simulations/plot_example.R

   1 source("RemembR/R/RemembeR.R")
   2 library(arrow)
   3 library(data.table)
   4 library(ggplot2)
   5 library(filelock)
   6 library(argparser)
   7
   8 source("summarize_estimator.R")
   9
  10
  11 parser <- arg_parser("Simulate data and fit corrected models.")
  12 parser <- add_argument(parser, "--infile", default="", help="name of the file to read.")
  13 parser <- add_argument(parser, "--remember-file", default="remembr.RDS", help="name of the remember file.")
  14 parser <- add_argument(parser, "--name", default="", help="The name to safe the data to in the remember file.")
  15 args <- parse_args(parser)
  16
  17
  18
  19 ## summarize.estimator <- function(df, suffix='naive', coefname='x'){
  20
  21 ##     part <- df[,c('N',
  22 ##                   'm',
  23 ##                   'Bxy',
  24 ##                   paste0('B',coefname,'y.est.',suffix),
  25 ##                   paste0('B',coefname,'y.ci.lower.',suffix),
  26 ##                   paste0('B',coefname,'y.ci.upper.',suffix),
  27 ##                   'y_explained_variance',
  28 ##                   'Bzx',
  29 ##                   'Bzy',
  30 ##                   'accuracy_imbalance_difference'
  31 ##                   ),
  32 ##                with=FALSE]
  33
  34 ##     true.in.ci <- as.integer((part$Bxy >= part[[paste0('B',coefname,'y.ci.lower.',suffix)]]) & (part$Bxy <= part[[paste0('B',coefname,'y.ci.upper.',suffix)]]))
  35 ##     zero.in.ci <- as.integer(0 >= part[[paste0('B',coefname,'y.ci.lower.',suffix)]]) & (0 <= part[[paste0('B',coefname,'y.ci.upper.',suffix)]])
  36 ##     bias <- part$Bxy - part[[paste0('B',coefname,'y.est.',suffix)]]
  37 ##     sign.correct <- as.integer(sign(part$Bxy) == sign(part[[paste0('B',coefname,'y.est.',suffix)]]))
  38
  39 ##     part <- part[,':='(true.in.ci = true.in.ci,
  40 ##                        zero.in.ci = zero.in.ci,
  41 ##                        bias=bias,
  42 ##                        sign.correct =sign.correct)]
  43
  44 ##     part.plot <- part[, .(p.true.in.ci = mean(true.in.ci),
  45 ##                           mean.bias = mean(bias),
  46 ##                           mean.est = mean(.SD[[paste0('B',coefname,'y.est.',suffix)]]),
  47 ##                           var.est = var(.SD[[paste0('B',coefname,'y.est.',suffix)]]),
  48 ##                           est.upper.95 = quantile(.SD[[paste0('B',coefname,'y.est.',suffix)]],0.95,na.rm=T),
  49 ##                           est.lower.95 = quantile(.SD[[paste0('B',coefname,'y.est.',suffix)]],0.05,na.rm=T),
  50 ##                           N.sims = .N,
  51 ##                           p.sign.correct = mean(as.integer(sign.correct & (! zero.in.ci))),
  52 ##                           variable=coefname,
  53 ##                           method=suffix
  54 ##                           ),
  55 ##                       by=c("N","m",'y_explained_variance','Bzx', 'Bzy', 'accuracy_imbalance_difference')
  56 ##                       ]
  57
  58 ##     return(part.plot)
  59 ## }
  60
  61 build_plot_dataset <- function(df){
  62
  63     x.true <-  summarize.estimator(df, 'true','x')
  64
  65     z.true <-  summarize.estimator(df, 'true','z')
  66
  67     x.naive <- summarize.estimator(df, 'naive','x')
  68
  69     z.naive <- summarize.estimator(df,'naive','z')
  70
  71     x.feasible <- summarize.estimator(df, 'feasible', 'x')
  72
  73     z.feasible <- summarize.estimator(df, 'feasible', 'z')
  74
  75     x.amelia.full <- summarize.estimator(df, 'amelia.full', 'x')
  76
  77     z.amelia.full <- summarize.estimator(df, 'amelia.full', 'z')
  78
  79     x.mecor <- summarize.estimator(df, 'mecor', 'x')
  80
  81     z.mecor <- summarize.estimator(df, 'mecor', 'z')
  82
  83     x.mecor <- summarize.estimator(df, 'mecor', 'x')
  84
  85     z.mecor <- summarize.estimator(df, 'mecor', 'z')
  86
  87     x.mle <- summarize.estimator(df, 'mle', 'x')
  88
  89     z.mle <- summarize.estimator(df, 'mle', 'z')
  90
  91     x.zhang <- summarize.estimator(df, 'zhang', 'x')
  92
  93     z.zhang <- summarize.estimator(df, 'zhang', 'z')
  94
  95     x.gmm <- summarize.estimator(df, 'gmm', 'x')
  96
  97     z.gmm <- summarize.estimator(df, 'gmm', 'z')
  98
  99     accuracy <- df[,mean(accuracy)]
 100     plot.df <- rbindlist(list(x.true,z.true,x.naive,z.naive,x.amelia.full,z.amelia.full,x.mecor, z.mecor, x.gmm, z.gmm, x.feasible, z.feasible,z.mle, x.mle, x.zhang, z.zhang, x.gmm, z.gmm),use.names=T)
 101     plot.df[,accuracy := accuracy]
 102     plot.df <- plot.df[,":="(sd.est=sqrt(var.est)/N.sims)]
 103     return(plot.df)
 104 }
 105
 106
 107 sims.df <- read_feather(args$infile)
 108 print(unique(sims.df$N))
 109
 110 # df <- df[apply(df,1,function(x) !any(is.na(x)))]
 111
 112 if(!('Bzx' %in% names(sims.df)))
 113     sims.df[,Bzx:=NA]
 114
 115 if(!('accuracy_imbalance_difference' %in% names(sims.df)))
 116     sims.df[,accuracy_imbalance_difference:=NA]
 117
 118 unique(sims.df[,'accuracy_imbalance_difference'])
 119
 120 change.remember.file(args$remember_file, clear=TRUE)
 121 #plot.df <- build_plot_dataset(df[accuracy_imbalance_difference==0.1][N==700])
 122 plot.df <- build_plot_dataset(sims.df)
 123
 124 remember(plot.df,args$name)
 125
 126 set.remember.prefix(gsub("plot.df.","",args$name))
 127
 128 remember(median(sims.df$cor.xz),'med.cor.xz')
 129 remember(median(sims.df$accuracy),'med.accuracy')
 130 remember(median(sims.df$accuracy.y0),'med.accuracy.y0')
 131 remember(median(sims.df$accuracy.y1),'med.accuracy.y1')
 132 remember(median(sims.df$fpr),'med.fpr')
 133 remember(median(sims.df$fpr.y0),'med.fpr.y0')
 134 remember(median(sims.df$fpr.y1),'med.fpr.y1')
 135 remember(median(sims.df$fnr),'med.fnr')
 136 remember(median(sims.df$fnr.y0),'med.fnr.y0')
 137 remember(median(sims.df$fnr.y1),'med.fnr.y1')
 138
 139 remember(median(sims.df$cor.resid.w_pred),'cor.resid.w_pred')
 140
 141 #ggplot(df,aes(x=Bxy.est.mle)) + geom_histogram() + facet_grid(accuracy_imbalance_difference ~ Bzy)
 142
 143 ## ## ## df[gmm.ER_pval<0.05]
 144
 145 ## plot.df.test <- plot.df[,':='(method=factor(method,levels=c("Naive","Multiple imputation", "Multiple imputation (Classifier features unobserved)","Regression Calibration","2SLS+gmm","Bespoke MLE", "Feasible"),ordered=T),
 146 ##                                    N=factor(N),
 147 ##                                    m=factor(m))]
 148
 149 ## plot.df.test <- plot.df.test[(variable=='x') & (method!="Multiple imputation (Classifier features unobserved)")]
 150 ## p <- ggplot(plot.df.test, aes(y=mean.est, ymax=mean.est + var.est/2, ymin=mean.est-var.est/2, x=method))
 151 ## p <- p + geom_hline(data=plot.df.test, mapping=aes(yintercept=0.1),linetype=2)
 152
 153 ## p <- p + geom_pointrange() + facet_grid(N~m,as.table=F,scales='free') + scale_x_discrete(labels=label_wrap_gen(4))
 154 ## print(p)
 155
 156 ## plot.df.test <- plot.df[,':='(method=factor(method,levels=c("Naive","Multiple imputation", "Multiple imputation (Classifier features unobserved)","Regression Calibration","2SLS+gmm","Bespoke MLE", "Feasible"),ordered=T),
 157 ##                                    N=factor(N),
 158 ##                                    m=factor(m))]
 159
 160 ## plot.df.test <- plot.df.test[(variable=='z') & (method!="Multiple imputation (Classifier features unobserved)")]
 161 ## p <- ggplot(plot.df.test, aes(y=mean.est, ymax=mean.est + var.est/2, ymin=mean.est-var.est/2, x=method))
 162 ## p <- p + geom_hline(data=plot.df.test, mapping=aes(yintercept=-0.1),linetype=2)
 163
 164 ## p <- p + geom_pointrange() + facet_grid(m~N,as.table=F,scales='free') + scale_x_discrete(labels=label_wrap_gen(4))
 165 ## print(p)
 166
 167
 168 ## x.mle <- df[,.(N,m,Bxy.est.mle,Bxy.ci.lower.mle, Bxy.ci.upper.mle, y_explained_variance, Bzx, Bzy, accuracy_imbalance_difference)]
 169 ## x.mle.plot <- x.mle[,.(mean.est = mean(Bxy.est.mle),
 170 ##                        var.est = var(Bxy.est.mle),
 171 ##                        N.sims = .N,
 172 ##                        variable='z',
 173 ##                        method='Bespoke MLE'
 174 ##                        ),
 175 ##                     by=c("N","m",'y_explained_variance', 'Bzx', 'Bzy','accuracy_imbalance_difference')]
 176
 177 ## z.mle <- df[,.(N,m,Bzy.est.mle,Bzy.ci.lower.mle, Bzy.ci.upper.mle, y_explained_variance, Bzx, Bzy, accuracy_imbalance_difference)]
 178
 179 ## z.mle.plot <- z.mle[,.(mean.est = mean(Bzy.est.mle),
 180 ##                        var.est = var(Bzy.est.mle),
 181 ##                        N.sims = .N,
 182 ##                        variable='z',
 183 ##                        method='Bespoke MLE'
 184 ##                        ),
 185 ##                     by=c("N","m",'y_explained_variance','Bzx')]
 186
 187 ## plot.df <- z.mle.plot
 188 ## plot.df.test <- plot.df[,':='(method=factor(method,levels=c("Naive","Multiple imputation", "Multiple imputation (Classifier features unobserved)","Regression Calibration","2SLS+gmm","Bespoke MLE", "Feasible"),ordered=T),
 189 ##                                    N=factor(N),
 190 ##                                    m=factor(m))]
 191
 192 ## plot.df.test <- plot.df.test[(variable=='z') & (m != 1000) & (m!=500) & (method!="Multiple imputation (Classifier features unobserved)")]
 193 ## p <- ggplot(plot.df.test, aes(y=mean.est, ymax=mean.est + var.est/2, ymin=mean.est-var.est/2, x=method))
 194 ## p <- p + geom_hline(aes(yintercept=0.2),linetype=2)
 195
 196 ## p <- p + geom_pointrange() + facet_grid(m~Bzx, Bzy,as.table=F) + scale_x_discrete(labels=label_wrap_gen(4))
 197 ## print(p)
 198
 199
 200 ## ## ggplot(plot.df[variable=='x'], aes(y=mean.est, ymax=mean.est + var.est/2, ymin=mean.est-var.est/2, x=method)) + geom_pointrange() + facet_grid(-m~N) + scale_x_discrete(labels=label_wrap_gen(10))
 201
 202 ## ## ggplot(plot.df,aes(y=N,x=m,color=p.sign.correct)) + geom_point() + facet_grid(variable ~ method) + scale_color_viridis_c(option='D') + theme_minimal() + xlab("Number of gold standard labels") + ylab("Total sample size")
 203
 204 ## ## ggplot(plot.df,aes(y=N,x=m,color=abs(mean.bias))) + geom_point() + facet_grid(variable ~ method) + scale_color_viridis_c(option='D') + theme_minimal() + xlab("Number of gold standard labels") + ylab("Total sample size")