My sister called from her trusted car repair shop. Her car needed a new transmission and fuel pump. She had a ‘95 Ford Escort; major problems like that meant her car was toast. It had been troublesome for months, and was now truly dead.
The car’s demise had left my sister, her husband, and their 2 year old daughter without transport. Worse, they had a 20+ mile commute to work, didn’t have time off, and would be fired if they couldn’t get to work at a moment’s notice.
I had less than 72 hours to find a replacement car. I spoke with my sister about what features she cared about.
“Operating cost” was the cost to run the car each year: repairs, insurance, and gas.
The goal became a reliable, fuel-efficient car. For under $5,000.
I had done research into how to buy a car using data. An hour of sleuthing on Craigslist and AutoTrader revealed that vehicles this cheap are 9+ years old and have 100K+ miles on them. Many seemed of dubious reliability.
Unfortunately there was no way to know the reliability of a car from its description. That suggested there were both ripoffs and deals in the listings. This was an information asymmetry problem. The seller had close to perfect knowledge and the buyer had very little.
Internet sleuthing led me to FleetBusiness, which reported how long different brands last before they die (are junked). I also found TrueDelta, which had reports from car owners about what car repairs, mileage and cost.
I quickly graphed the FleetBusinss data showing scrap rates of different brands of car. I conflated age with mileage/wear; I couldn’t avoid the confound given the time.
The car brand that died off quickest was Suzuki. The brands that died off slowest were Toyota, Honda and Subaru. The die-off rate was not a straight line…it was an S-shape, like the continuous normal distribution. Looking at the scrap rate per year, I saw a roughly normal distribution:
I also saw that cars died off en masse at 10-20 years old. The cars I was looking at were the worst possible age. The odds were good the car I purchased would die within the next 5-10 years.
I didn’t care about total car death rates; I cared about how long a car would last if I purchased it at a certain age. The cars I was looking at purchasing were 10-13 years old. Any cars that died before that age were not for sale so I could exclude that percentage.
I subtracted the 10-year death rate from each year’s death rate:
The most reliable brands to buy at 10 years’ age were Honda and Toyota followed by Chrysler. I picked 6 reliable models:
I had added 2 Hyundai models, the Elantra and Sonata, because I had heard their later-generation models were well built. This was not data-driven and stupid.
I started gathering data from car listings (for-sale ads). My goal was to have enough listings that I could be confident there were a few good deals. I excluded any listing that didn’t have the car’s year, model, mileage, and price listed. After a few hours I had 117 cars listed.
The biggest cost of owning a car is depreciation: the difference between your purchase price and what you sell it for. The best way to reduce that cost is to buy it as cheaply as possible and keep it running for as long as possible.
I didn’t really care about how many miles a car had or its age. I wanted a car with as many miles remaining as possible. I needed to find out how long each car model would last. If you buy a car with 125K miles on it, there’s a big difference between a car that will last 200K miles and 150K miles. The 200K car will go 300% as far.
A car’s mileage is more important its age. For the same price, which is a better car?
The 8-year-old car is a better choice. I guessed that mileage was roughly 5X more important than age. I knew maintenance costs would increase exponentially as mileage and age increased. I puzzled out an equation to compute a ‘reliability score’ for each car.
Score = fnNormalize ( Age ^ 1.2 ) * 20% + fnNormalize ( Mileage ^ 1.4 ) * 80%
The ratio of this score to the price is the ‘value score’. Cars with higher value scores were better deals:
I saw that, roughly, better-quality cars were more expensive. However, it was a general trend and not a clearly linear relationship. There were ripoffs (in the upper left, with smaller points) and great potential deals (in the lower right, with larger points).
Now I had a shopping list: the 5 cars with the highest value scores. I quickly called the top 3 car owners to ask if the cars were still on sale and available for a test drive. After a nap.
I met with my sister and family for lunch. Afterwards we went to see the #3 car at a nearby dealership. We went prepared with a to-check list.
The test drive was illuminating: the car was junk. The brakes barely worked, the fan belt made a loud whistling sound, the window seals were crumbling and the lowest gear didn’t work…in an automatic. We left in a hurry.
The owner of the first car on my list called back, saying that the car had just sold. Darn.
The owner of the car #2 called: the car was still for sale! I guessed the car would sell quickly and arranged for a test drive that evening.
I called car owner #4 on my list, leaving another message. I wasn’t confident car #2 would work out so I wanted to keep looking. After another break.
I wasn’t hopeful of finding a good car after that first test drive, and was surprised when this car handled well. The engine, brakes, steering, and lights all worked perfectly. A roller-coaster route through West Seattle found no issues.
The owner had copies of their maintenance logs for 3 years and seemed trustworthy. We quickly made plans for my trusted mechanic to look over the car the next day. I would buy the car if it passed a mechanical inspection.
A Google search of the seller’s name, email address and phone number confirmed her identity. Her name matched her online photo and she had enough of an Internet presence to be legit. This felt vaguely creepy, but also prudent given the money involved.
I arranged for an afternoon test-drive with car #5. I explained that I had a better lead that may not pan out; the seller was very understanding.
The mechanic confirmed car #2 was in good working condition except the it burned some oil when accelerating. Some hasty Internet searches suggested this was not unusual for old Toyota Corollas and didn’t mean the engine was toast. Success!
My sister and her family arrived at the auto repair shop, we signed all the paperwork, paid the seller, and went our separate ways.
Schools produce data. Thus far we have looked at school quality, and its relationship to house prices. What else affects the quality of school? School size? Parents’ education? Teacher salaries? Gender ratios? Teacher ratings?
Can we use this knowledge to make better decisions?
First, look at data. A common challenge in data analysis is knowing the appropriate methods to use. There are thousands of statistical and machine learning algorithms, and none of them are appropriate for all situations. Let’s see some density plots of different variables:
What do we see? High schools appear to come in two different size ranges. The cost of living has a positive skew. School budgets are relatively similar.
None of these attributes follow a ‘normal’ (Gaussian) distribution of values. The most common statistical methods are therefore inappropriate: mean, standard deviation, and Pearson’s correlation.
Let’s see which variables correlate well with High Achiever %. We can’t use Pearson’s correlation here, so we’ll use Spearman’s Rho instead.
Some variables have a strong correlation, such as parents’ education levels. Others don’t, such as the age of the school neighborhood.
We must remember that correlation does not imply causation. However, the lack of a correlation does imply a lack of causation. If two variables are not correlated then there a causal relationship is unlikely, if not impossible.
Let’s use a combination of factors to predict the quality of school. We’ll use a machine learning algorithm, Random Forests, an ensemble method of decision trees with good behavior for this class of problem.
A few lines of code produce a model to predict school quality. Let’s see how closely the predicted (green) values compare against the actual (red) values:
Good, but not perfect. It is impressive that we can predict school quality using only demographic and budget data. After all, we don’t have any information about teacher skill, different teaching techniques, or the students themselves.
Let’s plot the residual (actual - predicted), values, in black, against the predicted numbers, in green. The residual numbers contain variation that is not explained by demographics or budgets. They suggest that unobserved variables are involved, like teacher skill or student aptitude.
Analysis thus far suggests school quality is influenced by demographics and budgets. These are factors a school cannot control.
Let’s use that to our advantage. Groups of schools with similar demographics will still have different levels of quality. They can learn from each other, since schools in similar situations have done well.
Let’s group schools together using another algorithm, k-means. It belongs to a set of machine learning techniques for clustering.
The k-means algorithm has a limitation: it groups together objects by first computing the Euclidean distance between each. Euclidean distance treats each variable as equally important, which is not true for this example. We will compensate for this by ‘weighting’ each variable according to its importance, which the random forest algorithm has already calculated.
We can then assign a group to each school, clustering together similar schools into the same group.
Let’s look at the High Achiever % of each group.
As we can see, each group has some spread of values. This is encouraging, because these schools could learn from each other. High schools of similar size, with similar parental situations and similar demographics are likely to have relevant advice for each other.
This leads me to a triumvirate of conclusions:
Schools/Educators: You aren’t alone. There are many other schools just like yours, doing better and worse. Ask for help in finding them. Then learn from them, and teach in turn. Identify best and worst practices specific to your situation. Spread that information around. The more schools that participate, the more everyone benefits.
Parents: Teachers and principals have limits to their ability to educate children, and it depends on many things. You have a much more profound impact on your children’s education than their teachers do. You can’t dodge that responsibility, so don’t try. And for goodness sake, be careful about what data you pay attention to. Good test and SAT scores don’t predict whether your child will be happy or successful in 10 years.
Analysts/Data Scientists: Like all professionals, we have a choice of what to do with our talent. Each of us has access to huge amounts of data. We can use our skill and energy to make the world a better place. I encourage everyone to try.Permalink