Category: Analysis

Anything that I have tested empirically, or explain with detailed calculations, or just generally bother to get all the facts straight.

Post #1830A: Leaf extras-for-stat-nerds

 

The statistically savvy among you might have noticed something odd about the graph at the end of the last post.

When I let unconstrained math (via Excel) determine the best straight-line fit to these data points, it appears to tell me that these cars lost about 4% of battery capacity for every 10K miles.

But …

But what that doesn’t do is guarantee that with zero miles, I predict that you’ll have 100% of original capacity.

And that matters, because it looks like a lot of these cars must have lost a lot of range early on, that is, at low mileage.  And that loss of battery range doesn’t get factored into the 4% per 10K range loss estimate.

So this is a rare instance of a straight-line fit for which you are justified in “setting the intercept” manually, rather than letting an unconstrained least-squares fit to the data do it.  By definition, the line has to pass through zero miles matching 100% of original range.

Look what happens when I do that:

On average, pinning the linear trend to pass through 0 miles = 100% of original range, range loss is more like 8% per 10K miles. 

The whole cluster of dots is quite “low” on that graph, so to speak.  Those cars on average followed a path of losing a lot of range early on.  And having the loss taper off to a mere 4% per year.

My guess is that exponential decay is the line you’d like to fit, for something wasting away.  But I can’t seem to get Excel to give a trend passing through the required point, with the required shape.  So I freehanded what I think the Conventional Wisdom says about the time-path of battery loss.

No matter how you cut it, the actual observed battery loss over this range amounts to much more than 4%/10K miles.  The mid-point of the fitted line sits around a 40% loss over 5K miles.  That’s the 8%/year calculated when the regression line was pinned at 100% capacity at zero miles.

Why the high range losses?  Is this just a market for lemons, or are the dashboard estimates biased?

George Akerlof, economist, once wrote a piece whose title began “The Market for ‘Lemons’: ”  The paper is pretty deep, but the takeaway is pretty simple.  In this case, it boils down to: People sell their cars when those cars are lemons, not when they are peaches.

That’s one obvious explanation of why these losses appear far larger-than-expected.  (Where, lurking in the background is the idea that EVs lose, oh, like two percent a year, maybe, based on looking at a few graphs in the past.)  The point is, maybe most of these cars appear in the used car listings precisely because they had above-average battery capacity losses.  And that’s why their prior owners sold them.  And I’m seeing them.

The other possibility is that the mileage estimated from the dashboard readout is substantially biased downwards.  (I know it has a high variance, as it depends on recent driving style.  Leaf aficionados refer to it as the “guess gauge”.  That should just add noise, not bias, I think.)

There is one element of bias in what I did in my calculation, in that the range bars are each about 8 percent thick.  For example, when I saw four range bars, my calculation assumed the car was (4/12 =) 33% charged.  But if four bars is really just 3.5 bars, on average, then that charge level is actually just a shade over 28%, on average.  I would then triple that error in arriving at 100% of charge, and so end up understating total range by a factor of about 15%.

Adjusting for that  would require multiplying my range estimate by (1/.85 =) about 18%.  The upshot is that what looked like a 60-mile range based on the dashboard could plausibly be a 70-mile range in reality.

That said, I’m not sure this materially changes the situation.  No matter how I slice it, the range of the car I’m interested is far lower than the Recurrent.com estimate that I had been looking at.

Finally, I can’t fully discount that the losses observed in this low-end used car sample are typical of a random sample of Leafs.  But if it were, I doubt there’d be many Leaf fans out there.

On net, I think the explanation is that I’m looking at a market for lemons.  By looking only at the low-cost end of the market, and only looking at what people are trying to sell, I’m probably looking at a fairly biased sample of all Leafs.

Unfortunately, that’s the sample of cars I’m buying from, if I continue down this path.  Perhaps time for a re-think.

Post #1828: Factors affecting the price of a used Nissan Leaf

 

The single biggest way to save money on a used Nissan Leaf?  Don’t buy from Carvana.

That’s what the data say.  Other results are in the RESULTS section below.

Hedonic price function

I didn’t make up the term.  A hedonic price function is the method economists use to parse out the factors that contribute to the market value of an object.

Basically, you take what everybody knows — newer cars cost more, high-mileage cars cost less — and you quantify it.  You calculate a data-based estimate of just how much those factors affect the price of the object in question.

Probably the single most important use of this approach is in the U.S. Consumer Price Index.  That index is no longer a simple average of observed prices for various goods.  That’s because the qualities of the underlying goods keep changing.  Take computers for example.  A new PC is typically far more capable than a ten-year-old PC.  If you simply lump those two together as “PCs”, you ignore the increase in the value of what the consumer is buying.  So, the Bureau of Labor Statistics uses hedonic price estimates to determine how much “better” newer models are, compared to older models, and adjusts the prices of newer models (down) accordingly.  The other way to look at that is that this is how Uncle Sam systematically understates the true rate of inflation.  And, accordingly, saves money on all of its payments linked to inflation, such as Social Security payments.

Practically speaking, I “ran a regression”.  Which is, at root, just an average.  Instead of taking a simple average price, it’s an adjusted average.  It adjusts the average asking price for the factors that you have identified (e.g., age, mileage). 

The nice thing about it is that it gives you the independent effect of each factor.  So, newer used cars also tend to be lower-mileage used cars.  To what extent is their higher price for late-model cars due to low age, versus low mileage?  In theory, a regression parses that out for you.

Regressions, much like the Magic 8 Ball from which they were derived, sometimes give you a non-committal answer.  By that I mean that there’s so little systematic impact of some factor that you can’t really say whether or not you’re just picking up random associations.  The ones above that are egregiously so, I’ve put in faint gray type.  You should ignore any estimates that don’t pass a standard hurdle for “statistical significance” this way.

(I hope it goes without saying that regressions may sometimes give a wrong answer, and may sometimes give a right answer.  Again, much like the Magic 8 Ball.)

Results

By far, the most important factor for saving money on a used Leaf is not to buy it from Carvana.  Averaged across all the cars in the dataset, buying from Carvana added $3K to the price of the car.  I don’t think that’s news, as I vaguely recall that they had a reputation for being a convenient-but-relatively-costly source for used cars.

But just as interesting are the things that don’t matter.

The trim level or options level had no (statistically significant) effect on the price.  This, I think, also matches conventional wisdom.  As stated by my mother-in-law: Buy a loaded car because you want to drive a loaded car.  Don’t buy it for the additional resale value, because there isn’t any.  (But keep in mind, the effect of difference in range is accounted for elsewhere, and so differences in battery capacity across trim lines should not, in theory, contribute to the pure effect of the trim level.)

Curiously, history of having been in an accident also did not much matter.  I have no explanation for that, except that maybe most of those aren’t major accidents?

Now we get to the big three:  Year, Mileage, and Range.  To interpret these, you need to know that the average asking price in this dataset was about $15,000.

All other things equal, each year that a used Leaf ages, as a used car, it loses about $900 in value, or about 6% of its value.  See the note below on comparing this to the depreciation of a brand-new Leaf.  It’s lower, but not vastly lower, than the depreciation of a new car.  I think this matches our received folklore about car values.  The fastest reduction in value that you will experience is in the moment you drive your new car off the dealer’s lot.

Mileage hardly matters. Every additional 10K miles on the odometer only drops the value of the car by about $200, or just over 1%.  I think that’s quite  different from what you see with gas cars, and speaks to the robustness of the motors and drive train.  It appears that the typical buyer isn’t concerned with expensive repairs for a fairly high-mileage used Leaf.

Alternatively, these cars may simply have such low average miles, that mileage had not yet begun to matter.  Median mileage was just 36K, median age was about five years.  The low average mileage is a function of both low car age and few miles per year, consistent with a car in which long-distance travel was likely awkward for most.)

Slice that either way you like.  The fact is, odometer miles didn’t much matter.

Oddly, battery range only matters a bit.  This analysis was restricted to models that had an estimate of the actual remaining range, via some remote monitoring service that tests battery life.  Every 10 miles of additional battery range was only worth about $430, or maybe 3%.   I’m guessing that people interested in a used Leaf are probably coming from the same place that I am, in that they plan to use it as a “city car”.  So extra range is nice, but not a deal-breaker, as long as it’ll get you around town for the next few years.

I’m not sure the same would hold true for the price of a new EV, where the typical buyer may want to be able to take long trips in the car.  There, I’d expect to see a higher value on longer range.

Final takeaways.

First, I’m not going to buy from Carvana.  As if I ever were.

Second, on paper, as new cars, Leafs appear to have suffered from shockingly steep depreciation.  So much so that it gave me pause.  How could a car that lost value that quickly be a good buy as a used car?

But I think that’s largely an artifact of the (then) $7500 tax credit.  A person buying a $32500 Leaf actually only paid $25,000 for it, net of the tax credit.  If I factor the tax credit in, depreciation on a new Leaf that appeared to be in excess of 12%/year falls to a far more reasonable 8% per year.  Or just modestly higher than the rate of depreciation shown among the sample of used Leafs.

Third, at the price range I’m considering (~$10K), it probably won’t cost me much to buy the model with the bigger battery/higher range (SV) instead of the base model (S, SL).  The additional 30 miles of range (100 vs 70) ought to cost me just over an additional $1200.  And that additional range should be cheap insurance against the future degradation of the lithium-ion battery.

Finally, as long as the battery is in good shape, I probably shouldn’t worry too much about the mileage.  Or, at least, that’s what the market is telling me.  Buyers seemed to be almost indifferent to the miles on the odometer.  To me, that crowd-sources the conclusion that, battery aside, these cars don’t tend to wear out within the range of mileages typically found for used Leafs.

Epilogue

A wise philospher once said, “You cannot step into the same used-car query twice.”  I thought about adding more information (e.g., Carmax as seller).  But I “keypunched” the cars in the order they showed when I ran my first query.  To get back to those Edmunds.com search results, I’d have to re-run that query.  That could easily give me a slightly different set of cars, depending.  Which means the existing dataset might not synch up with the new data, as punched.  Not without extraordinary measures.

In this case, the juice is not worth the squeeze.  I’ll let this be the end of it.

Post 1826: Used Leaf versus remaining alternatives

 

I’m still working on getting something to replace the car I got rid of a year ago.  For use around town, when it’s inconvenient to borrow my wife’s car.

Bottom line:  By the process of calculatus eliminatus, my best and most realistic option is a well-used Nissan Leaf.  The 30% Federal tax rebate, on the purchase price of this qualifying used EV, is just gravy.

Beyond the items already discussed and dismissed (e-bike, e-motorcycle), the remaining options for local transport in my area seem to be:

  • Used Nissan Leaf
  • Uber or similar
  • Public transportation

In a nutshell, near as I can figure:

  • Used Leaf, convenient, $0.55/mile.
  • Uber, just about as convenient, $4/mile.
  • Public transport, inconvenient, not great in all weather, $1/mile.

Continue reading Post 1826: Used Leaf versus remaining alternatives

Post #1820: Dribs and drabs of Town of Vienna historical data.

 

Today I stumbled across the dollar value of the 1961 Town of Vienna operating budget, in an old Town of Vienna newsletter.  It seemed small to me, even after considering inflation.  So I decided to compare a few key statistics for the Town of Vienna, 1960 (ish) versus 2023 (ish).  And, in fact, it was small.

N.B., one U.S. dollar, in 1960, was worth just over $10, in current (July 2023) currency.  Actual silver coinage (90% silver coins) did not disappear from U.S. circulation until 1964.

Town of Vienna, VA:  2023 versus 1960

Population:  43% increase

  • Per the 1960 Census:  11,500
  • Per the 2020 Census:  16,500

Dwelling units: 100% increase:

  • Per the 1960 Census:  2,750
  • Per the 2020 Census:  5,600 (est).

Persons per household:  26% decrease.

  • Per the 1960 Census:  4.1
  • Circa the 2021 Census:  3.03

N.B. 2 x .74 = 1.48, so even though the data above come from different, independent data sources, the math very nearly reproduces the actual increase in population (44%, not 48%) over the period.

Median house price:  Roughly five-fold increase.

  • Per the 1960 Census: $18,400
  • The 1960 price in 2023 dollars:  $189,000
  • 2023 median (estimated), all houses: $900,000.
  • 2023 median, listed for sale:  $1,030,000

Town of Vienna operating budget:  Roughly 10-fold increase.

  • Per the June 1961 TOV newsletter:  $462,000.
  • The 1960 cost in 2023 dollars:  $4,800,000.
  • The 2023-24 actual cost:  $50,000,000

Most of the operating revenue for the Town comes from real estate taxes.  To reconcile the 10-fold growth in house prices, and the five-fold growth in the cost of government, you have to know that the tax rate per $100 of market value fell by roughly 50% over this period.  In 1961, it was $1.35 per $100, assessed at 32% of market value, or (1.35 x .32 =) 0.42 per $100 of market value.  That, from the June 1961 Town newsletter.  Currently, the rate is just over $0.20 per $100 of market value, per the 2023-24 Town budget.

References:

1960 Census of Population and Housing:  https://usa.ipums.org/usa/voliii/pubdocs/1960/pubvols1960.shtml

2020 Census:  https://www.census.gov/quickfacts/fact/table/viennatownvirginia/PST045222

Town of Vienna 1961 newsletter:  https://www.viennava.gov/engagement-central/newsroom/vienna-voice-town-newsletter

Town of Vienna operating budget:  https://www.viennava.gov/your-government/town-budget

Inflation calculator:  https://www.usinflationcalculator.com/

Post #1813: Won’t be hiring ChatGPT as my AI staff writer, either.

 

AI: It’ll always give you an answer.  It might give you a very convincing answer.  And it might sometimes even give you the right answer.  But the last part is far from guaranteed.

Inconceivable!

In my ongoing quest to automate this blog, I threw my test questions at ChatGPT.  These questions are:

  1. Explain why ceiling fans are more efficient than box fans.
  2. Explain why box fans are more efficient than ceiling fans.

I already know the answer.  Ceiling fans are more efficient, and I already worked out why, in a post that I have currently withdrawn.  To move the same volume of air per minute, a small fan has to move the air much faster, and that eats up energy, compared to moving a larger volume of air, more slowly.  Ceiling fans are more efficient at moving a given volume of air simply because they are larger, so they can move that required volume of air at a much slower speed.

To save you the reading, ChatGPT produces nicer text, and more of it.  But at root, the answers had the same problems as Bard.ai.

  • Just like Bard.ai, ChatGPT will do its best to give you the answer you’re looking for.  Even if you’re wrong, as in the second question above.
  • Just like Bard.ai, ChatGPT conflates the definition of efficiency (CFM/watt) with the reason for efficiency.
  • Just like Bard.ai, I get a lot of extraneous stuff that I now realize probably came from various ads and marketing materials.  Aerodynamic blade shapes and whatnot.

Even when I tried to point ChatGPT in the exact direction I wanted — explain the physics of why ceiling fans are more efficient in terms of CFM/watt — I got back nothing useful.

Worse, when I gave it the “wrong ” question, and specifically worked it in terms of CFM/watt, damned if ChatGPT didn’t spit back the wrong answer, forcefully and clearly worded in terms of CFM/watt.

Worse yet, when I gave it a neutrally-worded question, it did in fact pick ceiling fans over box fans.  But the explanations were just so much nice-sounding nonsense.  It never did get down to the basic physics of it.

But even worse, when I gave it a different, simpler wording, of that neutrally-worded question (“Which is more efficient …”), it gave me the wrong answer.  And backed that up with nonsense reasoning.  (e.g., Box fans are more efficient in terms of CFM/watt because they are compact and portable.)  That’s the last fan question entry below.  So not only did it seem to give me either answer (ceiling is more efficient than box, box is more efficient than ceiling), but which answer I got appeared to depend on irrelevant and arbitrary changes in the wording of the question.

Finally, I gave it a couple of questions for which a) I have a well-researched and definitive answer, and b) for which I am a national expert.  Does the hospice benefit reduce Medicare spending?

The plain answer to that is no, but you have a whole lot of people who wish the answer were “yes”, and a whole lot of bad research that seems to give them that answer.  (Along with considerable research supporting the answer of “no”.) So when I asked ChatGPT, twice, what I got back was a highly selective reading of the scholarly literature, and the usual pat answers about why hospice saves Medicare money.

(It doesn’t.  Among other things, a third of Medicare hospice outlays are for those who are outside of their last year of life.  Other than my own study, typical research methodologies exclude these costs.)

My take on it at this point is that if you want bland filler, that does in fact say some stuff, about some defined topic, these writing AIs are fine.

If you want the contents to be factually correct?  Not so much.  The answer you get depends heavily on the question you asked.   Seemingly neutral changes to a neutral question got ChatGPT to flip its answer.  And for a separate question where I have expert knowledge, ChatGPT presented a selective reading of the literature that more-or-less mimicked the for-profit industry’s party line.

To me that says that you really can’t ask these writing AIs for reasoned analysis.  Or right and wrong.  At best, what you’re really asking them is one of two things:

  • Give me some arguments that support my point of view.
  • What does the majority of text, that has been input to this AI, say about a particular topic.

You may think you’re asking it to provide you with some true information.  But at root, what you’re actually doing is a lot closer to one of the two bullet points above.

This, by the way, dovetails with a friend’s report of AIs that had been trained to be racist, by groups of individuals who chatted with the AI using racist language.  The AI was just a reflection of what it had been fed.

If you are looking for any sort of deep explanation, you’re out of luck.  Even if you know what that deep explanation is.  In terms of explaining why ceiling fans are more efficient, none of these went much beyond the tautology of telling me that they are more efficient because they move more CFM/watt.  Which is, in this context, the definition of efficiency.

I wonder if this explains what I see on blog sites when I’m researching a topic in gardening.  On any given topic, a lot of sites will reproduce some bland version of the standard talking points.  Many of them don’t seem to have first-hand knowledge of the topic.  And for sure, almost none do what I do, and actually test things to see if they are true or not.  Maybe a lot of what I’ve been seeing is based on AI-generated text.  A lot of that text has the same bland, one-hand-or-other-hand style that seems to be the default style of ChatGPT.

In short, decent writing?  Yes.  Good sense?  No.

But, in hindsight, these are billed as writing AIs.  They write.  That’s what they do.  Plausibly, you could use them to generate multiple answers to a question, and you’d pick the one you liked.  But in no sense that I see can you trust them  them to answer a question, in terms of sorting fact from fiction.

The might get it right.  They might not.  They might waffle.  For sure, if you ask them the same question, enough ways, you might get back an answer that you like.

If the AI manages to hit the nail on the head, it provides you with cheap, well-composed text.  But it’s going to be up to you to judge whether the AI nailed it, or flubbed.

It’ll always give you an answer.  It might give you a very convincing answer.  And it might sometimes even give you the right answer.  But the last part is far from guaranteed. 

It boils down to this.  In trying to find an artificial intelligence to write my blog, I keep using “AI”, but I do not think “AI” means what you think it means.  For sure, AI writers don’t do what I had hoped they would do.

Here’s how it went.

Explain why ceiling fans are more efficient than box fans.

ChatGPT response, emphasis mine

Ceiling fans are generally considered more efficient than box fans due to several factors:
  1. Air circulation: Ceiling fans are mounted on the ceiling, allowing them to circulate air throughout the entire room. They create a downward airflow that creates a wind chill effect, making you feel cooler. This circulation helps maintain a consistent temperature in the room and reduces the reliance on air conditioning or other cooling systems. On the other hand, box fans are typically placed on the floor or a surface, and their airflow is limited to the immediate area in front of them.
  2. Coverage area: Ceiling fans have larger blades and are positioned overhead, allowing them to cover a larger area compared to box fans. They can effectively cool an entire room by distributing air more evenly. In contrast, box fans are smaller and have a limited range of coverage. They are better suited for localized cooling or for smaller spaces.
  3. Energy consumption: Ceiling fans are generally more energy-efficient compared to box fans. They consume less power to provide the same level of cooling due to their design and the way they circulate air. Ceiling fans typically use around 15-90 watts of electricity, while box fans can consume 50-200 watts depending on their size and settings. This lower energy consumption translates into cost savings and reduced environmental impact.
  4. Integration with HVAC systems: Ceiling fans can complement the existing HVAC (heating, ventilation, and air conditioning) system in a room. By using a ceiling fan in conjunction with an air conditioner, you can set the thermostat at a slightly higher temperature while still feeling comfortable due to the cooling effect of the fan. This can lead to significant energy savings by reducing the load on the air conditioning system. Box fans do not integrate with HVAC systems in the same way.
  5. Convenience and space-saving: Ceiling fans are fixed in place and do not require additional floor or surface space, making them more convenient and space-saving compared to box fans. They are also operated by a switch or remote control, allowing for easy adjustment of speed and direction. Box fans, on the other hand, need to be placed in a suitable location, and their portability can be a disadvantage in terms of convenience and aesthetics.
It's important to note that the efficiency and effectiveness of fans can vary based on the specific models, settings, and conditions. However, in general, ceiling fans tend to provide more efficient and widespread cooling compared to box fans.

Continue reading Post #1813: Won’t be hiring ChatGPT as my AI staff writer, either.