update real data examples code and rerun project.

[ml_measurement_error_public.git] / civil_comments / 02_iv_example.R

diff --git a/civil_comments/02_iv_example.R b/civil_comments/02_iv_example.R
new file mode 100644 (file)
index 0000000..ba5ab12
--- /dev/null
+++ b/civil_comments/02_iv_example.R
@@ -0,0 +1,107 @@
+source("../simulations/RemembR/R/RemembeR.R")
+change.remember.file("iv_perspective_example.RDS")
+
+source('load_perspective_data.R')
+source("../simulations/measerr_methods.R")
+
+remember(accuracies, "civil_comments_accuracies")
+remember(f1s, "civil_comments_f1s")
+remember(positive_cases, "civil_comments_positive_cases")
+remember(proportions_cases, "civil_comments_proportions_cases")
+remember(cortab, "civil_comments_cortab")
+remember(nrow(df), 'n.annotated.comments')
+# for reproducibility
+set.seed(1)
+
+## another simple enough example: is P(toxic | funny and white) > P(toxic | funny nand white)? Or, are funny comments more toxic when people disclose that they are white?
+
+compare_iv_models <-function(pred_formula, outcome_formula, proxy_formula, truth_formula, df, sample.prop, sample.size, remember_prefix){
+
+    if(is.null(sample.prop)){
+        sample.prop <- sample.size / nrow(df)
+    }
+    if(is.null(sample.size)){
+        sample.size <- nrow(df) * sample.prop
+    }
+
+    remember(sample.size, paste0(remember_prefix, "sample.size"))
+    remember(sample.prop, paste0(remember_prefix, "sample.prop"))
+    remember(pred_formula, paste0(remember_prefix, "pred_formula"))
+    remember(outcome_formula, paste0(remember_prefix, 'outcome_formula'))
+    remember(proxy_formula, paste0(remember_prefix, 'proxy_formula'))
+    remember(truth_formula, paste0(remember_prefix, 'truth_formula'))
+
+    pred_model <- glm(pred_formula, df, family=binomial(link='logit'))
+    remember(coef(pred_model), paste0(remember_prefix, "coef_pred_model"))
+    remember(diag(vcov((pred_model))), paste0(remember_prefix, "se_pred_model"))
+
+    coder_model <- glm(outcome_formula, df, family=binomial(link='logit'))
+    remember(coef(coder_model), paste0(remember_prefix, "coef_coder_model"))
+    remember(diag(vcov((coder_model))), paste0(remember_prefix, "se_coder_model"))
+
+    df_measerr_method <- copy(df)[sample(1:.N, sample.size), toxicity_coded_1 := toxicity_coded]
+    df_measerr_method <- df_measerr_method[,toxicity_coded := toxicity_coded_1]
+    sample_model <- glm(outcome_formula, df_measerr_method, family=binomial(link='logit'))
+    remember(coef(sample_model), paste0(remember_prefix, "coef_sample_model"))
+    remember(diag(vcov((sample_model))), paste0(remember_prefix, "se_sample_model"))
+
+    measerr_model <- measerr_mle(df_measerr_method, outcome_formula, outcome_family=binomial(link='logit'), proxy_formula=proxy_formula, proxy_family=binomial(link='logit'),truth_formula=truth_formula, truth_family=binomial(link='logit'))
+
+    inv_hessian = solve(measerr_model$hessian)
+    stderr = diag(inv_hessian)
+    remember(stderr, paste0(remember_prefix, "measerr_model_stderr"))
+    remember(measerr_model$par, paste0(remember_prefix, "measerr_model_par"))
+}
+
+## print("running first iv example")
+
+## sample.prop <- 0.05
+
+## compare_iv_models(white ~ toxicity_pred*funny,
+##                   outcome_formula = white ~ toxicity_coded*funny,
+##                   proxy_formula = toxicity_pred ~ toxicity_coded*funny*white,
+##                   truth_formula = toxicity_coded ~ 1,
+##                   df=df,
+##                   sample.prop=sample.prop,
+##                   remember_prefix='cc_ex_tox.funny.white')
+
+
+
+pred_formula <- race_disclosed ~ likes * toxicity_pred
+outcome_formula <- race_disclosed ~ likes * toxicity_coded
+proxy_formula <- toxicity_pred ~ toxicity_coded * race_disclosed * likes
+truth_formula <- toxicity_coded ~ 1
+
+print("running first example")
+
+compare_iv_models(pred_formula = pred_formula,
+                  outcome_formula = outcome_formula,
+                  proxy_formula = proxy_formula,
+                  truth_formula = truth_formula,
+                  df=df,
+                  sample.prop=0.01,
+                  sample.size=NULL,
+                  remember_prefix='cc_ex_tox.likes.race_disclosed')
+
+print("running second example")
+
+compare_iv_models(pred_formula = pred_formula,
+                  outcome_formula = outcome_formula,
+                  proxy_formula = proxy_formula,
+                  truth_formula = truth_formula,
+                  df=df,
+                  sample.prop=NULL,
+                  sample.size=10000,
+                  remember_prefix='cc_ex_tox.likes.race_disclosed.medsamp')
+
+print("running third example")
+
+compare_iv_models(pred_formula = race_disclosed ~ likes * toxicity_pred,
+                  outcome_formula = race_disclosed ~ likes * toxicity_coded,
+                  proxy_formula = toxicity_pred ~ toxicity_coded + race_disclosed,
+                  truth_formula = toxicity_coded ~ 1,
+                  df=df,
+                  sample.prop=0.05,
+                  sample.size=NULL,
+                  remember_prefix='cc_ex_tox.likes.race_disclosed.largesamp')
+