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, seed, 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, 0.5)
40 ystar <- Bzy * z + Bxy * x
41 y <- rbinom(N,1,plogis(ystar))
43 # glm(y ~ x + z, family="binomial")
45 df <- data.table(x=x,y=y,ystar=ystar,z=z)
48 df <- df[sample(nrow(df), m), y.obs := y]
50 df <- df[, y.obs := y]
57 accuracy_imbalance_ratio <- (prediction_accuracy + accuracy_imbalance_difference/2) / (prediction_accuracy - accuracy_imbalance_difference/2)
59 # this works because of conditional probability
60 accuracy_z0 <- prediction_accuracy / (pz*(accuracy_imbalance_ratio) + (1-pz))
61 accuracy_z1 <- accuracy_imbalance_ratio * accuracy_z0
69 acc_z0 <- plogis(0.7*scale(yz0) + qlogis(accuracy_z0))
70 acc_z1 <- plogis(1.3*scale(yz1) + qlogis(accuracy_z1))
72 w0z0 <- (1-yz0)**2 + (-1)**(1-yz0) * acc_z0
73 w0z1 <- (1-yz1)**2 + (-1)**(1-yz1) * acc_z1
75 w0z0.noisy.odds <- rlogis(nz0,qlogis(w0z0))
76 w0z1.noisy.odds <- rlogis(nz1,qlogis(w0z1))
77 df[z==0,w:=plogis(w0z0.noisy.odds)]
78 df[z==1,w:=plogis(w0z1.noisy.odds)]
80 df[,w_pred:=as.integer(w > 0.5)]
82 print(mean(df[y==0]$y == df[y==0]$w_pred))
83 print(mean(df[y==1]$y == df[y==1]$w_pred))
84 print(mean(df$w_pred == df$y))
89 parser <- arg_parser("Simulate data and fit corrected models")
90 parser <- add_argument(parser, "--N", default=1000, help="number of observations of w")
91 parser <- add_argument(parser, "--m", default=500, help="m the number of ground truth observations")
92 parser <- add_argument(parser, "--seed", default=17, help='seed for the rng')
93 parser <- add_argument(parser, "--outfile", help='output file', default='example_2.feather')
94 parser <- add_argument(parser, "--y_explained_variance", help='what proportion of the variance of y can be explained?', default=0.005)
95 parser <- add_argument(parser, "--prediction_accuracy", help='how accurate is the predictive model?', default=0.73)
96 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)
98 args <- parse_args(parser)
105 df <- simulate_data(args$N, args$m, B0, Bxy, Bzy, args$seed, args$prediction_accuracy, args$accuracy_imbalance_difference)
107 result <- list('N'=args$N,'m'=args$m,'B0'=B0,'Bxy'=Bxy,'Bzy'=Bzy, 'seed'=args$seed, 'y_explained_variance'=args$y_explained_variance, 'prediction_accuracy'=args$prediction_accuracy, 'accuracy_imbalance_difference'=args$accuracy_imbalance_difference)
109 outline <- run_simulation_depvar(df, result, outcome_formula = y ~ x + z, proxy_formula = w_pred ~ y*x + y*z + z*x)
111 outfile_lock <- lock(paste0(args$outfile, '_lock'),exclusive=TRUE)
113 if(file.exists(args$outfile)){
114 logdata <- read_feather(args$outfile)
115 logdata <- rbind(logdata,as.data.table(outline),fill=TRUE)
117 logdata <- as.data.table(outline)
121 write_feather(logdata, args$outfile)