-library(data.table)
-library(MASS)
-
-scores <- fread("perspective_scores.csv")
-scores <- scores[,id:=as.character(id)]
-
-df <- fread("all_data.csv")
-
-# only use the data that has identity annotations
-df <- df[identity_annotator_count > 0]
-
-(df[!(df$id %in% scores$id)])
-
-df <- df[scores,on='id',nomatch=NULL]
-
+set.seed(1111)
+source('load_perspective_data.R')
## how accurate are the classifiers?
## the API claims that these scores are "probabilities"
return (2 * precision * recall ) / (precision + recall)
}
-df[, ":="(identity_attack_pred = identity_attack_prob >=0.5,
- insult_pred = insult_prob >= 0.5,
- profanity_pred = profanity_prob >= 0.5,
- severe_toxicity_pred = severe_toxicity_prob >= 0.5,
- threat_pred = threat_prob >= 0.5,
- toxicity_pred = toxicity_prob >= 0.5,
- identity_attack_coded = identity_attack >= 0.5,
- insult_coded = insult >= 0.5,
- profanity_coded = obscene >= 0.5,
- severe_toxicity_coded = severe_toxicity >= 0.5,
- threat_coded = threat >= 0.5,
- toxicity_coded = toxicity >= 0.5
- )]
-
-
## toxicity is about 93% accurate, with an f1 of 0.8
## identity_attack has high accuracy 97%, but an unfortunant f1 of 0.5.
## what's correlated with toxicity_error ?
df <- df[,approved := rating == "approved"]
+df <- df[,white := white > 0.5]
cortab <- cor(df[,.(toxicity_error,
identity_error,
cortab['identity_error',]
cortab['white',]
-glm(white ~ toxicity_coded + psychiatric_or_mental_illness, data = df, family=binomial(link='logit'))
+cortab <- cor(df[,.(toxicity_error,
+ identity_error,
+ toxicity_coded,
+ funny,
+ approved,
+ sad,
+ wow,
+ likes,
+ disagree,
+ gender_disclosed,
+ sexuality_disclosed,
+ religion_disclosed,
+ race_disclosed,
+ disability_disclosed)])
+
+
+## here's a simple example, is P(white | toxic and mentally ill) > P(white | toxic or mentally ill). Are people who discuss their mental illness in a toxic way more likely to be white compared to those who just talk about their mental illness or are toxic?
+summary(glm(white ~ toxicity_coded*psychiatric_or_mental_illness, data = df, family=binomial(link='logit')))
+
+summary(glm(white ~ toxicity_pred*psychiatric_or_mental_illness, data = df, family=binomial(link='logit')))
+
+summary(glm(white ~ toxicity_coded*male, data = df, family=binomial(link='logit')))
+
+summary(glm(white ~ toxicity_pred*male, data = df, family=binomial(link='logit')))
-glm(white ~ toxicity_pred + psychiatric_or_mental_illness, data = df, family=binomial(link='logit'))
+summary(glm(toxicity_coded ~ white*psychiatric_or_mental_illness, data = df, family=binomial(link='logit')))
-m1 <- glm.nb(funny ~ (male + female + transgender + other_gender + heterosexual + bisexual + other_sexual_orientation + christian + jewish + hindu + buddhist + atheist + other_religion + asian + latino + other_race_or_ethnicity + physical_disability + intellectual_or_learning_disability + white + black + psychiatric_or_mental_illness)*toxicity_coded, data = df)
+summary(glm(toxicity_pred ~ white*psychiatric_or_mental_illness, data = df, family=binomial(link='logit')))
+
+
+## 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?
+
+summary(glm(toxicity_pred ~ funny*white, data=df, family=binomial(link='logit')))
+summary(glm(toxicity_coded ~ funny*white, data=df, family=binomial(link='logit')))
+
+source("../simulations/measerr_methods.R")
+
+saved_model_file <- "measerr_model_tox.eq.funny.cross.white.RDS"
+overwrite_model <- TRUE
+
+# it works so far with a 20% and 15% sample. Smaller is better. let's try a 10% sample again. It didn't work out. We'll go forward with a 15% sample.
+df_measerr_method <- copy(df)[sample(1:.N, 0.05 * .N), toxicity_coded_1 := toxicity_coded]
+df_measerr_method <- df_measerr_method[,toxicity_coded := toxicity_coded_1]
+summary(glm(toxicity_coded ~ funny*white, data=df_measerr_method[!is.na(toxicity_coded)], family=binomial(link='logit')))
+
+if(!file.exists(saved_model_file) || (overwrite_model == TRUE)){
+ measerr_model <- measerr_mle_dv(df_measerr_method,toxicity_coded ~ funny*white,outcome_family=binomial(link='logit'), proxy_formula=toxicity_pred ~ toxicity_coded*funny*white)
+saveRDS(measerr_model, saved_model_file)
+} else {
+ measerr_model <- readRDS(saved_model_file)
+}
-m2 <- glm.nb(funny ~ (male + female + transgender + other_gender + heterosexual + bisexual + other_sexual_orientation + christian + jewish + hindu + buddhist + atheist + other_religion + asian + latino + other_race_or_ethnicity + physical_disability + intellectual_or_learning_disability + white + black + psychiatric_or_mental_illness)*toxicity_pred, data = df)
+inv_hessian <- solve(measerr_model$hessian)
+se <- diag(inv_hessian)
+lm2 <- glm.nb(funny ~ (male + female + transgender + other_gender + heterosexual + bisexual + other_sexual_orientation + christian + jewish + hindu + buddhist + atheist + other_religion + asian + latino + other_race_or_ethnicity + physical_disability + intellectual_or_learning_disability + white + black + psychiatric_or_mental_illness)*toxicity_pred, data = df)
m3 <- glm.nb(funny ~ (male + female + transgender + other_gender + heterosexual + bisexual + other_sexual_orientation + christian + jewish + hindu + buddhist + atheist + other_religion + asian + latino + other_race_or_ethnicity + physical_disability + intellectual_or_learning_disability + white + black + psychiatric_or_mental_illness)*toxicity, data = df)