Information on the internet was born free, but now lives everywhere in walled gardens. Blogging sometimes feels like a throwback to an earlier era. So many newer platforms have eclipsed blogs in popularity, almost all of which are harder to search and discover. Facebook was walled off from the beginning, Twitter is becoming more so. Podcasts and video tend to be open in theory, but hard to search as most lack transcripts. Longer-form writing is increasingly hidden behind paywalls on news sites and Substack. People have complained for years that Google search is getting worse; there are many reasons for this, like a complacent company culture and the cat-and-mouse game with SEO companies, but one is this rising tide of content that is harder to search and link.
To me part of the value of blogging is precisely that it remains open in an increasingly closed world. Its influence relative to the rest of the internet has waned since its heydey in ~2009, but most of this is due to how the rest of the internet has grown explosively at the expense of the real world; in absolute terms the influence of blogging remains high, and perhaps rising.
The closing internet of late 2023 will not last forever. Like so much else, AI is transforming it, for better and worse. AI is making it cheap and easy to produce transcripts of podcasts and videos, making them more searchable. Because AI needs large amounts of text to train models, text becomes more valuable. Open blogs become more influential because they become part of the training data for AI; because of what we have written here, AI will think and sound a little bit more like us. I think this is great, but others have the opposite reaction. The New York Times is suing to exclude their data from training AIs, and to delete any models trained with it. Twitter is becoming more closed partly in an attempt to limit scraping by AIs.
So AI leads to human material being easier for search engines to index, and some harder; it also means there will be a flood of AI-produced material, mostly low-quality, clogging up search results. The perpetual challenge of search engines putting relevant, high-quality results first will become much harder, a challenge which AI will of course be set to solve. Search engines already have surprisingly big problems with not indexing writing at all; searching for a post on my old blog with exact quotes and not finding it made me realize Google was missing some posts there, and Bing and DuckDuckGo were missing all of them. While we’re waiting for AI to solve and/or worsen this problem, Gwern has a great page of tips on searching for hard-to-find documents and information, both the kind that is buried deep down in Google and the kind that is not there at all.
Today I’ll go into more detail on several measures of the labor force, but I won’t only compare it to 2019. I’ll compare it to all available data. And the sum total of the data suggests the 2023 was one of the best years for the US labor market on record. Note: December 2023 data isn’t available until January 5th, so I’m jumping the gun a little bit. I’m going to assume December looks much like November. We can revisit in 2 weeks if that was wrong.
The Unemployment Rate has been under 4% for the entire year. The last time this happened (date goes back to 1948) was 1969, though 2022 and 2019 were both very close (just one month at 4%). In fact, the entire period from 1965-1969 was 4% or less, though following January 1970 there wasn’t single month under 4% under the year 2000!
Like GDP, the Unemployment Rate is one of the broadest and most widely used macro measures we have, but they are also often criticized for their shortcomings, as I wrote in an April 2023 post.
With that in mind, let’s look to some other measures of the labor market.
The Differences-in-Differences literature has blown up in the past several years. “Differences-in-Differences” refers to a statistical method that can be used to identify causal relationships (DID hereafter). If you’re interested in using the new methods in Stata, or just interested in what the big deal is, then this post is for you.
First, there’s the basic regression model where we have variables for time, treatment, and a variable that is the product of both. It looks like this:
The idea is that that there is that we can estimate the effect of time passing separately from the effect of the treatment. That allows us to ‘take out’ the effect of time’s passage and focus only on the effect of some treatment. Below is a common way of representing what’s going on in matrix form where the estimated y, yhat, is in each cell.
Each quadrant includes the estimated value for people who exist in each category. For the moment, let’s assume a one-time wave of treatment intervention that is applied to a subsample. That means that there is no one who is treated in the initial period. If the treatment was assigned randomly, then β=0 and we can simply use the differences between the two groups at time=1. But even if β≠0, then that difference between the treated and untreated groups at time=1 includes both the estimated effect of the treatment intervention and the effect of having already been treated prior to the intervention. In order to find the effect of the intervention, we need to take the 2nd difference. δ is the effect of the intervention. That’s what we want to know. We have δ and can start enacting policy and prescribing behavioral changes.
Easy Peasy Lemon Squeezy. Except… What if the treatment timing is different and those different treatment cohorts have different treatment effects (heterogeneous effects)?* What if the treatment effects change over time the longer an individual is treated (dynamic effects)**? Further, what if the there are non-parallel pre-existing time trends between the treated and untreated groups (non-parallel trends)?*** Are there design changes that allow us to estimate effects even if there are different time trends?**** There’re more problems, but these are enough for more than one blog post.
For the moment, I’ll focus on just the problem of non-parallel time trends.
What if untreated and the to-be-treated had different pre-treatment trends? Then, using the above design, the estimated δ doesn’t just measure the effect of the treatment intervention, it also detects the effect of the different time trend. In other words, if the treated group outcomes were already on a non-parallel trajectory with the untreated group, then it’s possible that the estimated δ is not at all the causal effect of the treatment, and that it’s partially or entirely detecting the different pre-existing trajectory.
Below are 3 figures. The first two show the causal interpretation of δ in which β=0 and β≠0. The 3rd illustrates how our estimated value of δ fails to be causal if there are non-parallel time trends between the treated and untreated groups. For ease, I’ve made β=0 in the 3rd graph (though it need not be – the graph is just messier). Note that the trends are not parallel and that the true δ differs from the estimated delta. Also important is that the direction of the bias is unknown without knowing the time trend for the treated group. It’s possible for the estimated δ to be positive or negative or zero, regardless of the true delta. This makes knowing the time trends really important.
STATA Implementation
If you’re worried about the problems that I mention above the short answer is that you want to install csdid2. This is the updated version of csdid & drdid. These allow us to address the first 3 asterisked threats to research design that I noted above (and more!). You can install these by running the below code:
program fra syntax anything, [all replace force] local from "https://friosavila.github.io/stpackages" tokenize `anything' if "`1'`2'"=="" net from `from' else if !inlist("`1'","describe", "install", "get") { display as error "`1' invalid subcommand" } else { net `1' `2', `all' `replace' from(`from') } qui:net from http://www.stata.com/ end fra install fra, replace fra install csdid2 ssc install coefplot
Once you have the methods installed, let’s examine an example by using the below code for a data set. The particulars of what we’re measuring aren’t important. I just want to get you started with the an application of the method.
local mixtape https://raw.githubusercontent.com/Mixtape-Sessions use `mixtape'/Advanced-DID/main/Exercises/Data/ehec_data.dta, clear qui sum year, meanonly replace yexp2 = cond(mi(yexp2), r(max) + 1, yexp2)
The csdid2 command is nice. You can use it to create an event study where stfips is the individual identifier, year is the time variable, and yexp2 denotes the times of treatment (the treatment cohorts).
The above output shows us many things, but I’ll address only a few of them. It shows us how treated individuals differ from not-yet treated individuals relative to the time just before the initial treatment. In the above table, we can see that the pre-treatment average effect is not statistically different from zero. We fail to reject the hypothesis that the treatment group pre-treatment average was identical to the not-yet treated average at the same time period. Hurrah! That’s good evidence for a significant effect of our treatment intervention. But… Those 8 preceding periods are all negative. That’s a little concerning. We can test the joint significance of those periods:
estat event, revent(-8/-1)
Uh oh. That small p-value means that the level of the 8 pretreatment periods significantly deviate from zero. Further, if you squint just a little, the coefficients appear to have a positive slope such that the post-treatment values would have been positive even without the treatment if the trend had continued. So, what now?
Wouldn’t it be cool if we knew the alternative scenario in which the treated individuals had not been treated? That’s the standard against which we’d test the observed post-treatment effects. Alas, we can’t see what didn’t happen. BUT, asserting some premises makes the job easier. Let’s say that the pre-treatment trend, whatever it is, would have continued had the treatment not been applied. That’s where the honestdid stata package comes in. Here’s the installation code:
local github https://raw.githubusercontent.com net install honestdid, from(`github'/mcaceresb/stata-honestdid/main) replace honestdid _plugin_check
What does this package do? It does exactly what we need. It assumes that the pre-treatment trend of the prior 8 periods continues, and then tests whether one or more post-treatment coefficients deviate from that trend. Further, as a matter of robustness, the trend that acts as the standard for comparison is allowed to deviate from the pre-treatment trend by a multiple, M, of the maximum pretreatment deviations from trend. If that’s kind of wonky – just imagine a cone that continues from the pre-treatment trend that plots the null hypotheses. Larger M’s imply larger cones. Let’s test to see whether the time-zero effect significantly differs from zero.
What does the above table tell us? It gives us several values of M and the confidence interval for the difference between the coefficient and the trend at the 95% level of confidence. The first CI is the original time-0 coefficient. When M is zero, then the null assumes the same linear trend as during the pretreatment. Again, M is the ratio by which maximum deviations from the trend during the pretreatment are used as the null hypothesis during the post-treatment period. So, above, we can see that the initial treatment effect deviates from the linear pretreatment trend. However, if our standard is the maximum deviation from trend that existed prior to the treatment, then we find that the alpha is just barely greater than 0.05 (because the CI just barely includes zero).
That’s the process. Of course, robustness checks are necessary and there are plenty of margins for kicking the tires. One can vary the pre-treatment periods which determine the pre-trend, which post-treatment coefficient(s) to test, and the value of M that should be the standard for inference. The creators of the honestdid seem to like the standard of identifying the minimum M at which the coefficient fails to be significant. I suspect that further updates to the program will come along that spits that specific number out by default.
I’ve left a lot out of the DID discussion and why it’s such a big deal. But I wanted to share some of what I’ve learned recently with an easy-to-implement example. Do you have questions, comments, or suggestions? Please let me know in the comments below.
The above code and description is heavily based on the original author’s support documentation and my own Statalist post. You can read more at the above links and the below references.
*Sun, Liyang, and Sarah Abraham. 2021. “Estimating Dynamic Treatment Effects in Event Studies with Heterogeneous Treatment Effects.” Journal of Econometrics, Themed Issue: Treatment Effect 1, 225 (2): 175–99. https://doi.org/10.1016/j.jeconom.2020.09.006.
**Sant’Anna, Pedro H. C., and Jun Zhao. 2020. “Doubly Robust Difference-in-Differences Estimators.” Journal of Econometrics 219 (1): 101–22. https://doi.org/10.1016/j.jeconom.2020.06.003.
***Callaway, Brantly, and Pedro H. C. Santa Anna. 2021. “Difference-in-Differences with Multiple Time Periods.” Journal of Econometrics, Themed Issue: Treatment Effect 1, 225 (2): 200–230. https://doi.org/10.1016/j.jeconom.2020.12.001.
****Rambachan, Ashesh, and Jonathan Roth. 2023. “A More Credible Approach to Parallel Trends.” The Review of Economic Studies 90 (5): 2555–91. https://doi.org/10.1093/restud/rdad018.
Lately many journalists and folks on X/Twitter have pointed out a seeming disconnect: by almost any normal indicator, the US economy is doing just fine (possibly good or great). But Americans still seem dissatisfied with the economy. I wanted to put all the data showing this disconnect into one post.
In particular, let’s make a comparison between November 2019 and November 2023 economic data (in some cases 2019q3 and 2023q3) to see how much things have changed. Or haven’t changed. For many indicators, it’s remarkable how similar things are to probably the last month before anyone most normal people ever heard the word “coronavirus.”
First, let’s start with “how people think the economy is doing.” Here’s two surveys that go back far enough:
The University of Michigan survey of Consumer Sentiment is a very long running survey, going back to the 1950s. In November 2019 it was at roughly the highest it had ever been, with the exception of the late 1990s. The reading for 2023 is much, much lower. A reading close to 60 is something you almost never see outside of recessions.
The Civiqs survey doesn’t go back as far as the Michigan survey, but it does provide very detailed, real-time assessments of what Americans are thinking about the economy. And they think it’s much worse than November 2019. More Americans rate the economy as “very bad” (about 40%) than the sum of “fairly good” and “very good” (33%). The two surveys are very much in alignment, and others show the same thing.
The government continues to be great at collecting data but not so good at sharing it in easy-to-use ways. That’s why I’ve been on a quest to highlight when independent researchers clean up government datasets and make them easier to use, and to clean up such datasets myself when I see no one else doing it; see previous posts on State Life Expectancy Data and the Behavioral Risk Factor Surveillance System.
National Health Expenditure Accounts Historical State Data: The original data from the Centers for Medicare and Medicaid Services on health spending by state and type of provider are actually pretty good as government datasets go: they offer all years (1980-2020) together in a reasonable format (CSV). But it comes in separate files for overall spending, Medicare spending, and Medicaid spending; I merge the variables from all 3 into a single file, transform it from a “wide format” to a “long format” that is easier to analyze in Stata, and in the “enhanced” version I offer inflation-adjusted versions of all spending variables. Excel and Stata versions of these files, together with the code I used to generate them, are here.
A warning to everyone using the data, since it messed me up for a while: in the documentation provided by CMMS, Table 3 provides incorrect codes for most variables. I emailed them about this but who knows when it will get fixed. My version of the data should be correct now, but please let me know if you find otherwise. You can find several other improved datasets, from myself and others, on my data page.
We study whether people will pay for a fact-check on AI writing. ChatGPT can be very useful, but human readers should not trust every fact that it reports. Yesterday’s post was about ChatGPT writing false things that look real.
The reason participants in our experiment might pay for a fact-check is that they earn bonus payments based on whether they correctly identify errors in a paragraph. If participants believe that the paragraph does not contain any errors, they should not pay for a fact-check. However, if they have doubts, it is rational to pay for a fact-check and earn a smaller bonus, for certain.
Abstract: We explore whether people trust the accuracy of statements produced by large language models (LLMs) versus those written by humans. While LLMs have showcased impressive capabilities in generating text, concerns have been raised regarding the potential for misinformation, bias, or false responses. In this experiment, participants rate the accuracy of statements under different information conditions. Participants who are not explicitly informed of authorship tend to trust statements they believe are human-written more than those attributed to ChatGPT. However, when informed about authorship, participants show equal skepticism towards both human and AI writers. There is an increase in the rate of costly fact-checking by participants who are explicitly informed. These outcomes suggest that trust in AI-generated content is context-dependent.
Our original hypothesis was that people would be more trusting of human writers. That turned out to be only partially true. Participants who are not explicitly informed of authorship tend to trust statements they believe are human-written more than those attributed to ChatGPT.
We presented information to participants in different ways. Sometimes we explicitly told them about authorship (informed treatment) and sometimes we asked them to guess about authorship (uninformed treatment).
This graph (figure 5 in our paper) shows that the overall rate of fact-checking increased when subjects were given more explicit information. Something about being told that a paragraph was written by a human might have aroused suspicion in our participants. (The kids today would say it is “sus.”) They became less confident in their own ability to rate accuracy and therefore more willing to pay for a fact-check. This effect is independent of whether participants trust humans more than AI.
We are thinking of fact-checking as often a good thing, in the context of our previous work on ChatGPT hallucinations. So, one policy implication is that certain types of labels can cause readers to think critically. For example, Twitter labels automated accounts so that readers know when content has been chosen or created by a bot.
Suggested Citation: Buchanan, Joy and Hickman, William, Do People Trust Humans More Than ChatGPT? (November 16, 2023). GMU Working Paper in Economics No. 23-38, Available at SSRN: https://ssrn.com/abstract=4635674
Citation: Buchanan, J., Hill, S., & Shapoval, O. (2024). ChatGPT Hallucinates Non-existent Citations: Evidence from Economics. The American Economist. 69(1), 80-87 https://doi.org/10.1177/05694345231218454
Blog followers will know that we reported this issue earlier with the free version of ChatGPT using GPT-3.5 (covered in the WSJ). We have updated this new article by running the same prompts through the paid version using GPT-4. Did the problems go away with the more powerful LLM?
The error rate went down slightly, but our two main results held up. It’s important that any fake citations at all are being presented as real. The proportion of nonexistent citations was over 30% with GPT-3.5, and it is over 20% with our trial of GPT-4 several months later. See figure 2 from our paper below for the average accuracy rates. The proportion of real citations is always under 90%. GPT-4, when asked about a very specific narrow topic, hallucinates almost half of the citations (57% are real for level 3, as shown in the graph).
The second result from our study is that the error rate of the LLM increases significantly when the prompt is more specific. If you ask GPT-4 about a niche topic for which there is less training data, then a higher proportion of the citations it produces are false. (This has been replicated in different domains, such as knowledge of geography.)
What does Joy Buchanan really think?: I expect that this problem with the fake citations will be solved quickly. It’s very brazen. When people understand this problem, they are shocked. Just… fake citations? Like… it printed out reference for papers that do not actually exist? Yes, it really did that. We were the only ones who quantified and reported it, but the phenomenon was noticed by millions of researchers around the world who experimented with ChatGPT in 2023. These errors are so easy to catch that I expect ChatGPT will clean up its own mess on this particular issue quickly. However, that does not mean that the more general issue of hallucinations is going away.
Not only can ChatGPT make mistakes, as any human worker can mess up, but it can make a different kind of mistake without meaning to. Hallucinations are not intentional lies (which is not to say that an LLM cannot lie). This paper will serve as bright clear evidence that GPT can hallucinate in ways that detract from the quality of the output or even pose safety concerns in some use cases. This generalizes far beyond academic citations. The error rate might decrease to the point where hallucinations are less of a problem than the errors that humans are prone to make; however, the errors made by LLMs will always be of a different quality than the errors made by a human. A human research assistant would not cite nonexistent citations. LLM doctors are going to make a type of mistake that would not be made by human doctors. We should be on the lookout for those mistakes.
ChatGPT is great for some of the inputs to research, but it is not as helpful for original scientific writing. As prolific writer Noah Smith says, “I still can’t use ChatGPT for writing, even with GPT-4, because the risk of inserting even a small number of fake facts… “
I still can't use ChatGPT for writing, even with GPT-4, because the risk of inserting even a small number of fake facts or bad interpretations into a blog post is unacceptable, meaning that it requires so much time to fact-check that it doesn't save effort.
Some economists love to write about sports because they love sports. Others love to write about sports because the data are so good compared to most other facets of the economy. What other industry constantly releases film of workers doing their jobs, and compiles and shares exhaustive statistics about worker performance?
To take an extreme example, suppose an average high-school athlete got thrown into a professional football or basketball game; a fan asked to evaluate them could probably figure out that they don’t belong there within minutes, or perhaps even just by glancing at them and seeing they are severely undersized. But what if an average high school coach were called up to coach at the professional level? How long would it take for a casual observer to realize they don’t belong? You might be able to observe them mismanaging games within a few weeks, but people criticize professional coaches for this all the time too; I think you couldn’t be sure until you see their record after a season or two. Even then it is much less certain than for a player- was their bad record due to their coaching, or were they just handed a bad roster to work with?
The sports economics literature seems to confirm my intuition that coaches are difficult to evaluate. This is especially true in football, where teams generally play fewer than 20 games in a season; a general rule of thumb in statistics is that you need at least 20 to 25 observations for statistical tests to start to work. This accords with general practice in the NFL, where it is considered poor form to fire a coach without giving him at least one full season. One recent article evaluating NFL coaches only tries to evaluate those with at least 3 seasons. If the article is to be believed, it wasn’t until 2020 that anyone published a statistical evaluation of NFL defensive coordinators, despite this being considered a vital position that is often paid over a million dollars a year:
The third quarter ‘All Transaction’ housing price data was just released this week. These numbers are interesting for a few of reasons. One reason is that home prices are a big component of our cost of living. Higher home prices are relevant to housing affordability. This week’s release is especially interesting because it’s starting to look like the Fed might be pausing its year 18-month streak of interest rates hikes. In case you don’t know, higher interest rates increase the cost of borrowing and decrease the price that buyers are willing to pay for a home. Nationally, we only had one quarter of falling home prices in late 2022, but the recent national growth rate in home prices is much slower than it was in 2021 through mid-2022.
Do you remember when there were a bunch of stories about remote workers and early retirees fleeing urban centers in the wake of Covid? We stopped hearing that story so much once interest rates started rising. The inflection point in the data was in Q2 of 2022. After that, price growth started slowing with the national average home price up 6.5%. But the national average masks some geographic diversity.
In the October 1980 Presidential debate, Ronald Reagan famously asked that question to the American voters. His next sentence made it clear he was talking about the relationship between prices and wages, or what economists call real wages: “is it easier for you to go and buy things in the stores than it was four years ago?”
Reagan was a master of political rhetoric, so it’s not surprising that many have tried to copy his question in the years since 1980. For example, Romney and Ryan tried to use this phrase in their 2012 campaign against Obama. But it’s a good question to ask! While the President may have less control over the economy than some observers think, the economy does seem to be a key factor in how voters decide (for example, Ray Fair has done a pretty good job of predicting election outcomes with a few major economic variables).
Voters in 2024 will probably be asking themselves a similar question, and both parties (at least for now) seem to be actively encouraging voters to make such a comparison. We still have 12 months of economic data to see before we can really ask the “4 years” question, but how would we answer that question right now? Here’s probably the best approach to see if people are “better off” in terms of being able to “go and buy things at the stores”: inflation-adjusted wages. This chart presents average wages for nonsupervisory workers, with two different inflation adjustments, showing the change over a 4-year time period.
Economist Writing Every Day 