PyMC Labs转发了
?????????????? ???? ?? ?????? ?????????? ?????????????? ??????????? ?????????? ???????????????? ?????? ?????? ???? ?????? ???????? “????????????” ???? ?????????????? ?????? ?????? ?????????? ?????????? ?? ?? I wanted to try out the new CausalPy package for causal inference and was brainstorming interesting research questions to apply it to. My inspiration came from a recent wave of LI posts in my feed, all sharing a chart showing a drop in Stack Overflow traffic after ChatGPT’s release in November 2022. ?? Given that many studies - including one we conducted internally at Sanofi - show that one of the most frequent use cases for GenAI in the business world is helping with email writing, I got curious about whether a similar pattern might show up in Google Trends for searches like “How to write an email.” ?? For the analysis, I used an interrupted time series method, which examines the effect of an intervention by comparing time series data before and after the intervention that happened at a known point in time, allowing us to assess any shifts in level or trend. ?? As the resulting charts show, there does seem to be a noticeable drop in searches on how to write emails following ChatGPT’s official release. In fact, the shift is so pronounced that it’s clear even with a quick eyeballing analysis. Sure, we can speculate about other factors to be involved - maybe we're sending fewer emails because we’re relying more on platforms like Slack or Teams - but as other available stats suggest, we can only wish that were the case ?? Either way, CausalPy turned out to be a super easy-to-use and user-friendly tool. I'm looking forward to using it in my future projects. Kudos to the entire CausalPy team! ?? To replicate the analysis, you can find the code on my blog here: https://lnkd.in/eauD3TV8 #causalinference #causalpy #genai #emailcommunication #python
A fun fact here is that at the end of each year the people knows how to write an email, and starting the new year they have to ask again ??????, just kidding. Thanks for this analysis Ludek Stehlik, Ph.D. !
Thanks for sharing this interesting content! Looking forward to seeing what other insights CausalPy can bring to light in your future work!
this analysis only counts if you use ChatGPT's data analysis tool to do it :). Given it's a unique package, that's probably not likely...(my frustration with using ChatGPT for coding...)
I'd suggest validating Causalpy results by first generating a data set using a known generation process you can control, and then measuring the degree to which it can identify the intended behavior. For example, I tried this with DoWhy/EconML and was shocked at not being able to replicate the capability of lmer()/glmer(). Trust (but verify!)
Very cool Ludek - looking forward to seeing the application!
Cool post! Both the hypothesis and the package. Putting aside the confounding factors involved as you mentioned. Another problem is the short history, which makes conclusions rather random and thin. For example: Is this the only time such a break occurred since the beginning of email and search adoption?
Nice!
Great example - your interrupted time series is especially important now that we have ubiquitous unobtrusive psychometrics based on digital exhaust where these powerful evaluative methods become much easier to deploy and automate with excellent statistical power from repeated measurements
Investigative Data Scientist / Operations Researcher / Mindset Challenger - PhD / MBA / CAP
4 个月I’m still stuck at needing help to write an email. Really??!!