1 ### EXAMPLE 2_b: demonstrates how measurement error can lead to a type sign error in a covariate
2 ### What kind of data invalidates fong + tyler?
3 ### This is the same as example 2, only instead of x->k we have k->x.
4 ### Even when you have a good predictor, if it's biased against a covariate you can get the wrong sign.
5 ### Even when you include the proxy variable in the regression.
6 ### But with some ground truth and multiple imputation, you can fix it.
16 library(predictionError)
17 options(amelia.parallel="no",
21 source("simulation_base.R")
24 ### we want to estimate x -> y; x is MAR
25 ### we have x -> k; k -> w; x -> w is used to predict x via the model w.
26 ### A realistic scenario is that we have an NLP model predicting something like "racial harassment" in social media comments
27 ### The labels x are binary, but the model provides a continuous predictor
30 #### how much power do we get from the model in the first place? (sweeping N and m)
33 ## one way to do it is by adding correlation to x.obs and y that isn't in w.
34 ## in other words, the model is missing an important feature of x.obs that's related to y.
35 simulate_latent_cocause <- function(N, m, B0, Bxy, Bgy, Bkx, Bgx, seed){
38 ## the true value of x
40 g <- rbinom(N, 1, 0.5)
42 # make w and y dependent
45 xprime <- Bgx * g + rnorm(N,0,1)
47 k <- Bkx*xprime + rnorm(N,0,1.5) + 1.1*Bkx*u
49 x <- as.integer(logistic(scale(xprime)) > 0.5)
51 y <- Bxy * x + Bgy * g + rnorm(N, 0, 1) + B0 + u
53 df <- data.table(x=x,k=k,y=y,g=g)
55 w.model <- glm(x ~ k,df, family=binomial(link='logit'))
58 df <- df[sample(nrow(df), m), x.obs := x]
60 df <- df[, x.obs := x]
65 w <- predict(w.model, df) + rnorm(N, 0, 1)
68 df[,':='(w=w, w_pred = as.integer(w>0.5),u=u)]
72 ## simulate_latent_cocause_2 <- function(N, m, B0, Bxy, Bgy, Bkx, Bgx, seed){
75 ## ## the true value of x
77 ## g <- rbinom(N, 1, 0.5)
79 ## # make w and y dependent
82 ## xprime <- Bgx * g + rnorm(N,0,1)
84 ## k <- Bkx*xprime + rnorm(N,0,3)
86 ## x <- as.integer(logistic(scale(xprime+0.3)) > 0.5)
88 ## y <- Bxy * x + Bgy * g + rnorm(N, 0, 1) + B0 + u
90 ## df <- data.table(x=x,k=k,y=y,g=g)
92 ## w.model <- glm(x ~ k, df, family=binomial(link='logit'))
95 ## df <- df[sample(nrow(df), m), x.obs := x]
97 ## df <- df[, x.obs := x]
100 ## w <- predict(w.model,data.frame(k=k)) + u
101 ## ## y = B0 + B1x + e
103 ## df[,':='(w=w, w_pred = as.integer(w>0.5),u=u)]
108 schennach <- function(df){
110 fwx <- glm(x.obs~w, df, family=binomial(link='logit'))
111 df[,xstar_pred := predict(fwx, df)]
112 gxt <- lm(y ~ xstar_pred+g, df)
117 parser <- arg_parser("Simulate data and fit corrected models")
118 parser <- add_argument(parser, "--N", default=5000, help="number of observations of w")
119 parser <- add_argument(parser, "--m", default=200, help="m the number of ground truth observations")
120 parser <- add_argument(parser, "--seed", default=432, help='seed for the rng')
121 parser <- add_argument(parser, "--outfile", help='output file', default='example_2.feather')
122 args <- parse_args(parser)
131 outline <- run_simulation(simulate_latent_cocause(args$N, args$m, B0, Bxy, Bgy, Bkx, Bgx, args$seed)
132 ,list('N'=args$N,'m'=args$m,'B0'=B0,'Bxy'=Bxy,'Bgy'=Bgy, 'Bkx'=Bkx, 'Bgx'=Bgx, 'seed'=args$seed))
134 outfile_lock <- lock(paste0(args$outfile, '_lock'),exclusive=TRUE)
135 if(file.exists(args$outfile)){
136 logdata <- read_feather(args$outfile)
137 logdata <- rbind(logdata,as.data.table(outline))
139 logdata <- as.data.table(outline)
143 write_feather(logdata, args$outfile)