X-Git-Url: https://code.communitydata.science/ml_measurement_error_public.git/blobdiff_plain/d8bc08f18f8c2128369ee959196e0e6080a11689..fa05dbab6bd2c5db6ed4eccf38cff03bb4fd6683:/civil_comments/01_dv_example.R

diff --git a/civil_comments/01_dv_example.R b/civil_comments/01_dv_example.R
new file mode 100644
index 0000000..4092243
--- /dev/null
+++ b/civil_comments/01_dv_example.R
@@ -0,0 +1,54 @@
+source('load_perspective_data.R')
+source("../simulations/measerr_methods.R")
+source("../simulations/RemembR/R/RemembeR.R")
+
+change.remember.file("dv_perspective_example.RDS")
+
+# for reproducibility
+set.seed(1111)
+
+## 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_dv_models <-function(pred_formula, outcome_formula, proxy_formula, df, sample.prop, remember_prefix){
+    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.prop * .N), 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_dv(df_measerr_method, outcome_formula, outcome_family=binomial(link='logit'), proxy_formula=proxy_formula, proxy_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 example")
+
+compare_dv_models(pred_formula = toxicity_pred ~ funny*white,
+                  outcome_formula = toxicity_coded ~ funny*white, proxy_formula,
+                  proxy_formula = toxicity_pred ~ toxicity_coded*funny*white,
+                  df=df,
+                  sample.prop=0.01,
+                  remember_prefix='cc_ex_tox.funny.white')
+
+
+print("running second example")
+
+compare_dv_models(pred_formula = toxicity_pred ~ likes+race_disclosed,
+                  outcome_formula = toxicity_coded ~ likes + race_disclosed, proxy_formula,
+                  proxy_formula = toxicity_pred ~ toxicity_coded*likes*race_disclosed,
+                  df=df,
+                  sample.prop=0.01,
+                  remember_prefix='cc_ex_tox.funny.race_disclosed')
+