1 ### EXAMPLE 1: demonstrates how measurement error can lead to a type sign error in a covariate
2 ### What kind of data invalidates fong + tyler?
3 ### Even when you have a good predictor, if it's biased against a covariate you can get the wrong sign.
4 ### Even when you include the proxy variable in the regression.
5 ### But with some ground truth and multiple imputation, you can fix it.
15 library(predictionError)
16 options(amelia.parallel="no",
20 source("simulation_base.R")
23 ### we want to estimate x -> y; x is MAR
24 ### we have x -> k; k -> w; x -> w is used to predict x via the model w.
25 ### A realistic scenario is that we have an NLP model predicting something like "racial harassment" in social media comments
26 ### The labels x are binary, but the model provides a continuous predictor
29 #### how much power do we get from the model in the first place? (sweeping N and m)
32 ## one way to do it is by adding correlation to x.obs and y that isn't in w.
33 ## in other words, the model is missing an important feature of x.obs that's related to y.
34 simulate_data <- function(N, m, B0, Bxy, Bgy, seed, y_explained_variance=0.025, prediction_accuracy=0.73, accuracy_imbalance_difference=0.3){
36 # make w and y dependent
37 g <- rbinom(N, 1, 0.5)
38 x <- rbinom(N, 1, 0.5)
40 y.var.epsilon <- (var(Bgy * g) + var(Bxy *x) + 2*cov(Bgy*g,Bxy*x)) * ((1-y_explained_variance)/y_explained_variance)
41 y.epsilon <- rnorm(N, sd = sqrt(y.var.epsilon))
42 y <- Bgy * g + Bxy * x + y.epsilon
44 df <- data.table(x=x,y=y,g=g)
47 df <- df[sample(nrow(df), m), x.obs := x]
49 df <- df[, x.obs := x]
56 accuracy_imbalance_ratio <- (prediction_accuracy + accuracy_imbalance_difference/2) / (prediction_accuracy - accuracy_imbalance_difference/2)
58 # this works because of conditional probability
59 accuracy_g0 <- prediction_accuracy / (pg*(accuracy_imbalance_ratio) + (1-pg))
60 accuracy_g1 <- accuracy_imbalance_ratio * accuracy_g0
67 dfg0 <- dfg0[sample(ng0, (1-accuracy_g0)*ng0), w_pred := (w_pred-1)**2]
68 dfg1 <- dfg1[sample(ng1, (1-accuracy_g1)*ng1), w_pred := (w_pred-1)**2]
70 df <- rbind(dfg0,dfg1)
72 w <- predict(glm(x ~ w_pred,data=df,family=binomial(link='logit')),type='response')
73 df <- df[,':='(w=w, w_pred = w_pred)]
77 parser <- arg_parser("Simulate data and fit corrected models")
78 parser <- add_argument(parser, "--N", default=5000, help="number of observations of w")
79 parser <- add_argument(parser, "--m", default=200, help="m the number of ground truth observations")
80 parser <- add_argument(parser, "--seed", default=432, help='seed for the rng')
81 parser <- add_argument(parser, "--outfile", help='output file', default='example_2.feather')
82 parser <- add_argument(parser, "--y_explained_variance", help='what proportion of the variance of y can be explained?', default=0.01)
83 parser <- add_argument(parser, "--prediction_accuracy", help='how accurate is the predictive model?', default=0.73)
84 parser <- add_argument(parser, "--accuracy_imbalance_difference", help='how much more accurate is the predictive model for one class than the other?', default=0.3)
86 args <- parse_args(parser)
92 df <- simulate_data(args$N, args$m, B0, Bxy, Bgy, args$seed, args$y_explained_variance, args$prediction_accuracy, args$accuracy_imbalance_difference)
94 result <- list('N'=args$N,'m'=args$m,'B0'=B0,'Bxy'=Bxy,'Bgy'=Bgy, 'seed'=args$seed, 'y_explained_variance'=args$y_explained_variance, 'prediction_accuracy'=args$prediction_accuracy, 'accuracy_imbalance_difference'=args$accuracy_imbalance_difference)
96 outline <- run_simulation_depvar(df=df, result)
99 outfile_lock <- lock(paste0(args$outfile, '_lock'),exclusive=TRUE)
100 if(file.exists(args$outfile)){
101 logdata <- read_feather(args$outfile)
102 logdata <- rbind(logdata,as.data.table(outline))
104 logdata <- as.data.table(outline)
108 write_feather(logdata, args$outfile)