RUC course project: Advanced Statistical Analysis in Spring 2023.
data.h5 and vtb.csv are from https://github.com/wandleshen/vup-data-analysis (data.csv is translated from data.h5 and I added some features on vtb.csv to construct vtb2.csv.)
Code is contained in vup.Rmd and the output file is vup.html.
My presentation slides is presentation.pdf and final report is analysis_report.pdf. acc_form.docx is JASA-required author contributions checklist (ACC) form.
Thanks for your reading! If your time is limited, just read analysis_report.pdf!
Select 50 different virtual hosts and get revenue and engagement data for their last 50 live streams. Their names, projects or unions they belong to, and user IDs are all manually entered into a csv file for subsequent data grabbing.
- Python package:
pandas version 1.4.4 (optional, just to translate data.h5 to data.csv)
- R packages:
reticulate version 1.28 (optional, just for python interface), GGalary version 2.1.2, pheatmap version 2.1.2, ggplot2 version 3.4.1, grf version 2.2.1, lubridate version 1.9.2, marginaleffects version 0.11.1, randomForest version 4.7-1.1, reshape2 version 1.4.4, MatchIt version 4.5.3, dyplr version 2.3.1, extrafont version 0.11
I do data analysis within Rmarkdown file vup.Rmd.
First read data with python virtual environment in ananconda. (You maybe need to modify this part by your environment.) Then use R to do other things: EDA, random forest, regression adjustment, causal forest and matching method with both regression adjustment and causal forest.
I will show both regression adjustment and causal forest with(out) matching method. The results of different methods are quite similar.
Overall, more viewer exposure can bring more benefit. While if conditioning on viewer exposure, affilliation of vup still makes about 20% difference but whether starting live on weekends or not makes no difference.
So many thanks! Thank https://github.com/wandleshen/ for his data. Thank https://weibo.com/u/7743187451 for her support. Thank Prof. Ma for his kindness.