Our company is founded on the belief that data can be used to manage people better. Following the recent furore surrounding Amazon’s treatment of their workers, we thought it would be interesting to see whether their dissatisfaction is measurable using publicly available data. Of course, if Jeff would like us to analyse his internal data, we’d be more than happy to help, but until then we thought we’d see what we can do anyway.
Caveat: The following analysis is primarily intended as a proof of concept, and a bit of fun. It’s probably not going to be accepted into the Journal of Managerial Psychology anytime soon.
Where did the data come from?
We downloaded 1.2 million tweets from approximately 1000 Twitter profiles belonging to engineers and developers from Google, Apple, Microsoft, Amazon, and Facebook. These tweets were then analysed for sentiment using a modified version of PredictionIO’s sentiment analyser. We’ve included a detailed description of how the data was collected at the end of the article.
So – are Amazon engineers unhappy?
Not if the sentiment of their tweets is anything to go by, at least compared to other tech engineers. Of course they could all be unhappy, but I’ll save that for another blog post. In fact, Amazon engineers had the highest percentage of strongly positive tweets out of the whole group. Although these might appear to be small differences (e.g. 9.29% vs. 8%) it’s a very large sample size (these are statistically significant differences) and you wouldn’t expect a huge variation in data such as this.
Percentage of tweets, by company, that were strongly positive.
Company mentions as a measure of loyalty. AKA Dogfooding.
With all the tweet data, we could also measure how often employees mentioned their employer. Interestingly Amazon also came out near top on this measure, marginally beaten by Microsoft, which by some metrics does have the most loyal employees compared to the other four. One caveat here is Apple, which is hard to measure due its wide variety of product names – Apple, Mac, iOS, iPhone etc – so this is likely an underestimate.
Percentage of tweets, by company, that mentioned their employer.
What apps are engineers tweeting about? Not Foursquare!
It didn’t come as a surprise to us, and probably won’t to you either, but the data confirmed Foursquare (we also included Swarm, not that it made a difference!) has had a massive fall from grace. Vine, Instagram, and Snapchat seem pretty much on par with each other, with Periscope making a late impact. The huge spike in Instagram tweets at the beginning of 2012 was the Facebook acquisition.
Percentage of tweets that mentioned a particular app.
What technology are engineers tweeting about? React.
We also thought it would be interesting to see what technologies the engineers were tweeting about. Ruby fading away did not come as much of a surprise, but the decline in mentions of mongoDB was unexpected. React appears to be tearing away, eating into Angular’s mindshare, at least in this sample.
Percentage of tweets that mentioned a particular technology.
What clients are they using?
Finally we looked at how the mix of Twitter clients differed across the companies. No prize for guessing that Android is not particularly popular in Cupertino. The level of iOS usage at Google was quite surprising though, although in their defence Google does have to develop apps for iOS as well. Interestingly there were basically zero Apple Mac users in our sample of Microsofties, which must make developing OS X apps quite hard.
Twitter client usage by company.
Having numbers on Amazon in isolation wouldn’t really tell us much; we needed some cohorts to benchmark against. The obvious choice were the other members of Eric Schmidt’s big four – Google, Apple, and Facebook – with Microsoft thrown in as a wildcard. We also wanted an apples to apples comparison – it wouldn’t be accurate to compare 100 Google account managers to 100 Facebook engineers – so we limited the cohorts to technical roles only, e.g. those mentioning engineer, developer etc.
We then took these criteria and ran them against the Twitter User API, searching for profiles that matched terms such as “developer @google” and “facebook engineer” and so on. This gave us a big (over 5k profiles) but messy list, thanks to the proliferation of SEO consultants and Social Media experts that stuff these terms in their bios. To get a usable list we were going to need to do quite a bit of post-processing.
We started by removing anyone that matched the obvious terms; SEO, consultant, agency etc. Then we removed profiles that matched “previously at Google” or “prior; Facebook” to try and ensure we had only current job roles. For good measure we then discarded anyone below 100 followers, and anyone who hadn’t tweeted in the last month. This took the sample down to approximately 1k profiles, which we eyeballed to double check they were good matches, removing any bad ones we’d missed along the way. At the end we were left with a solid sample of between 100 and 200 profiles per company.
With the profiles chosen, we then set to work retrieving their tweets. Over the course of a couple of days (to avoid hitting the rate limit) we pulled up to 2000 tweets per profile, most recent first. By the end we had 1.2 million tweets to play with!
Sentiment for each tweet was then scored using a modified version of PredictionIO’s sentiment analyser. We certainly owe PredicitionIO a beer or two for providing such an excellent (and free) product.
It’s early days for this kind of analysis, and it would be a bit of a stretch to draw any definite conclusions, but one thing’s for certain; data is becoming increasingly abundant, and it can and will be used for many purposes, from measuring how often someone tweets about their employer to predicting this season’s new fashion.
Although this was a fun experiment, Peakon already has concrete ways of measuring employee engagement, retention, and culture. If you are interested in getting immediate, actionable, and data-driven insights into how you can improve your business, drop us an email!