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, Bzx, Bzy, seed, y_explained_variance=0.025, prediction_accuracy=0.73, accuracy_imbalance_difference=0.3){
36 # make w and y dependent
37 z <- rbinom(N, 1, 0.5)
38 x <- rbinom(N, 1, Bzx * z + 0.5)
40 y.var.epsilon <- (var(Bzy * z) + var(Bxy *x) + 2*cov(Bzy*z,Bxy*x)) * ((1-y_explained_variance)/y_explained_variance)
41 y.epsilon <- rnorm(N, sd = sqrt(y.var.epsilon))
42 y <- Bzy * z + Bxy * x + y.epsilon
44 df <- data.table(x=x,y=y,z=z)
47 df <- df[sample(nrow(df), m), x.obs := x]
49 df <- df[, x.obs := x]
53 ## accuracy_imbalance_ratio <- (prediction_accuracy + accuracy_imbalance_difference/2) / (prediction_accuracy - accuracy_imbalance_difference/2)
55 ## # this works because of conditional probability
56 ## accuracy_x0 <- prediction_accuracy / (px*(accuracy_imbalance_ratio) + (1-px))
57 ## accuracy_x1 <- accuracy_imbalance_ratio * accuracy_x0
61 ## nx1 <- nrow(df[x==1])
62 ## nx0 <- nrow(df[x==0])
67 # tranform yz0.1 into a logistic distribution with mean accuracy_z0
68 ## acc.x0 <- plogis(0.5*scale(yx0) + qlogis(accuracy_x0))
69 ## acc.x1 <- plogis(1.5*scale(yx1) + qlogis(accuracy_x1))
71 ## w0x0 <- (1-x0)**2 + (-1)**(1-x0) * acc.x0
72 ## w0x1 <- (1-x1)**2 + (-1)**(1-x1) * acc.x1
74 accuracy_imbalance_ratio <- (prediction_accuracy + accuracy_imbalance_difference/2) / (prediction_accuracy - accuracy_imbalance_difference/2)
76 # this works because of conditional probability
77 accuracy_z0 <- prediction_accuracy / (pz*(accuracy_imbalance_ratio) + (1-pz))
78 accuracy_z1 <- accuracy_imbalance_ratio * accuracy_z0
80 z0x0 <- df[(z==0) & (x==0)]$x
81 z0x1 <- df[(z==0) & (x==1)]$x
82 z1x0 <- df[(z==1) & (x==0)]$x
83 z1x1 <- df[(z==1) & (x==1)]$x
85 yz0x0 <- df[(z==0) & (x==0)]$y
86 yz0x1 <- df[(z==0) & (x==1)]$y
87 yz1x0 <- df[(z==1) & (x==0)]$y
88 yz1x1 <- df[(z==1) & (x==1)]$y
90 nz0x0 <- nrow(df[(z==0) & (x==0)])
91 nz0x1 <- nrow(df[(z==0) & (x==1)])
92 nz1x0 <- nrow(df[(z==1) & (x==0)])
93 nz1x1 <- nrow(df[(z==1) & (x==1)])
98 # tranform yz0.1 into a logistic distribution with mean accuracy_z0
99 acc.z0x0 <- plogis(0.5*scale(yz0x0) + qlogis(accuracy_z0))
100 acc.z0x1 <- plogis(0.5*scale(yz0x1) + qlogis(accuracy_z0))
101 acc.z1x0 <- plogis(1.5*scale(yz1x0) + qlogis(accuracy_z1))
102 acc.z1x1 <- plogis(1.5*scale(yz1x1) + qlogis(accuracy_z1))
104 w0z0x0 <- (1-z0x0)**2 + (-1)**(1-z0x0) * acc.z0x0
105 w0z0x1 <- (1-z0x1)**2 + (-1)**(1-z0x1) * acc.z0x1
106 w0z1x0 <- (1-z1x0)**2 + (-1)**(1-z1x0) * acc.z1x0
107 w0z1x1 <- (1-z1x1)**2 + (-1)**(1-z1x1) * acc.z1x1
109 ##perrorz0 <- w0z0*(pyz0)
110 ##perrorz1 <- w0z1*(pyz1)
112 w0z0x0.noisy.odds <- rlogis(nz0x0,qlogis(w0z0x0))
113 w0z0x1.noisy.odds <- rlogis(nz0x1,qlogis(w0z0x1))
114 w0z1x0.noisy.odds <- rlogis(nz1x0,qlogis(w0z1x0))
115 w0z1x1.noisy.odds <- rlogis(nz1x1,qlogis(w0z1x1))
117 df[(z==0)&(x==0),w:=plogis(w0z0x0.noisy.odds)]
118 df[(z==0)&(x==1),w:=plogis(w0z0x1.noisy.odds)]
119 df[(z==1)&(x==0),w:=plogis(w0z1x0.noisy.odds)]
120 df[(z==1)&(x==1),w:=plogis(w0z1x1.noisy.odds)]
122 df[,w_pred:=as.integer(w > 0.5)]
123 print(mean(df[z==0]$x == df[z==0]$w_pred))
124 print(mean(df[z==1]$x == df[z==1]$w_pred))
125 print(mean(df$w_pred == df$x))
129 parser <- arg_parser("Simulate data and fit corrected models")
130 parser <- add_argument(parser, "--N", default=1400, help="number of observations of w")
131 parser <- add_argument(parser, "--m", default=500, help="m the number of ground truth observations")
132 parser <- add_argument(parser, "--seed", default=50, help='seed for the rng')
133 parser <- add_argument(parser, "--outfile", help='output file', default='example_2.feather')
134 parser <- add_argument(parser, "--y_explained_variance", help='what proportion of the variance of y can be explained?', default=0.01)
135 parser <- add_argument(parser, "--prediction_accuracy", help='how accurate is the predictive model?', default=0.73)
136 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)
137 parser <- add_argument(parser, "--Bzx", help='Effect of z on x', default=0.3)
138 parser <- add_argument(parser, "--Bzy", help='Effect of z on y', default=-0.3)
141 args <- parse_args(parser)
148 df <- simulate_data(args$N, args$m, B0, Bxy, args$Bzx, Bzy, args$seed, args$y_explained_variance, args$prediction_accuracy, args$accuracy_imbalance_difference)
150 result <- list('N'=args$N,'m'=args$m,'B0'=B0,'Bxy'=Bxy, Bzx=args$Bzx, 'Bzy'=Bzy, 'seed'=args$seed, 'y_explained_variance'=args$y_explained_variance, 'prediction_accuracy'=args$prediction_accuracy, 'accuracy_imbalance_difference'=args$accuracy_imbalance_difference, error='')
152 outline <- run_simulation(df, result, outcome_formula=y~x+z, proxy_formula=w_pred~x+z+y+x:y, truth_formula=x~z)
154 outfile_lock <- lock(paste0(args$outfile, '_lock'),exclusive=TRUE)
155 if(file.exists(args$outfile)){
156 logdata <- read_feather(args$outfile)
157 logdata <- rbind(logdata,as.data.table(outline), fill=TRUE)
159 logdata <- as.data.table(outline)
163 write_feather(logdata, args$outfile)