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){
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]
52 ## probablity of an error is correlated with y
53 p.correct <- plogis(y_bias*scale(y) + qlogis(prediction_accuracy))
55 acc.x0 <- p.correct[df[,x==0]]
56 acc.x1 <- p.correct[df[,x==1]]
58 df[x==0,w:=rlogis(.N,qlogis(1-acc.x0))]
59 df[x==1,w:=rlogis(.N,qlogis(acc.x1))]
61 df[,w_pred := as.integer(w>0.5)]
63 print(mean(df[z==0]$x == df[z==0]$w_pred))
64 print(mean(df[z==1]$x == df[z==1]$w_pred))
65 print(mean(df$w_pred == df$x))
66 print(mean(df[y>=0]$w_pred == df[y>=0]$x))
67 print(mean(df[y<=0]$w_pred == df[y<=0]$x))
72 parser <- arg_parser("Simulate data and fit corrected models")
73 parser <- add_argument(parser, "--N", default=1000, help="number of observations of w")
74 parser <- add_argument(parser, "--m", default=500, help="m the number of ground truth observations")
75 parser <- add_argument(parser, "--seed", default=51, help='seed for the rng')
76 parser <- add_argument(parser, "--outfile", help='output file', default='example_2.feather')
77 parser <- add_argument(parser, "--y_explained_variance", help='what proportion of the variance of y can be explained?', default=0.01)
78 parser <- add_argument(parser, "--prediction_accuracy", help='how accurate is the predictive model?', default=0.73)
79 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)
80 parser <- add_argument(parser, "--Bzx", help='Effect of z on x', default=0.3)
81 parser <- add_argument(parser, "--Bzy", help='Effect of z on y', default=-0.3)
82 parser <- add_argument(parser, "--Bxy", help='Effect of z on y', default=0.3)
83 parser <- add_argument(parser, "--proxy_formula", help='formula for the proxy variable', default="w_pred~x*y")
84 parser <- add_argument(parser, "--y_bias", help='coefficient of y on the probability a classification is correct', default=-0.75)
86 args <- parse_args(parser)
95 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)
97 ## df.pc <- df[,.(x,y,z,w_pred)]
98 ## # df.pc <- df.pc[,err:=x-w_pred]
99 ## pc.df <- pc(suffStat=list(C=cor(df.pc),n=nrow(df.pc)),indepTest=gaussCItest,labels=names(df.pc),alpha=0.05)
102 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,error='')
104 outline <- run_simulation(df, result, outcome_formula=y~x+z, proxy_formula=as.formula(args$proxy_formula), truth_formula=x~z)
106 outfile_lock <- lock(paste0(args$outfile, '_lock'),exclusive=TRUE)
107 if(file.exists(args$outfile)){
108 logdata <- read_feather(args$outfile)
109 logdata <- rbind(logdata,as.data.table(outline), fill=TRUE)
111 logdata <- as.data.table(outline)
115 write_feather(logdata, args$outfile)