISON SUMMARY\n")
cat(rep("=", 80), "\n")
# Create model comparison table
comparison_table <- data.frame(
Model = model_names,
R_Squared = numeric(length(models)),
Adj_R_Squared = numeric(length(models)),
AIC = numeric(length(models)),
BIC = numeric(length(models)),
Significant = character(length(models)),
stringsAsFactors = FALSE
)
for (i in seq_along(models)) {
summary_model <- model_summaries[[i]]$summary
comparison_table$R_Squared[i] <- summary_model$r.squared
comparison_table$Adj_R_Squared[i] <- summary_model$adj.r.squared
comparison_table$AIC[i] <- model_summaries[[i]]$aic
comparison_table$BIC[i] <- model_summaries[[i]]$bic
# Check if predictor is significant (p < 0.05)
p_value <- summary_model$coefficients[2, 4]
comparison_table$Significant[i] <- ifelse(p_value < 0.05, "YES", "NO")
}
print(comparison_table)
# =============================================================================
# RECOMMENDATIONS
# =============================================================================
cat("\n", rep("=", 80), "\n")
cat("RECOMMENDATIONS FOR ASSUMPTION VIOLATIONS\n")
cat(rep("=", 80), "\n")
cat("\n1. NORMALITY VIOLATIONS:\n")
cat(" - If violated: Consider transforming the dependent variable (log, sqrt, Box-Cox)\n")
cat(" - Alternative: Use robust regression methods or bootstrapping\n")
cat("\n2. HOMOSCEDASTICITY VIOLATIONS:\n")
cat(" - If violated: Use weighted least squares or robust standard errors\n")
cat(" - Alternative: Transform the dependent variable or use heteroscedasticity-consistent standard errors\n")
cat("\n3. INDEPENDENCE VIOLATIONS:\n")
cat(" - If violated: Check for clustering or repeated measures structure\n")
cat(" - Alternative: Use mixed-effects models or clustered standard errors\n")
cat("\n4. INFLUENTIAL OBSERVATIONS:\n")
cat(" - If present: Examine these cases for data entry errors\n")
cat(" - Consider: Running analysis with and without influential cases\n")
cat(" - Alternative: Use robust regression methods\n")
cat("\n5. LINEARITY VIOLATIONS:\n")
cat(" - If violated: Add polynomial terms or use splines\n")
cat(" - Alternative: Transform predictors or use non-parametric methods\n")
cat("\n", rep("=", 80), "\n")
cat("ANALYSIS COMPLETE\n")
cat(rep("=", 80), "\n")