In hnguyen1174/online_visits_assessment: To assess the effectiveness of online visit on booking metrics

1. Load Libraries

library(tidyverse)
library(here)
library(lubridate)
library(janitor)
library(zoo)
library(ggthemes)

library(forecast)
library(stargazer)
library(data.table)
library(tableone)
library(lattice)
library(pwr)
library(rcompanion)
library(scales)
library(plm)
library(readxl)
library(MatchIt)
library(lfe)
library(Synth)
library(gsynth)
library(panelView)
library(CausalImpact)
library(optmatch)

library(zeallot)

devtools::load_all()

2. Load and Process Data

c(metrics, property) %<-% get_property_data()

3. Matching - Aggregated Data

Features to match (these are the features that can potentially explain why a property was chosen for the online visit option): avg_review_rating_num: average rating of the property number_of_reviews_on_the_property_num: number of reviews a property has instant_booking_enabled: whether a property has the instant booking feature enabled premier_partner_badge: whether a property is from a premier partner maximum_occupancy: maximum occupancy of the property bedrooms: number of bedrooms bathrooms: number of bathrooms number_of_photos_of_the_property_on_the_property_page: number of photos on the property's page bookings_per_day: number of bookings per day (before 2019-01-01) booking_amount_per_day: amount of booking dollar per day (before 2019-01-01) booking_requests_per_day: number of booking request per day (before 2019-01-01) property_page_views_per_visit_per_day: number of page views per visit per day (before 2019-01-01) * displays_in_searchpage_per_day: number of times the property got displayed in the search page (before 2019-01-01)

metrics_agg <- metrics %>% 
  filter(date < as.Date('2019-01-01')) %>% 
  group_by(property_id) %>% 
  summarize(bookings_per_day = mean(bookings, na.rm = TRUE),
            booking_amount_per_day = mean(booking_amount, na.rm = TRUE),
            booking_requests_per_day = mean(booking_requests, na.rm = TRUE),
            property_page_views_per_visit_per_day = mean(property_page_views_visits, na.rm = TRUE),
            displays_in_searchpage_per_day = mean(number_of_times_the_property_got_displayed_in_search_page, na.rm = TRUE)) %>% 
  ungroup()

property_to_match <- property %>% 
  left_join(metrics_agg, by = 'property_id') %>% 
  dplyr::select(
    property_id, 
    location,
    avg_review_rating_num, 
    number_of_reviews_on_the_property_num,
    instant_booking_enabled,
    premier_partner_badge, 
    maximum_occupancy,
    bedrooms,
    bathrooms,
    number_of_photos_of_the_property_on_the_property_page,
    treat, 
    bookings_per_day, 
    booking_amount_per_day, 
    booking_requests_per_day, 
    property_page_views_per_visit_per_day,
    displays_in_searchpage_per_day
  ) %>% 
  drop_na() %>% 
  mutate_at(vars(-location), as.numeric)

Visualizing common support between the online visit group and non-online visit group:

Logistic regression model (only two features are statistically significant): avg_review_rating_num: average ratings property_page_views_per_visit_per_day: number of page views per visit per day (before 2019-01-01)

propensity_model <- glm(treat ~ avg_review_rating_num + number_of_reviews_on_the_property_num + instant_booking_enabled + premier_partner_badge + maximum_occupancy + bedrooms + bathrooms + number_of_photos_of_the_property_on_the_property_page + bookings_per_day + booking_amount_per_day + booking_requests_per_day + property_page_views_per_visit_per_day + displays_in_searchpage_per_day,
                  family = 'binomial',
                  data = property_to_match)

summary(propensity_model)

visualize_common_support(propensity_model)

Matching with optimal algorithm:

property_match <- matchit(treat ~ avg_review_rating_num + number_of_reviews_on_the_property_num + instant_booking_enabled + premier_partner_badge + maximum_occupancy + bedrooms + bathrooms + number_of_photos_of_the_property_on_the_property_page + bookings_per_day + booking_amount_per_day + booking_requests_per_day + property_page_views_per_visit_per_day + displays_in_searchpage_per_day,
                          method = 'optimal', 
                          data = property_to_match,
                          replace = TRUE, 
                          # Discard properties outside of common support region
                          discard = 'both')

summary(property_match)

property_matched <- match.data(property_match)
dim(property_matched)

Perform the t-test on variables selected for the propensity score matching procedure (the expectation is that they are not too different between the control and treatment group):

continuous_vars <- c('avg_review_rating_num', 
                     'number_of_reviews_on_the_property_num',
                     'instant_booking_enabled',
                     'premier_partner_badge', 
                     'maximum_occupancy',
                     'bedrooms',
                     'bathrooms',
                     'number_of_photos_of_the_property_on_the_property_page',
                     'bookings_per_day', 
                     'booking_amount_per_day', 
                     'booking_requests_per_day', 
                     'property_page_views_per_visit_per_day',
                     'displays_in_searchpage_per_day')

map(continuous_vars, vectorize_ttest_for_property, property_matched)

df_for_test <- property_matched %>% 
  left_join(metrics, by = 'property_id') %>% 
  filter(date > as.Date('2018-12-31')) %>% 
  group_by(property_id, treat) %>% 
  summarize(bookings_per_prop = mean(bookings, na.rm = TRUE),
            booking_amount_per_prop = mean(booking_amount, na.rm = TRUE),
            booking_requests_per_prop = mean(booking_requests, na.rm = TRUE),
            property_page_views_per_visit_per_prop = mean(property_page_views_visits, na.rm = TRUE),
            displays_in_searchpage_per_prop = mean(number_of_times_the_property_got_displayed_in_search_page, na.rm = TRUE)) %>% 
  ungroup()

Result: bookings_per_prop: stat sig (-0.0059) booking_amount_per_prop: not stat sig booking_requests_per_prop: stat sig (-0.0075) property_page_views_per_visit_per_prop: (-1.27991) * displays_in_searchpage_per_prop: not stat sig

# bookings_per_prop
# booking_amount_per_prop
# booking_requests_per_prop
# property_page_views_per_visit_per_prop
# displays_in_searchpage_per_prop

t.test(displays_in_searchpage_per_prop ~ treat, data = df_for_test,
       var.equal = FALSE, alternative = 'greater')

4. Matching - By Location

4.1. Florida

property_to_match_FL <- property_to_match %>% 
  filter(location == 'Destin, Florida')

propensity_model_FL <- glm(treat ~ avg_review_rating_num + number_of_reviews_on_the_property_num + instant_booking_enabled + premier_partner_badge + maximum_occupancy + bedrooms + bathrooms + number_of_photos_of_the_property_on_the_property_page + bookings_per_day + booking_amount_per_day + booking_requests_per_day + property_page_views_per_visit_per_day + displays_in_searchpage_per_day,
                  family = 'binomial',
                  data = property_to_match_FL)

visualize_common_support(propensity_model_FL)

property_match_FL <- matchit(treat ~ avg_review_rating_num + number_of_reviews_on_the_property_num + instant_booking_enabled + premier_partner_badge + maximum_occupancy + bedrooms + bathrooms + number_of_photos_of_the_property_on_the_property_page + bookings_per_day + booking_amount_per_day + booking_requests_per_day + property_page_views_per_visit_per_day + displays_in_searchpage_per_day,
                          method = 'optimal', 
                          data = property_to_match_FL,
                          replace = TRUE, 
                          discard = 'both')
property_matched_FL <- match.data(property_match_FL)

df_for_test_FL <- property_matched_FL %>% 
  left_join(metrics, by = 'property_id') %>% 
  filter(date > as.Date('2018-12-31')) %>% 
  group_by(property_id, treat) %>% 
  summarize(bookings_per_prop = mean(bookings, na.rm = TRUE),
            booking_amount_per_prop = mean(booking_amount, na.rm = TRUE),
            booking_requests_per_prop = mean(booking_requests, na.rm = TRUE),
            property_page_views_per_visit_per_prop = mean(property_page_views_visits, na.rm = TRUE),
            displays_in_searchpage_per_prop = mean(number_of_times_the_property_got_displayed_in_search_page, na.rm = TRUE)) %>% 
  ungroup()

Results for Florida: bookings_per_prop: not stat sig booking_amount_per_prop: not stat sig booking_requests_per_prop: not stat sig property_page_views_per_visit_per_prop: not stat sig * displays_in_searchpage_per_prop: not stat sig

# bookings_per_prop
# booking_amount_per_prop
# booking_requests_per_prop
# property_page_views_per_visit_per_prop
# displays_in_searchpage_per_prop

t.test(displays_in_searchpage_per_prop ~ treat, data = df_for_test_FL,
       var.equal = FALSE, alternative = 'greater')

4.2. Nevada

property_to_match_NV <- property_to_match %>% 
  filter(location == 'Vegas, Nevada')

propensity_model_NV <- glm(treat ~ avg_review_rating_num + number_of_reviews_on_the_property_num + instant_booking_enabled + premier_partner_badge + maximum_occupancy + bedrooms + bathrooms + number_of_photos_of_the_property_on_the_property_page + bookings_per_day + booking_amount_per_day + booking_requests_per_day + property_page_views_per_visit_per_day + displays_in_searchpage_per_day,
                  family = 'binomial',
                  data = property_to_match_NV)

visualize_common_support(propensity_model_NV)

property_match_NV <- matchit(treat ~ avg_review_rating_num + number_of_reviews_on_the_property_num + instant_booking_enabled + premier_partner_badge + maximum_occupancy + bedrooms + bathrooms + number_of_photos_of_the_property_on_the_property_page + bookings_per_day + booking_amount_per_day + booking_requests_per_day + property_page_views_per_visit_per_day + displays_in_searchpage_per_day,
                          method = 'optimal', 
                          data = property_to_match_NV,
                          replace = TRUE, 
                          discard = 'both')
property_matched_FL <- match.data(property_match_NV)

df_for_test_NV <- property_matched_FL %>% 
  left_join(metrics, by = 'property_id') %>% 
  filter(date > as.Date('2018-12-31')) %>% 
  group_by(property_id, treat) %>% 
  summarize(bookings_per_prop = mean(bookings, na.rm = TRUE),
            booking_amount_per_prop = mean(booking_amount, na.rm = TRUE),
            booking_requests_per_prop = mean(booking_requests, na.rm = TRUE),
            property_page_views_per_visit_per_prop = mean(property_page_views_visits, na.rm = TRUE),
            displays_in_searchpage_per_prop = mean(number_of_times_the_property_got_displayed_in_search_page, na.rm = TRUE)) %>% 
  ungroup()

Results for Nevada: bookings_per_prop: not stat sig booking_amount_per_prop: not stat sig booking_requests_per_prop: not stat sig property_page_views_per_visit_per_prop: not stat sig * displays_in_searchpage_per_prop: not stat sig

# bookings_per_prop
# booking_amount_per_prop
# booking_requests_per_prop
# property_page_views_per_visit_per_prop
# displays_in_searchpage_per_prop

t.test(displays_in_searchpage_per_prop ~ treat, data = df_for_test_NV,
       var.equal = FALSE, alternative = 'greater')

4.3. Texas

property_to_match_TX <- property_to_match %>% 
  filter(location == 'Austin, Texas')

propensity_model_TX <- glm(treat ~ avg_review_rating_num + number_of_reviews_on_the_property_num + instant_booking_enabled + premier_partner_badge + maximum_occupancy + bedrooms + bathrooms + number_of_photos_of_the_property_on_the_property_page + bookings_per_day + booking_amount_per_day + booking_requests_per_day + property_page_views_per_visit_per_day + displays_in_searchpage_per_day,
                  family = 'binomial',
                  data = property_to_match_TX)

visualize_common_support(propensity_model_TX)

property_match_TX <- matchit(treat ~ avg_review_rating_num + number_of_reviews_on_the_property_num + instant_booking_enabled + premier_partner_badge + maximum_occupancy + bedrooms + bathrooms + number_of_photos_of_the_property_on_the_property_page + bookings_per_day + booking_amount_per_day + booking_requests_per_day + property_page_views_per_visit_per_day + displays_in_searchpage_per_day,
                          method = 'optimal', 
                          data = property_to_match_TX,
                          replace = TRUE, 
                          discard = 'both')
property_matched_TX <- match.data(property_match_TX)

df_for_test_TX <- property_matched_TX %>% 
  left_join(metrics, by = 'property_id') %>% 
  filter(date > as.Date('2018-12-31')) %>% 
  group_by(property_id, treat) %>% 
  summarize(bookings_per_prop = mean(bookings, na.rm = TRUE),
            booking_amount_per_prop = mean(booking_amount, na.rm = TRUE),
            booking_requests_per_prop = mean(booking_requests, na.rm = TRUE),
            property_page_views_per_visit_per_prop = mean(property_page_views_visits, na.rm = TRUE),
            displays_in_searchpage_per_prop = mean(number_of_times_the_property_got_displayed_in_search_page, na.rm = TRUE)) %>% 
  ungroup()

Results for Texas: bookings_per_prop: stat sig (-0.00871577) booking_amount_per_prop: not stat sig booking_requests_per_prop: stat sig (-0.00996454) property_page_views_per_visit_per_prop: not stat sig * displays_in_searchpage_per_prop: not stat sig

# bookings_per_prop
# booking_amount_per_prop
# booking_requests_per_prop
# property_page_views_per_visit_per_prop
# displays_in_searchpage_per_prop

t.test(displays_in_searchpage_per_prop ~ treat, data = df_for_test_TX,
       var.equal = FALSE, alternative = 'greater')

hnguyen1174/online_visits_assessment documentation built on Dec. 20, 2021, 4:46 p.m.