Windypundit

Classical liberalism, criminal laws, the war on drugs, economics, free speech, technology, photography, and whatever else comes to mind.

Main Menu

The FACES of Justice

by Leave a Comment

The ACLU and several other public interest groups have filed a brief with the Florida Supreme Court urging them to look into the use of automated facial recognition in criminal proceedings.

In 2015, two undercover cops purchased $50 of crack cocaine from a Black man on a Jacksonville street. Instead of arresting him on the spot, one officer used an old phone to snap several photos of him. Trying to be discrete, the officer took the photos while holding the phone to his ear and pretending to be on a call. Needless to say, the resulting photos were not headshot quality.

Later, after failing to identify the man themselves, the officers emailed the photos to a crime analyst, who used a statewide face recognition system to see if one of the photos looked like any mugshots in the county’s database. The program spit out several possible matches, the first of which was Willie Lynch, who was soon arrested and charged with the drug sale.

The ACLU argues that this is a problem:

If the government uses an error-prone face recognition system to identify you as the perpetrator of a crime, you have a constitutional right to probe its accuracy before you are convicted based on its results. But amazingly, a Florida appeals court disagrees.

If a cop stops you and makes you breath into an Intoxilyzer 9000, the results issued by that machine can be used in court as evidence of your blood alcohol content. On the other hand, if the cop whips out a cheap personal-use portable breathalyzer and you blow into it, the results aren’t admissible in court because the device hasn’t been approved for evidential use. (This is a simplification, I’m not a lawyer, don’t blame me if you get convicted, etc.) However, the cop can use the results of that non-evidential test to decide if he wants to go through the trouble of getting you to take an official test.

I can’t quite tell from the descriptions of the Florida case whether the FACES facial recognition system was used more like the Intoxilyzer 9000 or the cheap portable breath tester. That is, were the results of the facial recognition program entered into evidence — “We know it was the defendant because the facial recognition software says so” — or did the prosecutor have the cops make the ID — “We recognize this guy as the individual we bought the drugs from” — with the facial recognition software being the tool they used to find his identity (name, address, etc.) so they could visit him and identify him as the drug dealer themselves?

If the former, if facial recognition was used as evidence, then I think it would be pretty clear that the technology can be challenged by the defense, as breathalyzer evidence often is. Given that the county’s use of facial recognition software came as a surprise to many criminal defense lawyers in the area, I suspect that it was standard practice for the prosecutor to not reveal that law enforcement agencies were regularly using facial recognition technology in investigations.

Even if the facial recognition results were not introduced as evidence, the ACLU argues they should have been made available to the defense as Brady material:

If any of this information had come from a human witness—or, in the case of the analyst’s suggestive submission to the investigators, a lineup—it would clearly be Brady material. For example, FACES identified Mr. Lynch and several other people with similar confidence as the perpetrator. Had an eyewitness done so, the state would be unquestionably obligated to disclose the identification of the alternate suspects.

[citations omitted]

That makes a lot of sense to me. There is evidence that a witness’s ability to accurately identify a suspect by looking a photo is impaired if the photo is presented improperly. Showing only one photo is highly suggestive, and the image in the photo is likely to supplant the witness’s actual memory. So the defense needs to know the details of how the defendant was identified if they are to put on an effective defense.

That said, as a software engineer, I’m uncomfortable with how far the ACLU wants to take this:

Prosecutorial misconduct and police adoption of face recognition technology are dangerous, and the ACLU has been pushing to halt both. Until that happens, prosecutors must give defendants full access to information about the algorithm used against them in places where face recognition technology has already been deployed. This includes the underlying model, training data, computer code, explanatory documentation, and any other results from which the final, reported result was chosen. Any validation studies should also be available as well as the opportunity to question the people who use and created the system.

Where I come from, turning over the source code to a software product is huge. The intellectual property value of even a moderately-sized software system could run into the millions of dollars. And once the source code gets into a courtroom, it’s not unheard of for it to be released to the public:

In July 2016, Judge Valerie Caproni of the Southern District of New York determined in U.S. v. Johnson that the source code of the Forensic Statistical Tool, a genotyping software, “is ‘relevant … [and] admissible’” at least during a Daubert hearing—a pretrial hearing where the admissibility of expert testimony is challenged. Caproni provided a protective order at that time.

This week, Caproni lifted that order after the investigative journalism organization ProPublica filed a motion arguing that there was a public interest in the code.

If anyone wants to know what that source code looks like, it’s available here. As I write this, people have made 49 additional public copies. It’s copyrighted, of course, but I have no idea what the limits are on information released through the courts like this.

In addition, I’m not real thrilled about the idea that defense lawyers could routinely subpoena me every time some police agency uses something I wrote.

That said, I can certainly see where defense lawyers are coming from. When you’re fighting a DUI charge where an Intoxilyzer 5000 was used, it makes sense to want to know how an Intoxilyzer 5000 works, and that includes information about the source code for its software. That same thinking applies if your client was fingered by a FACES hit. It just kind of makes sense that software created for use by law enforcement should be subject to discovery and examination as part of any criminal proceeding resulting from its use.

But what about general purpose software? What if the forensic analyst used Photoshop to clean up the phone image before sending it to the facial recognition system? If lawyers can get the software for the the facial recognition system, can they also force Adobe to turn over Photoshop source code that cost them hundreds of millions of dollars to develop? That seems insane.

And what happens next time? Is there some sort of precedent established? Or does the software owner have to keep giving it up in court? And what about maintenance? I could see a defense lawyer demanding to examine the FACES source code, only to have the prosecution argue that the court had found the FACES software acceptable in a previous case, to which the defense lawyer responds, “That was version 3.5.4, this case is about version 3.6.2.”

I sort of assume that the legal system has some of this figured out already, because we’ve been using software systems for a long time, and the issue doesn’t seem to come up often. When prosecutors want call detail records for a phone admitted into evidence, the defense doesn’t usually get to examine the phone company’s billing software that was used to produce the record. Or in the case discussed above, nobody seems to be asking the cellphone manufacturer for source code for the software running on the phone that created the image.

In 2012, a defendant in a cold case was prosecuted on a DNA hit that was characterized as “1.62 quintillion times more probable than a coincidental match to an unrelated black person.” The defense wanted to make the state prove that those insanely high odds were accurate, so they requested the source code to the DNA matching algorithm.

The trial court determined that Chubbs was entitled to examine the source code under protective order.

This decision was overturned on appeal in 2015. The appeals court said that Chubbs’ stated reasons to access the source code, even under protective order, did not outweigh trade secret protections. Further, as the court writes, “access to TrueAllele’s source code is not necessary to judge the software’s reliability,” because validation studies and expert testimony are sufficient to make that determination.

That sounds like a reasonable conclusion. I’m a software engineer, and if you asked me to evaluate the accuracy of a piece of scientific software, I would be far more interested in the testing methodology than the source code. After all, the reason the software industry does so much testing is because you can’t judge the quality of software by code inspection alone.

Still, it’s not as if DNA software is perfect. I seem to recall (I can’t find a link) that at least one DNA matching program was found to have inaccurately coded population statistics, causing it to miscalculate the odds of a random match. That is exactly the sort of thing the defense team would like to discover, because it’s also exactly the sort of thing that gets innocent people convicted.

Green New Disaster

by 2 Comments

I’ve been meaning to write something about the “Green New Deal” for a while now, but it’s been such a vague concept that I wasn’t sure how to address it. Should I write about the letter to Congress signed by hundreds of environmental and liberal-leaning groups? Or maybe the Sunrise Movement’s grand strategy? It would be pointless to write about some fringe version of the movement, but I had no idea where the center was.

Thankfully, House Representative Alexandria Ocasio-Cortez has produced a reasonably fixed target in the form of a non-binding resolution. Well, it’s sort of a fixed target. You see there was this FAQ, and then it got taken down and disavowed by Ocasio-Cortez, even though it was on her own website, and now she’s claiming that some of the copies of the resolution floating around are drafts, or even fakes… So I’ve gone through this post and checked all the quotes against the version on Congress’s website, and I think I got everything right. (I have in some quoted sections modified the text of the resolution to clarify the numbering system.)

The Green New Deal is not the first proposal I’ve seen to combine a response to the climate crisis with a jobs program (and other liberal causes), so it’s about what I was expecting.

Resolved, That it is the sense of the House of Representatives that—

(1) it is the duty of the Federal Government to create a Green New Deal—

(A) to achieve net-zero greenhouse gas emissions through a fair and just transition for all communities and workers;

And a bunch of other things. There’s a schedule too, with a deadline based on climate predictions from last October’s IPCC report:

(2) the goals described in subparagraphs (A) through (E) of paragraph (1) (referred to in this resolution as the “Green New Deal goals”) should be accomplished through a 10-year national mobilization (referred to in this resolution as the “Green New Deal mobilization”) that will require the following goals and projects

I’m don’t have time to go over everything the GND proposes to accomplish in 10 years, but the main approach to stopping greenhouse gas emissions seems to be phasing out fossil fuels:

(2)(C) meeting 100 percent of the power demand in the United States through clean, renewable, and zero-emission energy sources, including—
(i) by dramatically expanding and upgrading renewable power sources; and
(ii) by deploying new capacity;

If that was the only goal of the Green New Deal, it would still be a very ambitious proposal. As of last year, the United States was generating about about 85,000 GWh of electrical power each year through solar power. The United States uses about 4 million GWh of electricity per year, so the Green New Deal requires us to increase solar power generation 47-fold in ten years.

To get an idea of the scale of the project, this handy table suggests that the best case 2-axis concentrated photovoltaic technology will require 2.8 acres of land per GWh of energy per year. A back-of-the-envelope calculation shows we’ll need about 17,500 square miles of solar power facilities — the size of New Hampshire and Massachusetts combined — to meet all of our current electrical needs.

That’s not completely impossible. We’ve already covered 6 times that much space with cities and towns, and we add almost 1600 square miles of urban area every year, so if we stopped all new urban construction and retasked everyone involved to building solar photovoltaic arrays, we could have the generating capacity finished in a little over a decade.

(Or maybe Ocasio-Cortez can get President Trump to back her plan by promising to build the solar power collectors in a 9-mile wide strip along the entire 2000 mile U.S.-Mexico border and call it the Trump Solar Wall. It sounds like something he would like.)

On the other hand, the Green New Deal seems to have something else in mind for our construction industry:

(2)(E) upgrading all existing buildings in the United States and building new buildings to achieve maximum energy efficiency, water efficiency, safety, affordability, comfort, and durability, including through electrification;

Megan McArdle at the Washington Post explains why this is nuts:

Homeowners tend to start renovations hoping to make their houses greener and more energy efficient, if only for reasons of parsimony. They quickly discover that what they lightheartedly imagined to be a few minor upgrades are in fact massive expenses. All they wanted was energy-efficient windows, better insulation, a tankless water heater, a radiant system to replace their (dry, noisy, inefficient) HVAC and . . . dear God, did the contractor misplace a decimal point? We’re updating a modest row house, not building Versailles!

Ceilings and walls must be removed, then replaced; fancy new equipment installed; ductwork rerouted; slabs perhaps broken and repoured; ancillary systems brought up to modern housing codes. Much labor is required; much homeowner patience and cash, too. And the older the building, the more expensive and disruptive the process, since decades of obsolete junk have to be ripped out, with all appropriate environmental remediation measures taken along the way.

Yet Ocasio-Cortez, or someone in her office, apparently thought those repairs could be made to every building in the United States within a decade.

As if that weren’t enough, the Green New Deal is includes massive changes to our transportation infrastructure:

(2)(H) overhauling transportation systems in the United States to remove pollution and greenhouse gas emissions from the transportation sector as much as is technologically feasible, including through investment in—
(i) zero-emission vehicle infrastructure and manufacturing;
(ii) clean, affordable, and accessible public transit; and
(iii) high-speed rail;

My gut reaction is that if we’re going to have zero-emission vehicles that don’t emit greenhouse gases, what is the benefit of having public transit and high-speed rail? You can’t get less than zero emissions. I guess the argument is that public transit and high speed rail will reduce the cost of converting the American vehicle stock to zero-emissions because we won’t need so many individual vehicles.

But building additional public transit in our crowded metropolitan regions is expensive, and high-speed rail is damned near impossible to build for anything like a reasonable cost. The closest thing we have to high-speed rail in this country, the Acela Express, takes 6.5 hours to make the 457 mile trip from Boston, MA to Washington, DC., which is not very fast by high-speed rail standards. The proposed bullet train in California, which is now projected to cost more than double its original estimates, appears to be in the process of being cancelled.

And if we’re going to use zero-emission vehicles, that means replacing internal combustion engines with electric motors, which will require an immense charging infrastructure and a lot more electrical generation capacity. The solar power collection facilities would have to be even more gigantic than those discussed above.

Meanwhile, even as we’re upgrading every building in the country, re-engineering our ground transportation systems, and building a solar collection array the size of Denmark, the Green New Deal would also make big changes to our agricultural industry:

(2)(G) working collaboratively with farmers and ranchers in the United States to remove pollution and greenhouse gas emissions from the agricultural sector as much as is technologically feasible, including—
(i) by supporting family farming;
(ii) by investing in sustainable farming and land use practices that increase soil health; and
(iii) by building a more sustainable food system that ensures universal access to healthy food;

This goes way beyond fighting global warming, to include support for family farming, improving soil health, and something called “access to healthy food,” which makes me suspicious. (Why not just “food”?) I have the feeling “healthy food” means food other than what people have shown they want to eat. Or maybe it means “organic” food. Combined with a move to small family farms, organic growing would reduce the efficiency of our farm system, requiring us to place more land in service for growing food.

Here are a couple more items:

(2)(D) building or upgrading to energy-efficient, distributed, and ‘‘smart’’ power grids, and working to ensure affordable access to electricity;

(2)(J) removing greenhouse gases from the atmosphere and reducing pollution, including by restoring natural ecosystems through proven low-tech solutions that increase soil carbon storage, such as preservation and afforestation;

I’m probably forgetting something, but so far Ocasio-Cortez wants to generate all electricity from renewable sources, upgrade the power grid, upgrade all buildings, replace most cars and trucks, build high-speed rail everywhere, make big changes to our agricultural industry, and plant big forests. All at once, all in about 10 years.

At least those projects are arguably somewhat related to energy efficiency and greenhouse gases. However, the Green New Deal goes on and on through a shopping list of liberal/progressive/leftist policies:

(1)(B) to create millions of good, high-wage jobs and ensure prosperity and economic security for all people of the United States;

(2)(K) restoring and protecting threatened, endangered, and fragile ecosystems through locally appropriate and science-based projects that enhance biodiversity and support climate resiliency;

(2)(L) cleaning up existing hazardous waste and abandoned sites, ensuring economic development and sustainability on those sites;

(4)(C) providing resources, training, and high-quality education, including higher education, to all people of the United States, with a focus on frontline and vulnerable communities, so that all people of the United States may be full and equal participants in the Green New Deal mobilization;

(4)(E) directing investments to spur economic development, deepen and diversify industry and business in local and regional economies, and build wealth and community ownership, while prioritizing high-quality job creation and economic, social, and environmental benefits in frontline and vulnerable communities, and deindustrialized communities, that may otherwise struggle with the transition away from greenhouse gas intensive industries;

(4)(F) ensuring the use of democratic and participatory processes that are inclusive of and led by frontline and vulnerable communities and workers to plan, implement, and administer the Green New Deal mobilization at the local level;

(4)(G) ensuring that the Green New Deal mobilization creates high-quality union jobs that pay prevailing wages, hires local workers, offers training and advancement opportunities, and guarantees wage and benefit parity for workers affected by the transition;

(4)(H) guaranteeing a job with a family-sustaining wage, adequate family and medical leave, paid vacations, and retirement security to all people of the United States;

(4)(I) strengthening and protecting the right of all workers to organize, unionize, and collectively bargain free of coercion, intimidation, and harassment;

(4)(J) strengthening and enforcing labor, workplace health and safety, antidiscrimination, and wage and hour standards across all employers, industries, and sectors;

(4)(K) enacting and enforcing trade rules, procurement standards, and border adjustments with strong labor and environmental protections—
(i) to stop the transfer of jobs and pollution overseas; and
(ii) to grow domestic manufacturing in the United States;

(4)(N) ensuring a commercial environment where every businessperson is free from unfair competition and domination by domestic or international monopolies; and

(4)(O) providing all people of the United States with—
(i) high-quality health care;
(ii) affordable, safe, and adequate housing;
(iii) economic security; and
(iv) access to clean water, clean air, healthy and affordable food, and nature.

My God, that list is amazing in its scope. You may or may not think that cleaning up hazardous waste sites, empowering unions, and providing family medical leave are good ideas, but they have nothing to do with climate change. Neither do universal higher education, community business ownership, or trade restrictions. (The resolution is so full of progressive wet dream policies that I’m kind of surprised no one slipped in a clause about abortion or Palestine.)

In 2010, Naomi Klein published The Shock Doctrine: The Rise of Disaster Capitalism, in which she argued that the evil advocates of free-market capitalism have been taking advantage of national disasters in places like Latin America and Eastern Europe to spread their ideology. She called this “disaster capitalism.”

I bring that up because the Green New Deal is basically disaster socialism: The authors are using the climate crisis to stir people to action, to justify a massive government response, but then they are cramming a crapload of unrelated policies and social programs into their proposed response.

This is, by their own standards, reprehensible.

The opening sections of the Green New Deal resolution emphasize the dire consequences of continuing global warming. I can’t vouch for all the details, but I certainly agree in principle: Global warming is a crisis on a planetary scale. We’re already seeing evidence of adverse effects, and if CO2 levels continue to rise while we do nothing, the resulting climate changes will affect the lives of billions of people for thousands of years.

The authors of the Green New Deal argue that solving this crisis demands “a new national, social, industrial, and economic mobilization on a scale not seen since World War II.” Global warming is as bad as a world war, they say, and it needs a world-war-style response.

So why does so much of their proposed have nothing to do with global warming? Why is their response larded full of liberal social goals? If they are sincere in their belief that global warming requires a massive response, why are they siphoning off vast resources for other projects? If the fight against global warming is a world war, then they are war profiteers. If we are in a war against global warming, then they are committing treason by impairing our fight against it.

Or perhaps they aren’t treasonous. Perhaps they are merely liars, exaggerating the dangers of global warming to manipulate the public. Perhaps Alexandria Ocasio-Cortez is spreading climate alarmism for the same reason Donald Trump spreads alarmist stories about immigrant caravans. Perhaps global warming is just a myth that leftists use to push their agenda.

I don’t actually believe global warming is a myth. But after reading the Green New Deal, I find it easier to be sympathetic with people who believe it is, because it must seem awful convenient to them that the biggest proposal to fight global warming…is a giant wish-list of progressive policies.

And I find myself wondering…does Alexandria Ocasio-Cortez believe in global warming? She seems quite happy to use it as an excuse for everything in the Green New Deal, but if she really believed global warming was a disaster, wouldn’t she have proposed a better solution?

Model Justice — Starting Over

by 2 Comments

It’s been a while since I wrote about my Model Justice project, which is an attempt to build a software model of certain aspects of the criminal justice system. In my last post on the subject, I explained why I was abandoning the Insight Maker systems modeling environment in favor of the Python general-purpose programming language. Since then, I’ve been slowly climbing the learning curve, and even though I haven’t quite caught up to the Insight Maker based implementation, I think I’m at a point where I can begin talking about it.

My goal with this Python rewrite is not to build a single monolithic criminal justice simulation program. As I’ve emphasized from the start, this project is about learning and exploration, and I haven’t yet learned enough to build an all-in-one solution like that. Rather, I’m writing a collection of Python components — classes in software terminology — that I can assemble and configure in various ways to build a variety of models.

For my purposes, the basic unit of simulation is the Case class, which represents a single criminal case, consisting of a single criminal charge against a single defendant, making its way through the criminal justice system. Each Case entering the system has two basic parameters which describe it:

  • prob_convict: A number between 0 and 1 inclusive, representing the probability of a conviction if the case goes to trial.
  • sentence_if_convicted: A number indicating the length of the sentence the judge will impose if the defendant is convicted by the trial. This is a unitless value, but I tend to think of it as months.

Cases are grouped together by the Docket class. Each Docket represents a group of cases that are being considered at the same time. The simulation creates one Docket for each iteration of the simulation by calling a DocketFactory class, which itself uses a CaseFactory class to create each of the Cases it puts in the Docket. The CaseFactory class, in turn, uses a random number generation class to generate the prob_convict value for each case. (The sentence_if_convicted value is currently hardcoded at 60 until I implement a better way to generate it.) The DocketFactory class itself uses a random number generator to decide how many Cases to put in each Docket it creates.

(I’ve included three random number generator classes, each providing a different distribution, any of which can be used by a factory class.)

Currently, the simulation is so simple that this grouping of Cases into Dockets has no effect on the outcome, but in the future I hope to use the Docket class to implement the court system’s resource limits and decision making horizon. For example, if there are only enough courtrooms to hold 20 trials per simulation cycle, but 30 cases are on the docket, then the simulated decision makers will have to resolve the remaining 10 cases by some other means, such as plea bargaining or dismissal.

It would have been simpler to have the main simulation code create a Case or a Docket whenever it needed one, but then I’d have to change the main code each time I wanted to run a different simulation with different parameters. Since I’d rather keep the core simulation code stable, I put the code for Case and Docket creation in corresponding CaseFactory and DocketFactory classes that the simulation can invoke when it needs a Case or a Factory. This allows me to change the parameters of the simulation by changing the factory classes without having to change any of the other code in the simulation. This is called a factory pattern, and it’s a good way to write flexible software.

At the top level of the simulation is the Experiment class, which actually runs the simulation engine. Constructing a new Experiment takes three parameters:

  • A DocketFactory object which it uses to create dockets and cases,
  • a PleaBargainingStrategy object which specifies how plea bargaining decisions are made, and
  • a Trial object which simulates the trial outcome.

I’ve already discussed the DocketFactory, but the other two classes are new.

The PleaBargainingStrategy object processes each Docket of cases and figures out which cases will settle with plea bargains and which ones will go on to trial. In the future, I hope to have subclasses that implement a reasonably plausible model of the plea bargain process, but for now all I have is PleaBargainingAtRandom, which implements a dumb algorithm that chooses what to do with each case entirely at random. E.g. if you set its prob_plea parameter to 0.9, the strategy will randomly generate a plea deal 90% of the time. The deal will always be for a sentence that is the mathematically expected length: If prob_convict on a case specifies a 30% chance of conviction, and the sentence if convicted is 60 months (which it always is right now), then the plea deal will be for a sentence of 0.3 * 60 = 18 months.

The Trial class is even simpler: It generates a uniform random number between 0 and 1, and if it’s less than prob_convict the defendant is found guilty and serves the full sentence. Otherwise the defendant is acquitted and does no time.

Those are the pieces that make up a simulation of how cases work through the criminal court system. To show how the pieces come together, let’s walk through a script that runs an experiment.

(If you’re a Python user, you can follow along by installing the crimjustsimpy package in your Python environment. If you are reading this some time after I published it, I may have released newer versions which work differently, so to get the package this post is talking about, make sure to specify version 1.0.5. If you’re interested, the source code for this project is also available on GitHub, and you can specifically view release 1.0.5.)

The sample script starts by defining some parameters at the top:

AVG_CASES_PER_INTERVAL = 100
MAX_CASES_PER_INTERVAL = 300

The number of cases to be generated for each docket is specified by a Poisson distribution, which is a common distribution used to estimate the arrival rate for random independent events, where each one has the same probability of occurring over a given time period. It’s commonly used to simulate things like orders arriving at a warehouse or cars arriving at a gas station. In this case, we’re assuming an average arrival rate of 100 cases per simulation interval.

Although large numbers (relative to the average) would be very rare, the Poisson distribution places no upper limit on the number of cases that can arrive during the interval. To avoid insane outliers, we’re limiting arrivals to a maximum of 300. That’s probably not necessary, because you can see in this histogram of a sample of 100,000 dockets shows, it never gets close to 300:

The next three lines describe the probability distribution from which probabilities of conviction will be drawn.

MIN_PROB_CONVICT = 0.05
MEAN_PROB_CONVICT = 0.4
MAX_PROB_CONVICT = 0.95

In my first attempt, I used a normal distribution with upper and lower limits. This time around, I’m using a beta distribution. I chose it because it has a nice shape, similar to the normal curve, but it also generates values in a strict range of 0 to 1 without the need to impose external cutoffs. As you can see, the configured range only runs from 0.05 to 0.95, so neither a conviction nor an acquittal is ever a certainty. The beta variate is internally scaled to fit withing these limits. Moreover, the beta distribution appears to be a good choice for representing a probability distribution from which to draw probabilities, which is exactly what I’m using it for. 

The resulting histogram of conviction probabilities looks like this:

The next parameter just tells the simulation what probability to use when deciding whether a case will end in a plea deal. This example uses 80%:

PROB_PLEA=0.8

Finally, we specify the number of iterations over which the model will cycle.

ITERATIONS = 240

After that, the script begins constructing and assembling the parts of the experiment. First, it creates a random number generator object that uses the Beta distribution to generate the conviction probabilities for the cases:

convict_gen = cj.RandomScaledBetaProb(
    shape=10.0,
    lower=MIN_PROB_CONVICT,
    middle=MEAN_PROB_CONVICT,
    upper=MAX_PROB_CONVICT)

RandomScaledBetaProb is a wrapper around the Python library function that generates the Beta distribution, specifying the bounds and the mean of the generated values. The shape parameter controls how tightly the numbers cluster around the mean or stretch out to the edges. I just picked 10 because it gives a nice looking curve. It should probably be a parameter in the future.

The next line of the script creates a CaseFactory from the conviction probability generator:

case_factory = cj.CaseFactory(convict_gen=convict_gen)

(If you’re unfamiliar with Python named parameters, that probably looks a little confusing. In “convict_gen=convict_gen“, the first “convict_gen” is the name of the CaseFactory parameter that receives the conviction probability generator, and the second “convict_gen” is a reference to the variable that received the conviction probability generator created in the previous code snippet. This is a common coding convention.)

The next two lines do the same thing for the DocketFactory class. First create the object that generates the arrival rates for the dockets over the Poisson distribution, then create the DocketFactory itself, passing it the CaseFactory we created above and the arrival rate generator created here.

arrival_gen = cj.RandomPoissonBounded(
    mean=AVG_CASES_PER_INTERVAL,
    upper=MAX_CASES_PER_INTERVAL)
docket_factory = cj.DocketFactory(
    case_factory=case_factory,
    arrival_gen=arrival_gen)

The plea bargaining strategy is pretty simple, because it only takes the single value representing the probability of handling the case with a plea deal.

plea_bargaining = cj.PleaBargainingAtRandom(prob_plea=PROB_PLEA)

The trial is even simpler:

trial = cj.Trial()

Now we can assemble all these pieces to build an Experiment object:

experiment = cj.Experiment(
    docket_factory=docket_factory,
    trial=trial,
    plea_bargaining=plea_bargaining)

That brings us to the meat of the simulation: Running the experiment.

data = experiment.run(ITERATIONS)

It’s just one line, but it does all the work of creating the dockets filled with cases, resolving some of them with plea deals, resolving the rest with trial, and gathering up the results. With only 240 dockets and and 100 cases per docket, this runs in less than a second on my computer. In the future, with more cases and more complicated resolution logic, it will probably run a lot longer.

The results of the simulation run are returned in an ExperimentData object that contains all the Docket and Case objects with the results of all the cases. There’s nothing to stop us from traversing the Case objects directly, but there’s a better way:

df = data.casesToDataFrame()

This converts the Case data to a DataFrame, which is a highly-optimized standard Python data structure for handling tabular data. This will make it easier to process the data in the future, because lots of Python code packages understand how to work with DataFrames. For example, the PyCharm development environment I use can render DataFrames as tables:

This shows the output from a simulation run for each case. The fields are defined as follows:

  • id: A unique integer ID assigned to each case.
  • docketId: A unique integer ID indicating the docket the case belonged to.
  • pConvict: The probability of conviction of the case when created.
  • maxSentence: The maximum possible sentence if convicted.
  • plead: True if the case was handled with a plea deal.
  • tried: True if the case was decided with a trial.
  • acquitted: True if the defendant was acquitted at trial. Always False if the case didn’t go to trial.
  • convicted: True if the defendant was convicted at trial. Always False if the case didn’t go to trial.
  • guilty: True if the defendant was convicted at trial or if the defendant took a plea deal.
  • sentence: Actual resulting sentence. Always 0 if the defendant was acquitted.

We could plot this data using standard Python graphics packages such as matplotlib and seaborn, but the crimjustsimpy package also includes a Visualization class that collects a handful of convenience methods for creating plots from the DataFrame. For example, this pie chart showing the ratio of cases that plead out vs. going to trial:

This is how the trials turned out:

And this is a breakdown showing all three possible case results: Acquitted at trial, convicted at trial, and plead out:

None of that is particularly surprising. We set up the simulation to take a plea deal 80% of the time, and it did. We set it up for 40% of trials to result in convictions, and they did. We’re just confirming that the simulation ran the way we wanted it to.

It’s slightly more interesting to plot a histogram of the lengths of the resulting sentences for all the cases:

That looks a little weird, but it makes sense. The hump in the middle is showing the sentences from the plea deals, all of which are for the maximum sentence weighted by the probability of conviction, so they follow the same beta distribution hump as the probability of conviction. The bars at the ends are all the cases that went to trial, because they ended with either no sentence or the maximum sentence. The difference between the plea hump and the conviction bar on the right is what defense lawyers call the “trial tax.” It’s the difference between what happens when a defendant admits guilt, and what happens when he makes the government prove it.

That’s about all I’ve got done so far, other than some unit tests and a few sample scripts, which leaves me a lot more to do.

The previous version of this model used an agent-based systems model that tracked the changing state of the cases over time as they worked their way through the system. That was interesting to see, and it gave me the simulated steady-state size of the prison population, which seems like an important metric for studying criminal justice.

I’d like to put that output capability back into the model, but I’m not sure if I should keep a running track of the state of each case like last time, or if I should just run a post-pass to compute time series data for the prison population from the case results.

I’d also like to be able to see intermediate results from the simulation, either by having it generate usable output as it runs, or by being able to run it for a few iterations and then stop to examine what it’s doing. This suggests that keeping a running state will work better than a post-pass analysis.

In addition, if I can stop and start the simulation, I’d also like to be able to change the parameters while it’s stopped, so I can model changes to the system. This suggests I need to break down the Experiment class into at least two parts: The simulation engine, which can be modified along the way, and the simulation data, which persists throughout the simulation run.

Of course, none of that helps me with what I really want to look at, which is the plea bargaining decision. I have some ideas about that, but this post is long enough.

As always, any suggestions would be appreciated.

State of the Union 2019

by Leave a Comment

I skipped reviewing the State of the Union speech last year, so I thought I’d give it a try this year, since I have to be awake for something else anyway.

The official White House transcript wasn’t up yet when I started this, so I’m using the Vox rush transcript of President Trump’s notes. It’s a little messy, so I’ve cleaned it up a little here and there and reformatted it for readability. I also removed some of the applause notes, since they only get in the way..

Thank you very much madam speaker, vice president, members of Congress, the first lady of the United States, and my fellow Americans. We meet tonight at a moment of unlimited potential, as we begin a new Congress, I stand here ready to work with you to achieve historic breakthroughs for all Americans. Millions of our fellow citizens are watching us, gathered in this great chamber, hoping we will govern not as two parties, but as one nation.

And so begins the usual Call to Unity. The people in power always like to call for unity, because “unity” sounds like a good thing. It’s a positive note. But what they always mean is unity behind their agenda.

[…] This year, America will recognize two important anniversaries that show us the majesty of America’s vision and the power of American pride.

In June we marked 75 years since the start of what general Dwight Eisenhower called “the great crusade,” the allied liberation of Europe in world war II. On d-day, June 6, 1944, 15,000 young American men jumped from the skies, and 60,000 more stormed in from the sea to save our civilization from tyranny.

At this point, he introduces three people who were involved in that action. I’ve always found the presence of special guests to be cheap trick to bring drama to a tedious speech. I’m pretty sure we have Bill Clinton to blame for this. In any case, I’m going to skip the guests.

An amazing quality of life for all of our citizens is within reach. We can make our communities safer, our families stronger, our culture richer, our faith deeper, and our middle class and more prosperous than ever before. But we must reject the politics of revenge, resistance and retribution, and embrace the boundless potential of cooperation, compromise, and the common good.

Am I crazy, or is “revenge, resistance and retribution” a reference to the Mueller investigation and they way his opponents call themselves “the resistance”? Weird thing to put in the State of the Union speech if it is…

Together we can break decades of political stalemate. We can bridge all divisions, heal old wounds, build new coalitions, forge new solutions, and unlock the extraordinary promise of America’s future. The decision is ours to make. We must choose between greatness or gridlock, results or resistance, vision or vengeance, incredible progress or pointless destruction. Tonight, I ask you to choose greatness.

Yeah, no. Like “unity” calls for “greatness” always sound like a good thing, but “national greatness” has been a Republican talking point for decades. It’s their go-to justification for giant, expensive government initiatives. I’ll take gridlock please.

Over the last two years, my administration has moved with a urgency and historic speed to confront problems neglected by leaders of both parties over many decades. In just over two years since the election, we have launched an unprecedented economic boom, a boom that has rarely been seen before. There has been nothing like it. We have created 5.3 million new jobs, and importantly, added 600,000 new manufacturing jobs, something which almost everyone said was impossible to do, but the fact is, we are just getting started.

The boom has been going on for close to 10 years. Obama claimed credit for it, and now Trump is claiming credit for it. But it’s mostly the recovery from the Great Recession that hit at the end of the Bush presidency.

We aren’t just getting started. We are wrapping up. The recovery is about over, unemployment is at its lowest in decades, and even labor force participation seems to be recovering. I wouldn’t be surprised to see slower growth and a bit of a downturn in some economic numbers as the economy runs out of room for the boom. (For the record, that wouldn’t be Trump’s fault any more than the recovery is.)

Wages are rising at the fastest pace in decades, and growing for blue-collar workers, who I promise to fight for. They are growing faster than anyone else. Nearly 5 million Americans have been lifted off food stamps. The U.S. Economy is growing almost twice as fast today as when I took office, and we are considered far and away the hottest economy anywhere in the world. Not even close. Unemployment has reached the lowest rate in over half a century. African-American, Hispanic American, and Asian-American unemployment have all reached their lowest levels ever recorded.

Remember all this stuff about how great the economy is. It will come up later.

[…] And to give critically ill patients access to life-saving cures, we passed very importantly, Right to Try.

I probably ought to give Trump and the Republicans credit for Right to Try. I think letting critically ill people try experimental therapies is a good idea, providing they are properly informed of the risks. But the devil is in the details, and I just don’t know enough about the details of this legislation to be sure.

My administration has cut more regulations in a short period of time than any other administration during its entire tenure. companies are coming back to our country in large numbers, thanks to our historic reductions in taxes and regulations. And we have unleashed a revolution in American energy. The United States is now the number one producer of oil and natural gas anywhere in the world. And now, for the first time in 65 years, we are a net exporter of energy.

This is another thing that Trump has inherited. The trend towards increasing energy independence has been going on for at least a decade, driven by technology and the economics of fossil fuels.

After 24 months of rapid progress, our economy is the envy of the world. Our military is the most powerful on Earth, by far, and America — America is again winning each and every day. Members of Congress, the state of our union is strong.

Is Trump taking credit for American military might? That’s something else he’s inherited. The United States has been the undisputed military heavyweight champion since the old Soviet Union collapsed. And we were probably #1 even when they were at the top of their game.

Our country is vibrant and our economy is thriving like never before. Friday it was announced we added another 304,000 jobs last month alone, almost double the number expected. An economic miracle is taking place in the United States, and the only thing that can stop it are foolish wars, politics, or ridiculous, partisan investigations. If there is going to be peace in legislation, there cannot be war and investigation. It just does not work that way.

Yeah, he’s saying the Mueller investigation is a threat to American greatness…

[…] And just weeks ago, all parties united for groundbreaking criminal justice reform. They said it couldn’t be done.

Kudos for criminal justice reform. It wasn’t a lot, but it wasn’t nothing.

Last year, I heard through friends, the story of Alice Johnson. I was deeply moved. In 1997, Alice was sentenced to life in prison as a first time, nonviolent drug offender. Over the next 22 years, she became a prison minister, inspiring others to choose a better path. She had a big impact on that prison population and far beyond. Alice’s story underscores the disparities and unfairness that can exist in criminal sentencing, and the need to remedy this total injustice. She served almost that 22 years, and had expected to be in prison for the remainder of her life. In June, I commuted Alice’s sentence.

Kudos again. And Kudos to Kim Kardashian West for being one of Trump’s friends who brought this to his attention. Seriously.

Inspired by stories like Alice’s, my administration worked closely with members of both parties to sign the First Step Act into law. Big deal. That is a big deal.

This legislation reformed sentencing laws that have wrongly and disproportionately harmed the African-American community. The first step act gives nonviolent offenders the chance to reenter society as productive, law-abiding citizens. Now, states across the country are following our lead. America is a nation that believes in redemption.

That was a nice bit. And now we come to the alarmism part of the speech…

Now, Republicans and Democrats must join forces again to confront an urgent national crisis. Congress has 10 days left to pass a bill that will fund our government, protect our homeland, and secure our very dangerous southern border. Now is the time for Congress to show the world that America is committed to ending illegal immigration and putting the ruthless coyotes, cartels, drug dealers, and human traffickers out of business.

As we speak, large, organized caravans are on the March to the United States. We have just heard that Mexican cities, in order to remove the illegal immigrants from their communities, are getting trucks and buses to bring them up to our country in areas where there is little border protection. I have ordered another 3750 troops to our southern border to prepare for this tremendous onslaught.

Oh my God. “Tremendous onslaught!” We’re talking about, what? Maybe ten thousand people total in these caravans? Something like a million people cross our borders every day. We have individual office buildings that hold as many people as are in these caravans. Seven times that many people were at the Superbowl. This is not a threat.

I want people to come into our country in the largest numbers ever,

— this is not shown by anything else you’ve ever said or done —

Meanwhile, working-class Americans are left to pay the price for mass illegal immigration, reduced jobs, lower wages, overburdened schools, hospitals that are so crowded you can’t get in, increased crime, and a depleted social safety net.

Almost none of that is true.

One in three women is sexually assaulted on the long journey north. Smugglers use migrant children as human pawns to exploit our laws and gain access to our country. Human traffickers and sex traffickers take advantage of the wide-open areas between our ports of entry to smuggle thousands of young girls and women into the United States and to sell them into prostitution and modern-day slavery.

All of this happens because our immigration laws make it so hard for people to get in legally. (Except for the sex trafficking, which almost never happens in real life.) Make immigration simple and easy, and these problems will go away.

Tens of thousands of innocent Americans are killed by lethal drugs that cross our border and flood into our cities, including meth, heroin, cocaine, and fentanyl.

This has nothing to do with immigration. It’s just how prohibition works. Force a product onto the black market and quality control suffers, often with deadly results. Also, those drugs cross the border because Americans want them. Nobody is forcing Americans to use drugs.

The savage gang MS 13 now operates in at least 20 different American states and they almost all come through our southern border.

Mara Salvatrucha actually started here in the United States and spread to other countries in the Americas later, but go ahead…

Just yesterday, an MS-13 gang member was taken into custody for a fatal shooting on a subway platform in New York City. We are removing these gang members by the thousands, but until we secure our border, they are going to keep streaming right back in. Year after year, countless Americans are murdered by criminal illegal aliens.

Year after year, Americans are murdered by other Americans too. Basically, if you live in the U.S., you’re really only likely to be murdered by someone else who lives in the U.S., which includes illegal immigrants and gang members. There’s no evidence that immigrants (legal or not) are more murderous than Americans in general.

We are joined tonight by one of those law enforcement heroes, ice special agent Elvin Hernandez. […] Thanks to his work work and that his incredible colleagues, more than 300 women and girls have been rescued from the horror of this terrible situation and more than 1500 sadistic traffickers have been put behind bars.

There’s no way those numbers are right. I don’t have sources handy, but 300 is probably more than half the total number of actual sex trafficking cases filed in the entire U.S. in a year. I doubt he could find that many real trafficking victims even if he worked for a decade. Even the massive multi-agency National Johns Suppression sting operations sometimes come up completely empty when it comes to human trafficking. I suspect someone is counting as human trafficking things that aren’t really human trafficking.

This is a smart, strategic, see-through steel barrier — not just a simple concrete wall. It will be deployed in the areas identified by border agents as having the greatest need, and as these agents will tell you, where walls go up, illegal crossings go way down. San Diego used to have the most illegal border crossings in the country. In response, a strong security wall was put in place. This powerful barrier almost completely ended illegal crossings. The border city of El Paso, Texas, used to have extremely high rates of violent crime — one of the highest in the country, and considered one of our nation’s most dangerous cities. Now, immediately upon its building, with a powerful barrier in place, El Paso is one of our safest cities. Simply put, walls work and walls save lives.

Walls work locally. If you put up a wall in an area where people have been walking through, they will stop walking through in that area. They can just cross someplace else. For under $500, they can fly to Canada and cross into the United States across the longest undefended border in the world. Or they can enter the U.S. on a tourist visa and never leave.

(I’m pretty sure El Paso was always a fairly safe city, but I’m too lazy to look it up.)

(Also, how is that steel-slat fence supposed to stop drug smuggling? I’m no expert, but those gaps look large enough to pass bricks of heroin between them.)

To build on our incredible economic success, one priority is paramount — reversing decades of calamitous trade policies. So bad. We are now making it clear to China that after years of targeting our industries, and stealing our intellectual property, the theft of American jobs and wealth has come to an end.

They aren’t stealing our jobs and wealth. They are trading with us. We buy their stuff because we want to. Saying that China is stealing from us is saying we’re too stupid to know what’s good for us.

Therefore, we recently imposed tariffs on $250 billion of Chinese goods

Which will almost certainly result in higher costs to Americans. You’ve raised taxes on Americans.

[…] Our new U.S.-Mexico-Canada agreement — or USMCA — will replace NAFTA and deliver for American workers like they have not had delivered to for a long time. I hope you can pass the USMCA into law so we can bring back our manufacturing jobs in greater numbers, expand American agriculture, protect intellectual property, and ensure that more cars are proudly stamped with four beautiful words — made in the U.S.A.

Why? Seriously, just a few minutes ago, Trump was taking credit for how great the economy was. Why do we need more jobs — especially manufacturing and agricultural jobs, both of which are highly automated — when we’re all working already?

Tonight, I am also asking you to pass the United States reciprocal trade act, so that if another country places an unfair tariff on an American product, we can charge them the exact same tariff on the same product that they sell to us.

I’m not sure what the Reciprocal Trade Act does, but I’m pretty sure countries don’t import the same products that they export. That’s not how trade works.

Both parties should be able to unite for a great rebuilding of America’s crumbling infrastructure. I know that the Congress is eager to pass an infrastructure bill — and I am eager to work with you on legislation to deliver new and important infrastructure investment, including investments in the cutting edge industries of the future.

Repairing infrastructure is something we need to do. On the other hand, I don’t trust the government to be able to recognize “cutting edge industries of the future,” and I don’t want them playing their investment hunches with my tax dollars.

We should also require drug companies, insurance companies, and hospitals to disclose real prices to foster competition and bring costs down.

That would be a good idea. But…the only reason healthcare providers get away with not providing prices is because the market is so broken that the people buying these goods and services don’t care about the prices. Requiring providers to better disclose prices won’t fix the market failure.

Under my administration, we will never apologize for advancing America’s interests. For example, decades ago the United States entered into a treaty with Russia in which we agreed to limit and reduce our missile capabilities. While we followed the agreement and the rules to the letter, Russia repeatedly violated its terms. It has been going on for many years. That is why I announced that the United States is officially withdrawing from the intermediate-range nuclear forces treaty, or inf treaty.

We really have no choice. Perhaps we can negotiate a different agreement, adding China and others, or perhaps we can’t – in which case, we will outspend and out-innovate all others by far.

In other words, we will start a nuclear arms race. Sigh.

As part of a bold new diplomacy, we continue our historic push for peace on the Korean peninsula. Our hostages have come home, nuclear testing has stopped, and there has not been a missile launch in 15 months. If I had not been elected president of the United States, we would right now, in my opinion, be in a major war with North Korea.

This is total fantasy bullshit.

I’m going to skip over all the stuff about our troops overseas because I don’t know enough about foreign policy in those areas to say anything useful.

That brings us to the close:

What will we do with this moment? How will we be remembered? I ask the men and women of this Congress, look at the opportunities before us. Our most thrilling achievements are still ahead. Our most exciting journeys still await. Our biggest victories are still to come. We have not yet begun to dream.

We must choose whether we are defined by our differences — or whether we dare to transcend them. We must choose whether we will squander our inheritance — or whether we will proudly declare that we are Americans. We do the incredible. We defy the impossible. We conquer the unknown. This is the time to re-ignite the American imagination. This is the time to search for the tallest summit, and set our sights on the brightest star.

This is the time to rekindle the bonds of love and loyalty and memory that link us together as citizens, as neighbors, as patriots. This is our future — our fate — and our choice to make. I am asking you to choose greatness.

Please no. When powerful national leaders set their sights on achieving their grand dreams of greatness, the rest of us suffer — always in our treasures, sometimes with our blood.

Racist Algorithms

by 1 Comment

So the other day, according to someone named Ryan Saavedra, newly elected congresswoman Alexandria Ocasio-Cortez said that algorithms are racist:

Socialist Rep. Alexandria Ocasio-Cortez (D-NY) claims that algorithms, which are driven by math, are racist

— Ryan Saavedra @RealSaavedra

Ocasio-Cortez says a lot of crazy things, but this is not as crazy as it sounds. Obviously, algorithms cannot be racist in the way that humans are racist — they don’t have feelings about people of different races — but there are applications of algorithms that can fairly be described as racially biased.

The word “algorithm” has a traditional meaning in computer science: It’s the rules a program must follow to manipulate data to solve a problem. These are usually very concrete, expressed in code (or pseudocode) in exacting detail. For example, here’s an algorithm I threw together in the C# programming language to solve the problem of finding all the prime numbers in a range from lo to hi:

private static List<int> GetPrimes(int lo, int hi)
{
    var primes = new List<int> {2};
    for (int i = 3; i <= hi; i+=2) {
        var maxFactor = Math.Sqrt(i);
        foreach (int factor in primes) {
            if (factor > maxFactor) break;
            if (i % factor == 0) {
                goto not_prime;
            }
        }
        primes.Add(i);
        not_prime:;
    }
    return primes.Where(p => lo <= p && p <= hi).ToList();
}

You can use this algorithm to find out that the range 100 to 200 contains the prime numbers 101, 103, 107, 109, 113, 127, 131, 137, 139, 149, 151, 157, 163, 167, 173, 179, 181, 191, 193, 197, and 199. If you want to give it a test drive you can run it on .NET Fiddle. (The algorithm is not designed to be efficient for really big integer inputs, so it’s probably best not to set lo and hi greater than about 10 million or the algorithm might run too long.)

This is where Saavedra sort of has a point, because this is not a racist algorithm, and there’s no way it could be. But that’s not because it’s “driven by math.” This algorithm can’t be racist because racism is about people, and this algorithm is not about people.

Once you add people into the mix, it’s not hard to create a racist algorithm. For example, banks use algorithms to calculate a score indicating whether they should loan someone money for a home mortgage. It would be easy to modify the algorithm to, say, reduce an applicant’s score by 10% if they’re black.

You could probably make a pedantic argument that “racist” is the wrong word to describe this algorithm because the algorithm doesn’t feel hate toward black people, but we can safely say that it prejudges black people and puts them at a disadvantage. I think we can fairly call that racist.

In this day and age, no U.S. bank would do anything so blatantly discriminatory. But that doesn’t mean the algorithm couldn’t still function in a racist manner.

Consider the basic question the bank is trying to answer: If we loan the applicant our money, will they pay it back? (There’s more to loan analysis than that, but it’s certainly one of the key questions.) That’s a difficult question to answer for one very obvious reason: It’s about the future, and the future hasn’t happened yet. Whether someone will pay back a loan is essentially unknowable at the time you make the decision to lend them money. But there are a lot of clues.

For example, you can look at the borrower’s history of making payments on other loans (i.e. their credit history). Unlike their future payments, their history of past payments is available to us. And it makes sense that someone who has paid their debts in the past will continue to pay their debts in the future. Their past loan payment patterns predict their future loan payment patterns.

A person’s income is another good predictor of future debt payments, because a person’s capacity to make loan payments is constrained by their wealth and income. The source of their income could be important too, because some jobs are more at risk of extended unemployment than others, and it doesn’t matter how good their credit history is, or how much they earn on the job, if they could lose their job and not find another quickly.

Furthermore, since these are mortgage loans, there will be a lot of information about the property that secures the loan, such as the value of the house, the size of the house, how many bedrooms it has, construction details, and where it’s located. These could all turn out to be statistically significant predictors of future loan repayment.

To build an algorithm that takes these things into account, data scientists would begin by gathering a massive database of historic loan data, including all the information I’ve been describing plus information about how well the loan performed (was it paid back, were payments late, etc.). They would then try to build a mathematical model that predicts loan performance based on what was known about the loan at the time of the application.

The goal is to find a model that does a good job at predicting loan performance for loans in the database, and then apply it to new loans to try to predict how they will do. This is how we can make predictions about whether people will pay back loans in the unknowable future, using data that that we have right now in the present.

Statisticians have been doing things like this for years — fitting models to data and then using the models to make predictions. (In some sense, every expansion of scientific knowledge takes this form.) In recent years, this has expanded into into the fields of data science, predictive analytics, and machine learning, and there’s a whole world of software tools to make the job easier.

Getting back to our mortgage loan model, a few potential sources of racial bias have already begun to creep in. To start with, take a look at this map of where people of different races live in my hometown of Chicago:

(You can see a scrollable map like this for the whole United States at the Cooper Center demographics research group at Virginia University.)

The map shows that people of the same race tend to clump together. So if you know where a house is located, you have a pretty good idea of the race of the people living there. This might be a problem for our loan payment model if (1) it predicts that loans on homes in certain locations have higher default rates, and (2) those locations tend to be occupied by racial minorities.

In data science terms, the home’s location would then be serving as a proxy for the race of the homeowner. Even though we have deliberately chosen not to incorporate the borrower’s race in our loan prediction model, we’ve incorporated another variable that correlates to race, which has the same effect. We have, perhaps quite unintentionally, introduced race as a factor in our loan scoring.

Some of the other model variables may also be race proxies, such as the amount and source of income. Black households have, on average, lower incomes than white households, so by denying loans to people with low incomes, banks are disproportionately denying loans to black people. And certain jobs are heavily segregated. So if a loan applicant lists their employment as “veterinarian,” there’s a 95% chance they’re white. A loan approval process that favors veterinarians may also favor white people.

On the other hand, banks are in the loan business because they want to make money, so not making profitable loans because the borrowers are black would be a business mistake. In fact, most studies of loan default rates find that blacks and whites are equally likely to default on a home loan, which implies that banks are not holding blacks to a higher standard or giving out loans to less qualified white people.

That we can check the accuracy of the loan approval algorithm against real world results is very important, because it allows us to correct errors. That’s possible because the loan default rate is negative feedback, which means it can be used to push the algorithm back in the right direction. So if the default rate is too high for, say, mortgages on townhouses, the bank can tune the algorithm to be more strict about approving loans to buy townhouses.

Unfortunately, not all models are so self-correcting. For example, a law enforcement agency’s predictive policing algorithm might notice that crime is rising in a particular neighborhood and recommend that more police officers be assigned there. If the crimes monitored by the algorithm are violent ones like homicide and assault, the additional police may help reduce them.

But if the crimes being monitored include nuisance crimes such as vagrancy, public drinking, or panhandling (perhaps under a “broken windows” theory of policing), it’s likely that the majority of these crimes have been missed or ignored by the police, meaning the the reported rate is only a small fraction of the true rate. In that case, there are plenty of unreported crimes to discover, so if police departments put more effort into it, they will discover more crimes.

This creates a pernicious positive feedback loop. It becomes possible for a few racist cops to over-aggressively arrest a lot of people in predominantly black neighborhoods, producing an apparent increase in the crime rate for those neighborhoods. That could trigger the predictive policing algorithm to assign more officers into the area, where those officers will make even more arrests for even more nuisance crimes. That will drive up the reported crime rate again, which could cause the predictive policing algorithm to assign even more police, and so on. In this way, positive feedback can cause an algorithm to capture and amplify the racism of the people in the system.

These effects can go on to affect other algorithms. For example, an algorithm for a recidivism model — such as those used for sentencing and parole decisions — may take into account the likelihood that number of arrests a person has on their record, or even the number of people in that person’s neighborhood who are criminals, both of which could artificially inflated by runaway positive feedback in the predictive policing algorithm. Those arrests could also get plugged into analytics tools for evaluating job applications, leading to rejections of people with criminal records.

Speaking of job applications, it can be useful to score applicants by where they live, since people with longer commutes are more likely to quit sooner. That sounds like a neutral decision, unless the company is located in a predominantly white area, in which case the geographic scoring will have the effect of preferring white people.

These days, machine learning is all the rage. So if a company has a historic database of 100,000 job applications which lead to 10,000 people being hired, a data scientist might feed that database to a machine learning system in the hope that it will generate an algorithmic model that will mimic the hiring decisions of the company’s managers. If the algorithm can narrow down the applicant field by rejecting 80,000 applications before any human sees them, that will speed up the hiring process a lot. Of course, if the hiring managers were racially biased, the machine learning model could learn to mimic their bias.

This sort of thing happens a lot, and professional data scientist Cathy O’Neil wrote a pretty decent book called Weapons of Math Destruction about the ways which big data modeling can go wrong, including some of the ways models can capture and replicate racial bias. We can make philosophical arguments about whether this is really racist, but it’s not how we want our algorithms to work.

css.php