simulations/irr_dv_simulation_base.R

   1 library(matrixStats) # for numerically stable logsumexps
   2
   3 options(amelia.parallel="no",
   4         amelia.ncpus=1)
   5 library(Amelia)
   6
   7 source("pl_methods.R")
   8 source("measerr_methods_2.R") ## for my more generic function.
   9
  10 run_simulation_depvar <- function(df, result, outcome_formula = y ~ x + z, coder_formulas = list(y.obs.0 ~ 1, y.obs.1 ~ 1), proxy_formula = w_pred ~ y.obs.1+y.obs.0+y){
  11     (accuracy <- df[,mean(w_pred==y)])
  12     result <- append(result, list(accuracy=accuracy))
  13     (error.cor.z <- cor(df$x, df$w_pred - df$z))
  14     (error.cor.x <- cor(df$x, df$w_pred - df$y))
  15     (error.cor.y <- cor(df$y, df$y - df$w_pred))
  16     result <- append(result, list(error.cor.x = error.cor.x,
  17                                   error.cor.z = error.cor.z,
  18                                   error.cor.y = error.cor.y))
  19
  20     model.null <- glm(y~1, data=df,family=binomial(link='logit'))
  21     (model.true <- glm(y ~ x + z, data=df,family=binomial(link='logit')))
  22     (lik.ratio <- exp(logLik(model.true) - logLik(model.null)))
  23
  24     true.ci.Bxy <- confint(model.true)['x',]
  25     true.ci.Bzy <- confint(model.true)['z',]
  26
  27
  28     result <- append(result, list(lik.ratio=lik.ratio))
  29
  30     result <- append(result, list(Bxy.est.true=coef(model.true)['x'],
  31                                   Bzy.est.true=coef(model.true)['z'],
  32                                   Bxy.ci.upper.true = true.ci.Bxy[2],
  33                                   Bxy.ci.lower.true = true.ci.Bxy[1],
  34                                   Bzy.ci.upper.true = true.ci.Bzy[2],
  35                                   Bzy.ci.lower.true = true.ci.Bzy[1]))
  36
  37     (model.naive <- lm(y~w_pred+z, data=df))
  38
  39     naive.ci.Bxy <- confint(model.naive)['w_pred',]
  40     naive.ci.Bzy <- confint(model.naive)['z',]
  41
  42     result <- append(result, list(Bxy.est.naive=coef(model.naive)['w_pred'],
  43                                   Bzy.est.naive=coef(model.naive)['z'],
  44                                   Bxy.ci.upper.naive = naive.ci.Bxy[2],
  45                                   Bxy.ci.lower.naive = naive.ci.Bxy[1],
  46                                   Bzy.ci.upper.naive = naive.ci.Bzy[2],
  47                                   Bzy.ci.lower.naive = naive.ci.Bzy[1]))
  48
  49
  50
  51     loa0.feasible <- glm(y.obs.0 ~ x + z, data = df[!(is.na(y.obs.0))], family=binomial(link='logit'))
  52
  53     loa0.ci.Bxy <- confint(loa0.feasible)['x',]
  54     loa0.ci.Bzy <- confint(loa0.feasible)['z',]
  55
  56     result <- append(result, list(Bxy.est.loa0.feasible=coef(loa0.feasible)['x'],
  57                                   Bzy.est.loa0.feasible=coef(loa0.feasible)['z'],
  58                                   Bxy.ci.upper.loa0.feasible = loa0.ci.Bxy[2],
  59                                   Bxy.ci.lower.loa0.feasible = loa0.ci.Bxy[1],
  60                                   Bzy.ci.upper.loa0.feasible = loa0.ci.Bzy[2],
  61                                   Bzy.ci.lower.loa0.feasible = loa0.ci.Bzy[1]))
  62
  63
  64     ## df.loa0.mle <- copy(df)
  65     ## df.loa0.mle[,y:=y.obs.0]
  66     ## loa0.mle <- measerr_mle_dv(df.loa0.mle, outcome_formula=outcome_formula, proxy_formula=proxy_formula)
  67     ## fisher.info <- solve(loa0.mle$hessian)
  68     ## coef <- loa0.mle$par
  69     ## ci.upper <- coef + sqrt(diag(fisher.info)) * 1.96
  70     ## ci.lower <- coef - sqrt(diag(fisher.info)) * 1.96
  71
  72     ## result <- append(result, list(Bxy.est.loa0.mle=coef['x'],
  73     ##                               Bzy.est.loa0.mle=coef['z'],
  74     ##                               Bxy.ci.upper.loa0.mle = ci.upper['x'],
  75     ##                               Bxy.ci.lower.loa0.mle = ci.lower['x'],
  76     ##                               Bzy.ci.upper.loa0.mle = ci.upper['z'],
  77     ##                               Bzy.ci.lower.loa0.mle = ci.upper['z']))
  78
  79     loco.feasible <- glm(y.obs.0 ~ x + z, data = df[(!is.na(y.obs.0)) & (y.obs.1 == y.obs.0)], family=binomial(link='logit'))
  80
  81     loco.feasible.ci.Bxy <- confint(loco.feasible)['x',]
  82     loco.feasible.ci.Bzy <- confint(loco.feasible)['z',]
  83
  84     result <- append(result, list(Bxy.est.loco.feasible=coef(loco.feasible)['x'],
  85                                   Bzy.est.loco.feasible=coef(loco.feasible)['z'],
  86                                   Bxy.ci.upper.loco.feasible = loco.feasible.ci.Bxy[2],
  87                                   Bxy.ci.lower.loco.feasible = loco.feasible.ci.Bxy[1],
  88                                   Bzy.ci.upper.loco.feasible = loco.feasible.ci.Bzy[2],
  89                                   Bzy.ci.lower.loco.feasible = loco.feasible.ci.Bzy[1]))
  90
  91
  92     ## df.double.proxy <- copy(df)
  93     ## df.double.proxy <- df.double.proxy[,y.obs:=NA]
  94     ## df.double.proxy <- df.double.proxy[,y:=NA]
  95
  96     ## double.proxy.mle <- measerr_irr_mle_dv(df.double.proxy, outcome_formula=y~x+z, outcome_family=binomial(link='logit'), coder_formulas=list(y.obs.0 ~ y),  proxy_formula=w_pred ~ y.obs.0 + y, proxy_family=binomial(link='logit'))
  97     ## print(double.proxy.mle$hessian)
  98     ## fisher.info <- solve(double.proxy.mle$hessian)
  99     ## coef <- double.proxy.mle$par
 100     ## ci.upper <- coef + sqrt(diag(fisher.info)) * 1.96
 101     ## ci.lower <- coef - sqrt(diag(fisher.info)) * 1.96
 102
 103     ## result <- append(result, list(Bxy.est.double.proxy=coef['x'],
 104     ##                               Bzy.est.double.proxy=coef['z'],
 105     ##                               Bxy.ci.upper.double.proxy = ci.upper['x'],
 106     ##                               Bxy.ci.lower.double.proxy = ci.lower['x'],
 107     ##                               Bzy.ci.upper.double.proxy = ci.upper['z'],
 108     ##                               Bzy.ci.lower.double.proxy = ci.lower['z']))
 109
 110
 111     df.triple.proxy <- copy(df)
 112     df.triple.proxy <- df.triple.proxy[,y.obs:=NA]
 113     df.triple.proxy <- df.triple.proxy[,y:=NA]
 114
 115     triple.proxy.mle <- measerr_irr_mle_dv(df.triple.proxy, outcome_formula=outcome_formula, outcome_family=binomial(link='logit'), coder_formulas=coder_formulas, proxy_formula=proxy_formula, proxy_family=binomial(link='logit'))
 116     print(triple.proxy.mle$hessian)
 117     fisher.info <- solve(triple.proxy.mle$hessian)
 118     print(fisher.info)
 119     coef <- triple.proxy.mle$par
 120     ci.upper <- coef + sqrt(diag(fisher.info)) * 1.96
 121     ci.lower <- coef - sqrt(diag(fisher.info)) * 1.96
 122
 123     result <- append(result, list(Bxy.est.triple.proxy=coef['x'],
 124                                   Bzy.est.triple.proxy=coef['z'],
 125                                   Bxy.ci.upper.triple.proxy = ci.upper['x'],
 126                                   Bxy.ci.lower.triple.proxy = ci.lower['x'],
 127                                   Bzy.ci.upper.triple.proxy = ci.upper['z'],
 128                                   Bzy.ci.lower.triple.proxy = ci.lower['z']))
 129
 130     ## df.loco.mle <- copy(df)
 131     ## df.loco.mle[,y.obs:=NA]
 132     ## df.loco.mle[(y.obs.0)==(y.obs.1),y.obs:=y.obs.0]
 133     ## df.loco.mle[,y.true:=y]
 134     ## df.loco.mle[,y:=y.obs]
 135     ## print(df.loco.mle[!is.na(y.obs.1),mean(y.true==y,na.rm=TRUE)])
 136     ## loco.mle <- measerr_mle_dv(df.loco.mle, outcome_formula=outcome_formula, proxy_formula=proxy_formula)
 137     ## fisher.info <- solve(loco.mle$hessian)
 138     ## coef <- loco.mle$par
 139     ## ci.upper <- coef + sqrt(diag(fisher.info)) * 1.96
 140     ## ci.lower <- coef - sqrt(diag(fisher.info)) * 1.96
 141
 142     ## result <- append(result, list(Bxy.est.loco.mle=coef['x'],
 143     ##                               Bzy.est.loco.mle=coef['z'],
 144     ##                               Bxy.ci.upper.loco.mle = ci.upper['x'],
 145     ##                               Bxy.ci.lower.loco.mle = ci.lower['x'],
 146     ##                               Bzy.ci.upper.loco.mle = ci.upper['z'],
 147     ##                               Bzy.ci.lower.loco.mle = ci.lower['z']))
 148
 149
 150
 151     ## my implementatoin of liklihood based correction
 152     mod.zhang <- zhang.mle.dv(df.loco.mle)
 153     coef <- coef(mod.zhang)
 154     ci <- confint(mod.zhang,method='quad')
 155
 156     result <- append(result,
 157                      list(Bxy.est.zhang = coef['Bxy'],
 158                           Bxy.ci.upper.zhang = ci['Bxy','97.5 %'],
 159                           Bxy.ci.lower.zhang = ci['Bxy','2.5 %'],
 160                           Bzy.est.zhang = coef['Bzy'],
 161                           Bzy.ci.upper.zhang = ci['Bzy','97.5 %'],
 162                           Bzy.ci.lower.zhang = ci['Bzy','2.5 %']))
 163
 164
 165
 166     print(df.loco.mle)
 167
 168     # amelia says use normal distribution for binary variables.
 169     tryCatch({
 170         amelia.out.k <- amelia(df.loco.mle, m=200, p2s=0, idvars=c('y','ystar','w','y.obs.1','y.obs.0','y.true'))
 171         mod.amelia.k <- zelig(y.obs~x+z, model='ls', data=amelia.out.k$imputations, cite=FALSE)
 172         (coefse <- combine_coef_se(mod.amelia.k, messages=FALSE))
 173         est.x.mi <- coefse['x','Estimate']
 174         est.x.se <- coefse['x','Std.Error']
 175         result <- append(result,
 176                          list(Bxy.est.amelia.full = est.x.mi,
 177                               Bxy.ci.upper.amelia.full = est.x.mi + 1.96 * est.x.se,
 178                               Bxy.ci.lower.amelia.full = est.x.mi - 1.96 * est.x.se
 179                               ))
 180
 181         est.z.mi <- coefse['z','Estimate']
 182         est.z.se <- coefse['z','Std.Error']
 183
 184         result <- append(result,
 185                          list(Bzy.est.amelia.full = est.z.mi,
 186                               Bzy.ci.upper.amelia.full = est.z.mi + 1.96 * est.z.se,
 187                               Bzy.ci.lower.amelia.full = est.z.mi - 1.96 * est.z.se
 188                               ))
 189
 190     },
 191     error = function(e){
 192         message("An error occurred:\n",e)
 193         result$error <- paste0(result$error,'\n', e)
 194     })
 195     ## mle.irr <- measerr_irr_mle( df, outcome_formula = outcome_formula, rater_formula = rater_formula, proxy_formula=proxy_formula, truth_formula=truth_formula)
 196
 197     ## fisher.info <- solve(mle.irr$hessian)
 198     ## coef <- mle.irr$par
 199     ## ci.upper <- coef + sqrt(diag(fisher.info)) * 1.96
 200     ## ci.lower <- coef - sqrt(diag(fisher.info)) * 1.96
 201
 202     ## result <- append(result,
 203     ##                  list(Bxy.est.mle = coef['x'],
 204     ##                       Bxy.ci.upper.mle = ci.upper['x'],
 205     ##                       Bxy.ci.lower.mle = ci.lower['x'],
 206     ##                       Bzy.est.mle = coef['z'],
 207     ##                       Bzy.ci.upper.mle = ci.upper['z'],
 208     ##                       Bzy.ci.lower.mle = ci.lower['z']))
 209
 210     return(result)
 211
 212 }