Created
January 21, 2020 21:38
-
-
Save nstrayer/7af117b3b423335fc73f7fc51b9cf530 to your computer and use it in GitHub Desktop.
R Script to simulate missing not at random data and look at performance of different imputation strategies.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
library(tidyverse) | |
n <- 150 | |
sensitivity_threshold <- 5 | |
data <- tibble( | |
a = rgamma(n = n, shape = 5, rate = 0.5), | |
b = rgamma(n = n, shape = a/2, rate = 0.5) | |
) | |
generate_missing_data <- function(i){ | |
data_w_missing <- data %>% | |
mutate(obs = 1:n()) %>% | |
pivot_longer(c(a, b), names_to = "variable") %>% | |
mutate( | |
distance_from_threshold = sensitivity_threshold - value, | |
prob_of_missing = ifelse(distance_from_threshold > 0, | |
distance_from_threshold/sensitivity_threshold, | |
0), | |
missing = as.logical(rbinom(n = n(), size = 1, prob = prob_of_missing)), | |
value = ifelse(missing, NA, value) | |
) %>% | |
pivot_wider( | |
id_cols = obs, | |
names_from = variable, | |
values_from = value | |
) %>% | |
mutate(type = "no impute") | |
bind_rows( | |
data_w_missing, | |
mutate(data_w_missing, | |
a = ifelse(is.na(a), 0, a), | |
b = ifelse(is.na(b), 0, b), | |
type = "impute zero"), | |
mutate(data_w_missing, | |
a = ifelse(is.na(a), mean(a, na.rm = TRUE), a), | |
b = ifelse(is.na(b), mean(b, na.rm = TRUE), b), | |
type = "impute mean"), | |
mutate(data_w_missing, | |
a = ifelse(is.na(a), median(a, na.rm = TRUE), a), | |
b = ifelse(is.na(b), median(b, na.rm = TRUE), b), | |
type = "impute median"), | |
mutate(data_w_missing, | |
a = ifelse(is.na(a), min(a, na.rm = TRUE), a), | |
b = ifelse(is.na(b), min(b, na.rm = TRUE), b), | |
type = "impute min") | |
) %>% | |
mutate(i = i) | |
} | |
num_sims <- 100 | |
correlations <- 1:num_sims %>% | |
map_dfr(generate_missing_data) %>% | |
group_by(type, i) %>% | |
summarise( | |
pearson = cor(a,b, use = "complete.obs"), | |
spearman = cor(a,b, method = "spearman", use = "complete.obs") | |
) %>% | |
pivot_longer(c(pearson, spearman), names_to = "correlation_type") | |
true_correlations <- tibble( | |
correlation_type = c("pearson", "spearman"), | |
truth = c(cor(data$a, data$b), cor(data$a, data$b, method = "spearman"))) | |
ggplot(correlations, aes(x = value)) + | |
geom_density(fill = "steelblue", color = "steelblue") + | |
facet_grid(type~correlation_type) + | |
geom_vline( | |
data = true_correlations, | |
aes(xintercept = truth), | |
color = "black" | |
) + | |
labs(title = "Distribution of correlations by imputation type", | |
subtitle = "Missingness structure is related to distance from minimum threashold\nVertical line is true correlation value", | |
x = "correlation") | |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment