X-Git-Url: https://code.communitydata.science/ml_measurement_error_public.git/blobdiff_plain/f8f58301e0285118f7b669a96ed9367a9914ba02..b8d2048cc5338fbd872b55029c3e5d01c739a397:/simulations/02_indep_differential.R

diff --git a/simulations/02_indep_differential.R b/simulations/02_indep_differential.R
index 5a7784b..4e3a132 100644
--- a/simulations/02_indep_differential.R
+++ b/simulations/02_indep_differential.R
@@ -31,77 +31,148 @@ source("simulation_base.R")
 
 ## one way to do it is by adding correlation to x.obs and y that isn't in w.
 ## in other words, the model is missing an important feature of x.obs that's related to y.
-simulate_data <- function(N, m, B0, Bxy, Bgy, Bkx, Bgx, seed, xy.explained.variance=0.01, u.explained.variance=0.1){
+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,Px=0.5,accuracy_imbalance_difference=0.3){
     set.seed(seed)
+    # make w and y dependent
+    z <- rnorm(N,sd=0.5)
+    x <- rbinom(N, 1, plogis(Bzx * z + qlogis(Px)))
 
-    ## the true value of x
+    y.var.epsilon <- (var(Bzy * z) + var(Bxy *x) + 2*cov(Bzy*z,Bxy*x)) * ((1-y_explained_variance)/y_explained_variance)
+    y.epsilon <- rnorm(N, sd = sqrt(y.var.epsilon))
+    y <- Bzy * z + Bxy * x + y.epsilon
+    
+    df <- data.table(x=x,y=y,z=z)
 
-    g <- rbinom(N, 1, 0.5)
+    if(m < N){
+        df <- df[sample(nrow(df), m), x.obs := x]
+    } else {
+        df <- df[, x.obs := x]
+    }
 
-    # make w and y dependent
-    u <- rnorm(N,0,)
+    ## probablity of an error is correlated with y
+    ## pz <- mean(z)
+    ## accuracy_imbalance_ratio <- (prediction_accuracy + accuracy_imbalance_difference/2) / (prediction_accuracy - accuracy_imbalance_difference/2)
 
-    xprime <- Bgx * g + rnorm(N,0,1) 
+    ## # this works because of conditional probability
+    ## accuracy_z0 <- prediction_accuracy / (pz*(accuracy_imbalance_ratio) + (1-pz))
+    ## accuracy_z1 <- accuracy_imbalance_ratio * accuracy_z0
 
-    k <- Bkx*xprime + rnorm(N,0,1.5) + 1.1*Bkx*u
+    ## z0x0 <- df[(z==0) & (x==0)]$x
+    ## z0x1 <- df[(z==0) & (x==1)]$x
+    ## z1x0 <- df[(z==1) & (x==0)]$x
+    ## z1x1 <- df[(z==1) & (x==1)]$x
 
-    x <- as.integer(logistic(scale(xprime)) > 0.5)
+    ## yz0x0 <- df[(z==0) & (x==0)]$y
+    ## yz0x1 <- df[(z==0) & (x==1)]$y
+    ## yz1x0 <- df[(z==1) & (x==0)]$y
+    ## yz1x1 <- df[(z==1) & (x==1)]$y
 
-    y <-  Bxy * x  + Bgy * g + B0 + u + rnorm(N, 0, 1)
+    ## nz0x0 <- nrow(df[(z==0) & (x==0)])
+    ## nz0x1 <- nrow(df[(z==0) & (x==1)])
+    ## nz1x0 <- nrow(df[(z==1) & (x==0)])
+    ## nz1x1 <- nrow(df[(z==1) & (x==1)])
 
-    df <- data.table(x=x,k=k,y=y,g=g)
+    ## yz1 <- df[z==1]$y 
+    ## yz1 <- df[z==1]$y 
 
-    w.model <- glm(x ~ k,df, family=binomial(link='logit')) 
+    ## # tranform yz0.1 into a logistic distribution with mean accuracy_z0
+    ## acc.z0x0 <- plogis(0.5*scale(yz0x0) + qlogis(accuracy_z0))
+    ## acc.z0x1 <- plogis(0.5*scale(yz0x1) + qlogis(accuracy_z0))
+    ## acc.z1x0 <- plogis(1.5*scale(yz1x0) + qlogis(accuracy_z1))
+    ## acc.z1x1 <- plogis(1.5*scale(yz1x1) + qlogis(accuracy_z1))
 
-    if( m < N){
-        df <- df[sample(nrow(df), m), x.obs := x]
-    } else {
-        df <- df[, x.obs := x]
-    }
+    ## w0z0x0 <- (1-z0x0)**2 + (-1)**(1-z0x0) * acc.z0x0
+    ## w0z0x1 <- (1-z0x1)**2 + (-1)**(1-z0x1) * acc.z0x1
+    ## w0z1x0 <- (1-z1x0)**2 + (-1)**(1-z1x0) * acc.z1x0
+    ## w0z1x1 <- (1-z1x1)**2 + (-1)**(1-z1x1) * acc.z1x1
 
-    df[, x.obs := x.obs]
+    ## ##perrorz0 <- w0z0*(pyz0)
+    ## ##perrorz1 <- w0z1*(pyz1)
 
-    w <- predict(w.model, df) + rnorm(N, 0, 1)
-    ## y = B0 + B1x + e
+    ## w0z0x0.noisy.odds <- rlogis(nz0x0,qlogis(w0z0x0))
+    ## w0z0x1.noisy.odds <- rlogis(nz0x1,qlogis(w0z0x1))
+    ## w0z1x0.noisy.odds <- rlogis(nz1x0,qlogis(w0z1x0))
+    ## w0z1x1.noisy.odds <- rlogis(nz1x1,qlogis(w0z1x1))
 
-    df[,':='(w=w, w_pred = as.integer(w>0.5),u=u)]
-    return(df)
-}
+    ## df[(z==0)&(x==0),w:=plogis(w0z0x0.noisy.odds)]
+    ## df[(z==0)&(x==1),w:=plogis(w0z0x1.noisy.odds)]    
+    ## df[(z==1)&(x==0),w:=plogis(w0z1x0.noisy.odds)]    
+    ## df[(z==1)&(x==1),w:=plogis(w0z1x1.noisy.odds)]    
 
-schennach <- function(df){
+    ## df[,w_pred:=as.integer(w > 0.5)]
+    ## print(mean(df[z==0]$x == df[z==0]$w_pred))
+    ## print(mean(df[z==1]$x == df[z==1]$w_pred))
+    ## print(mean(df$w_pred == df$x))
 
-    fwx <- glm(x.obs~w, df, family=binomial(link='logit'))
-    df[,xstar_pred := predict(fwx, df)]
-    gxt <- lm(y ~ xstar_pred+g, df)
 
-}
+    resids <- resid(lm(y~x + z))
+    odds.x1 <- qlogis(prediction_accuracy) + y_bias*qlogis(pnorm(resids[x==1])) + z_bias * qlogis(pnorm(z[x==1],sd(z)))
+    odds.x0 <- qlogis(prediction_accuracy,lower.tail=F) + y_bias*qlogis(pnorm(resids[x==0])) + z_bias * qlogis(pnorm(z[x==0],sd(z)))
 
+    ## acc.x0 <- p.correct[df[,x==0]]
+    ## acc.x1 <- p.correct[df[,x==1]]
+
+    df[x==0,w:=plogis(rlogis(.N,odds.x0))]
+    df[x==1,w:=plogis(rlogis(.N,odds.x1))]
+
+    df[,w_pred := as.integer(w > 0.5)]
+
+
+    print(mean(df$w_pred == df$x))
+    print(mean(df[y>=0]$w_pred == df[y>=0]$x))
+    print(mean(df[y<=0]$w_pred == df[y<=0]$x))
+    return(df)
+}
 
 parser <- arg_parser("Simulate data and fit corrected models")
 parser <- add_argument(parser, "--N", default=5000, help="number of observations of w")
-parser <- add_argument(parser, "--m", default=200, help="m the number of ground truth observations")
-parser <- add_argument(parser, "--seed", default=432, help='seed for the rng')
+parser <- add_argument(parser, "--m", default=500, help="m the number of ground truth observations")
+parser <- add_argument(parser, "--seed", default=51, help='seed for the rng')
 parser <- add_argument(parser, "--outfile", help='output file', default='example_2.feather')
+parser <- add_argument(parser, "--y_explained_variance", help='what proportion of the variance of y can be explained?', default=0.1)
+parser <- add_argument(parser, "--prediction_accuracy", help='how accurate is the predictive model?', default=0.75)
+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)
+parser <- add_argument(parser, "--Bzx", help='Effect of z on x', default=0.3)
+parser <- add_argument(parser, "--Bzy", help='Effect of z on y', default=-0.3)
+parser <- add_argument(parser, "--Bxy", help='Effect of z on y', default=0.3)
+parser <- add_argument(parser, "--outcome_formula", help='formula for the outcome variable', default="y~x+z")
+parser <- add_argument(parser, "--proxy_formula", help='formula for the proxy variable', default="w_pred~y*z*x")
+parser <- add_argument(parser, "--y_bias", help='coefficient of y on the probability a classification is correct', default=-0.5)
+parser <- add_argument(parser, "--z_bias", help='coefficient of z on the probability a classification is correct', default=0)
+parser <- add_argument(parser, "--truth_formula", help='formula for the true variable', default="x~z")
+parser <- add_argument(parser, "--Px", help='base rate of x', default=0.5)
+
 args <- parse_args(parser)
 
 B0 <- 0
-Bxy <- 0.2
-Bgy <- 0
-Bkx <- 2
-Bgx <- 0
+Px <- args$Px
+Bxy <- args$Bxy
+Bzy <- args$Bzy
+Bzx <- args$Bzx
 
+if(args$m < args$N){
 
-outline <- run_simulation(simulate_data(args$N, args$m, B0, Bxy, Bgy, Bkx, Bgx, args$seed)
-                         ,list('N'=args$N,'m'=args$m,'B0'=B0,'Bxy'=Bxy,'Bgy'=Bgy, 'Bkx'=Bkx, 'Bgx'=Bgx, 'seed'=args$seed))
+    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)
 
-outfile_lock <- lock(paste0(args$outfile, '_lock'),exclusive=TRUE)
-if(file.exists(args$outfile)){
-    logdata <- read_feather(args$outfile)
-    logdata <- rbind(logdata,as.data.table(outline))
-} else {
-    logdata <- as.data.table(outline)
-}
+    ## df.pc <- df[,.(x,y,z,w_pred,w)]
+    ##                                     #    df.pc <- df.pc[,err:=x-w_pred]
+    ## pc.df <- pc(suffStat=list(C=cor(df.pc),n=nrow(df.pc)),indepTest=gaussCItest,labels=names(df.pc),alpha=0.05)
+    ## plot(pc.df)
+
+    result <- list('N'=args$N,'m'=args$m,'B0'=B0,'Bxy'=Bxy, 'Bzx'=args$Bzx, 'Bzy'=Bzy, 'Px'=Px, '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='')
+
+    outline <- run_simulation(df, result, outcome_formula=as.formula(args$outcome_formula), proxy_formula=as.formula(args$proxy_formula), truth_formula=as.formula(args$truth_formula))
+    
+   
+ outfile_lock <- lock(paste0(args$outfile, '_lock'),exclusive=TRUE)
+    if(file.exists(args$outfile)){
+        logdata <- read_feather(args$outfile)
+        logdata <- rbind(logdata,as.data.table(outline), fill=TRUE)
+    } else {
+        logdata <- as.data.table(outline)
+    }
 
-print(outline)
-write_feather(logdata, args$outfile)
-unlock(outfile_lock)
+    print(outline)
+    write_feather(logdata, args$outfile)
+    unlock(outfile_lock)
+}