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, Bzy, Bxz=0, seed=0, prediction_accuracy=0.73, z_bias=-0.75){
37 # make w and y dependent
39 x <- rbinom(N,1,plogis(Bxz*z))
41 ystar <- Bzy * z + Bxy * x + B0
42 y <- rbinom(N,1,plogis(ystar))
44 # glm(y ~ x + z, family="binomial")
46 df <- data.table(x=x,y=y,ystar=ystar,z=z)
49 df <- df[sample(nrow(df), m), y.obs := y]
51 df <- df[, y.obs := y]
54 odds.y1 <- qlogis(prediction_accuracy) + z_bias*df[y==1]$z
55 odds.y0 <- qlogis(prediction_accuracy,lower.tail=F) + z_bias*df[y==0]$z
57 df[y==0,w:=plogis(rlogis(.N,odds.y0))]
58 df[y==1,w:=plogis(rlogis(.N,odds.y1))]
60 df[,w_pred := as.integer(w > 0.5)]
62 print(mean(df[x==0]$y == df[x==0]$w_pred))
63 print(mean(df[x==1]$y == df[x==1]$w_pred))
64 print(mean(df$w_pred == df$y))
68 parser <- arg_parser("Simulate data and fit corrected models")
69 parser <- add_argument(parser, "--N", default=1000, help="number of observations of w")
70 parser <- add_argument(parser, "--m", default=500, help="m the number of ground truth observations")
71 parser <- add_argument(parser, "--seed", default=17, help='seed for the rng')
72 parser <- add_argument(parser, "--outfile", help='output file', default='example_4.feather')
73 parser <- add_argument(parser, "--prediction_accuracy", help='how accurate is the predictive model?', default=0.75)
74 ## parser <- add_argument(parser, "--z_bias_y1", help='how is the classifier biased when y = 1?', default=-0.75)
75 ## parser <- add_argument(parser, "--z_bias_y0", help='how is the classifier biased when y = 0 ?', default=0.75)
76 parser <- add_argument(parser, "--z_bias", help='how is the classifier biased?', default=-0.5)
77 parser <- add_argument(parser, "--Bxy", help='coefficient of x on y', default=0.7)
78 parser <- add_argument(parser, "--Bzy", help='coeffficient of z on y', default=-0.7)
79 parser <- add_argument(parser, "--Bzx", help='coeffficient of z on y', default=1)
80 parser <- add_argument(parser, "--B0", help='coeffficient of z on y', default=0)
81 parser <- add_argument(parser, "--outcome_formula", help='formula for the outcome variable', default="y~x+z")
82 parser <- add_argument(parser, "--proxy_formula", help='formula for the proxy variable', default="w_pred~y+z")
83 parser <- add_argument(parser, "--confint_method", help='method for approximating confidence intervals', default='quad')
84 args <- parse_args(parser)
92 df <- simulate_data(args$N, args$m, B0, Bxy, Bzx, Bzy, args$seed, args$prediction_accuracy, args$z_bias)
94 # result <- list('N'=args$N,'m'=args$m,'B0'=B0,'Bxy'=Bxy,'Bzy'=Bzy,'Bzx'=Bzx, 'seed'=args$seed, 'y_explained_variance'=args$y_explained_variance, 'prediction_accuracy'=args$prediction_accuracy, 'z_bias_y0'=args$z_bias_y0,'z_bias_y1'=args$z_bias_y1,'outcome_formula' = args$outcome_formula, 'proxy_formula' = args$proxy_formula)
95 result <- list('N'=args$N,'m'=args$m,'B0'=B0,'Bxy'=Bxy,'Bzy'=Bzy,'Bzx'=Bzx, 'seed'=args$seed, 'y_explained_variance'=args$y_explained_variance, 'prediction_accuracy'=args$prediction_accuracy, 'z_bias'=args$z_bias,'outcome_formula' = args$outcome_formula, 'proxy_formula' = args$proxy_formula, confint_method=args$confint_method)
97 outline <- run_simulation_depvar(df, result, outcome_formula = as.formula(args$outcome_formula), proxy_formula = as.formula(args$proxy_formula),confint_method=args$confint_method)
98 print(outline$error.cor.z)
100 outfile_lock <- lock(paste0(args$outfile, '_lock'),exclusive=TRUE)
102 if(file.exists(args$outfile)){
103 logdata <- read_feather(args$outfile)
104 logdata <- rbind(logdata,as.data.table(outline),fill=TRUE)
106 logdata <- as.data.table(outline)
110 write_feather(logdata, args$outfile)