Any time you have a benchmark that gets taken seriously, some people will start gaming the benchmark. Some famous examples in computing are the CPU benchmark specfp and video game benchmarks. With specfp, Sun managed to increase its score on 179.art (a sub-benchmark of specfp) by 12x with a compiler tweak that essentially re-wrote the benchmark kernel, which increased the Sun UltraSPARC’s overall specfp score by 20%. At times, GPU vendors have added specialized benchmark-detecting code to their drivers that lowers image quality during benchmarking to produce higher benchmark scores. Of course, gaming the benchmark isn't unique to computing and we see people do this in other fields. It’s not surprising that we see this kind of behavior since improving benchmark scores by cheating on benchmarks is much cheaper (and therefore higher ROI) than improving benchmark scores by actually improving the product.

As a result, I'm generally suspicious when people take highly specific and well-known benchmarks too seriously. Without other data, you don't know what happens when conditions aren't identical to the conditions in the benchmark. With GPU and CPU benchmarks, it’s possible for most people to run the standard benchmarks with slightly tweaked conditions. If the results change dramatically for small changes to the conditions, that’s evidence that the vendor is, if not cheating, at least shading the truth.

Benchmarks of physical devices can be more difficult to reproduce. Vehicle crash tests are a prime example of this -- they're highly specific and well-known benchmarks that use up a car for some test runs.

While there are multiple organizations that do crash tests, they each have particular protocols that they follow. Car manufacturers, if so inclined, could optimize their cars for crash test scores instead of actual safety. Checking to see if crash tests are being gamed with hyper-specific optimizations isn't really feasible for someone who isn't a billionaire. The easiest way we can check is by looking at what happens when new tests are added since that lets us see a crash test result that manufacturers weren't optimizing for just to get a good score.

While having car crash test results is obviously better than not having them, the results themselves don't tell us what happens when we get into an accident that doesn't exactly match a benchmark. Unfortunately, if we get into a car accident, we don't get to ask the driver of the vehicle we're colliding with to change their location, angle of impact, and speed, in order for the collision to comply with an IIHS, NHTSA, or *NCAP, test protocol.

For this post, we're going to look at IIHS test scores when they added the (driver side) small overlap and passenger side small overlap tests, which were added in 2012, and 2018, respectively. We'll start with a summary of the results and then discuss what those results mean and other factors to consider when evaluating car safety, followed by details of the methodology.

Results

The ranking below is mainly based on how well vehicles scored when the driver-side small overlap test was added in 2012 and how well models scored when they were modified to improve test results.

• Tier 1: good without modifications
  • Volvo
• Tier 2: mediocre without modifications; good with modifications
  • None
• Tier 3: poor without modifications; good with modifications
  • Mercedes
  • BMW
• Tier 4: poor without modifications; mediocre with modifications
  • Honda
  • Toyota
  • Subaru
  • Chevrolet
  • Tesla
  • Ford
• Tier 5: poor with modifications or modifications not made
  • Hyundai
  • Dodge
  • Nissan
  • Jeep
  • Volkswagen

These descriptions are approximations. Honda, Ford, and Tesla are the poorest fits for these descriptions, with Ford arguably being halfway in between Tier 4 and Tier 5 but also arguably being better than Tier 4 and not fitting into the classification and Honda and Tesla not really properly fitting into any category (with their category being the closest fit), but some others are also imperfect. Details below.

General commentary

If we look at overall mortality in the U.S., there's a pretty large age range for which car accidents are the leading cause of death. Although the numbers will vary depending on what data set we look at, when the driver-side small overlap test was added, the IIHS estimated that 25% of vehicle fatalities came from small overlap crashes. It's also worth noting that small overlap crashes were thought to be implicated in a significant fraction of vehicle fatalities at least since the 90s; this was not a novel concept in 2012.

Despite the importance of small overlap crashes, from looking at the results when the IIHS added the driver-side and passenger-side small overlap tests in 2012 and 2018, it looks like almost all car manufacturers were optimizing for benchmark and not overall safety. Except for Volvo, all carmakers examined produced cars that fared poorly on driver-side small overlap crashes until the driver-side small overlap test was added.

When the driver-side small overlap test was added in 2012, most manufacturers modified their vehicles to improve driver-side small overlap test scores. However, until the IIHS added a passenger-side small overlap test in 2018, most manufacturers skimped on the passenger side. When the new test was added, they beefed up passenger safety as well. To be fair to car manufacturers, some of them got the hint about small overlap crashes when the driver-side test was added in 2012 and did not need to make further modifications to score well on the passenger-side test, including Mercedes, BMW, and Tesla (and arguably a couple of others, but the data is thinner in the other cases; Volvo didn't need a hint).

Other benchmark limitations

There are a number of other areas where we can observe that most car makers are optimizing for benchmarks at the expensive of safety.

Gender, weight, and height

Another issue is crash test dummy overfitting. For a long time, adult NHSTA and IIHS tests used a 1970s 50%-ile male dummy, which is 5'9" and 171lbs. Regulators called for a female dummy in 1980 but due to budget cutbacks during the Reagan era, initial plans were shelved and the NHSTA didn't put one in a car until 2003. The female dummy is a scaled down version of the male dummy, scaled down to 5%-ile 1970s height and weight (4'11", 108lbs; another model is 4'11", 97lbs). In frontal crash tests, when a female dummy is used, it's always a passenger (a 5%-ile woman is in the driver's seat in one NHSTA side crash test and the IIHS side crash test). For reference, in 2019, the average weight of a U.S. adult male was 198 lbs and the average weight of a U.S. adult female was 171 lbs.

Using a 1970s U.S. adult male crash test dummy causes a degree of overfitting for 1970s 50%-ile men. For example, starting in the 90s, manufacturers started adding systems to protect against whiplash. Volvo and Toyota use a kind of system that reduces whiplash in men and women and appears to have slightly more benefit for women. Most car makers use a kind of system that reduces whiplash in men but, on average, has little impact on whiplash injuries in women.

It appears that we also see a similar kind of optimization for crashes in general and not just whiplash. We don't have crash test data on this, and looking at real-world safety data is beyond the scope of this post, but I'll note that, until around the time the NHSTA put the 5%-ile female dummy into some crash tests, most car manufacturers not named Volvo had a significant fatality rate differential in side crashes based on gender (with men dying at a lower rate and women dying at a higher rate).

Volvo claims to have been using computer models to simulate what would happen if women (including pregnant women) are involved in a car accident for decades.

Other crashes

Volvo is said to have a crash test facility where they do a number of other crash tests that aren't done by testing agencies. A reason that they scored well on the small overlap tests when they were added is that they were already doing small overlap crash tests before the IIHS started doing small overlap crash tests.

Volvo also says that they test rollovers (the IIHS tests roof strength and the NHSTA computes how difficult a car is to roll based on properties of the car, but neither tests what happens in a real rollover accident), rear collisions (Volvo claims these are especially important to test if there are children in the 3rd row of a 3-row SUV), and driving off the road (Volvo has a "standard" ditch they use; they claim this test is important because running off the road is implicated in a large fraction of vehicle fatalities).

If other car makers do similar tests, I couldn't find much out about the details. Based on crash test scores, it seems like they weren't doing or even considering small overlap crash tests before 2012. Based on how many car makers had poor scores when the passenger side small overlap test was added in 2018, I think it would be surprising if other car makers had a large suite of crash tests they ran that aren't being run by testing agencies, but it's theoretically possible that they do and just didn't include a passenger side small overlap test.

Caveats

We shouldn't overgeneralize from these test results. As we noted above, crash test results test very specific conditions. As a result, what we can conclude when a couple new crash tests are added is also very specific. Additionally, there are a number of other things we should keep in mind when interpreting these results.

Limited sample size

One limitation of this data is that we don't have results for a large number of copies of the same model, so we're unable to observe intra-model variation, which could occur due to minor, effectively random, differences in test conditions as well as manufacturing variations between different copies of same model. We can observe that these do matter since some cars will see different results when two copies of the same model are tested. For example, here's a quote from the IIHS report on the Dodge Dart:

The Dodge Dart was introduced in the 2013 model year. Two tests of the Dart were conducted because electrical power to the onboard (car interior) cameras was interrupted during the first test. In the second Dart test, the driver door opened when the hinges tore away from the door frame. In the first test, the hinges were severely damaged and the lower one tore away, but the door stayed shut. In each test, the Dart’s safety belt and front and side curtain airbags appeared to adequately protect the dummy’s head and upper body, and measures from the dummy showed little risk of head and chest injuries.

It looks like, had electrical power to the interior car cameras not been disconnected, there would have been only one test and it wouldn't have become known that there's a risk of the door coming off due to the hinges tearing away. In general, we have no direct information on what would happen if another copy of the same model were tested.

Using IIHS data alone, one thing we might do here is to also consider results from different models made by the same manufacturer (or built on the same platform). Although this isn't as good as having multiple tests for the same model, test results between different models from the same manufacturer are correlated and knowing that, for example, a 2nd test of a model that happened by chance showed significantly worse results should probably reduce our confidence in other test scores from the same manufacturer. There are some things that complicate this, e.g., if looking at Toyota, the Yaris is actually a re-branded Mazda2, so perhaps that shouldn't be considered as part of a pooled test result, and doing this kind of statistical analysis is beyond the scope of this post.

Actual vehicle tested may be different

Although I don't think this should impact the results in this post, another issue to consider when looking at crash test results is how results are shared between models. As we just saw, different copies of the same model can have different results. Vehicles that are somewhat similar are often considered the same for crash test purposes and will share the same score (only one of the models will be tested).

For example, this is true of the Kia Stinger and the Genesis G70. The Kia Stinger is 6" longer than the G70 and a fully loaded AWD Stinger is about 500 lbs heavier than a base-model G70. The G70 is the model that IIHS tested -- if you look up a Kia Stinger, you'll get scores for a Stinger with a note that a base model G70 was tested. That's a pretty big difference considering that cars that are nominally identical (such as the Dodge Darts mentioned above) can get different scores.

Quality may change over time

We should also be careful not to overgeneralize temporally. If we look at crash test scores of recent Volvos (vehicles on the Volvo P3 and Volvo SPA platforms), crash test scores are outstanding. However, if we look at Volvo models based on the older Ford C1 platform¹, crash test scores for some of these aren't as good (in particular, while the S40 doesn't score poorly, it scores Acceptable in some categories instead of Good across the board). Although Volvo has had stellar crash test scores recently, this doesn't mean that they have always had or will always have stellar crash test scores.

Models may vary across markets

We also can't generalize across cars sold in different markets, even for vehicles that sound like they might be identical. For example, see this crash test of a Nissan NP300 manufactured for sale in Europe vs. a Nissan NP300 manufactured for sale in Africa. Since European cars undergo EuroNCAP testing (similar to how U.S. cars undergo NHSTA and IIHS testing), vehicles sold in Europe are optimized to score well on EuroNCAP tests. Crash testing cars sold in Africa has only been done relatively recently, so car manufacturers haven't had PR pressure to optimize their cars for benchmarks and they'll produce cheaper models or cheaper variants of what superficially appear to be the same model. This appears to be no different from what most car manufacturers do in the U.S. or Europe -- they're optimizing for cost as long as they can do that without scoring poorly on benchmarks. It's just that, since there wasn't an African crash test benchmark, that meant they could go all-in on the cost side of the cost-safety tradeoff².

This report compared U.S. and European car models and found differences in safety due to differences in regulations. They found that European models had lower injury risk in frontal/side crashes and that driver-side mirrors were designed in a way that reduced the risk of lane-change crashes relative to U.S. designs and that U.S. vehicles were safer in rollovers and had headlamps that made pedestrians more visible.

Non-crash tests

Over time, more and more of the "low hanging fruit" from crash safety has been picked, making crash avoidance relatively more important. Tests of crash mitigation are relatively primitive compared to crash tests and we've seen that crash tests had and have major holes. One might expect, based on what we've seen with crash tests, that Volvo has a particularly good set of tests they use for their crash avoidance technology (traction control, stability control, automatic braking, etc.), but "bar room" discussion with folks who are familiar with what vehicle safety tests are being done on automated systems seems to indicate that's not the case. There was a relatively recent recall of quite a few Volvo vehicles due to the safety systems incorrectly not triggering. I'm not going to tell the story about that one here, but I'll say that it's fairly horrifying and indicative of serious systemic issues. From other backchannel discussions, it sounds like BMW is relatively serious about the software side of safety, for a car company, but the lack of rigor in this kind of testing would be horrifying to someone who's seen a release process for something like a mainstream CPU.

Crash avoidance becoming more important might also favor companies that have more user-friendly driver assistance systems, e.g., in multiple generations of tests, Consumer Reports has given GM's Super Cruise system the highest rating while they've repeatedly noted that Tesla's Autopilot system facilitates unsafe behavior.

Scores of vehicles of different weights aren't comparable

A 2700lb subcompact vehicle that scores Good may fare worse than a 5000lb SUV that scores Acceptable. This is because the small overlap tests involve driving the vehicle into a fixed obstacle, as opposed to a reference vehicle or vehicle-like obstacle of a specific weight. This is, in some sense, equivalent to crashing the vehicle into a vehicle of the same weight, so it's as if the 2700lb subcompact was tested by running it into a 2700lb subcompact and the 5000lb SUV was tested by running it into another 5000 lb SUV.

How to increase confidence

We've discussed some reasons we should reduce our confidence in crash test scores. If we wanted to increase our confidence in results, we could look at test results from other test agencies and aggregate them and also look at public crash fatality data (more on this later). I haven't looked at the terms and conditions of scores from other agencies, but one complication is that the IIHS does not allow you to display the result of any kind of aggregation if you use their API or data dumps (I, time consumingly, did not use their API for this post because of that).

Using real life crash data

Public crash fatality data is complex and deserves its own post. In this post, I'll note that, if you look at the easiest relevant data for people in the U.S., this data does not show that Volvos are particularly safe (or unsafe). For example, if we look at this report from 2017, which covers models from 2014, two Volvo models made it into the report and both score roughly middle of the pack for their class. In the previous report, one Volvo model is included and it's among the best in its class, in the next, one Volvo model is included and it's among the worst in its class. We can observe this kind of variance for other models, as well. For example, among 2014 models, the Volkswagen Golf had one of the highest fatality rates for all vehicles (not just in its class). But among 2017 vehicles, it had among the lowest fatality rates for all vehicles. It's unclear how much of that change is from random variation and how much is because of differences between a 2014 and 2017 Volkswagen Golf.

Overall, it seems like noise is a pretty important factor in results. And if we look at the information that's provided, we can see a few things that are odd. First, there are a number of vehicles where the 95% confidence interval for the fatality rate runs from 0 to N. We should have pretty strong priors that there was no 2014 model vehicle that was so safe that the probability of being killed in a car accident was zero. If we were taking a Bayesian approach (though I believe the authors of the report are not), and someone told us that the uncertainty interval for the true fatality rate of a vehicle had a >= 5% of including zero, we would say that either we should use a more informative prior or we should use a model that can incorporate more data (in this case, perhaps we could try to understand the variance between fatality rates of different models in the same class and then use the base rate of fatalities for the class as a prior, or we could incorporate information from other models under the same make if those are believed to be correlated).

Some people object to using informative priors as a form of bias laundering, but we should note that the prior that's used for the IIHS analysis is not completely uninformative. All of the intervals reported stop at zero because they're using the fact that a vehicle cannot create life to bound the interval at zero. But we have information that's nearly as strong that no 2014 vehicle is so safe that the expected fatality rate is zero, using that information is not fundamentally different from capping the interval at zero and not reporting negative numbers for the uncertainty interval of the fatality rate.

Also, the IIHS data only includes driver fatalities. This is understandable since that's the easiest way to normalize for the number of passengers in the car, but it means that we can't possibly see the impact of car makers not improving passenger small-overlap safety until the passenger-side small overlap test was added in 2018, the result of lack of rear crash testing for the case Volvo considers important (kids in the back row of a 3rd row SUV). This also means that we cannot observe the impact of a number of things Volvo has done, e.g., being very early on pedestrian and then cyclist detection in their automatic braking system, adding a crumple zone to reduce back injuries in run-off-road accidients, which they observed often cause life-changing spinal injuries due to the impact from vehicles drop, etc.

We can also observe that, in the IIHS analysis, many factors that one might want to control for aren't (e.g., miles driven isn't controlled for, which will make trucks look relatively worse and luxury vehicles look relatively better, rural vs. urban miles driven also isn't controlled for, which will also have the same directional impact). One way to see that the numbers are heavily influenced by confounding factors is by looking at AWD or 4WD vs. 2WD versions of cars. They often have wildly different fatalty rates even though the safety differences are not very large (and the difference is often in favor of the 2WD vehicle). Some plausible causes of that are random noise, differences in who buys different versions of the same vehicle, and differences in how the vehicle are used.

If we'd like to answer the question "which car makes or models are more or less safe", I don't find any of the aggregations that are publicly available to be satisfying and I think we need to look at the source data and do our own analysis to see if the data are consistent with what we see in crash test results.

Conclusion

We looked at 12 different car makes and how they fared when the IIHS added small overlap tests. We saw that only Volvo was taking this kind of accident seriously before companies were publicly shamed for having poor small overlap safety by the IIHS even though small overlap crashes were known to be a significant source of fatalities at least since the 90s.

Although I don't have the budget to do other tests, such as a rear crash test in a fully occupied vehicle, it appears plausible and perhaps even likely that most car makers that aren't Volvo would have mediocre or poor test scores if a testing agency decided to add another kind of crash test.

Bonus: "real engineering" vs. programming

As Hillel Wayne has noted, although programmers often have an idealized view of what "real engineers" do, when you compare what "real engineers" do with what programmers do, it's frequently not all that different. In particular, a common lament of programmers is that we're not held liable for our mistakes or poor designs, even in cases where that costs lives.

Although automotive companies can, in some cases, be held liable for unsafe designs, just optimizing for a small set of benchmarks, which must've resulted in extra deaths over optimizing for safety instead of benchmark scores, isn't something that engineers or corporations were, in general, held liable for.

Bonus: reputation

If I look at what people in my extended social circles think about vehicle safety, Tesla has the best reputation by far. If you look at broad-based consumer polls, that's a different story, and Volvo usually wins there, with other manufacturers fighting for a distant second.

I find the Tesla thing interesting since their responses are basically the opposite of what you'd expect from a company that was serious about safety. When serious problems have occurred (with respect to safety or otherwise), they often have a very quick response that's basically "everything is fine". I would expect an organization that's serious about safety or improvement to respond with "we're investigating", followed by a detailed postmortem explaining what went wrong, but that doesn't appear to be Tesla's style.

For example, on the driver-side small overlap test, Tesla had one model with a relevant score and it scored Acceptable (below Good, but above Poor and Marginal) even after modifications were made to improve the score. Tesla disputed the results, saying they make "the safest cars in history" and implying that IIHS should be ignored because they have ulterior motives, in favor of crash test scores from an agency that is objective and doesn't have ulterior motives, i.e., the agency that gave Tesla a good score:

While IIHS and dozens of other private industry groups around the world have methods and motivations that suit their own subjective purposes, the most objective and accurate independent testing of vehicle safety is currently done by the U.S. Government which found Model S and Model X to be the two cars with the lowest probability of injury of any cars that it has ever tested, making them the safest cars in history.

As we've seen, Tesla isn't unusual for optimizing for a specific set of crash tests and achieving a mediocre score when an unexpected type of crash occurs, but their response is unusual. However, it makes sense from a cynical PR perspective. As we've seen over the past few years, loudly proclaiming something, regardless of whether or not it's true, even when there's incontrovertible evidence that it's untrue, seems to not only work, that kind of bombastic rhetoric appears to attract superfans who will aggressively defend the brand. If you watch car reviewers on youtube, they'll sometimes mention that they get hate mail for reviewing Teslas just like they review any other car and that they don't see anything like it for any other make.

Apple also used this playbook to good effect in the 90s and early '00s, when they were rapidly falling behind in performance and responded not by improving performance, but by running a series of ad campaigns saying that had the best performance in the world and that they were shipping "supercomputers" on the desktop.

Another reputational quirk is that I know a decent number of people who believe that the safest cars they can buy are "American Cars from the 60's and 70's that aren't made of plastic". We don't have directly relevant small overlap crash test scores for old cars, but the test data we do have on old cars indicates that they fare extremely poorly in overall safety compared to modern cars. For a visually dramatic example, see this crash test of a 1959 Chevrolet Bel Air vs. a 2009 Chevrolet Malibu.

Appendix: methodology summary

The top-line results section uses scores for the small overlap test both because it's the one where I think it's the most difficult to justify skimping on safety as measured by the test and it's also been around for long enough that we can see the impact of modifications to existing models and changes to subsequent models, which isn't true of the passenger side small overlap test (where many models are still untested).

For the passenger side small overlap test, someone might argue that the driver side is more important because you virtually always have a driver in a car accident and may or may not have a front passenger. Also, for small overlap collisions (which simulates a head-to-head collision where the vehicles only overlap by 25%), driver's side collisions are more likely than passenger side collisions.

Except to check Volvo's scores, I didn't look at roof crash test scores (which were added in 2009). I'm not going to describe the roof test in detail, but for the roof test, someone might argue that the roof test score should be used in conjunction with scoring the car for rollover probability since the roof test just tests roof strength, which is only relevant when a car has rolled over. I think, given what the data show, this objection doesn't hold in many cases (the vehicles with the worst roof test scores are often vehicles that have relatively high rollover rates), but it does in some cases, which would complicate the analysis.

In most cases, we only get one reported test result for a model. However, there can be multiple versions of a model -- including before and after making safety changes intended to improve the test score. If changes were made to the model to improve safety, the test score is usually from after the changes were made and we usually don't get to see the score from before the model was changed. However, there are many exceptions to this, which are noted in the detailed results section.

For this post, scores only count if the model was introduced before or near when the new test was introduced, since models introduced later could have design changes that optimize for the test.

Appendix: detailed results

On each test, IIHS gives an overall rating (from worst to best) of Poor, Marginal, Acceptable, or Good. The tests have sub-scores, but we're not going to use those for this analysis. In each sub-section, we'll look at how many models got each score when the small overlap tests were added.

Volvo

All Volvo models examined scored Good (the highest possible score) on the new tests when they were added (roof, driver-side small overlap, and passenger-side small overlap). One model, the 2008-2017 XC60, had a change made to trigger its side curtain airbag during a small overlap collision in 2013. Other models were tested without modifications.

Mercedes

Of three pre-existing models with test results for driver-side small overlap, one scored Marginal without modifications and two scored Good after structural modifications. The model where we only have unmodified test scores (Mercedes C-Class) was fully re-designed after 2014, shortly after the driver-side small overlap test was introduced.

As mentioned above, we often only get to see public results for models without modifications to improve results xor with modifications to improve results, so, for the models that scored Good, we don't actually know how they would've scored if you bought a vehicle before Mercedes updated the design, but the Marginal score from the one unmodified model we have is a negative signal.

Also, when the passenger side small overlap test was added, the Mercedes vehicles also generally scored Good. This is, indicating that Mercedes didn't only increase protection on the driver's side in order to improve test scores.

BMW

Of the two models where we have relevant test scores, both scored Marginal before modifications. In one of the cases, there's also a score after structural changes were made in the 2017 model (recall that the driver-side small overlap test was introduced in 2012) and the model scored Good afterwards. The other model was fully-redesigned after 2016.

For the five models where we have relevant passenger-side small overlap scores, all scored Good, indicating that the changes made to improve driver-side small overlap test scores weren't only made on the driver's side.

Honda

Of the five Honda models where we have relevant driver-side small overlap test scores, two scored Good, one scored Marginal, and two scored Poor. The model that scored Marginal had structural changes plus a seatbelt change in 2015 that changed its score to Good, other models weren't updated or don't have updated IIHS scores.

Of the six Honda models where we have passenger driver-side small overlap test scores, two scored Good without modifications, two scored Acceptable without modifications, and one scored Good with modifications to the bumper.

All of those models scored Good on the driver side small overlap test, indicating that when Honda increased the safety on the driver's side to score Good on the driver's side test, they didn't apply the same changes to the passenger side.

Toyota

Of the six Toyota models where we have relevant driver-side small overlap test scores for unmodified models, one score Acceptable, four scored Marginal, and one scored Poor.

The model that scored Acceptable had structural changes made to improve its score to Good, but on the driver's side only. The model was later tested in the passenger-side small overlap test and scored Acceptable. Of the four models that scored Marginal, one had structural modifications made in 2017 that improved its score to Good and another had airbag and seatbelt changes that improved its score to to Acceptable. The vehicle that scored Poor had structural changes made that improved its score to acceptable in 2014, followed by later changes that improved its score to Good.

There are four additional models where we only have scores from after modifications were made. Of those, one scored Good, one score Acceptable, one scored Marginal, and one scored Poor.

In general, changes appear to have been made to the driver's side only and, on introduction of the passenger side small overlap test, vehicles had passenger side small overlap scores that were the same as the driver's side score before modifications.

Ford

Of the two models with relevant driver-side small overlap test scores for unmodified models, one scored Marginal and one scored Poor. Both of those models were produced into 2019 and neither has an updated test result. Of the three models where we have relevant results for modified vehicles, two scored Acceptable and one score Marginal. Also, one model was released the year the small overlap test was introduced and one the year after; both of those scored Acceptable. It's unclear if those should be considered modified or not since the design may have had last-minute changes before release.

We only have three relevant passenger-side small overlap tests. One is Good (for a model released in 2015) and the other two are Poor; these are the two models mentioned above as having scored Marginal and Poor, respectively, on the driver-side small overlap test. It appears that the models continued to be produced into 2019 without safety changes. Both of these unmodified models were trucks and this isn't very unusual for a truck and is one of a number of reasons that fatality rates are generally higher in trucks -- until recently, many of them are based on old platforms that hadn't been updated for a long time.

Chevrolet

Of the three Chevrolet models where we have relevant driver-side small overlap test scores before modifications, one scored Acceptable and two scored Marginal. One of the Marginal models had structural changes plus a change that caused side curtain airbags to deploy sooner in 2015, which improved its score to Good.

Of the four Chevrolet models where we only have relevant driver-side small overlap test scores after the model was modified (all had structural modifications), two scored Good and two scored Acceptable.

We only have one relevant score for the passenger-side small overlap test, that score is Marginal. That's on the model that was modified to improve its driver-side small overlap test score from Marginal to Good, indicating that the changes were made to improve the driver-side test score and not to improve passenger safety.

Subaru

We don't have any models where we have relevant passenger-side small overlap test scores for models before they were modified.

One model had a change to cause its airbag to deploy during small overlap tests; it scored Acceptable. Two models had some kind of structural changes, one of which scored Good and one of which score Acceptable.

The model that had airbag changes had structural changes made in 2015 that improved its score from Acceptable to Good.

For the one model where we have relevant passenger-side small overlap test scores, the score was Marginal. Also, for one of the models with structural changes, it was indicated that, among the changes, were changes to the left part of the firewall, indicating that changes were made to improve the driver's side test score without improving safety for a passenger on a passenger-side small overlap crash.

Tesla

There's only one model with relevant results for the driver-side small overlap test. That model scored Acceptable before and after modifications were made to improve test scores.

Hyundai

Of the five vehicles where we have relevant driver-side small overlap test scores, one scored Acceptable, three scored Marginal, and one scored Poor. We don't have any indication that models were modified to improve their test scores.

Of the two vehicles where we have relevant passenger-side small overlap test scores for unmodified models, one scored Good and one scored Acceptable.

We also have one score for a model that had structural modifications to score Acceptable, which later had further modifications that allowed it to score Good. That model was introduced in 2017 and had a Good score on the driver-side small overlap test without modifications, indicating that it was designed to achieve a good test score on the driver's side test without similar consideration for a passenger-side impact.

Dodge

Of the five models where we have relevant driver-side small overlap test scores for unmodified models, two scored Acceptable, one scored Marginal, and two scored Poor. There are also two models where we have test scores after structural changes were made for safety in 2015; both of those models scored Marginal.

We don't have relevant passenger-side small overlap test scores for any model, but even if we did, the dismal scores on the modified models means that we might not be able to tell if similar changes were made to the passenger side.

Nissan

Of the seven models where we have relevant driver-side small overlap test scores for unmodified models, two scored Acceptable and five scored Poor.

We have one model that only has test scores for a modified model; the frontal airbags and seatbelts were modified in 2013 and the side curtain airbags were modified in 2017. The score afterward modifications was Marginal.

One of the models that scored Poor had structural changes made in 2015 that improved its score to Good.

Of the four models where we have relevant passenger-side small overlap test scores, two scored Good, one scored Acceptable (that model scored good on the driver-side test), and one score Marginal (that model also scored Marginal on the driver-side test).

Jeep

Of the two models where we have relevant driver-side small overlap test scores for unmodified models, one scored Marginal and one scored Poor.

There's one model where we only have test score after modifications; that model has changes to its airbags and seatbelts and it scored Marginal after the changes. This model was also later tested on the passenger-side small overlap test and scored Poor.

One other model has a relevant passenger-side small overlap test score; it scored Good.

Volkswagen

The two models where we have relevant driver-side small overlap test scores for unmodified models both scored Marginal.

Of the two models where we only have scores after modifications, one was modified 2013 and scored Marginal after modifications. It was then modified again in 2015 and scored Good after modifications. That model was later tested on the passenger side small-overlap test, where it scored Acceptable, indicating that the modifications differentially favored the driver's side. The other scored Acceptable after changes made in 2015 and then scored Good after further changes made in 2016. The 2016 model was later tested on the passenger-side small overlap test and scored Marginal, once again indicating that changes differentially favored the driver's side.

We have passenger-side small overlap test for two other models, both of which scored Acceptable. These were models introduced in 2015 (well after the introduction of the driver-side small overlap test) and scored Good on the driver-side small overlap test.

2021 update

The IIHS has released the first set of results for their new "upgraded" side-impact tests. They've been making noises about doing this for quite and have mentioned that in real-world data on (some) bad crashes, they've observed intrusion into the cabin that's significantly greater than is seen on their tests. They've mentioned that some vehicles do relatively well on on the new tests and some less well but haven't released official scores until now.

The results in the new side-impact tests are different from the results described in the posts above. So far, only small SUVs have had their results released and only the Mazda CX-5 has a result of "Good". Of the three manufacturers that did well on the tests describe in this post, only Volvo has public results and they scored "Acceptable". Some questions I have are:

• Will Volvo score better for their other vehicles (most of their vehicles are built on a different platform from the vehicle that has public results)?
• Will Volvo quickly update their vehicles to achieve the highest score on the test? Unlike a lot of other manufacturers, we don't have recent data from Volvo on how they responded to something like this because they didn't need to update their vehicles to achieve the highest score on the last two new tests
• Will BMW and Mercedes either score well and the new test or quickly update their vehicles to score well once again?
• Will other Mazda vehicles also score well without updates?

2024 update

In a 2024 analysis of fatality rate per mile driven from 2018-2022, the worst car manufacturers were, starting from the worst, were Tesla, Kia, Buick, Dodge, and then Hyundai. Buick wasn't ranked in this post and Kia and Hyundai were considered equivalent, so of the four ranked makes, three had the worst score in this rating. And, as originally noted in the post, Tesla doesn't fit into the categorization very well and shows signs of being the worst for safety as well as signs of being perhaps average, and there are dimensions on which cars weren't ranked where Tesla seems to have very poor safety (ADAS / self-driving), so there's a strong case that Tesla should have also been put in the worst category.

Also note that none of the three manufactueres that were rated well even had a single car that made the list of models with the highest fatality rate per mile. But it's hard to say how much of this is about the car and how much is about other properties (such as how the car is used) since fatalities per mile are fairly strongly negatively correlated with car price and all three manufacturers are luxury brands that have well above average sale price. Luxury cars also tend to be larger and heavier than average and weight is also negatively correlated with fatalities per mile driven.

Over the time period ranked, Tesla appears to have had the highest average selling price (even higher than the three top ranked luxury brands) and also had well above median weight per vehicle, making Tesla an extreme outlier in fatalities per mile.

Appendix: miscellania

A number of name brand car makes weren't included. Some because they have relatively low sales in the U.S. are low and/or declining rapidly (Mitsubishi, Fiat, Alfa Romeo, etc.), some because there's very high overlap in what vehicles are tested (Kia, Mazda, Audi), and some because there aren't relevant models with driver-side small overlap test scores (Lexus). When a corporation owns an umbrella of makes, like FCA with Jeep, Dodge, Chrysler, Ram, etc., these weren't pooled since most people who aren't car nerds aren't going to recognize FCA, but may recognize Jeep, Dodge, and Chrysler.

If the terms of service of the API allowed you to use IIHS data however you wanted, I would've included smaller makes, but since the API comes with very restrictive terms on how you can display or discuss the data which aren't compatible with exploratory data analysis and I couldn't know how I would want to display or discuss the data before looking at the data, I pulled all of these results by hand (and didn't click through any EULAs, etc.), which was fairly time consuming, so there was a trade-off between more comprehensive coverage and the rest of my life.

Appendix: what car should I buy?

That depends on what you're looking for, there's no way to make a blanket recommendation. For practical information about particular vehicles, Alex on Autos is the best source that I know of. I don't generally like videos as a source of practical information, but car magazines tend to be much less informative than youtube car reviewers. There are car reviewers that are much more popular, but their popularity appears to come from having witty banter between charismatic co-hosts or other things that not only aren't directly related to providing information, they actually detract from providing information. If you just want to know about how cars work, Engineering Explained is also quite good, but the information there is generally practical.

For reliability information, Consumer Reports is probably your best bet (you can also look at J.D. Power, but the way they aggregate information makes it much less useful to consumers).

Thanks to Leah Hanson, Travis Downs, Prabin Paudel, Jeshua Smith, and Justin Blank for comments/corrections/discussion