1 ### EXAMPLE 2_b: demonstrates how measurement error can lead to a type sign error in a covariate
2 ### This is the same as example 2, only instead of x->k we have k->x.
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",
19 source("simulation_base.R")
22 ### we want to estimate x -> y; x is MAR
23 ### we have x -> k; k -> w; x -> w is used to predict x via the model w.
24 ### A realistic scenario is that we have an NLP model predicting something like "racial harassment" in social media comments
25 ### The labels x are binary, but the model provides a continuous predictor
28 #### how much power do we get from the model in the first place? (sweeping N and m)
30 logistic <- function(x) {1/(1+exp(-1*x))}
32 simulate_latent_cocause <- function(N, m, B0, Bxy, Bgy, Bkx, Bgx, seed){
35 ## the true value of x
37 g <- rbinom(N, 1, 0.5)
38 xprime <- Bgx * g + rnorm(N,0,1)
40 k <- Bkx*xprime + rnorm(N,0,3)
42 x <- as.integer(logistic(scale(xprime)) > 0.5)
44 y <- Bxy * x + Bgy * g + rnorm(N, 0, 2) + B0
45 df <- data.table(x=x,k=k,y=y,g=g)
48 df <- df[sample(nrow(df), m), x.obs := x]
50 df <- df[, x.obs := x]
53 w.model <- glm(x ~ k,df, family=binomial(link='logit'))
54 w <- predict(w.model,data.frame(k=k)) + rnorm(N,0,1)
57 df[,':='(w=w, w_pred = as.integer(w>0.5))]
61 schennach <- function(df){
63 fwx <- glm(x.obs~w, df, family=binomial(link='logit'))
64 df[,xstar_pred := predict(fwx, df)]
65 gxt <- lm(y ~ xstar_pred+g, df)
69 parser <- arg_parser("Simulate data and fit corrected models")
70 parser <- add_argument(parser, "--N", default=100, help="number of observations of w")
71 parser <- add_argument(parser, "--m", default=20, help="m the number of ground truth observations")
72 parser <- add_argument(parser, "--seed", default=4321, help='seed for the rng')
73 parser <- add_argument(parser, "--outfile", help='output file', default='example_2_B.feather')
74 args <- parse_args(parser)
84 outline <- run_simulation(simulate_latent_cocause(args$N, args$m, B0, Bxy, Bgy, Bkx, Bgx, args$seed)
85 ,list('N'=args$N,'m'=args$m,'B0'=B0,'Bxy'=Bxy,'Bgy'=Bgy, 'Bkx'=Bkx, 'Bgx'=Bgx, 'seed'=args$seed))
87 outfile_lock <- lock(paste0(args$outfile, '_lock'),exclusive=TRUE)
88 if(file.exists(args$outfile)){
89 logdata <- read_feather(args$outfile)
90 logdata <- rbind(logdata,as.data.table(outline))
92 logdata <- as.data.table(outline)
96 write_feather(logdata, args$outfile)
99 ## Ns <- c(1e6, 5e4, 1000)
100 ## ms <- c(100, 250, 500, 1000)
103 ## rowssets <- list()
104 ## library(doParallel)
105 ## options(mc.cores = parallel::detectCores())
106 ## cl <- makeCluster(20)
107 ## registerDoParallel(cl)
109 ## ## library(future)
111 ## ## plan(multiprocess,workers=40,gc=TRUE)
117 ## new.rows <- foreach(iter=seeds, .combine=rbind, .packages = c('mecor','Amelia','Zelig','predictionError','data.table'),
118 ## .export = c("run_simulation","simulate_latent_cocause","logistic","N","m","B0","Bxy","Bgy","Bkx","Bgx")) %dopar%
120 ## {run_simulation(simulate_latent_cocause(N, m, B0, Bxy, Bgy, Bkx, Bgx, iter)
121 ## ,list('N'=N,'m'=m,'B0'=B0,'Bxy'=Bxy,'Bgy'=Bgy, 'Bkx'=Bkx, 'Bgx'=Bgx, 'seed'=iter))}
122 ## rowsets <- append(rowssets, list(data.table(new.rows)))
126 ## ## rows <- append(rows, list(future({run_simulation(simulate_latent_cocause(N, m, B0, Bxy, Bgy, Bkx, Bgx, seed)
128 ## ,list(N=N,m=m,B0=B0,Bxy=Bxy,Bgy=Bgy, Bkx=Bkx, Bgx=Bgx, seed=seed))w},
129 ## packages=c('mecor','Amelia','Zelig','predictionError'),
133 ## df <- rbindlist(rowsets)
135 ## write_feather(df,"example_2B_simulation.feather")
137 ## run_simulation <- function(N, m, B0, Bxy, Bgy, Bkx, Bgx, seed){
139 ## df <- simulate_latent_cocause(N, m, B0, Bxy, Bgy, Bkx, Bgx, seed)
141 ## result <- append(result, list(N=N,
149 ## (accuracy <- df[,.(mean(w_pred==x))])
150 ## result <- append(result, list(accuracy=accuracy))
152 ## (model.true <- lm(y ~ x + g, data=df))
153 ## true.ci.Bxy <- confint(model.true)['x',]
154 ## true.ci.Bgy <- confint(model.true)['g',]
156 ## result <- append(result, list(Bxy.est.true=coef(model.true)['x'],
157 ## Bgy.est.true=coef(model.true)['g'],
158 ## Bxy.ci.upper.true = true.ci.Bxy[2],
159 ## Bxy.ci.lower.true = true.ci.Bxy[1],
160 ## Bgy.ci.upper.true = true.ci.Bgy[2],
161 ## Bgy.ci.lower.true = true.ci.Bgy[1]))
163 ## (model.feasible <- lm(y~x.obs+g,data=df))
165 ## feasible.ci.Bxy <- confint(model.feasible)['x.obs',]
166 ## result <- append(result, list(Bxy.est.feasible=coef(model.feasible)['x.obs'],
167 ## Bxy.ci.upper.feasible = feasible.ci.Bxy[2],
168 ## Bxy.ci.lower.feasible = feasible.ci.Bxy[1]))
170 ## feasible.ci.Bgy <- confint(model.feasible)['g',]
171 ## result <- append(result, list(Bgy.est.feasible=coef(model.feasible)['g'],
172 ## Bgy.ci.upper.feasible = feasible.ci.Bgy[2],
173 ## Bgy.ci.lower.feasible = feasible.ci.Bgy[1]))
175 ## (model.naive <- lm(y~w+g, data=df))
177 ## naive.ci.Bxy <- confint(model.naive)['w',]
178 ## naive.ci.Bgy <- confint(model.naive)['g',]
180 ## result <- append(result, list(Bxy.est.naive=coef(model.naive)['w'],
181 ## Bgy.est.naive=coef(model.naive)['g'],
182 ## Bxy.ci.upper.naive = naive.ci.Bxy[2],
183 ## Bxy.ci.lower.naive = naive.ci.Bxy[1],
184 ## Bgy.ci.upper.naive = naive.ci.Bgy[2],
185 ## Bgy.ci.lower.naive = naive.ci.Bgy[1]))
188 ## ## multiple imputation when k is observed
189 ## ## amelia does great at this one.
190 ## amelia.out.k <- amelia(df, m=200, p2s=0, idvars=c('x','w_pred'),noms=c("x.obs","g"),lgstc=c('w'))
191 ## mod.amelia.k <- zelig(y~x.obs+g, model='ls', data=amelia.out.k$imputations, cite=FALSE)
192 ## (coefse <- combine_coef_se(mod.amelia.k, messages=FALSE))
194 ## est.x.mi <- coefse['x.obs','Estimate']
195 ## est.x.se <- coefse['x.obs','Std.Error']
196 ## result <- append(result,
197 ## list(Bxy.est.amelia.full = est.x.mi,
198 ## Bxy.ci.upper.amelia.full = est.x.mi + 1.96 * est.x.se,
199 ## Bxy.ci.lower.amelia.full = est.x.mi - 1.96 * est.x.se
202 ## est.g.mi <- coefse['g','Estimate']
203 ## est.g.se <- coefse['g','Std.Error']
205 ## result <- append(result,
206 ## list(Bgy.est.amelia.full = est.g.mi,
207 ## Bgy.ci.upper.amelia.full = est.g.mi + 1.96 * est.g.se,
208 ## Bgy.ci.lower.amelia.full = est.g.mi - 1.96 * est.g.se
211 ## ## What if we can't observe k -- most realistic scenario. We can't include all the ML features in a model.
212 ## amelia.out.nok <- amelia(df, m=200, p2s=0, idvars=c("x","k"), noms=c("x.obs",'g'),lgstc = c("w"))
213 ## mod.amelia.nok <- zelig(y~x.obs+g, model='ls', data=amelia.out.nok$imputations, cite=FALSE)
214 ## (coefse <- combine_coef_se(mod.amelia.nok, messages=FALSE))
216 ## est.x.mi <- coefse['x.obs','Estimate']
217 ## est.x.se <- coefse['x.obs','Std.Error']
218 ## result <- append(result,
219 ## list(Bxy.est.amelia.nok = est.x.mi,
220 ## Bxy.ci.upper.amelia.nok = est.x.mi + 1.96 * est.x.se,
221 ## Bxy.ci.lower.amelia.nok = est.x.mi - 1.96 * est.x.se
224 ## est.g.mi <- coefse['g','Estimate']
225 ## est.g.se <- coefse['g','Std.Error']
227 ## result <- append(result,
228 ## list(Bgy.est.amelia.nok = est.g.mi,
229 ## Bgy.ci.upper.amelia.nok = est.g.mi + 1.96 * est.g.se,
230 ## Bgy.ci.lower.amelia.nok = est.g.mi - 1.96 * est.g.se
233 ## p <- v <- train <- rep(0,N)
237 ## df <- df[order(x.obs)]
242 ## # gmm gets pretty close
243 ## (gmm.res <- predicted_covariates(y, x, g, w, v, train, p, max_iter=100, verbose=FALSE))
245 ## result <- append(result,
246 ## list(Bxy.est.gmm = gmm.res$beta[1,1],
247 ## Bxy.ci.upper.gmm = gmm.res$confint[1,2],
248 ## Bxy.ci.lower.gmm = gmm.res$confint[1,1]))
250 ## result <- append(result,
251 ## list(Bgy.est.gmm = gmm.res$beta[2,1],
252 ## Bgy.ci.upper.gmm = gmm.res$confint[2,2],
253 ## Bgy.ci.lower.gmm = gmm.res$confint[2,1]))
256 ## mod.calibrated.mle <- mecor(y ~ MeasError(w, reference = x.obs) + g, df, B=400, method='efficient')
257 ## (mod.calibrated.mle)
258 ## (mecor.ci <- summary(mod.calibrated.mle)$c$ci['x.obs',])
259 ## result <- append(result, list(
260 ## Bxy.est.mecor = mecor.ci['Estimate'],
261 ## Bxy.upper.mecor = mecor.ci['UCI'],
262 ## Bxy.lower.mecor = mecor.ci['LCI'])
265 ## (mecor.ci <- summary(mod.calibrated.mle)$c$ci['g',])
267 ## result <- append(result, list(
268 ## Bxy.est.mecor = mecor.ci['Estimate'],
269 ## Bxy.upper.mecor = mecor.ci['UCI'],
270 ## Bxy.lower.mecor = mecor.ci['LCI'])