Data

Median Family Income in US States, 2022

Jeremy HorpedahlLeave a comment

Last week I wrote about median income in the US, and how it had declined since 2019 and 2021 through 2022 (inflation adjusted, of course). The big story is that median income (both for households and families) has been falling in recent years. While there are some silver linings when looking at subgroups, such as Black families, the overall data isn’t good.

But while that is true for the US overall, it’s not true for every state. In fact, it’s not even true for most states! From 2019 to 2022, there were 29 states that saw their median family incomes rise! That’s adjusted for inflation (I’m using the C-CPI-U, which is Census’s preferred inflation measure for this data). The income data in this post all comes from the Census ACS 1-year estimates.

Here’s a map showing the states that had increases in median family income (green) and those that had decreases (in red). (This is my first time experimenting with Datawrapper maps, feedback appreciated!)

Some states had pretty robust growth, with New Mexico and Arizona leading the way with around 5 percent growth. There is substantial variation across US states, including with big declines like Wyoming at -5 percent, and Oklahoma and Illinois are -3 percent.

A few weeks ago I also wrote about the richest and poorest MSAs in the US. But what about the richest and poorest states in the US? The following map shows that data.

The immediate fact which will jump out at you is that the lowest income US states are almost all located in the South. This will probably not surprise most of us, although it probably is a bit surprising since the data is adjusted for differences in the cost of living (using the BEA RPP data). Even after making these adjustments, the South is still clearly the poorest region (and it definitely was the poorest without the adjustments).

Among the higher income states, they are distributed pretty well across the rest of the non-South. There are 16 states (plus DC) that have median family incomes over $100,000 (again, cost of living adjusted), and while many of these are in New England and the Mid-Atlantic, there area still a few in the Midwest, Great Plains, and the West. Utah and New Jersey have similar incomes, as do Virginia and Rhode Island.

The highest income states are Massachusetts and Connecticut, with over $112,000 in median family income, while the lowest are Mississippi and West Virginia, both under $78,000. Median family income in Massachusetts is 46 percent higher than Mississippi. And that’s after adjusting for differences in the cost of living.

Median Income Is Down Again. Are There Any Silver Linings in the Data?

Jeremy Horpedahl1 Comment

This week the Census Bureau released their annual update on “Income, Poverty and Health Insurance Coverage in the United States.” This release is always exciting for researchers, because it involves as massive release of data based on a fairly large (75,000 household) sample with detailed questions about income and related matters. For non-specialists, it also generates some of the most commonly used national data on income and poverty. Have you heard of the poverty rate? It’s from this data. How about median household income? Also from this data.

I’ll focus on income data in this post, though there is a lot you could say about poverty and health insurance too. The headline result on median income is, once again, a dismal one. Whether you look at median household income (very commonly reported, even though I don’t like this measure) or median family income (which I prefer), both are down from 2021 to 2022 when adjusted for inflation. Both are still down noticeably from the pre-pandemic high in 2019 (though both are also above 2018 — we aren’t quite back to the Great Depression or Dark Ages, folks!).

These headline results are bad. There is no way to sugarcoat or “on the other hand” those results. And these results are probably more robust and representative than other measures of average or median earnings, since they aren’t subject to “composition effects” — when those with zero wages in one period don’t show up in the data. I will note that these results are for 2022, and we are highly likely to see a turnaround when we get the 2023 data in about a year (inflation has slowed to less than wage growth in 2023).

But given that obviously bad headline result, was there any good data? As I mentioned above, a ton of data, sliced many different ways, is released with this report. Some of it also gives us consistent data back decades, in some cases to the 1940s. What else can we learn from this data release?

Median Income by Race

When we look at median income by race, there are a few silver linings. The headline data from Census tells us that only the drop in household income for White, Non-Hispanics was statistically significant. For other races and ethnicities, the changes were not statistically significant from 2021 to 2022 — and some of those changes were actually positive. We shouldn’t dismiss White, Non-Hispanics — they are the largest racial/ethnic group! — but it is useful to look at others.

Black household and families are the most interesting to look at in more detail, especially because they are the poorest large racial group in the US. Black household and family income increased from 2021 to 2022, although the increase was small enough that we can’t say it is statistically significant (remember, this is a sample, not the universe of the decennial Census).

But what’s more important is that median Black household income is now at the highest level it has ever been (adjusted for inflation, as always). Median Black household income is about $1,000, or around 2 percent higher than in 2019 — the peak date for overall median income. Two percent growth over 3 years is nothing to shout from the rooftops, but it is very different from White, Non-Hispanic households, which are down over 6 percent since 2019.

Median black family income is roughly flat since 2019, but it is up about 1.5 percent in the past year — not quite as robust, but still better than the overall numbers.

Historical Income Data

The other silver lining I always like to mention is the long-run historical data. This data often gets overlooked in the obsessive focus on the most recent changes, so it’s useful to sit back and look at how far we have come. Let’s start where we just left off, with Black families. I wrote a post back in February about Black family income, which had data current through 2021, but it’s useful once again to look at the data with another year (plus they have updated the inflation adjustments for 2000 onward).

The chart shows the percent of Black families that are in three income groups, using total money income data. The data is adjusted for inflation. The progress is dramatic. In 1967, the first year available, half of Black families had incomes under $35,000. By 2022 that number had been cut in half to just one quarter of families (the 2022 number is the lowest on record, even beating 2019). Twenty-five percent is still very high, especially when compared to White, Non-Hispanics (it’s about 12 percent), but it’s still massive progress. It’s even a 10-percentage point drop from just 10 years ago. And Black families haven’t just moved up a little bit: the “middle class” group (between $35,000 and $100,000) has been pretty stable in the mid-40 percentages, while the number of rich (over $100,000) Black families has grown dramatically, from just 5 percent to over 30 percent.

We saw earlier that progress for White, Non-Hispanics has stumbled in the past 3 years, but the long run data is much more optimistic (this data starts in 1972).

The progress here should be evident too, but let me highlight one thing for emphasis: as far back as 1999, the largest of these three groups was the “rich” (over $100,000 group). And since 2017, the upper income group has been the majority, with median White Non-Hispanic family income surpassing $100,000 in 2017, up from $70,000 at the beginning of the series in the early 1970s (all inflation adjusted, of course).

The next question I often get with this historical data is: How much of this increase is due to the rise of two-income households. Well, this same data release allows us to look at that data too! This final chart shows median family income for families with either one or two earners (there are families with zero earners or more than two, but these two categories make up the bulk of families). This data is pretty cool because it goes all the way back to 1947.

This chart doesn’t look so good for one-earner families. After growing along with two-earner families in the 1950s and 1960s, it basically stagnates from the early 1970s until the late 2010s. Then you get a little growth. Not good!

I think more investigation is needed here, but the share of families that have two earners has grown dramatically, from 26 percent of families in 1947 to 42 percent in 2022. Single earner families shrunk from 59 percent to 31 percent, and dual-income families have been the most common family type since the late 1960s. There are some important compositional differences here in what types of families only have one earner. If we imagine some alternate history where, by law, only one spouse was allowed to work, certainly the single earner line would have risen more. And many of the single earner families today are single mothers, who for a variety of reasons have much lower earning potential than the fathers heading married couples in the 1950s and 1960s. So the numbers aren’t perfectly comparable.

Still, even for single earner families, real median income has more than doubled since 1947 — though most of that growth had happened by the early 1970s.

As we make our way through a challenging economic time following the pandemic and 2 years of unusually high inflation, hopefully we can look forward to a future of resuming the upward trajectory of incomes for all kinds of families.

Generative AI Nano-Tutorial

Scott Buchanan4 Comments

Everyone who has not been living under a rock this year has heard the buzz around ChatGPT and generative AI. However, not everyone may have clear definitions in mind, or understanding of how this stuff works.

Artificial intelligence (AI) has been around in one form or another for decades. Computers have long been used to analyze information and come up with actionable answers. Classically, computer output has been in the form of numbers or graphical representation of numbers. Or perhaps in the form of chess moves, beating all human opponents since about 2000.

Generative AI is able to “generate” a variety of novel content, such as images, video, music, speech, text, software code and product designs, with quality which is difficult to distinguish from human-produced content. This mimicry of human content creation is enabled by having the AI programs analyze reams and reams of existing content (“training data”), using enormous computing power.

I wanted to excerpt here a fine article I just saw which is informative on this subject. Among other things, it lists some examples of gen-AI products, and describes the “transformer” model that underpins many of these products. I skipped the section of the article that discusses the potential dangers of gen-AI (e.g., problems with false “hallucinations”), since that topic has been treated already in this blog.

Between this article and the Wikipedia article on Generative artificial intelligence , you should be able to hold your own, or at least ask intelligent questions, when the subject next comes up in your professional life (which it likely will, sooner or later).

One technical point for data nerds is the distinction between “generative” and “discriminative” approaches in modeling. This is not treated in the article below, but see here.

All text below the line of asterisks is from Generative AI Defined: How it Works, Benefits and Dangers, by Owen Hughes, Aug 7, 2023.

*******************************************************

What is generative AI in simple terms?

Generative AI is a type of artificial intelligence technology that broadly describes machine learning systems capable of generating text, images, code or other types of content, often in response to a prompt entered by a user.

Generative AI models are increasingly being incorporated into online tools and chatbots that allow users to type questions or instructions into an input field, upon which the AI model will generate a human-like response.

How does generative AI work?

Generative AI models use a complex computing process known as deep learning to analyze common patterns and arrangements in large sets of data and then use this information to create new, convincing outputs. The models do this by incorporating machine learning techniques known as neural networks, which are loosely inspired by the way the human brain processes and interprets information and then learns from it over time.

To give an example, by feeding a generative AI model vast amounts of fiction writing, over time the model would be capable of identifying and reproducing the elements of a story, such as plot structure, characters, themes, narrative devices and so on.

……

Examples of generative AI

…There are a variety of generative AI tools out there, though text and image generation models are arguably the most well-known. Generative AI models typically rely on a user feeding it a prompt that guides it towards producing a desired output, be it text, an image, a video or a piece of music, though this isn’t always the case.

Examples of generative AI models include:

  • ChatGPT: An AI language model developed by OpenAI that can answer questions and generate human-like responses from text prompts.
  • DALL-E 2: Another AI model by OpenAI that can create images and artwork from text prompts.
  • Google Bard: Google’s generative AI chatbot and rival to ChatGPT. It’s trained on the PaLM large language model and can answer questions and generate text from prompts.
  • Midjourney: Developed by San Francisco-based research lab Midjourney Inc., this gen AI model interprets text prompts to produce images and artwork, similar to DALL-E 2.
  • GitHub Copilot: An AI-powered coding tool that suggests code completions within the Visual Studio, Neovim and JetBrains development environments.
  • Llama 2: Meta’s open-source large language model can be used to create conversational AI models for chatbots and virtual assistants, similar to GPT-4.
  • xAI: After funding OpenAI, Elon Musk left the project in July 2023 and announced this new generative AI venture. Little is currently known about it.

Types of generative AI models

There are various types of generative AI models, each designed for specific challenges and tasks. These can broadly be categorized into the following types.

Transformer-based models

Transformer-based models are trained on large sets of data to understand the relationships between sequential information, such as words and sentences. Underpinned by deep learning, these AI models tend to be adept at NLP [natural language processing] and understanding the structure and context of language, making them well suited for text-generation tasks. ChatGPT-3 and Google Bard are examples of transformer-based generative AI models.

Generative adversarial networks

GANs are made up of two neural networks known as a generator and a discriminator, which essentially work against each other to create authentic-looking data. As the name implies, the generator’s role is to generate convincing output such as an image based on a prompt, while the discriminator works to evaluate the authenticity of said image. Over time, each component gets better at their respective roles, resulting in more convincing outputs. Both DALL-E and Midjourney are examples of GAN-based generative AI models…

Multimodal models

Multimodal models can understand and process multiple types of data simultaneously, such as text, images and audio, allowing them to create more sophisticated outputs. An example might be an AI model capable of generating an image based on a text prompt, as well as a text description of an image prompt. DALL-E 2 and OpenAI’s GPT-4 are examples of multimodal models.

What is ChatGPT?

ChatGPT is an AI chatbot developed by OpenAI. It’s a large language model that uses transformer architecture — specifically, the “generative pretrained transformer”, hence GPT — to understand and generate human-like text.

What is Google Bard?

Google Bard is another example of an LLM based on transformer architecture. Similar to ChatGPT, Bard is a generative AI chatbot that generates responses to user prompts.

Google launched Bard in the U.S. in March 2023 in response to OpenAI’s ChatGPT and Microsoft’s Copilot AI tool. In July 2023, Google Bard was launched in Europe and Brazil.

…….

Benefits of generative AI

For businesses, efficiency is arguably the most compelling benefit of generative AI because it can enable enterprises to automate specific tasks and focus their time, energy and resources on more important strategic objectives. This can result in lower labor costs, greater operational efficiency and new insights into how well certain business processes are — or are not — performing.

For professionals and content creators, generative AI tools can help with idea creation, content planning and scheduling, search engine optimization, marketing, audience engagement, research and editing and potentially more. Again, the key proposed advantage is efficiency because generative AI tools can help users reduce the time they spend on certain tasks so they can invest their energy elsewhere. That said, manual oversight and scrutiny of generative AI models remains highly important.

Use cases of generative AI

Generative AI has found a foothold in a number of industry sectors and is rapidly expanding throughout commercial and consumer markets. McKinsey estimates that, by 2030, activities that currently account for around 30% of U.S. work hours could be automated, prompted by the acceleration of generative AI.

In customer support, AI-driven chatbots and virtual assistants help businesses reduce response times and quickly deal with common customer queries, reducing the burden on staff. In software development, generative AI tools help developers code more cleanly and efficiently by reviewing code, highlighting bugs and suggesting potential fixes before they become bigger issues. Meanwhile, writers can use generative AI tools to plan, draft and review essays, articles and other written work — though often with mixed results.

The use of generative AI varies from industry to industry and is more established in some than in others. Current and proposed use cases include the following:

  • Healthcare: Generative AI is being explored as a tool for accelerating drug discovery, while tools such as AWS HealthScribe allow clinicians to transcribe patient consultations and upload important information into their electronic health record.
  • Digital marketing: Advertisers, salespeople and commerce teams can use generative AI to craft personalized campaigns and adapt content to consumers’ preferences, especially when combined with customer relationship management data.
  • Education: Some educational tools are beginning to incorporate generative AI to develop customized learning materials that cater to students’ individual learning styles.
  • Finance: Generative AI is one of the many tools within complex financial systems to analyze market patterns and anticipate stock market trends, and it’s used alongside other forecasting methods to assist financial analysts.
  • Environment: In environmental science, researchers use generative AI models to predict weather patterns and simulate the effects of climate change

….

Generative AI vs. machine learning

As described earlier, generative AI is a subfield of artificial intelligence. Generative AI models use machine learning techniques to process and generate data. Broadly, AI refers to the concept of computers capable of performing tasks that would otherwise require human intelligence, such as decision making and NLP.

Machine learning is the foundational component of AI and refers to the application of computer algorithms to data for the purposes of teaching a computer to perform a specific task. Machine learning is the process that enables AI systems to make informed decisions or predictions based on the patterns they have learned.

( Again, to make sure credit goes where it is due, the text below the line of asterisks above was excerpted from Generative AI Defined: How it Works, Benefits and Dangers, by Owen Hughes).

Cool the Schools

James BaileyLeave a comment

Short post today because I’m busy watching my kids, who had their school canceled because of excessive heat, like many schools in Rhode Island today.

I thought this was a ridiculous decision until my son told me he heard from his teacher that his elementary school is the only one in town that has air conditioning for every classroom. Given that, the decision to cancel given the circumstances is at least reasonable, but the lack of AC is not.

It’s not just that hot classrooms are unpleasant for students and staff, or that sudden cancellations like this are a major burden for parents. Several economics papers have found that air conditioning significantly improves students’ learning as measured by test scores (though some find not). Park et al. (2020 AEJ: EP) find that:

Student fixed effects models using 10 million students who retook the PSATs show that hotter school days in the years before the test was taken reduce scores, with extreme heat being particularly damaging. Weekend and summer temperatures have little impact, suggesting heat directly disrupts learning time. New nationwide, school-level measures of air conditioning penetration suggest patterns consistent with such infrastructure largely offsetting heat’s effects. Without air conditioning, a 1°F hotter school year reduces that year’s learning by 1 percent.

This can actually be a bigger issue in somewhat Northern places like Rhode Island- we’re South enough to get some quite hot days, but North enough that AC is not ubiquitous. Data from the Park paper shows that New York and New England are actually some of the worst places for hot schools:

This is because of the lack of AC in the North:

The days are only getting hotter…. it’s time to cool the schools.

The Dodge Caravan, Quality Improvements, and Affordability

Jeremy HorpedahlLeave a comment

1996 was a big year for minivans. While modern minivans had been around for about a decade by that point, 1996 marked a turning point. That year Dodge introduced what is referred to as the “third generation” of its Caravan, and it won Motor Trend’s car of the year award. That’s the first, and only time, that a minivan ever won this award. If you drive a minivan today or see one on the road, you are seeing the look, style, and features that were first introduced in 1996 (interestingly, that year also seems to have marked the peak in sales for the Chrysler family of minivans).

If you wanted to buy the cheapest possible Dodge Caravan in 1996, you would have paid about $18,500. You could always pay more for more features, as with any car, but if you wanted this “car of the year,” and you wanted it new and cheap, that was what you paid.

Dodge continued to produce the Caravan for the US market until 2020, when it was discontinued in favor of other nameplates (though it still lived on in Canada). In 2020, the base model Caravan was about $29,000 (and now only available in the “Grand” version, an upgrade in 1996).

Oren Cass has used the prices of these two minivans to make a point about price indexes, quality adjustments, and affordability. If you look at the raw prices, clearly it is more expensive. But the consumer price index tells us that the price of new cars was flat between 1996 and 2020.

So what gives?

Continue reading →

Interpolation Vs Transition

Zachary BartschLeave a comment

Sometimes you read an academic article and the author fills in the data gaps with interpolation. That is, they assume some functional form of the data and then replace the missing values with the estimated ones. Often, lacking an informed opinion about functional form, authors will just linearly interpolate between the closest known values. Sometimes this method is OK. But sometimes we can do better.

Historical census data provides a good example because the frequency was only every ten years. Say that we want to know more about child migration patterns between 1850 and 1860. What happened in the intervening years? Who knows. Let’s look at the data.

Using data on individuals who have been linked across censuses allows us to fill in the gaps a little bit. For simplicity, let’s just look at whether a child migrant lived in an urban location and whether they lived on a farm. That means that there are 4 possible ways to describe their residence. Below is a summary of where children migrants lived at the age of zero in 1850 and where the same children lived a decade later at the age of ten in 1860 given that they moved counties.

When I’m the mean time did these children move from one place and to the other? We don’t know exactly. The popular answer is to say that they moved uniformly throughout the decade. That’s ‘fine’. But it assumes that the rate at which people departed places was rising and the rate at which they arrived places was falling. Maybe that’s true, but we don’t really know. Below-left is a graph that shows the linear interpolation.

The nice thing about linear interpolation is that everyone is accounted for at each point in time. The total number of people don’t rise or fall in the intervening interpolation period. But if we were to assume that children departed/arrived at each type of place at a constant rate (maybe a more reasonable assumption), then suddenly we lose track of people. That is, the sum of people dips below 100% as people depart faster than they arrive.

What’s the alternative to linear interpolation?

Continue reading →

Long Covid is Real in the Claims Data… But so is “Early Covid”?

James Bailey4 Comments

I’ve seen plenty of investigations of “Long Covid” based on surveys (ask people about their symptoms) or labs (x-ray the lungs, test the blood). But I just ran across a paper that uses insurance claims data instead, to test what happens to people’s use of medical care and their health spending in the months following a Covid diagnosis. The authors create some nice graphics showing that Long Covid is real and significant, in the sense that on average people use more health care for at least 6 months post-Covid compared to their pre-Covid baseline:

Source: Figure 5 of “Long-haul COVID: healthcare utilization and medical expenditures 6 months post-diagnosis“, BMC Health Services Research 2022, by Antonios M. Koumpias, David Schwartzman & Owen Fleming

The graph is a bit odd in that its scales health spending relative to the month after people are diagnosed with Covid. Their spending that month is obviously high, so every other month winds up being negative, meaning just that they spent less than the month they had Covid. But the key is, how much less? At baseline 6 months prior it was over $1000/month less. The second month after the Covid diagnosis it was about $800 less- a big drop from the Covid month but still spending $200+/month more than baseline. Each month afterwards the “recovery” continues but even by month 6 its not quite back to baseline. I’m not posting it because it looks the same, but Figure 4 of the paper shows the same pattern for usage of health care services. By these measures, Long Covid is both statistically and economically significant and it can last at least 6 months, though worried people should know that it tends to get better each month.

I was somewhat surprised at the size of this “post Covid” effect, but much more surprised at the size of the “pre Covid” or “early Covid” effect- the run-up in spending in the months before a Covid diagnosis. For the month immediately before, the authors have a good explanation, the same one I had thought of- people are often sick with Covid a couple days before they get tested and diagnosed:

There is a lead-up of healthcare utilization to the diagnosis date as illustrated by the relatively high utilization levels 30–1 days before diagnosis. This may be attributed to healthcare visits only days prior to the lab-confirmed infection to assess symptoms before the manifestation or clinical detection of COVID-19.

But what about the second month prior to diagnosis? People are spending almost $150/month more than at the 6-month-prior baseline and it is clearly statistically significant (confidence intervals of months t-6 and t-2 don’t overlap). The authors appear not to discuss this at all in the paper, but to me ignoring this lead-up is burying the lede. What is going on here that looks like “Early Covid”?

My guess is that people were getting sick with other conditions, and something about those illnesses (weakened immune system, more time in hospitals near Covid patients) made them more likely to catch Covid. But I’d love to hear actual evidence about this or other theories. The authors, or someone else using the same data, could test whether the types of health care people are using more of 2 months pre-diagnosis are different from the ones they use more of 2 months post-diagnosis. Doctors could weigh in on the immunological plausibility of the “weakened immune system” idea. Researchers could test whether they see similar pre-trends / “Early Covid” in other claims/utilization data; probably they have but if these pre-trends hold up they seem worthy of a full paper.

What are the Richest and Poorest MSAs in the US? Cost of Living Is Probably Less Important Than You Think

Jeremy Horpedahl2 Comments

Income varies a lot across the US. So does the cost of living. Does it mostly wash out when you adjust incomes for the costs of living? No, not even close. Apples-to-apples comparisons are always hard, but it’s still worth making comparisons.

Let’s use some data that Ryan Radia put together that I really like, for several reasons. He uses the 100 largest MSAs — these comprise about 2/3 of the US population. He uses median income, so outliers shouldn’t effect the income data. He uses median family income, since the more common median household income is, in my opinion, very difficult to interpret (5 college students living together are a household, and so is one elderly person living alone). And Ryan also limits it to non-elderly, married couples, and then separates the data by the employment status of each member of the couple.

As an illustration, let’s use the data for married couples with only one spouse working full-time (I have played around with the data for other working statuses, and the results are similar). Before adjusting for the cost of living, here are the top MSAs with the highest median incomes:

  1. San Jose, CA: $169,000
  2. San Francisco: $140,000
  3. Bridgeport–Stamford, CT: $130,000
  4. Seattle: $130,000
  5. Boston: $129,000
  6. Washington, DC: $123,000
  7. Hartford, CT: $110,000
  8. Oxnard–Thousand Oaks, CA: $107,390
  9. Austin: $105,420
  10. New York: $105,000
Continue reading →

Life Tables are Cool

Zachary BartschLeave a comment

Demography is cool generally, but life tables are really cool in their elegance. Don’t know what a life table is? Let me ‘splain.

A life table uses data from private or public death registers, or even genealogical records, to identify a variety of survival and death estimates. Briefly, the tables include for each age:

  • Probability of death in the next year
  • Probability of surviving to the age
  • The life expectancy

There is more in the tables, but these are the big items that people often want to know. All of the various table columns can be calculated from survival rates. The US government and the UN each has created many such tables for a variety of time, locations, and development details. For example, the earliest and most dependable one is from 1901 and includes separate tables by race, sex, migrant status, urbanity, and even for some specific states.

Continue reading →

The Least Terrible Car Safety Sites

James Bailey1 Comment

I’m looking for a new car now and would like to know what the safest reasonable option is. There are lots of ways to get some information about this, but none are very good.

The government provides safety ratings based on crash tests they perform. This is better than nothing but the crash tests only test certain things and don’t necessarily tell you how a car performs in the real world. They also have a habit of just giving their top rating (5 stars) to tons of vehicles so it doesn’t help you pick between them, and they only compare cars to other cars in the same “class”, ignoring that some classes are safer than others. On top of all the problems with the ratings themselves, they also don’t provide any lists of their ratings, instead making you search one car at a time.

Several other sites improve on the government ratings by using real-world data on how often cars actually crash (much of which comes from the government, which as usual is great at collecting data but not so great at presenting it in helpful user-friendly ways). The Auto Professor grades cars using real-world data but otherwise has the same problems as the government (NHTSA) site. Cars get letter grades rather than a rank or meaningful number, so it’s not actually clear which car is best, or how much better the good cars are than the average or bad cars. You can search the grades for one car at a time but they don’t just list the safest cars anywhere, including on their page labelled “safest cars list“.

The Insurance Institute for Highway Safety uses real world data and provides actual numbers of fatality rates for different vehicles. This is great because you don’t have the problem of “dozens of cars all have 5-star / A, which is best?” or the problem of “how much better is 5 star than 4 star, or A than B?”. But they don’t include data from the 2 most recent years, and they only post their ratings for a handful of cars. Not only do they not present a complete list, they seem to have no search function whatsoever for their real-world data (they do for their NHSTA-style crash test data). Some 3rd party sites seem to have posted more complete versions of their data, but it still doesn’t show data for most car models.

The least-terrible car safety site I have found is Real Safe Cars. The good: they use real-world safety data, they apply reasonable-sounding corrections and controls do it, they present meaningful quantitative measures like “vehicle lifetime fatality chance” and “vehicle lifetime injury chance”, and they present the data using both a search function and lists of “safest vehicles”. For 2020 you can see that the #1 car, the 2020 Audi e-tron Sportback, has a vehicle lifetime fatality chance of 0.0158%. Compare this to the #100 car, which is about average overall- the 2020 Acura TLX has a vehicle lifetime fatality chance of 0.0435% (almost 3x the safest). The site makes it hard to find the very worst car but near the bottom is the 2020 Hyundai Accent, which “has a vehicle lifetime fatality chance of 0.0744%”.

The lists could be better; the only list that includes all vehicle classes is restricted to only 2020 makes. Meanwhile when you search a car it ranks it only relative to cars in the same year, though you can make comparisons across years yourself using the quantitative “fatality chance” and “injury chance” measures. I’m not totally convinced of the ratings themselves, given how well many smaller sedans do. Their front page explains how taller cars are generally safer, but also lists the Mini Cooper as the #18 safest car of 2020 across all classes. But Real Safe Cars seems like the current best site to me (maybe I’m biased since one of its creators is an economics professor).

I hope these sites will address some of the weaknesses I identified here, though I’m not optimistic about most of them, because other than Real Safe Cars the “bad” decisions seem to be clearly driven by incentives like keeping car companies happy or SEO.

I also think there’s still room for another effort by economists or other quantitatively-skilled people to make another site. The underlying crash data is public and the statistical problems are not especially hard; I think a single economist could run the numbers in about the time it takes to write a typical economics paper (weeks to months for a 1st draft), and a decent website could be built off that quickly as well. You could probably make a decent amount of money off the site, though perhaps not if you do the right thing and publicly post all the data and code. Posting the data would make it easy for others to copy you and make their own sites. You could fight that with copyright, but given the huge public good aspect here and the lives at stake it might make more sense to get grant funding up front and then make the data and code totally public. A sane world would have done this already; NHTSA’s annual budget is over $1 billion, with $35 million of that going to research and analysis. I think any decent funder should be able to do at least as well as the sites above with under $200k, or anyone with good data chops could do it out of the goodness of their heart in a few months. I don’t have a few months right now but perhaps one of you could take this up or start applying for grants to do it.

For everyone who just wants to know about which cars are safe, for now I think Real Safe Cars is the best bet, though I’d also like to hear if you think I missed anything.