Kyle Carlson
@KyleCSN
Followers
722
Following
2K
Media
247
Statuses
1K
Experiments tech lead @stripe Previous: Taught experiments @USF_Economics Econ PhD @Caltech RA @BostonFed @UNIL
United States
Joined April 2013
@Empty_America If you have children, at a minimum, you really want a post-1950s home. Paint manufacturers started restricting lead content in interior paints. Source: https://t.co/lYy1GB3L79
0
2
1
Many families are locked into 3% mortgages and renovating their homes. Beware the risks. I know of 2 families that renovated recently and lead poisoned their young children. Here’s an example of a Victorian house that poisoned a whole family, pregnant mom, abortion, dogs, nanny.
1
0
1
@Empty_America Finnish schools with a no shoe rule have less airborn lead , cadmium and other metals. https://t.co/XBgapvNPTb
0
2
7
🧪 I'm excited to announce that I will teach a new course about experimentation (A/B tests) on the @get_sphere platform starting April 3. We will go deep into the nuts and bolts of engineering and analysis. See the link below for details. https://t.co/31yc9RcRqq
0
3
17
Retirement accounts do not count towards M2 money supply. So my anti inflation policy proposal is a temporary increase in after tax IRA and 401k contribution limits of $x0,000 per year.
0
0
3
Rather it’s a mechanically perfect layup when it could’ve been an outrageous slam dunk.
0
0
0
Hotels is a *natural* context for Trump but the voice is generic. The structure is good with the buildup of tension and Snoop accusing Trump of bias. Big miss: Snoop didn’t mention Holiday Inn or Versace “7 star” hotel. The dialogue is a missed layup.
1
0
0
ChatGPT is surprisingly good at translating the Trump style into a super niche context (Stata do file) including even the regression model. But the dialogue with Snoop is mixed…🧵
2
0
1
Heterogeneous treatment effects and doubly robust learning Doubly Robust Learning — econml 0.13.1 documentation https://t.co/PNh10lJGEg Susan Athey, "Machine Learning and Causal Inference for Policy Evaluation" - YouTube
1
0
0
Previous posts: Clustered standard errors https://t.co/nP9DnZ9lkC Random effects, fixed effects and power https://t.co/vKk2lHw1HL Hypothesis testing and decision performance in A/B tests
Decision strategies in experiments: *Flipping a coin* can outperform hypothesis testing. Why? Conventional alpha levels can be inappropriately conservative. Note: I don't recommend to actually flip a coin (but it might make you happier). Explanation of simulation in thread. 1/N
1
0
0
If enough people are interested I’ll write a new blog post with code. Options: 1/Doubly robust causal learning vs vanilla ML + econML 2/Surrogate index 3/Something else (suggest in replies) See thread for details and previous posts (so you know what you’ll get).
1
0
1
Anyone know good writing about API design for scientific computing and modeling? More content like below? Bonus if it discusses trade offs for combining learning and scoring in one class like scikit learn vs separate classes.
0
0
2
significance level = 5% power = 80%
What's a scam that's become so normalized that we don't even realize it's a scam anymore?
0
0
2
Brilliant move by the data science intern that created this model
0
0
4
It was too slow to do the simulations I wanted to run. But I had no problem with DoubleML (python) which is a similar technique and well integrated with sklearn.
0
0
3
The R ecosystem does itself a disservice with one-off packages that easily become unusably slow. I tried playing with tmle. The toy example in the docs took 12 minutes to run just increasing N to 10k. Not to pick on tmle. I’ve seen this in many other cases.
1
0
3
The coexistence of large data science ecosystems in both R and Python is a hugely costly coordination failure.
0
2
9