Category: Uncategorized

The dose-response work of Callaway, Goodman-Bacon, and Pedro Sant’Anna seems to be coming along nicely. If you haven’t had enough of the parallel trends assumption, get ready for the “strong” parallel trends assumption!

“𝘐𝘯 𝘵𝘩𝘪𝘴 𝘱𝘢𝘱𝘦𝘳, 𝘸𝘦 𝘥𝘪𝘴𝘤𝘶𝘴𝘴 𝘢𝘯 𝘢𝘭𝘵𝘦𝘳𝘯𝘢𝘵𝘪𝘷𝘦 𝘣𝘶𝘵 𝘵𝘺𝘱𝘪𝘤𝘢𝘭𝘭𝘺 𝘴𝘵𝘳𝘰𝘯𝘨𝘦𝘳 𝘢𝘴𝘴𝘶𝘮𝘱𝘵𝘪𝘰𝘯, 𝘸𝘩𝘪𝘤𝘩 𝘸𝘦 𝘤𝘢𝘭𝘭 𝘴𝘵𝘳𝘰𝘯𝘨 𝘱𝘢𝘳𝘢𝘭𝘭𝘦𝘭 𝘵𝘳𝘦𝘯𝘥𝘴. 𝘚𝘵𝘳𝘰𝘯𝘨 𝘱𝘢𝘳𝘢𝘭𝘭𝘦𝘭 𝘵𝘳𝘦𝘯𝘥𝘴 𝘰𝘧𝘵𝘦𝘯 𝘳𝘦𝘴𝘵𝘳𝘪𝘤𝘵𝘴 𝘵𝘳𝘦𝘢𝘵𝘮𝘦𝘯𝘵 𝘦𝘧𝘧𝘦𝘤𝘵 𝘩𝘦𝘵𝘦𝘳𝘰𝘨𝘦𝘯𝘦𝘪𝘵𝘺 𝘢𝘯𝘥 𝘫𝘶𝘴𝘵𝘪𝘧𝘪𝘦𝘴 𝘤𝘰𝘮𝘱𝘢𝘳𝘪𝘯𝘨 𝘥𝘰𝘴𝘦 𝘨𝘳𝘰𝘶𝘱𝘴. 𝘐𝘯𝘵𝘶𝘪𝘵𝘪𝘷𝘦𝘭𝘺, 𝘵𝘰 𝘣𝘦 𝘢 𝘨𝘰𝘰𝘥 𝘤𝘰𝘶𝘯𝘵𝘦𝘳𝘧𝘢𝘤𝘵𝘶𝘢𝘭, 𝘭𝘰𝘸𝘦𝘳-𝘥𝘰𝘴𝘦 𝘶𝘯𝘪𝘵𝘴 𝘮𝘶𝘴𝘵 𝘳𝘦𝘧𝘭𝘦𝘤𝘵 𝘩𝘰𝘸 𝘩𝘪𝘨𝘩𝘦𝘳-𝘥𝘰𝘴𝘦 𝘶𝘯𝘪𝘵𝘴’ 𝘰𝘶𝘵𝘤𝘰𝘮𝘦𝘴 𝘸𝘰𝘶𝘭𝘥 𝘩𝘢𝘷𝘦 𝘤𝘩𝘢𝘯𝘨𝘦𝘥 𝘸𝘪𝘵𝘩𝘰𝘶𝘵 𝘵𝘳𝘦𝘢𝘵𝘮𝘦𝘯𝘵 𝘢𝘯𝘥 𝘢𝘵 𝘵𝘩𝘦 𝘭𝘰𝘸𝘦𝘳 𝘭𝘦𝘷𝘦𝘭 𝘰𝘧 𝘵𝘩𝘦 𝘵𝘳𝘦𝘢𝘵𝘮𝘦𝘯𝘵. 𝘞𝘦 𝘴𝘩𝘰𝘸 𝘵𝘩𝘢𝘵 𝘸𝘩𝘦𝘯 𝘰𝘯𝘦 𝘰𝘯𝘭𝘺 𝘪𝘮𝘱𝘰𝘴𝘦𝘴 𝘵𝘩𝘦 “𝘴𝘵𝘢𝘯𝘥𝘢𝘳𝘥” 𝘱𝘢𝘳𝘢𝘭𝘭𝘦𝘭 𝘵𝘳𝘦𝘯𝘥𝘴 𝘢𝘴𝘴𝘶𝘮𝘱𝘵𝘪𝘰𝘯, 𝘤𝘰𝘮𝘱𝘢𝘳𝘪𝘴𝘰𝘯𝘴 𝘢𝘤𝘳𝘰𝘴𝘴 𝘵𝘳𝘦𝘢𝘵𝘮𝘦𝘯𝘵 𝘥𝘰𝘴𝘢𝘨𝘦𝘴 𝘢𝘳𝘦 “𝘤𝘰𝘯𝘵𝘢𝘮𝘪𝘯𝘢𝘵𝘦𝘥” 𝘸𝘪𝘵𝘩 𝘴𝘦𝘭𝘦𝘤𝘵𝘪𝘰𝘯 𝘣𝘪𝘢𝘴 𝘳𝘦𝘭𝘢𝘵𝘦𝘥 𝘵𝘰 𝘵𝘳𝘦𝘢𝘵𝘮𝘦𝘯𝘵 𝘦𝘧𝘧𝘦𝘤𝘵 𝘩𝘦𝘵𝘦𝘳𝘰𝘨𝘦𝘯𝘦𝘪𝘵𝘺. 𝘛𝘩𝘶𝘴, 𝘸𝘪𝘵𝘩𝘰𝘶𝘵 𝘢𝘥𝘥𝘪𝘵𝘪𝘰𝘯𝘢𝘭 𝘴𝘵𝘳𝘶𝘤𝘵𝘶𝘳𝘦, 𝘤𝘰𝘮𝘱𝘢𝘳𝘪𝘴𝘰𝘯 𝘢𝘤𝘳𝘰𝘴𝘴 𝘥𝘰𝘴𝘢𝘨𝘦𝘴 𝘮𝘢𝘺 𝘯𝘰𝘵 𝘪𝘥𝘦𝘯𝘵𝘪𝘧𝘺 𝘤𝘢𝘶𝘴𝘢𝘭 𝘦𝘧𝘧𝘦𝘤𝘵𝘴. 𝘛𝘩𝘦 𝘱𝘭𝘢𝘶𝘴𝘪𝘣𝘪𝘭𝘪𝘵𝘺 𝘰𝘧 𝘴𝘵𝘳𝘰𝘯𝘨 𝘱𝘢𝘳𝘢𝘭𝘭𝘦𝘭 𝘵𝘳𝘦𝘯𝘥𝘴 𝘥𝘦𝘱𝘦𝘯𝘥𝘴 𝘰𝘯 𝘵𝘩𝘦 𝘦𝘮𝘱𝘪𝘳𝘪𝘤𝘢𝘭 𝘤𝘰𝘯𝘵𝘦𝘹𝘵 𝘰𝘧 𝘵𝘩𝘦 𝘢𝘯𝘢𝘭𝘺𝘴𝘪𝘴, 𝘢𝘯𝘥 𝘸𝘦 𝘥𝘪𝘴𝘤𝘶𝘴𝘴 𝘴𝘰𝘮𝘦 𝘧𝘢𝘭𝘴𝘪𝘧𝘪𝘤𝘢𝘵𝘪𝘰𝘯 𝘴𝘵𝘳𝘢𝘵𝘦𝘨𝘪𝘦𝘴 𝘵𝘩𝘢𝘵 𝘤𝘢𝘯 𝘣𝘦 𝘶𝘴𝘦𝘥 𝘵𝘰 𝘢𝘴𝘴𝘦𝘴𝘴 𝘪𝘵.”

Source

2024 will be the year of lawsuits against generative AI companies. We’ve already had the GitHub Copilot case over assisted coding (https://lnkd.in/eH8Ap-eJ) and the Anthropic case over AI lyrics (https://lnkd.in/eY6UF9Cn). Now the Times joins the fray (https://lnkd.in/e8wHHMzx), and more are likely to follow.

So far, Gen AI companies have defended themselves by arguing fair use and transformative use – that their models create something substantially new and serve a different purpose than the original (thus doesn’t substitute the original, as in Google Books). But recent Supreme Court decisions such as Warhol v. Goldsmith made clear that transformative use claims face high bars.

What might come next?
– New business models for content licensing
– Restrictions on public access to some internal models
– Calls for updated copyright laws and content use regulations
– Technical solutions like attribution, data provenance, and content tagging
– What else?

In our latest short post at Data Duets, Duygu Dagli and I share a recent paper (and an R package) that offers a better solution.

How often do you log transform your data? How do you deal with zeros? Do you just add 1 or some other constant before taking the log?

In our latest short post, Gorkem Turgut (G.T.) Ozer and I share a recent paper (and an R package) that offers a better solution!

Source

The founding event of artificial intelligence as a field is considered to be the 1956 Dartmouth Workshop in Hanover, New Hampshire.

The proposal listed seven areas of focus for AI: automation of higher-level functions, language models, neural networks, computational efficiency, self-learning, abstraction and generalization from sensor data, and creativity.

These were all revolutionary ideas at the time (and still are), but the one that stands out to me the most is creativity:

“𝘈 𝘧𝘢𝘪𝘳𝘭𝘺 𝘢𝘵𝘵𝘳𝘢𝘤𝘵𝘪𝘷𝘦 𝘢𝘯𝘥 𝘺𝘦𝘵 𝘤𝘭𝘦𝘢𝘳𝘭𝘺 𝘪𝘯𝘤𝘰𝘮𝘱𝘭𝘦𝘵𝘦 𝘤𝘰𝘯𝘫𝘦𝘤𝘵𝘶𝘳𝘦 𝘪𝘴 𝘵𝘩𝘢𝘵 𝘵𝘩𝘦 𝘥𝘪𝘧𝘧𝘦𝘳𝘦𝘯𝘤𝘦 𝘣𝘦𝘵𝘸𝘦𝘦𝘯 𝘤𝘳𝘦𝘢𝘵𝘪𝘷𝘦 𝘵𝘩𝘪𝘯𝘬𝘪𝘯𝘨 𝘢𝘯𝘥 𝘶𝘯𝘪𝘮𝘢𝘨𝘪𝘯𝘢𝘵𝘪𝘷𝘦 𝘤𝘰𝘮𝘱𝘦𝘵𝘦𝘯𝘵 𝘵𝘩𝘪𝘯𝘬𝘪𝘯𝘨 𝘭𝘪𝘦𝘴 𝘪𝘯 𝘵𝘩𝘦 𝘪𝘯𝘫𝘦𝘤𝘵𝘪𝘰𝘯 𝘰𝘧 𝘢 𝘴𝘰𝘮𝘦 𝘳𝘢𝘯𝘥𝘰𝘮𝘯𝘦𝘴𝘴.”

Today, most generative AI models seem to follow this idea of injecting some randomness. But can a touch of randomness turn 𝘶𝘯𝘪𝘮𝘢𝘨𝘪𝘯𝘢𝘵𝘪𝘷𝘦 𝘤𝘰𝘮𝘱𝘦𝘵𝘦𝘯𝘵 𝘵𝘩𝘪𝘯𝘬𝘪𝘯𝘨 into 𝘤𝘳𝘦𝘢𝘵𝘪𝘷𝘪𝘵𝘺? Well, this 𝘤𝘭𝘦𝘢𝘳𝘭𝘺 𝘪𝘴 𝘢𝘯 𝘪𝘯𝘤𝘰𝘮𝘱𝘭𝘦𝘵𝘦 𝘤𝘰𝘯𝘫𝘦𝘤𝘵𝘶𝘳𝘦.

Randomness alone can’t make a model imaginative. Imagination requires an understanding of cause-effect relationships and counterfactual reasoning.

𝘈𝘯 𝘈𝘐 𝘮𝘰𝘥𝘦𝘭 𝘤𝘢𝘯 𝘱𝘦𝘳𝘧𝘦𝘤𝘵𝘭𝘺 𝘳𝘦𝘱𝘳𝘰𝘥𝘶𝘤𝘦 𝘵𝘩𝘦 𝘤𝘰𝘭𝘰𝘳𝘴 𝘰𝘧 𝘢 𝘴𝘶𝘯𝘴𝘦𝘵 𝘪𝘯 𝘱𝘪𝘹𝘦𝘭𝘴, 𝘺𝘦𝘵 𝘪𝘵 𝘸𝘰𝘶𝘭𝘥 𝘧𝘢𝘪𝘭 𝘵𝘰 𝘨𝘳𝘢𝘴𝘱 𝘵𝘩𝘦 𝘤𝘳𝘦𝘢𝘵𝘪𝘷𝘦 𝘴𝘱𝘢𝘳𝘬 𝘵𝘩𝘢𝘵 𝘵𝘶𝘳𝘯𝘴 𝘢 𝘮𝘦𝘳𝘦 𝘱𝘪𝘤𝘵𝘶𝘳𝘦 𝘪𝘯𝘵𝘰 𝘢 𝘸𝘰𝘳𝘬 𝘰𝘧 𝘢𝘳𝘵.

That’s why the more exciting potential today seems to lie in creative human input to a model, or in using the output of the model as input to creative human brain.

Source

Daniel Millimet and Marc F. Bellemare work on an interesting paper on the feasibility of assuming that fixed effects are fixed over long periods in causal inference models. They highlight an overlooked reality that fixed effects may fail to control for unobserved heterogeneity over long periods of time.

One lesson for causal identification using long panels is to think twice before assuming that fixed effects will take care of unobserved heterogeneity.

More on this is in our short post with Duygu Dagli at Data Duets. She uses rapid gentrification as an example. The short format is a new idea to post more often.

Source

This is the tag line on boomy.com, a generative AI-based music platform.

We have entered an era where everyone seems to be able to 𝘤𝘳𝘦𝘢𝘵𝘦 𝘰𝘳𝘪𝘨𝘪𝘯𝘢𝘭 𝘟𝘟𝘟 𝘪𝘯 𝘴𝘦𝘤𝘰𝘯𝘥𝘴, 𝘦𝘷𝘦𝘯 𝘪𝘧 𝘵𝘩𝘦𝘺’𝘷𝘦 𝘯𝘦𝘷𝘦𝘳 𝘤𝘳𝘦𝘢𝘵𝘦𝘥 𝘟𝘟𝘟 𝘣𝘦𝘧𝘰𝘳𝘦, a fact that leads to the need for a new vocabulary, since “original” is defined as

– created directly and personally by a particular artist
– not dependent on other people’s ideas

Meanwhile, “Boomy artists have created 18,047,099 original songs” and that 𝘴𝘰𝘶𝘯𝘥𝘴 great.

Imagination (of counterfactuals) draws on domain knowledge and creativity and is key to causal reasoning; it is also where humans continue to outperform algorithms. What about the role of data?

𝘖𝘯𝘦 𝘵𝘳𝘪𝘤𝘬 𝘪𝘴 𝘵𝘰 𝘶𝘴𝘦 𝘥𝘢𝘵𝘢 𝘢𝘴 𝘢 𝘩𝘦𝘭𝘱𝘧𝘶𝘭 𝘨𝘶𝘪𝘥𝘦, 𝘯𝘰𝘵 𝘵𝘩𝘦 𝘰𝘯𝘭𝘺 𝘨𝘶𝘪𝘥𝘦. 𝘐𝘵’𝘴 𝘭𝘪𝘬𝘦 𝘶𝘴𝘪𝘯𝘨 𝘢 𝘮𝘢𝘱. 𝘈 𝘮𝘢𝘱 𝘤𝘢𝘯 𝘵𝘦𝘭𝘭 𝘺𝘰𝘶 𝘸𝘩𝘦𝘳𝘦 𝘵𝘩𝘦 𝘴𝘵𝘳𝘦𝘦𝘵𝘴 𝘢𝘳𝘦, 𝘣𝘶𝘵 𝘪𝘵 𝘤𝘢𝘯’𝘵 𝘵𝘦𝘭𝘭 𝘺𝘰𝘶 𝘪𝘧 𝘵𝘩𝘦𝘳𝘦’𝘴 𝘢 𝘱𝘢𝘳𝘢𝘥𝘦. 𝘐𝘵 𝘤𝘢𝘯’𝘵 𝘵𝘦𝘭𝘭 𝘺𝘰𝘶 𝘸𝘩𝘢𝘵 𝘦𝘭𝘴𝘦 𝘮𝘪𝘨𝘩𝘵 𝘩𝘢𝘷𝘦 𝘣𝘦𝘦𝘯 𝘩𝘢𝘱𝘱𝘦𝘯𝘪𝘯𝘨 𝘦𝘪𝘵𝘩𝘦𝘳. 𝘠𝘰𝘶 𝘩𝘢𝘷𝘦 𝘵𝘰 𝘭𝘰𝘰𝘬 𝘶𝘱 𝘧𝘳𝘰𝘮 𝘵𝘩𝘦 𝘮𝘢𝘱 𝘢𝘯𝘥 𝘶𝘴𝘦 𝘺𝘰𝘶𝘳 𝘦𝘺𝘦𝘴 𝘢𝘯𝘥 𝘺𝘰𝘶𝘳 𝘣𝘳𝘢𝘪𝘯.

This map distorts the data on the relative positions of the continents, but it uses the data correctly on their location and size. It’s just as wrong as any other world map, but it makes one look up, think, and imagine.

What might have happened if
– Central America was in the arctic zone and Siberia was subtropical
– Cuba was off the east coast of Canada and the USA
– Japan was off the coasts of Portugal and Spain
– North Korea was part of South Korea
– Taiwan was next to France

Source

Gilbert may have a point here, except for the part about how markets work. This is a popular free market product with a growing number of customers. Today I heard that ChatGPT Plus now has a waiting list. I also find Gilbert’s statement a bit overdramatic, but that’s beside the point. LLMs are useful tools if their limitations are well defined, communicated, and acknowledged, and if the lingering issues of copyright and privacy are resolved.

But if not, and if we continue to treat a computer model as if it had some kind of consciousness and general intelligence, progress will be painful. I’ve expressed this concern several times, using Searle’s Chinese Room argument and pointing out the dangers of so-called “hallucinations” in the hands of well-meaning users who don’t always know what they don’t know.

Source

I started coding in C and used C# before switching to Java (those were the dark days). Today, R and Python make up my daily stack (and I dabble in Julia for fun). Now, in retrospect, one habit I wish the data science community had developed better is exception handling. More often than not, I come across code that is doomed to crash (hey, it’s called an exception for a reason).

The next time you code to analyze data, remember this video of division by zero. This is the pain and suffering your computer goes through when your denominator hits zero. This is also a reminder to appreciate (not complain about) the errors and warnings in R and Python (my former and future students, hear me?).

Received the video in personal communication and happy to add a source.

…because experts are more likely to know what they don’t know. When users don’t know what they don’t know, so-called “hallucinations” are less likely to be detected and this seems to be a growing problem, likely exacerbated by the Dunning-Kruger effect. Well, this is my take on them.

In the study cited in the article, several LLM models are asked to summarize news articles to measure how often they “hallucinated” or made up facts.

The LLM models showed different rates of “hallucination”, with OpenAI having the lowest (about 3%), followed by Meta (about 5%), Anthropic’s Claude 2 system (over 8%), and Google’s Palm chat with the highest (27%).

Source