Last active
January 27, 2019 23:14
-
-
Save juliasilge/cd046eaaa98417f32e05ada87b3fd1bc to your computer and use it in GitHub Desktop.
Views to answers on Stack Overflow questions
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
## this analysis assumes a dataframe `post_views` with columns: | |
## PostId | |
## CreationDate | |
## Tag | |
## AnswerCount | |
## ViewCount | |
library(tidyverse) | |
post_views %>% | |
distinct(PostId, .keep_all = TRUE) %>% | |
mutate(AnswerCount = as.factor(AnswerCount), | |
AnswerCount = fct_lump(AnswerCount), | |
AnswerCount = fct_recode(AnswerCount, | |
`More than 3` = "Other")) %>% | |
ggplot(aes(AnswerCount, ViewCount)) + | |
geom_boxplot() + | |
scale_y_log10() + | |
labs(x = "Answers per question", y = "Views per question", | |
title = "Answers and views on Stack Overflow questions", | |
subtitle = "There is an enormous amount of question-to-question variation") | |
simple_model <- post_views %>% | |
distinct(PostId, .keep_all = TRUE) %>% | |
lm(AnswerCount ~ ViewCount, data = .) | |
summary(simple_model) | |
model_with_time <- post_views %>% | |
distinct(PostId, .keep_all = TRUE) %>% | |
lm(AnswerCount ~ ViewCount + CreationDate, data = .) | |
summary(model_with_time) | |
model_no_intercept <- post_views %>% | |
distinct(PostId, .keep_all = TRUE) %>% | |
lm(AnswerCount ~ 0 + ViewCount, data = .) | |
summary(model_no_intercept) | |
log_model <- post_views %>% | |
distinct(PostId, .keep_all = TRUE) %>% | |
lm(AnswerCount ~ log10(ViewCount), data = .) | |
summary(log_model) | |
library(broom) | |
trained_models <- post_views %>% | |
replace_na(list(AnswerCount = 0)) %>% | |
add_count(Tag) %>% | |
filter(n > 1e4) %>% | |
nest(-Tag) %>% | |
mutate(Model = map(data, ~ lm(AnswerCount ~ log10(ViewCount), data = .))) | |
slopes <- trained_models %>% | |
unnest(map(Model, tidy)) %>% | |
filter(term == "log10(ViewCount)") | |
slopes | |
library(ggrepel) | |
median_slope <- slopes %>% pull(estimate) %>% median() | |
post_views %>% | |
count(Tag, sort = TRUE) %>% | |
inner_join(slopes) %>% | |
ggplot(aes(n, estimate, label = Tag)) + | |
geom_hline(yintercept = median_slope, | |
lty = 2, color = "gray70", size = 2, alpha = 0.8) + | |
geom_point() + | |
geom_text_repel(family = "IBMPlexSans-Medium") + | |
scale_x_log10() + | |
labs(x = "Number of questions", | |
y = "Slope (Number of answers per 10x increase in views)", | |
title = "Views and answers on Stack Overflow by tag", | |
subtitle = paste("The median increase in answers per 10x increase in views for this group of technologies is", round(median_slope, 2))) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Hello, saw this linked from SO.
Since you have a large number of tags, you might consider using a mixed effect model instead and fitting a single large model. Using lme4, it would look like
model <- lmer(AnswerCount ~ (1 + log10(ViewCount) | Tag), post_views)
Stan can provide the bayesian version of that model if you wanted to take that route.