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, y_bias=-0.8,z_bias=0,accuracy_imbalance_difference=0.3){
36 # make w and y dependent
38 x <- rbinom(N, 1, plogis(Bzx * z))
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]
52 ## probablity of an error is correlated with y
54 ## accuracy_imbalance_ratio <- (prediction_accuracy + accuracy_imbalance_difference/2) / (prediction_accuracy - accuracy_imbalance_difference/2)
56 ## # this works because of conditional probability
57 ## accuracy_z0 <- prediction_accuracy / (pz*(accuracy_imbalance_ratio) + (1-pz))
58 ## accuracy_z1 <- accuracy_imbalance_ratio * accuracy_z0
60 ## z0x0 <- df[(z==0) & (x==0)]$x
61 ## z0x1 <- df[(z==0) & (x==1)]$x
62 ## z1x0 <- df[(z==1) & (x==0)]$x
63 ## z1x1 <- df[(z==1) & (x==1)]$x
65 ## yz0x0 <- df[(z==0) & (x==0)]$y
66 ## yz0x1 <- df[(z==0) & (x==1)]$y
67 ## yz1x0 <- df[(z==1) & (x==0)]$y
68 ## yz1x1 <- df[(z==1) & (x==1)]$y
70 ## nz0x0 <- nrow(df[(z==0) & (x==0)])
71 ## nz0x1 <- nrow(df[(z==0) & (x==1)])
72 ## nz1x0 <- nrow(df[(z==1) & (x==0)])
73 ## nz1x1 <- nrow(df[(z==1) & (x==1)])
78 ## # tranform yz0.1 into a logistic distribution with mean accuracy_z0
79 ## acc.z0x0 <- plogis(0.5*scale(yz0x0) + qlogis(accuracy_z0))
80 ## acc.z0x1 <- plogis(0.5*scale(yz0x1) + qlogis(accuracy_z0))
81 ## acc.z1x0 <- plogis(1.5*scale(yz1x0) + qlogis(accuracy_z1))
82 ## acc.z1x1 <- plogis(1.5*scale(yz1x1) + qlogis(accuracy_z1))
84 ## w0z0x0 <- (1-z0x0)**2 + (-1)**(1-z0x0) * acc.z0x0
85 ## w0z0x1 <- (1-z0x1)**2 + (-1)**(1-z0x1) * acc.z0x1
86 ## w0z1x0 <- (1-z1x0)**2 + (-1)**(1-z1x0) * acc.z1x0
87 ## w0z1x1 <- (1-z1x1)**2 + (-1)**(1-z1x1) * acc.z1x1
89 ## ##perrorz0 <- w0z0*(pyz0)
90 ## ##perrorz1 <- w0z1*(pyz1)
92 ## w0z0x0.noisy.odds <- rlogis(nz0x0,qlogis(w0z0x0))
93 ## w0z0x1.noisy.odds <- rlogis(nz0x1,qlogis(w0z0x1))
94 ## w0z1x0.noisy.odds <- rlogis(nz1x0,qlogis(w0z1x0))
95 ## w0z1x1.noisy.odds <- rlogis(nz1x1,qlogis(w0z1x1))
97 ## df[(z==0)&(x==0),w:=plogis(w0z0x0.noisy.odds)]
98 ## df[(z==0)&(x==1),w:=plogis(w0z0x1.noisy.odds)]
99 ## df[(z==1)&(x==0),w:=plogis(w0z1x0.noisy.odds)]
100 ## df[(z==1)&(x==1),w:=plogis(w0z1x1.noisy.odds)]
102 ## df[,w_pred:=as.integer(w > 0.5)]
103 ## print(mean(df[z==0]$x == df[z==0]$w_pred))
104 ## print(mean(df[z==1]$x == df[z==1]$w_pred))
105 ## print(mean(df$w_pred == df$x))
108 resids <- resid(lm(y~x + z))
109 odds.x1 <- qlogis(prediction_accuracy) + y_bias*qlogis(pnorm(resids[x==1])) + z_bias * qlogis(pnorm(z[x==1],sd(z)))
110 odds.x0 <- qlogis(prediction_accuracy,lower.tail=F) + y_bias*qlogis(pnorm(resids[x==0])) + z_bias * qlogis(pnorm(z[x==0],sd(z)))
112 ## acc.x0 <- p.correct[df[,x==0]]
113 ## acc.x1 <- p.correct[df[,x==1]]
115 df[x==0,w:=plogis(rlogis(.N,odds.x0))]
116 df[x==1,w:=plogis(rlogis(.N,odds.x1))]
118 df[,w_pred := as.integer(w > 0.5)]
121 print(mean(df$w_pred == df$x))
122 print(mean(df[y>=0]$w_pred == df[y>=0]$x))
123 print(mean(df[y<=0]$w_pred == df[y<=0]$x))
127 parser <- arg_parser("Simulate data and fit corrected models")
128 parser <- add_argument(parser, "--N", default=5000, help="number of observations of w")
129 parser <- add_argument(parser, "--m", default=500, help="m the number of ground truth observations")
130 parser <- add_argument(parser, "--seed", default=51, help='seed for the rng')
131 parser <- add_argument(parser, "--outfile", help='output file', default='example_2.feather')
132 parser <- add_argument(parser, "--y_explained_variance", help='what proportion of the variance of y can be explained?', default=0.1)
133 parser <- add_argument(parser, "--prediction_accuracy", help='how accurate is the predictive model?', default=0.75)
134 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)
135 parser <- add_argument(parser, "--Bzx", help='Effect of z on x', default=0.3)
136 parser <- add_argument(parser, "--Bzy", help='Effect of z on y', default=-0.3)
137 parser <- add_argument(parser, "--Bxy", help='Effect of z on y', default=0.3)
138 parser <- add_argument(parser, "--outcome_formula", help='formula for the outcome variable', default="y~x+z")
139 parser <- add_argument(parser, "--proxy_formula", help='formula for the proxy variable', default="w_pred~y*z*x")
140 parser <- add_argument(parser, "--y_bias", help='coefficient of y on the probability a classification is correct', default=-0.5)
141 parser <- add_argument(parser, "--z_bias", help='coefficient of z on the probability a classification is correct', default=0)
142 parser <- add_argument(parser, "--truth_formula", help='formula for the true variable', default="x~z")
144 args <- parse_args(parser)
153 df <- simulate_data(args$N, args$m, B0, Bxy, Bzx, Bzy, args$seed, args$y_explained_variance, args$prediction_accuracy, y_bias=args$y_bias)
155 ## df.pc <- df[,.(x,y,z,w_pred,w)]
156 ## # df.pc <- df.pc[,err:=x-w_pred]
157 ## pc.df <- pc(suffStat=list(C=cor(df.pc),n=nrow(df.pc)),indepTest=gaussCItest,labels=names(df.pc),alpha=0.05)
160 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$accuracy_imbalance_difference, 'y_bias'=args$y_bias,'outcome_formula'=args$outcome_formula, 'proxy_formula'=args$proxy_formula,truth_formula=args$truth_formula, error='')
162 outline <- run_simulation(df, result, outcome_formula=as.formula(args$outcome_formula), proxy_formula=as.formula(args$proxy_formula), truth_formula=as.formula(args$truth_formula))
165 outfile_lock <- lock(paste0(args$outfile, '_lock'),exclusive=TRUE)
166 if(file.exists(args$outfile)){
167 logdata <- read_feather(args$outfile)
168 logdata <- rbind(logdata,as.data.table(outline), fill=TRUE)
170 logdata <- as.data.table(outline)
174 write_feather(logdata, args$outfile)