73 lines
4.1 KiB
R
73 lines
4.1 KiB
R
set.seed(1111)
|
|
source('load_perspective_data.R')
|
|
## how accurate are the classifiers?
|
|
|
|
## the API claims that these scores are "probabilities"
|
|
## say we care about the model of the classification, not the probability
|
|
|
|
## toxicity error is weakly correlated pearson's R = 0.1 with both "white" and "black".
|
|
|
|
## compare regressions with "white" or "black" as the outcome and "toxicity_coded" or "toxicity_pred" as a predictor.
|
|
## here's a great example with presumambly non-differential error: about what identities is toxicity found humorous?
|
|
## a bunch of stars reappear when you used the ground-truth data instead of the predictions.
|
|
## pro/con of this example: will have to implement family='poisson'.
|
|
## shouldn't be that bad though haha.
|
|
cortab['toxicity_error',]
|
|
cortab['toxicity_error','funny']
|
|
cortab['toxicity_coded',]
|
|
cortab['identity_error',]
|
|
cortab['white',]
|
|
|
|
## 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')))
|
|
|
|
summary(glm(toxicity_coded ~ white*psychiatric_or_mental_illness, data = df, family=binomial(link='logit')))
|
|
|
|
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)
|
|
}
|
|
|
|
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)
|
|
|
|
|
|
glm(white ~ disagree, data = df, family=binomial(link='logit'))
|
|
|
|
## example with differential error
|
|
|
|
glm(white ~ toxicity_coded + toxicity_error, data=df,family=binomial(link='logit'))
|
|
|
|
glm(toxicity_coded ~ white, data = df, family=binomial(link='logit'))
|
|
|
|
glm(toxicity_pred ~ white, data = df, family=binomial(link='logit'))
|