Prior to 2013, monthly synthetic opioid deaths averaged just a couple hundred a month. But in the years that followed, deaths increased sharply to nearly 7,000 at the peak in mid-2023. Then, suddenly, they started to plummet. By late 2024, monthly deaths had more than halved. As of September 2025, the most recent month for which data can be estimated, deaths have fallen to around 2,600, a drop of more than 60% from the peak.

This has not gone entirely unnoticed. In January, a study in Science documented a supply shock in the illicit fentanyl market: declining purity in seized drugs, fewer seizures overall, and even complaints on social media about fentanyl becoming harder to find. The decline showed up simultaneously in Canada, ruling out US-specific explanations like changes in policing or treatment access.

The most likely cause, according to the study, is that China cracked down on exports of fentanyl precursor chemicals, the raw ingredients that Mexican cartels use to manufacture the drug. That crackdown came out of sustained diplomatic pressure from the Biden administration, culminating in a formal agreement at the Biden-Xi summit in November 2023. As Jeffrey Prescott of the Carnegie Endowment wrote, “Foreign policy outcomes can be hard to measure. This one isn’t.”

The chart above extends through September 2025, using CDC provisional data adjusted for reporting lags. What the extra months reveal is that the decline has continued. The annualized death toll from synthetic opioids is now around 32,000, down from over 75,000 at the peak. To put that in perspective, a drop of over 40,000 annual deaths is comparable to the number of Americans who die annually in car accidents.

Whether this continues is an open question. The diplomatic arrangement that appears to have driven the decline is fragile. It depends on Chinese cooperation that could be withdrawn, and on a precursor supply chain that cartels may eventually route around. But for now, the data suggests one of the largest public policy successes since the development of the COVID-19 vaccines in 2020.

Plumbers aren’t alone. The chart below shows the change in relative wage position for a number of occupations.

The decline in relative wage standing occurred across many blue collar middle-skills occupations, including welders, electricians and CNC programmers. But it has also affected jobs with minimal training requirements, such as forklift drivers, meat cutters and general laborers. In a previous post, I noted the U.S. economy’s shift away from goods-producing industries, and toward services. The declines in relative wages for many occupations reflect that broader structural shift.

Exceptions to the declines in middle-skills wages

Several middle-skills jobs have maintained or improved their relative wage position in the past four decades. Caring occupations that provide healthcare and healthcare support have generally seen increases in their wages relative to the median since 1980. These occupations have strong prospects as the aging population drives continued demand for these services.

Other middle-skills occupations that have performed well include protective services occupations. Police officers and firefighters both earned wages 28% higher than the median wage in 1980, and both have increased their advantage slightly to 33% above the median in 2023.

The value of college

Looking beyond middle skills, it’s evident that jobs that require a college education tend to pay above the median wage, and that many of those jobs have actually seen increased wages relative to the median over the past four decades. To economists, the persistence of the college wage premium is no mystery. Wages reflect the supply and demand for labor, and the demand for college-educated labor has remained strong even as growth in college attainment has slowed. Some people argue that the college wage premium is not causal and reflects selection effects, but the evidence suggests otherwise. College causes substantial financial benefits to the students themselves, to society and even to government budgets.

To the general public, this might be surprising. Skepticism about the value of college has become broadly popular. People focus on rising sticker prices without realizing that expanding financial aid has kept the net cost of college attendance basically flat for many years. “Skip college, learn a trade” has become a popular mantra among adults advising younger generations. Mike Rowe and other commentators routinely disparage college education, while touting an always-approaching shortage of non-college-educated workers. Even advertising by Indiana’s public community college system subtly suggests that 4-year colleges are a bad deal:

Vocational education and training are good. There are many people who will benefit immensely from acquiring additional skills. As we’ve seen above, there are many middle-skills occupations that offer increased economic opportunities, such as in healthcare, policing and firefighting. Even the occupations that have declined in relative standing offer some financial benefits. Welders still earn more than warehouse workers.

But promoting vocational education and training by publicizing falsehoods about the value of 4-year degrees is a mistake. Recent research shows that community colleges provide significant benefits to people who would otherwise have only a high school diploma, but those benefits can be offset by worsened outcomes for people who are diverted from attending 4-year college in favor of 2-year college. Promoting vocational education and training in a way that undercuts attendance at 4-year institutions is misleading and potentially harmful to many people. It also threatens future economic prosperity by putting downward pressure on college attainment rates at a time when the evidence suggests the opposite is needed. Instead, we should understand that there are many middle-skills jobs that pay well and represent economic opportunities for people, while being truthful about which jobs are most likely to provide economic opportunity.

Code for this post is available on GitHub.

Given the lack of evidence for the President’s outlandish claims, administration officials critics quickly turned their focus to the historical accuracy of the monthly jobs numbers instead. In this process, a number of people have made errant claims about the data, including defenders of the Bureau’s work. In order to better understand the role and the value of the jobs numbers, we need to understand where these numbers come from.

The “jobs number” in the BLS monthly employment report comes from the Current Employment Statistics (CES) program, a.k.a. the payroll survey. The survey1 asks businesses to provide the number of employees on payroll, as well as their hours and wages. The headline jobs number produced by the program reflects the total number of net jobs gained or lost each month.

There is an initial estimate released for the prior month. This estimate is then revised in the two subsequent employment reports, based largely on additional responses to the survey. Later on, the series of jobs numbers is adjusted again, to align with actual jobs counts from the Bureau’s Quarterly Census of Employment and Wages (QCEW). This process, known as “benchmarking”, doesn’t do much to correct the monthly change estimates. It serves mainly to bring the total jobs levels back in line with reality.

So how good are the initial estimates from the CES? It’s an important question. Unfortunately both critics and defenders of the data are making some important errors in attempting to answer it. Here are three crucial facts about the jobs numbers:

1. Revision size is not an measure of accuracy.

Several attempts to both critique and defend the BLS jobs numbers treat the difference between the initial estimates and the revised estimates as a measure of the survey’s error. This is a mistake, as the revised estimate does not reflect the true change in employment, but is merely an updated estimate from the same program. Treating revisions as errors is analogous to assessing the accuracy of Nate Silver’s August election forecast by judging it against his October forecast, rather than judging it against the actual results of the election. As mentioned earlier, the true number of the jobs gained or lost in any given month is knowable. It is released several months later as part of the Bureau’s QCEW program, which is based on mandatory unemployment insurance filings.

2. The CES attempts to estimate the net monthly change in jobs, not the overall level of employment.

This error, coming mostly from defenders of the CES numbers, argues that the error in the jobs numbers is actually tiny by comparing the error size (typically around 100K) to the total jobs level (approaching 160M), rather than to the net change in jobs. This misunderstands the methodology and purpose of the program, which is explicitly designed to measure the change number, and only produces the jobs levels by applying the estimated change percentages to the previous month’s employment levels. As discussed above, the levels are ultimately anchored to previously-released QCEW counts.

Furthermore, total employment tends to change by less than 0.2% each month, so regularly achieving a small “error rate” in the estimate of total employment is an extremely low bar for a statistical project. In reality, the CES is designed to do something much more difficult, and more useful.

3. Accuracy is subjective. What matters is whether the estimates are useful.

Whether the estimates are considered accurate or not depends on whether the they contain useful information. In the case of the CES, larger errors might be offset by their timeliness. The estimates are typically released on the first Friday of the month following the month that is estimated. The true jobs numbers are not known until several months later, which is too late to be useful for many purposes, such as setting monetary policy.

My analysis below reveals that while the estimates might appear inaccurate based on the magnitude of the errors, they still contain useful information that would otherwise be unavailable so early.

What the Data Shows

The correct way to assess the initial CES estimates is to compare them to actual counts from QCEW2. Doing this, I find:

• CES initial estimates of over-the-month changes explain 65% of the variation in the actual employment changes, suggesting that they almost certainly contain useful information for policymakers.

• The unexplained variation is still significant. This means the initial estimates tend to be off by 126K jobs on average over the comparison period.

• The CES initial estimates tend to understate the magnitude of changes in employment by around 11.5%, so an initial estimate of 200K jobs gained actually predicts a gain of around 226K, on average.

• Furthermore, the estimates have minimal overall bias, tending to understate the monthly change in jobs by just 4.4K over the entire period.

All in all, this is a fairly impressive feat. BLS is reliably capturing information about monthly employment changes in almost real time. There might be some question about the appropriateness of how the news media reports on the monthly numbers. Economists have long warned that the numbers have large error ranges and suggested that the initial estimates receive disproportionate attention compared to the longer term trends and revised estimates. But the numbers still appear to be useful for policymakers.

Tracking the errors over time reveals some interesting trends. Specifically, the absolute error has tended to hover slightly above 100K, except for the period from 2020 to 2022, where it more than doubled. This is likely due to lower survey response rates that began during the COVID-19 pandemic. Since 2022, the error rates seem to have returned to normal.

Error direction

The net error of the CES initial estimates over longer periods such as presidential terms has remained small, as positive errors tend to cancel out negative ones over time. Under Bush and Obama, the CES initial estimates understated actual job growth by an average of 4.7K and 15.7K jobs per month, respectively. Under Trump’s first term, the net error was slightly negative, understating job growth by just 300 jobs per month. Under Biden, the net error was slightly positive at 9.3K jobs per month.

These are all small numbers considering the absolute magnitude of the net change has averaged 334K jobs per month from 2003 to 2024.

Private alternatives have less value

The main private alternative to the CES estimates is the ADP National Employment Report. This report is usually released two days prior to the BLS report, but only covers private payrolls. Evaluating these estimates against QCEW private employment suggests that they are not better than the CES estimates.

During months where both reports’ data are available, ADP estimates appear to be slightly inferior to the CES estimates in predicting the actual change in employment. It explains less variation in the monthly change (28%) than the CES (45%), and it overestimates the magnitude of the changes by 47%. Plus, the ADP estimates are only available for private employment and contains less detailed breakouts than the CES numbers.

But perhaps more importantly, privately-produced numbers lack the transparency and oversight that have driven the U.S. government to consistently produce high-quality data for the benefit of the public. The BLS and other statistical agencies have built up technical and bureaucratic barriers to political and other external influence. The President’s political attacks on the BLS threaten these safeguards, and subsequent claims about the accuracy of the data appear to be unfounded.

All of the code and data for this analysis are available here.

The yield curve, the difference between short- and long-term interest rates, has been a reliable indicator of future economic recessions for at least half a century. When the curve inverts, meaning short-term rates exceed long-term rates, this typically suggests that investors think the economy is nearing an economic downturn within the next year or two. But the last time the yield curve inverted was in April 2022, and as of July 2025, the U.S. economy has avoided recession for 38 months. This beats the 34 months between the 1998 yield curve inversion and the 2001 recession, making it the longest recession-free stretch following a yield curve inversion since 1978.

The yield curve wasn’t alone in predicting a recession. Following the yield curve inversion, a WSJ survey of economists rated a future recession as more likely than not for an entire year. A Bloomberg model declared a 100% chance of entering a recession before October 2023. The early warning Sahm indicator suggested that we’d entered a recession beginning in July 2024.

So far since 2022, unemployment has risen very slightly and leveled off around 4.2%. Payroll employment has increased consistently, continuing the second-longest streak of positive job growth since the series began in 1919. However, some cracks are beginning to show. July's jobs report suggests weaker growth in payrolls: The three-month average of net job growth has fallen to just 35,000.

Also troubling is that the integrity of future economic data is in jeopardy due to President Trump’s firing of BLS Commissioner Erika McEntarfer within hours of the report’s release. Trump accompanied the firing with groundless assertions that the numbers were “rigged” for political purposes. This politicization of economic data threatens the credibility of information that businesses, investors, and policymakers rely on for critical decisions.

Furthermore, Trump’s bizarre insistence on imposing constantly-changing tariff schedules on our trade partners and his continued obsession with deporting immigrants are likely to put downward pressure on the economy. Whether these policies are enough to push the economy into recession remains to be seen.

Unfortunately, this is not going to work. Manufacturing’s share of the U.S. economy has been falling since the 1940s. While the decline has slowed in the past decade, the shift has already devastated communities across the country that were previously dependent on manufacturing activity.

But protectionist policies won’t bring back the jobs that have been lost. The decline of the manufacturing sector isn’t a specific thing that happened just in the United States. It has happened in every developed country so far. The chart below shows how three main sectors of the economy evolve as economies become richer:

• Agriculture’s share of the economy drops as output increases

• Manufacturing’s share increases initially, but once an economy’s productivity exceeds around $40,000 (2017 USD), it begins to fall

• The services sector’s share increases continuously throughout the economy’s growth, becoming dominant after manufacturing’s decline

Importantly, these trends appear similar across all developed countries’ economies regardless of their specific timeline of economic development. In the U.S., the manufacturing sector’s peak occurred in the 1940s. In South Korea, it was around 1990. China’s manufacturing sector is currently in the process of leveling off.

Some refer to the process of how employment shifts between these three sectors as “tertiarization”, since economies invariably become dominated by the production of services, i.e. the tertiary sector. In a wealthier economy, demand shifts toward services such as healthcare, education, entertainment and leisure. Remaining manufacturing shifts away from labor-intensive goods and more towards products that require more specialized skills and technologies.

Since tertiarization appears to be an unavoidable part of an economic progress, attempting to turn back the clock with trade restrictions is unlikely to produce the desired results. Instead, economists recommend public investment in declining communities. A recent report from Tim Bartik at the Upjohn Institute suggests that the most impactful policies for struggling manufacturing-heavy communities include:

• Customized services for small and medium businesses, including manufacturing extension services and job training services

• Public spending on education, from preschool and K-12 to colleges and universities, as well as vocational education and job training programs for workers

• Public investment in infrastructure and increasing land supply for business and housing development

Less effective policies include broad tax cuts for business, targeted business attraction incentives and attempts to reduce workers’ wages. One of the best ways to fund investment in distressed communities is with broad-based tax increases, while one of the worst ways is to cut K-12 education spending.

Data and code available here.

First, the analysis script and summary data is available here.

Second, I've realized that there was a mode change that I failed to note in the initial post. The Monitoring the Future survey switched from paper forms to electronics tablets between 2018 and 2020. Although most users of this dataset ignore this change, its effects can be estimated.

Based on 2019 data, where both modes were tested together, the mode change could account for a 1.5 points of the 15.2 point drop in complete agreement on the equal pay question, and 5.6 of the 17.8 point drop on the equal opportunity question.

The critically-acclaimed Netflix show, Adolescence, has stoked public conversation about the problem of misogyny and violence among school-aged boys. It’s been discussed in op-eds, on morning shows and by the British Prime Minister. Many commentators accept the show’s fictional portrayal of a 14-year-old boy’s descent into a misogynist online subculture as representative of real social trends. But until now, there’s been little hard evidence of this trend, leaving room for some to claim that concern about increasing sexism among boys is just a moral panic.

Unfortunately, the trend appears to be real. The long-running Monitoring the Future study has been surveying 8th and 10th graders since 1991, and shows a sharp drop in the proportion of 8th and 10th graders that believe in gender equality in the last five years.

In 2018, 84% of 8th and 10th grade boys agreed that women should have the same job opportunities as men. But in the last five years, the number dropped to 72%. The proportion of boys who completely agreed (as opposed to “mostly agree”) saw an even steeper drop, from 63% to 45%.

The share of boys agreeing that women deserve equal pay also fell from 87% in 2018 to 79% in 2023. The share who agreed completely fell from 72% to 57%.

Is it social media?

Many commentators on the topic of misogyny among boys are quick to blame social media and internet subcultures such as the “manosphere” or “incels”. Some point to evidence that social media algorithms tend to amplify misogynistic content. This particular hypothesis is not confirmed by data from this survey.

In 2018, the survey started asking children about their time spent using social networks, watching videos and playing video games. But it asked only about the amount they spent on each activity, not about the nature of the content they consumed. The data shows that the sharper drops in the share of boys endorsing gender equality occurred in those who spend the least time using social networks.

It also seems that boys who spent the least amount of time watching videos experienced the largest decreases in support for gender equality.

Video games also appear unlikely as a culprit, as again, it was the non-gamers who had the largest decreases in beliefs in gender equality since 2018.

This certainly seems to undercut the popular thesis that social media and the internet are driving changing views on gender equality. It is still possible that these media formats play a role in changing views about gender equality, but it probably isn’t a simple function of the amount of time spent on them.

Social isolation and lack of romantic relationships

Is the decreasing popularity of gender equality related to social isolation? Maybe not in the way people might expect. Gender equality appears to have become less popular among both social and non-social boys, but the decrease appears to have been larger for boys who are more social, not less.

For decades, support for gender equality has been similar for boys who get together with friends informally at least once a week, and those who don’t. But in recent years, as both proportions declined, support among the more social group declined more.

A similar story emerges with respect to dating activity. Contrary to the hypothesis that misogyny is driven by romantic struggles, support for gender equality has typically been lower among boys who go on dates, and has decreased by a similar amount among boys who go on dates and those who don’t.

Absent fathers

One evergreen hypothesis regarding the behavior of boys is that it all relates to the absence of a father figure in the household. But views of gender equality have been almost identical between boys who did or didn’t have a father in the household for the past three decades. And that doesn’t seem to have changed during the current decline.

Education

In general, boys with less educated parents tend to endorse gender equality at lower rates. But this gap has actually closed in recent years, as the decline in egalitarian views has been steeper among boys with college-educated mothers than those with non-college-educated mothers.

Parenting styles

Maybe the communication styles of parents are important? Some boys report being able to discuss any problems they have with their parents, while others are not comfortable discussing everything with their parents. Endorsement of gender equality appears to have fallen regardless of whether boys feel comfortable talking to their parents about their problems.

Religion

Religion is one factor that actually shows some promise in explaining why belief in gender equality has lost ground with boys. Views of gender equality among boys who say religion is not important in their life have not decreased by nearly as much in the past five years. The share of boys agreeing completely that women deserve equal job opportunities fell 22 percentage points, from 60% in 2018 to 38% in 2023. Complete agreement with equal pay for women also fell by 21 percentage points, from 71% to 50%.

The fact that the popularity of gender equality plummeted so much among religious boys seems to be the main clue available in the survey about what factors might be driving this trend. Unfortunately, the survey contains little other information on religious or political beliefs of these 8th and 10th graders, and further research will be necessary to understand what is driving these trends.

Two recent attempts to defend Sanders' assertion come from Seth Ackerman at Jacobin and Matt Bruenig at People's Policy Project. While everyone ends up using the same dataset (the 2022 Survey of Consumer Finances), we also get a lot of stretching of definitions and chopping off sections of the dataset to try to get it to match Sanders’ stated estimate.

The Technical Questions

Not all of the definition-stretching and data filtering is unreasonable. I’ll accept the chosen metric of days of income in liquid savings. I’ll also accept restricting the universe to working-age Americans, since neither elderly people nor children tend to be depending on their own paychecks.

As far as how many days of income in liquid savings consitutes living paycheck-to-paycheck? There is really no need to fuss around on this one. Only 10% of American workers are paid less frequently than semimonthly, and the most common paycheck frequency is biweekly. Someone who is living paycheck-to-paycheck would have anywere from 0 to 14 days of liquid savings at any particular moment.

Ackerman's main move to get his number to 60% is to exclude business owners from the calculation. His reasoning is that these folks probably disregarded the survey instructions to exclude business accounts when listing personal assets, which causes their business wealth to be counted as personal savings. The editor at Jacobin doesn’t offer much evidence that this happened except that business owners are wealthy, so we just have to accept his theory that the survey fails to capture the true financial struggles of the capital-owning class. I won’t accept this because I think the real reason that business owners appear wealthy in the survey is that business owners tend to be wealthy. And even if business owners do sometimes intermingle personal and business accounts, those accounts still represent a liquidity source that provides financial security which is the entire point of this question.

The Answer

Of households where the head is age 25 to 64, an estimated 30% had less than 14 days of liquid savings in 2022. Moving the number of days in either direction shows the sensitivity of the estimate to the definition changes. About 20% had less than 7 days of savings, and 47% had less than 30 days of savings.

One thing that is striking is that these percentages were at historic lows in 2022. The financial cushions of American households have actually improved significantly over time, especially from 1995 to 2022. Ackerman suggests that things might have worsened significantly since 2022, which might be true. But until the next set of data comes out, it is still speculation. Certainly some excess savings from temporary pandemic assistance programs may have been drawn down, but also the economy has continued to expand continuously since 2022.

The Points We're Missing

This debate over precise percentages has probably obscured some more important points. First, having very little financial cushion is bad, so we should acknowledge when it improves. And since much of the improvement in 2022 might be attributed to policies Sanders has supported, such as pandemic assistance and the expanded child tax credit, he should take some credit for it.

But also, even at these historic lows, having only two weeks of savings is a precarious state, and 30% of American households are in it. There is no need to inflate this statistic at all. It is a bad number!

Sometimes one metric doesn’t capture everything that’s wrong. It’s fine to talk about more than one thing. Ackerman gets around to this at the end of his article when he addresses the relative inequality of improvements to people’s financial conditions. Many people are better off, but the biggest benefits have been to people who were already financially secure.

Since 1995, liquid savings at the 80th percentile increased from 57 days of income to 127 days. This dwarfs the increase at the median, from 15 days to 30 days. It doesn’t take any exaggeration to point out that there is tremendous inequality in financial security in the United States, and that the gains, while real, have not been evenly distributed.

• Computer and electronic products

• Pharmaceuticals & medicine

• Aerospace products and parts

While these sub-sectors don’t capture the entirety of high-tech manufacturing, they do represent a large portion of it. Using 2022 County Business Patterns data from the Census Bureau, I’ve created a map to summarize the current state of high-tech manufacturing employment in the U.S. (Note: I’m adding medical devices—NAICS 339112 surgical and medical instruments—in as well):

I was curious how this landscape has changed, so I went back 35 years to make a version for 19871. Click here to view an animated version. This comparison reveals that while traditional centers like Silicon Valley, Seattle, and Boston's Route 128 corridor remain important, the intensity of employment has generally decreased across most regions. Some areas have maintained their specializations—aerospace in Seattle and Wichita, pharmaceuticals in New Jersey, and electronics in Silicon Valley—but at lower overall employment levels. Pharmaceuticals and medical devices have increased in relative prominence since 1987.

National employment numbers confirm a large drop in employment in high-tech manufacturing. Since 1990, employment has fallen by nearly 1 million in these sub-sectors. Computers and electronics dropped by 850K while aerospace employment fell by nearly 300K. Only pharmaceuticals and medical devices gained employment, with a net increase of just 189K.

As a proportion of national employment, the decline has been even steeper. High-tech manufacturing employment dropped from 2.8% to 1.3%, a 50% drop in the share of total employment.

High-tech manufacturing’s portion of the economy has also been falling, from 4.8% in 1987 to just 2.6% in 2023.

While the decline of high-tech manufacturing is evident in both employment and output trends, some optimism has been sparked by 2022’s CHIPS and Science Act, which included billions in tax credits to promote semiconductor manufacturing located in the U.S. While the CHIPS Act seems to have led to a boom in factory construction for computer and electronic products, the effect on employment in the industry remains difficult to see. While the act’s employment effects might not be visible due to the fact that the semiconductor fabrication facilities have not opened yet, delays in opening due to workforce issues at TSMC’s Pheonix plant suggest potential staffing problems. President Trump’s funding freeze and preference for tarriffs over subsidies could throw even more uncertainty into the mix. Regardless, estimates of the employment effect of the CHIPS Act have ranged from just 36,300 to 56,000 jobs2, which would offset just 4-6% of the total declines over the past 35 years.

Indiana lawmakers are proposing to cut the potential duration of unemployment insurance (UI) benefits from 26 weeks to 14 weeks, following similar moves by several other states in recent years. While proponents argue this will encourage faster returns to work, recent research questions the size of the effects and suggests there are important downsides. Cutting unemployment benefits might achieve the goal of slightly faster returns to work, but the evidence suggests this might come at a significant cost to workers and to the broader economy.

Regular unemployment benefits in the United States are available for up to 26 weeks in most states. This is already shorter than the typical benefits duration in developed countries, though the U.S. does tend to extend benefits during recessions. A maximum duration of 14 weeks would be the 4th shortest of all 50 states.

Will Shorter Benefits Increase Employment?

The most common justification offered for shortening unemployment benefits duration is that it will increase job-finding rates and re-employment. Studies generally suggest that cutting unemployment benefits duration by one week reduces unemployment duration by about 0.1-0.2 weeks1, though these estimates are affected by significant publication bias2. Effects are likely to be smaller during periods of low unemployment than during recessions3, and effects will be even smaller in the U.S. where UI benefits duration is already relatively low4. For Indiana's proposed 12-week reduction, this suggests that workers might, at best, find jobs 1-2 weeks faster on average.

The Costs of Shorter Unemployment Benefits

But faster job-finding comes at a price. This is because unemployment benefits subsidize job search, which is a productive activity. Recent research has found that increasing unemployment benefits duration can lead to higher job quality and higher re-employment wages5. The additional liquidity provided by unemployment benefits allows workers more time to find positions that better match their skills and provide higher wages, especially for workers who might otherwise not have the funds for a longer job search6. The higher-quality matches between workers and jobs also mean that workers stay at their new jobs longer than they otherwise would7.

All of this matters not just for individual workers, but for the broader economy. When workers are forced into poor job matches, it can lead to higher turnover, lower productivity, and reduced economic growth.

Additional potential costs of shortened unemployment benefits include lower labor force participation8 and more poverty9. Workers who fail to find work before exhausting unemployment benefits become more likely to drop out of the labor force entirely. They also face higher poverty rates and require more support from public assistance programs such as SNAP and Medicaid.

Who Stands to Gain from Shorter Benefits?

When unemployment benefits expire quickly, workers face intense pressure to accept any available job, regardless of pay, working conditions, or fit with their skills. So if cutting benefits forces workers to accept jobs that are a worse fit, then who benefits? The main beneficiaries are likely to be employers in low-wage, high-turnover industries, who may find it easier and less costly to fill positions in an environment where workers have less flexibility and less bargaining power due to shortened benefits.

That’s why, in state after state, UI benefits cuts are being pushed by business lobbying groups and CEOs. In many cases, the groups backing these cuts are the same as the ones attempting to strip away workplace protections and roll back long-standing child labor laws.

Here’s a map of when American housing structures were built.

Subscribe now

Pre-1960 housing stock dominates the Northeast and parts of the Midwest, concentrated in what would become known as the Rust Belt. This aligns with these regions' early industrialization and urbanization.

More recent construction (1980-2023) is heavily concentrated in the Sun Belt, particularly in the Southeast. This tracks with significant population shifts toward these regions, driven by factors including lower cost of living and climate preferences.

The West presents a more fragmented pattern, with development largely confined to coastal areas and specific inland population centers, constrained by geography and water access. Still, areas of more recent construction appear in several mountain state metros as well as in the Pacific Northwest and California’s Central Valley.

But a more interesting finding from the survey is that for either party, the safety ratings are almost entirely unrelated to city homicide rates1.

Breaking out the safety ratings by party doesn’t change much. People’s perceptions of which cities are safe simply do not line up with the cities’ murder rates, regardless of political party. The cities considered least safe on the list include Detroit, Chicago, Los Angeles and New York. While Detroit has a high murder rate, Chicago is near the middle of the pack, just above Miami, which is rated as the fourth safest city on the list. Los Angeles and New York are near the bottom of the list in murder rate.

That these latter three cities are the most populous in the country might not be a concidence. If perceptions of crime are affected by consumption of crime-related news media, then it might follow that cities that recieve more news coverage—due to their size—would be percieved as more dangerous.

Note: I’m doing something usually frowned upon by people who publish crime stats: comparing crime rates across cities. There are various ways these comparisons can go awry, related to how data are reported for cities with many law enforcement agencies (the data here only include the largest agency).

Since the beginning of the COVID-19 pandemic, economists have debated the impact of the increase in remote work. Would remote work continue to be popular? And if so, how would it change labor markets or local economies? In May of 2020, 41 percent of employed people were working from home, up from just 5 percent in 20191. But by September 2022, this percent had fallen to just 10 percent, and could still be falling.

Remote work didn’t increase uniformly for all workers during the pandemic. Workers with more years of education teleworked at much higher rates. Rates of working from home peaked at 76 percent for workers with a graduate degree, and 61 percent for those with a bachelor’s degree. For those with a high school education or less, only 16 percent worked from home during the pandemic’s peak. By September, rates of working from home had fallen substantially. Just 18 percent of those with a graduate degree and 16 percent of those with a bachelor’s degree still work from home. This is still a substantial increase from the pre-pandemic work-from-home rate of 7 percent for college graduates. Of workers at or below the high school level, only 5 percent work from home—close to the pre-pandemic rate of 3.5 percent.

Breaking down work-from-home rates by industry gives a fuller picture of how remote work has evolved throughout the pandemic. The rate of working from home was highest in the education industry, at 79 percent in May of 2020. But this high rate has not persisted, as both schools at all levels—K-12 and post-secondary institutions—prioritized a return to in-person learning in 2022. Only 8% of workers in the education industry still work from home in September 2022.

Three industries where working from home reached high levels during the pandemic, and have remained well above pre-pandemic rates, are professional, scientific, and technical services; finance and insurance; and business services. These industries peaked at around 70 percent working from home, and rates remain around 22 to 26 percent in recent months. These were already industries with the highest work-from-home rates in 2019.

But perhaps if working from home is more popular among younger workers, it will catch on over time? Unfortunately, this doesn’t seem to be the case. While telecommuting due to the pandemic was slightly more common among 35- to 44-year-olds, current rates of working from home are level from age 25 to 64, at 11 percent.

I should note that the way these estimates are constructed shouldn’t be expected to produce exact estimates. I am simply adding the estimate of teleworking due to the pandemic from the Current Population Survey (CPS) to the work-from-home estimate from the 2019 American Community Survey (ACS). This might be inaccurate, especially early in the pandemic when there was considerable unemployment. There is also a risk of underestimating working from home if workers no longer attribute their working from home in 2022 to COVID-19. On the other hand, there is a risk of overestimating work-from-home jobs if people attribute them to COVID-19 in cases where they would have existed anyway. According the the 2021 ACS estimates, the overall work-from-home estimate was 18 percent. This is slightly lower than the 2019 ACS estimate plus the CPS estimate of teleworking for 2021 (21 percent). This suggests that this method is reasonably accurate, and might even overestimate working from home.

In response, several pundits and economists claimed that this increase actually just represents disabled people finding employment at higher rates, due to tightening labor markets and new remote work options.

Note that the claim in this tweet is incorrect. The New York Fed’s report doesn’t say that. You might not guess it from his description, but the title of the report is, “Long COVID Appears to Have Led to a Surge of the Disabled in the Workplace”, and its thesis is that COVID-19 was a mass disabling event and accommodations will be necessary to help those who are now working with long-term COVID-19 symptoms.

It seems strange to me that people are arguing that the increase in disabled workers doesn’t reflect an increase in disability, since there has clearly been a large increase in self-reported disability. But few of the experts who ridiculed Ding’s tweet seemed concerned with actually disentangling the two factors. So I’ve just gone ahead and done it, decomposing the change in disabled working-age adults, by employment status, between 2019Q3 and 2022Q3.

To calculate composition effect, I simply apply 2019 employment rates by age, sex and disability status to the 2022 population. This estimates the amount of change that is caused by changing composition of the working-age population.

I find that overall, there was an increase of 1.3 million disabled people in the 18 to 64 age group. This increase contributed 504 thousand people to the increase in disabled workers, and 792 thousand to disabled non-workers. Higher employment rates in 2022 can be credited with shifting 426 thousand disabled non-workers to workers. In other words, increased disability accounts for 53.6 percent of the increase in disabled workers, while higher employment rates account for 45.3 percent. Population growth accounts for the remaining 1.1 percent increase.

It is important to note that if the increase in reported disabilities is composed of less severe disabilities, such that people are more likely to be employed, my estimate will understate its contribution to the increase in disabled workers. This seems probable.

A different analysis of this question by researchers at the Brookings Institution attributes the entire increase in labor force participation among disabled people to increased rates of disability, and none to other factors. It looks to me that the main difference between my analysis and theirs is that they use the 2017-2019 trend in labor force participation as a counterfactual, rather than just the 2019 level. That might be reasonable, but I'm not sure that I could justify it.

The strange thing about this whole business is that it is obviously true that self-reported disability has increased by a lot among working-age Americans. That seems to be Ding's main point, but the chart he chose can't quite prove it because it only shows workers. Whether disabled workers are now employed at higher rates doesn't affect the overall increase in disability, it just changes how they break down by employment status. While I've been clear that the increase in self-reported disability is tricky to interpret, I find it odd how many economists seem eager to ignore this trend altogether. It's great that remote work and tight labor markets might have benefits for disabled people. But pointing that out doesn't require waving away another trend.

Many people will point out that 42 is actually a small number compared to 128,961, the total number of K-12 schools in the U.S. In one case, a “risk communication consultant” noted that the daily risk of a child dying in a school shooting is just 1 in 614,000,000, despite the fact that pretty much no one uses daily risk of death for any other causes. These attempts to downplay the impact of shootings fail to address the basic concern of most parents, which is simply, “what is the chance that my child will experience a school shooting at some point in their schooling?”

The first step to answering this question is recognizing that school shootings affect far more people than those who are killed. It goes without saying that the lives lost in these events are the largest of the costs they impose. But there are also broader impacts to consider. Our objectives for school safety can and should go beyond our children’s physical survival.

Many students have lost classmates, experienced trauma or lost a sense of control and security. Studies on survivors of school shootings find that they have increased dropout rates and absenteeism1, worsened academic performance2 and increased risky behavior3. Longer-term studies find worsened education and employment outcomes4, worsened health and well-being5 and higher risk of dying by accident or by suicide6. Some effects even extend to the children of those affected7. Many of these effects are not small, and they are not all restricted to the “mass shootings” that get the most media attention.

According to data collected by the Washington Post, of the 42 school shootings in 2021, the median number of students enrolled during a school shooting was 882. The total number of K-12 students who experienced a shooting in their school in 2021 was nearly 44,000.

The second step to quantifying the risk is accounting for the fact that students spend around 13 years in K-12 education. Extrapolating the data to an entire K-12 experience suggests that, if 2021 rates persist, 1 in 97 American children will experience a school shooting at some point in their schooling8. That’s probably a higher number than most people expect. In other developed countries, the percent of children expected to experience combat situations in order to attain a high school education is zero.

As with many things in American life, the risk of experiencing a school shooting varies by race and ethnicity. Among white students, the risk is 1 in 190. Among black students, it is 5 times as high, at 1 in 38. For Hispanics, it is 1 in 90. And for Asians it is 1 in 290.

The mass shootings that generate national attention are not as common as other shootings. Mass shootings are expected to impact 1 in 1200 students at some point in their schooling, and are more commonly experienced by white students (1 in 740). Targeted shootings, on the other hand, will impact 1 in 180 students, with a disproportionate impact on black students (1 in 65).

The number of Americans age 18 to 54 reporting serious difficulty remembering or concentrating has been increasing throughout the COVID-19 pandemic. Prior to the pandemic, about 2.7% of this age group reported this disability. In September of 2022, that percentage has reached 3.4%.

Over 80% of Americans are estimated to have had COVID-19 at least once. The CDC recognizes “brain fog”, difficulty thinking or concentrating, as one of the potential long-term effects of COVID-19. So it seems plausible that this increase in self-reported cognitive impairments is caused by COVID-19 infections.

Other types of disability have not increased as much. Disabilities related to hearing, seeing or doing errands are slightly higher than immediately before the pandemic, but within the range of past estimates. There has been no increase in reports of disabilities related to mobility or personal self-care.

One curious feature of the increase in reported cognitive disability is that it only shows up among younger adults. Among people age 55 and over, there has been no increase1.

The increase in these symptoms is concerning. But there are questions about how to interpret self-reported cognitive impairment. There is often a disconnect between self-reported, perceived cognitive impairment and objective measures of cognitive ability. Many studies2 find a relationship between perceived cognitive impairment and symptoms of depression, but a weak relationship with objective cognitive function. But this also means that it is also possible that cognitive impairment is under-reported, since many people who undergo cognitive decline might not recognize it enough to report it on these surveys. COVID-19 has already been found to lead to increased cognitive deficits and dementia, but unlike the survey results reported here, the increase occurs mostly among older adults.

But all this speculation raises a question: What is it that makes a Wordle answer hard to guess? Uncommon words? Words with uncommon letters? Or perhaps other factors related to the how the game is played?

A Twitter bot created by Kevin O'Connor offers some clues. The bot has collected and tabulated data on Wordle scores based on Twitter posts since early January.

Because very few people fail at a given Wordle (typically around 3%), I use the percent of people who use all six guesses, whether they win (6/6) or fail (X/6) as a proxy for a word's difficulty. This percentage varies quite a bit, suggesting that some answers are definitely harder than others. The hardest answers so far include SWILL (40% using six guesses), PROXY (32%), SHAKE (30%) and ELDER (29%), while the easiest include THORN (4%), POINT (5%), PANIC (6%) and DRINK (6%).

Uncommon words

Some have suggested that uncommon words are harder to guess. But the correlation between difficulty and how common the words are only accounts for about 10% of the variance in difficulty.

Other factors

Many attempts to identify the best first guess focus on letter frequencies. If people prioritize more common letters in their guesses it might make words with uncommon letters harder to guess. I calculate Scrabble scores as a summary measure of letter frequencies in each word.

Another possibility is that words with a letter appearing more than once are tricky for guessers since they require foregoing the opportunity to learn about a new letter in order to win.

Finally, another challenge players can face is knowing four letters, but having many options for the remaining letter. This can be the case with a pattern like S?ILL, which represents many potential answers: SHILL, SKILL, SPILL, STILL, SWILL.

I estimate a linear model of these factors on the percent of people who use all six guesses, and they all appear to matter somewhat, altogether accounting for 49% of the variance. A word having two of the same letter results in a 10 percentage point increase in the percent of people using all six guesses. A word having three other common words spelled differently by only one letter implies a 7 percentage point increase. A seven-point increase in Scrabble score (the difference between an X and an E tile) increases the percent of six-guess players by 6 points.

The obscurity of the word appears less important. A word that is 10 times less common than another word (a Zipf score decrease of 1) is only associated with a 1 percentage point increase in people using six guesses.

Biased reporting

One potential problem with the data being used here is that it is self-reported. People are probably more likely to tweet their scores when they do well, and less likely to do so when they do poorly. That means there is probably more variation in people scores than is apparent in the Twitter data.

I test this by decomposing the trend in the total number of players into its seasonal component (day of week effects), its overall trend and the remaining variation. The remaining variation appears to be negatively correlated with the difficulty of the word, suggesting that people do in fact choose to report more often when the words are easier to guess.

An even stronger relationship is apparent between reporting and doing well (winning in 3 guesses or fewer).

This probably shouldn't impact the overall conclusions of what makes words difficult to guess. If anything, it makes it slightly harder to explain why some words are harder than others, since it reduces the variation between easy and hard words in the reported data.

Data Availability

Data used for this post is available for download here.

Indiana public schools that started the school year with a face mask mandate are reporting fewer COVID-19 cases than those that did not enact mask mandates. While reported cases spiked in both groups in late August, the gap between mandate and non-mandate schools remained. A third group of school districts began the school year without a mask mandate but enacted one in the first few weeks of school. The overall infection rate in these school districts was higher at the beginning of the school year, but has since fallen below the rate at districts without a mandate.

School mask mandates and community spread

Not only did school districts with mask mandates have lower COVID-19 infection rates in schools, they also have lower rates of community spread than areas without school mask mandates. An analysis by ZIP code shows a similar pattern to the school district analysis, with cases increasing in both groups, but staying lower in areas with mask mandates.

This ZIP code-level analysis also allows inclusion of other data, such as vaccination rates. Below, the areas are categorized into three roughly equal groups by the percent of population fully vaccinated in early September:

• Low vax (40% or less)

• Mid vax (40% to 50%)

• High vax (over 50%)

The story doesn't change. In each vaccination category, areas with mask mandates still have lower infection rates than those without mandates.

What drives mask mandate adoption?

School policies have changed rapidly in the first few weeks of school. Just 29 percent students in Indiana public school corporations began under a mask mandate. But as of September 18th, 65 percent was under a mask mandate. Heated debates about mask mandate policies broke out across the state at school board meetings in August.

Politics appears to be a major factor in whether a school started with or eventually added a mask mandate. In school districts that voted for Hillary Clinton in 2016 (detailed 2020 data is not yet available), 83 percent of students started with a mandate, rising to 98 percent weeks later. The percentage in districts where Donald Trump won at least 70 percent of the vote started at 2 percent and has risen to just 29 percent. The prevalence of mask mandates where Trump received between 50 percent and 70 percent of the vote saw a dramatic increase from 22 percent early in the school year, to 77 percent by September 18th.

This association between politics and public health measures is unfortunate, as many people turn to disinformation instead of expertise. But it also makes it difficult to measure the impact of policies like school mask mandates. Generally, it suggests that school districts without mask mandates might differ from those with mandates in many ways that would also influence the spread of COVID-19. Of districts that started without mask mandates, 81 percent are rural, while only 19 percent of districts that opened with mandates are rural. Rates of educational attainment also differ between the two groups: the rate of college graduates in mandate districts is 31%, compared to 20 percent in districts with no mandate. This difference in educational attainment could reflect further differences in social structure and occupational exposure to infection.

In the chart below, it is apparent that school districts with mask mandates tend to have lower COVID-19 infection rates, but it is not clear that there is a difference between mandate and non-mandate districts that are similar politically. School districts that changed mask mandate policies after school started are excluded, since policy changes during the school year are more likely to be reacting to infection rates.

In both analyses of school cases and community spread, the main barrier to making conclusions about the effectiveness of mask mandates is that areas with mask mandates are different in many ways from areas without them. Other recent studies of school mask mandates (one of public schools in two Arizona counties, and another nationwide county-level study) come to similar conclusions, but suffer from similar limitations. The best-quality evidence for the effectiveness of mask-wearing on COVID-19 spread comes from a randomized trial in Bangladesh, which showed that an increase in mask-wearing from 13 to 42 percent caused a 12 percent decrease in the prevalence of COVID-19 symptoms, and a 9 percent decrease in COVID-19 symptomatic seroprevalence.

Data Availability

Data used for this post is available for download here.

I surveyed users of Reddit.com who live in Indiana, and the results suggest that most Indy residents have not seen a single cicada near their home this year. Most of these people expected to see cicadas, and many of them describe their reaction to this year's periodical brood as "disappointed." So what happened? Are the cicadas dying out, or is their geographic range changing? Did someone invent the idea of cicadas swarming Indianapolis, somehow convincing Indy residents that this was actually a thing that happens every 17 years?

Survey results

There were 93 responses to the survey with usable location data within the state of Indiana. Of 13 Hoosiers located north of the Indianapolis area (Marion County plus neighboring counties), none reported seeing a single cicada. Of the 14 to the south of the city, 71 percent reported seeing cicadas. In the Indy area, just 11 of 66 respondents (17 percent) saw a single cicada this year.

A closer look at the geography of reported cicada sightings shows that most of the sightings in the Indy area were in the northeast part of the city near the Geist reservior and Fort Harrison State Park, as well as a couple in the suburbs of Carmel and Fishers. While this probably confirms the presence of some Brood X cicadas in Indianapolis, it doesn't quite live up to the predictions of a cicada invasion of the sort many experienced in Washington D.C.

This is confirmed in responses to the survey that indicate that 86 percent of Indy residents (and 100 percent of Hoosiers north of Indy) saw fewer cicadas than they expected to see. In constrast, only half of Hoosiers living south of Indy saw fewer cicadas than expected, and over a third actually saw more than they expected to see.

The two dominant sentiments expressed by respondents in Indianapolis and to its north were disappointment and relief, suggesting that many had been excited about the predicted cicada swarm, while others had been less enthusiastic. South of Indy, the most common reactions were positive, with a minority expressing disgust or disappointment.

What influenced people's expectations of Brood X in 2021? Since the brood emerges every 17 years, one might expect that in areas where Brood X was prevalent, long-time residents would based their expectations on memories of the previous invasions. Only 42 percent of Indy residents indicate that their own experiences influenced their expectations for the 2021 emergence, compared to 79 percent of residents of southern Indiana. Basing expectations on news media and other sources was much more common in Indianapolis (74 percent) and northern Indiana (77 percent) than in sourthern Indiana (50 percent). This suggests that where personal and community experiences with Brood X cicadas were less common, the news media might have filled in the gap, raising expectations where cicadas were not actually common in recent history.

News coverage of Brood X

Various news outlets reported on Brood X cicadas this spring, and many published maps of the brood's range. Several of these outlets used maps based on county-level data from the U.S. Forest Service, including the Washington Post, New York Times, Indy Star, CBS News, PBS, Vox and USA Today. On the surface, there's not much reason to doubt this data. It comes from an official source and it seems to be based on previous research. But the data it uses is actually somewhat old, and not very reliable for predicting cicada geography today. It is based largely on the work of late 19th century entomologists, C.V. Riley and Charles Marlatt, recoded to match today's counties.

Newer research from University of Connecticut entomologist John R. Cooley, using modern technology for field mapping and crowd-sourcing, demonstrates that Marlatt's methods were not very precise and often led to overestimating the range of periodical cicadas. Combined with the fact that cicadas might have actually been more widely distributed a century ago than they are today, the data from the U.S. Forest Service does seem to overestimate today's Brood X range. Large portions of Indiana, Ohio and Pennsylvania are included in the U.S. Forest Service map, but do not appear in Cooley's data.

There do appear to be a couple reports of cicadas in northeast Marion County in 2004—consistent with the survey data showing a similar cluster in 2021—but the data generally does not indicate many cicadas in central Indiana or to its north.

Cooley and other cicada researchers are currently mapping the 2021 Brood X emergence, and crowd-sourced reports are again part of the strategy. You can download the Cicada Safari app to submit pictures or mating calls to have your reports verified and added to the Cicada Safari map. Perhaps with better data, we'll all have more realistic expectations for the next emergence in 2038.

Although Republican presidential candidates have benefited from the electoral college in recent years—2 of their last 3 election winners lost the popular vote—there is nothing about the electoral college that specifically favors Republicans. Its effects are largely random, and can be expected to change over time. One illustration of how arbitrary these effects are is that a state's status as a swing state can often be eliminated by moving a few counties into a bordering state, instantly devaluing the value of its residents' votes. It would only take a couple of these changes to shift the advantage of the electoral college to the Democratic party.

For 2016, researchers Pierre-Antoine Kremp and Andrew Gelman used polling data to estimate the probability of casting the decisive vote for voters in every state. This measure of how your vote matters, plotted below, highlights the main feature of the electoral college: Some people's votes matter and some simply don't. Note that the scale used below is logarithmic: Doubling the space between two states actually represents a ten-fold increase in the difference in voter power.

In terms of vote power, the median voter lived in Illinois or Texas in 2016. Voters in New Hampshire and Colorado were 1000x more likely to affect the election results than the median voter. Voters in Nevada, Wisconsin, and Pennsylvania had 500x as much power as the median voter. Voters with the least power lived in the solidly Republican states of Wyoming and Oklahoma, with 1/30th the power of the median voter, and 1/30,000th the power of voters in New Hampshire and Colorado.

Arguments about the merits of the electoral college often turn to how it affects certain categories of voters. Proponents of the electoral college sometimes argue that voters in cities should have less power than voters in rural areas. Opponents of the system argue that it privileges white voters over minorities. But are these thing true?

If we consider voters who live in states with 100x the power of the median voter to be "high-power voters,"" we can simply calculate the percentage of high-power voters in each demographic group across the country. Overall, these people make up 27.7% of the citizen voting-age population.

The incidence of high-power voters is highest (29.6%) among white non-Hispanics, and slightly less among black non-Hispanics (27.1%). Those disadvantaged by the electoral college are Hispanics (21.8%) and those who fall into other racial categories (19.0%). The disadvantage to Hispanic voters is likely due to the high percentage of Hispanics in non-swing states of Texas and California. This could change if Texas were to become a swing state, as some predict.

Breaking out the civilian voting-age population by census tract density tests the hypothesis that the electoral college reduces the voting power of urban voters. Here I am using the density cutoffs used by Jed Kolko here, as well as an additional cutoff for suburban tracts used by David Montgomery's Congressional Density Index. This reveals that only 21.9% of voters in urban areas are high-powered voters. So it is true—at least for the moment—that the electoral college puts urban voters at a disadvantage.

As noted earlier, a lot of this is very tentative, and is based on which states happened to be split most evenly in 2016. Which states will have the most powerful voters is subject to change, and there's not really anything meaningful driving those changes. It's mostly just a matter of the combinations of voters that happen to fall within state borders on election day.

Note: Source code for calculations and tabulations can be found here.