[ View menu ]

November 19, 2014

Do NYC cab drivers quit too early when it rains?

Filed in Ideas ,Research News
Subscribe to Decision Science News by Email (one email per week, easy unsubscribe)

A BIG DATA ANALYSIS WITH NEW YORK CITY CAB DATA

nycc

Students of judgment and decision making (aka behavioral economics) are familiar with the idea that cab drivers work until they hit an income target and then quit, ignoring opportunities to make more money on especially profitable days, such as when it rains. They probably get the idea from hearing people talk about the paper Labor Supply of New York City Cabdrivers: One Day at a Time. We say “hearing people talk about the paper” because that paper does not say that cab drivers make more money when it rains and is otherwise quite cautious (e.g., it concludes “because evidence of negative labor supply responses to transitory wage changes is so much at odds with conventional economic wisdom, these results should be treated with caution.”)

A working paper and presentation by Princeton economist Henry Farber looked at recently released data on every cab ride taken in New York City from 2009-2013 (about 900 million trips, which he down-samples). Some of Farber’s conclusions are:

  • Increase in demand and reduction in supply make it difficult to find a taxi in the rain.
  • But wage is no higher when it rains.
  • Lower supply is not the result of drivers stopping after having reached their target.
  • Lower supply is result of less pleasant driving in the rain.

ABSTRACT

In a seminal paper, Camerer, Babcock, Loewenstein, and Thaler (1997) find that the wage elasticity of daily hours of work New York City (NYC) taxi drivers is negative and conclude that their labor supply behavior is consistent with target earning (having reference dependent preferences). I replicate and extend the CBLT analysis using data from all trips taken in all taxi cabs in NYC for the five years from 2009-2013. The overall pattern in my data is clear: drivers tend to respond positively to unanticipated as well as anticipated increases in earnings opportunities. This is consistent with the neoclassical optimizing model of labor supply and does not support the reference dependent preferences model.

I explore heterogeneity across drivers in their labor supply elasticities and consider whether new drivers differ from more experienced drivers in their behavior. I find substantial heterogeneity across drivers in their elasticities, but the estimated elasticities are generally positive and only rarely substantially negative. I also find that new drivers with smaller elasticities are more likely to exit the industry while drivers who remain learn quickly to be better optimizers (have positive labor supply elasticities that grow with experience).

REFERENCE
Farber, Henry S. (2014). Why You Can’t Find a Taxi in the Rain and Other Labor Supply Lessons from Cab Drivers. National Bureau of Economic Research Working Paper Series No. 20604 http://www.nber.org/papers/w20604

H/T Eric Jaffe http://www.citylab.com/weather/2014/10/why-new-yorkers-cant-find-a-taxi-when-it-rains/381652/
Photo credit: https://www.flickr.com/photos/chrisschoenbohm/6186211082/

November 14, 2014

What size will you be after you lose weight?

Filed in Ideas ,R ,Research News
Subscribe to Decision Science News by Email (one email per week, easy unsubscribe)

REDDITORS’ BEFORE AND AFTER MEASUREMENTS ANALYZED

reddit_before_after

Click to enlarge

How many pounds do you need to lose in order to reduce your waistline by one inch? How many kilos do you need to lose to reduce your waistline by one centimeter?

We wanted to find out. We were having trouble finding published data (though we are expecting some soon), so we turned to Reddit, where the progresspics subreddit contains people’s before-and-after weight change stories. Most posts contain only pictures, but if you do some web scraping, you can find cases in which people post their before-and-after waist measurements.

We found 46 such cases, typed them up, ran them through R, tidyr, dplyr, and ggplot and made the picture above.

Multiple regression tells us that on average, for every 8.5 pounds lost, people dropped an inch off their waist. (And for every 1.5 kilograms lost, people dropped a centimeter off their waist.)

Every 10 pounds lost was accompanied by 1.18 inches of waistline reduction. (Every 5 kg lost was accompanied by 3.33 cm of waistline reduction.)

The picture is a bit rosier for those who were losing smaller amounts (under 55 pounds or 25 kg): They only had to lose 6.1 pounds to lose an inch (or 1.1 kg to lose a centimeter).

Want to see the data split out by gender? Voila:

reddit_before_after_gender

Click to enlarge

Want to make this graph yourself? OK.

Why am I doing this? Hal are following up our face morphing stuff with body morphing stuff.

November 6, 2014

When to fly to get there on time? Six million flights analyzed.

Filed in Encyclopedia ,Ideas ,R
Subscribe to Decision Science News by Email (one email per week, easy unsubscribe)

EVERY U.S. FLIGHT IN 2013 ANALYZED

Flight_Delays_By_Hour_DelayType
(click to enlarge)

If you read Decision Science News, you are probably interested in decision making, you probably fly a lot, and you probably like making decisions about flying.

Data of the type the U.S. Government provides enable us to predict how delayed we will be when we fly at various hours of the day.

To make the plot above, we analyzed every single flight in the United States in 2013 for which there were Bureau of Transportation Statistics data. Filtering out flights between midnight and 6AM that leaves us with a little over six million flights (6,283,085 flights, to be precise). The BTS defines delay as the difference between the time the plane actually arrived and the time listed in the computerized reservation system. Many flights got in early, but because we’re just interested in delays (not speedups), we negative delays with zeroes.

What do we learn?

The later you leave, the greater the average delay you will face until around 6PM when things flatten out and 10PM when we see benefits in leaving later. It makes sense that delays increase as the day goes on because, we understand, the primary cause of delays is waiting for the plane to arrive from another city. The first flights out in the morning don’t have this problem.

About 60% of flights had no delay at all (3,726,061/6,283,085 or 59.3% to be precise). This has something to do with padding the expected arrival times in the computerized reservation system. Hence all the “negative” delays.

Leaving at 11PM gives you the same delay as leaving at 11AM. Miracle of miracles. Want a rule of thumb? Try not to leave between 11AM and 11PM.

The arrival and departure curves are quite similar. To save space, we’ll only look at departure delays from here on.

Now, you may be thinking “20 minutes delay if you depart at the worst possible time? That’s not such a big deal.” But remember, these are averages and 60% of the time there will be zero delay. To show you how bad things can get, here we plot the 95th and 75th percentiles of the delay distribution:

Flight_Delays_By_Hour_95thIf you leave at the worst time of day,  1 time in 4 you’ll be delayed more than 20 minutes, and 1 time in 20  you’ll be delayed more than an hour and a half!

Do different airports have differing delay patterns? One might expect them to due to weather, total number of flights, longitude and the like. We isolate the ten airports with the most passenger traffic below:

Flight_Delays_By_Hour_Airport_Top5

Flight_Delays_By_Hour_Airport_6to10

In an early analysis, we thought we’d discovered something pretty cool about day of the week effects. We had chosen two months at random and noticed certain days were predictably worse than others. But then, when we looked at two different months, different days emerged as the worst ones. Digging deeper, we found that the day-0f-week effects are attributable mostly to rather random events which change from month to month. Here we look at median (not mean) delays on every day of 2013. Each panel represents one month.

Flight_Delays_By_Departure_Date

The big spike on April 18, 2013? Five inches of rain in Chicago. December 9th, 2013? Delays are mostly due to winter weather in Texas. These little bumps can really alter the day-of-week findings.

Bon voyage!

R-code, as usual, for those who want it. To get the flight data, just go to … aw heck, I’ll be nice and let you download my cleaned up copy (25 Mb)

This is our first use of Hadley Wickham’s tidyr package. We like it!

ADDENDUM

1. We just learned of some extensive analyses pre-2009 flight data you might find interesting. See the FlowingData blog post. The supplemental information in this paper has some interesting analysis of flight delays. For example, hub airports tend to have a lot of outbound delays because they hold planes when an incoming flight is late. This leads to a lot of arrival delays at non-hub airports. See wicklin-supplemental.pdf page 7.

2. Poking around at this link, we were above to find somewhat steady day of week patterns in this poster which draws on multi-year data.

October 29, 2014

Getting old in baseball

Filed in Ideas ,R
Subscribe to Decision Science News by Email (one email per week, easy unsubscribe)

BATTING AND AGING

BA_by_Age_smooth

With baseball’s World Series drawing to a close, we thought we’d get in one last 2014 post on the US national pastime.

Keeping up with our aging theme, we’ll look at what happens to players’ batting averages as they age. We use the Lahman package in R, which has data from 1871 to 2013. We take the set of players who played in the majors for at least two years and look at the mean batting average at every age.

The green line (above, with smoothed plots, below with raw results with standard error bars) shows this basic result. Pro baseball players have their highest averages just over age 30. The area of the circles is proportional to the number of observations in that point.

When you look at results like those in the green line, however, you must stop to consider that the players who show up in the graph only tell part of the story. At a given age, there were other players who are not plotted because they were cut from the team years before (often due to their poor batting performance).

To illustrate this, at each age, I plot in the blue line the batting average of players who are in their last year of major league play. As one would expect, batting averages are low the year before players disappear from the major leagues. In the red line, we see the performance at each age of players who are not in their last year. For this subset of the data, peak batting average occurs at age 36 and the maximum is a bit flatter.

What is up with the increase in the blue line? The increasing trend is present even if you exclude the first two unusually low points. We are no experts on baseball (or sports of any kind) and are open to suggestions.

One thing to keep in mind is that people whose last year was at age 20 probably only played 2 years (I only considered players who played at least 2 years), while people whose last year was age 40 probably played about 20 years.

BA_by_Age

As usual, those who want to reproduce this in R are welcome to do so.

October 21, 2014

The October 2014 SJDM Newsletter is ready for download

Filed in SJDM ,SJDM-Conferences
Subscribe to Decision Science News by Email (one email per week, easy unsubscribe)

SOCIETY FOR JUDGMENT AND DECISION MAKING NEWSLETTER

 

The quarterly Society for Judgment and Decision Making newsletter is ready for download from the SJDM site.

http://sjdm.org/newsletters/

This issue includes the 2014 conference program.

While we have your attention

  • Don’t forget to vote in the 2014 SJDM elections. Polls close on October 26.
  • Don’t forget to register for the conference, which takes place November 21-24th in Long Beach, CA! Information at http://sjdm.org

Enjoy!

Dan Goldstein
Your Decision Science News / SJDM Newsletter Editor

October 15, 2014

Rules of thumb to predict how long you will live

Filed in Encyclopedia ,Ideas ,R
Subscribe to Decision Science News by Email (one email per week, easy unsubscribe)

HEURISTICS FOR ESTIMATING LIFE EXPECTANCY

linear

Life expectancy: The predicted average number of years remaining in a person’s life.

We once learned from a doctor a rule of thumb for predicting how much longer a person will live (i.e., the life expectancy). The doctor’s heuristic was:

(100 minus the patient’s age) divided by 2

We wanted to see how accurate this rule was, so we downloaded life expectancy data from the US government and plotted the model’s predictions against the official estimates of life expectancy. See above. The black dots are the government’s calculated life expectancies at every age from 30 to 110. The doctor’s model is in blue. It’s pretty good in the 65 to 95 age range.  The doctor worked in a nursing home. The heuristic fit the environment.That said, the doctor’s rule lousy outside that age range. And of course it assumes people will die by 100.

The doctor’s heuristic is a simple linear model. How well does simple linear regression do? We solved for it and plotted it in red above. We’ll see later how they compare in error, but it’s safe to say they’re both pretty lousy.

Let’s fit some better models. Forget survival models. Too hard for mortals to apply. Looking at the life expectancy curve, it seems like a polynomial and a two-part linear function would do a good job. They do.

fitted_nonlinear

However, our goal is to get something that someone could do in their head. Something like the doctor’s heuristic, but smarter.

We came up with two candidates.

1) The heuristic bi-linear model. We made this by making the best bi-linear model a bit simpler to apply.
It goes:

If you’re under 85, your life expectancy is 72 minus 80% of your age.
Otherwise it’s 22 minus 20% of your age

2) The 50-15-5 model. This one asks you to remember some key values and then to interpolate between those values. It goes:

The life expectancies of 30, 70, 90 and 110 year olds are about 50, 15, 5, and 0.
Go forth and interpolate!

Here is the performance of the heuristic models:

heuristic_nonlinear

Pretty awesome!

It’s not a horserace without some measure of accuracy. Below we plot the mean absolute deviation for all the models. Except for the linear models, the heuristics make estimates that are off by less than one year on average. That said, one needs to understand that one’s life expectancy is just the best guess, but there’s a lot of variation around that best guess. The 90% confidence interval around my estimated age at death spans roughly 40 years!

mad_predictions

The lesson is, linear fits to life expectancy are bad. Everything else we tested was pretty good. The heuristic bi-linear is especially easy to remember and do in your head.

Can you come up with better heuristics? Here’s some R code to see if you can:

For fun, check out this sweet plot of how the mean absolute error changes as you vary the cut point in the best bi-linear model. The cut point is 30+”cut” in the graph, so we cut around age 80.

ABSDEV_minimizing

ADDENDUM: Dean Foster just stopped by my desk and promoted

October 7, 2014

Part 2 of Who We Are: Society for Judgment and Decision Making (SJDM)

Filed in R ,SJDM
Subscribe to Decision Science News by Email (one email per week, easy unsubscribe)

WHERE THE SJDM MEMBERS ARE FROM IN THE WORLD

us.member.map.s

Last week, we looked at where the Society for Judgment and Decision Making members were from in terms of academic areas. This week we look at where they’re from geographically.

To start, we note that most members (1195/1714 or 70%) are from the USA.

USAvRest.s

Inside the USA, there are members in 45 states, as seen at the top of the page. The states with more than 5 members are these:

MembersByState.s2

Outside the US, we have members in North America, Europe, Asia, Oceania, South America and Africa.

MembersByRegion.s

And those regions comprise 39 countries. Here are the countries with more than 5 members.

MembersByCountry.s

Now, you’re probably wanting to reproduce these graphs. Or, if you’re like most people, you aren’t.

You’ll need directory.csv.gz, state_table.csv.gz, and regions.csv.gz. And the code below.

H/T to @winston_chang’s R Graphics Cookbook, from which I borrowed a code snippet or two.

October 1, 2014

Society for Judgment and Decision Making: Who Are We (Part 1)

Filed in R ,SJDM
Subscribe to Decision Science News by Email (one email per week, easy unsubscribe)

ANALYSIS OF SJDM MEMBERS’ DEPARTMENTS

MembersByArea.s2

Over the next two posts, we’ll look at the membership of SJDM, the Society for Judgment and Decision Making.

Next week, we’ll break down the membership by region, country, and state.

This week, we look at where our members come from in terms of departments. To make the plot above, we had to do a fair amount of recoding, so results are approximate.

R code is below. You can download some anonymized member information here. Save it to your R working directory to proceed.

R CODE TO REPRODUCE THIS FIGURE

September 17, 2014

Nudge yourself thin

Filed in Books
Subscribe to Decision Science News by Email (one email per week, easy unsubscribe)

SLIM BY DESIGN: BRIAN WANSINK’S NEW BOOK

sbd

Brian Wansink’s new book Slim by Design: Mindless Eating Solutions for Everyday Life is out and is on one of our favorite topics: Nudges for eating better.

We just ordered one!

September 11, 2014

What makes a good academic conference?

Filed in Ideas ,R ,SJDM-Conferences
Subscribe to Decision Science News by Email (one email per week, easy unsubscribe)

WHAT WE LOOK FOR WHEN WE DECIDE

cities.c.s

We have attended hundreds of academic conferences over the years.

We’ve seen all kinds of formats.

We’ve heard many opinions about what’s good and bad.

Here’s what we look for in an academic conference.

  • Small – Fewer than 400 attendees is ideal. At large conferences, people get overwhelmed, retreat into their cliques and tend not to meet anyone new. At smaller conferences, people get to know the other attendees. After attending a small conference for a few years, you get to know a substantial proportion of the attendees. Of course, our choosing to attend small conferences only makes them bigger, but we can’t be bothered to think that many steps ahead.
  • Few parallel sessions – We look for conferences that have at most 3 parallel sessions. If you sit in the back of the room at a big conference and look on, you’ll notice that people spend amazing amount of  time looking through the program trying to decide what to go to next.  Big conferences suffer from the cable TV problem: 57 channels and nothing on. With three tracks of quality content, people will spend less time choosing and will tend to choose well even if they choose arbitrarily (like following the person you were talking to at the coffee break).
  • Selective – Selectivity predicts quality. Life’s too short to sit through bad talks. We find that conferences with acceptance rates under 30% tend to be better. Organizers should use poster sessions to enable people to come even if their presentations are not accepted.
  • Short talks – With short talks, those who are interested in the research can always get more info later, and those who aren’t interested in the research don’t have to suffer. We like the 20-20-3 model: 20-minute talks (at maximum), 20 talks per track (at maximum), and three tracks (at maximum).
  • Plenary and Presidential addresses take priority – If deciding whether to accept more talks or have more plenary / presidential addresses, go for the later. The big talks often have the big ideas, and give all the attendees a common experience to talk about during the conference. Having fewer talk slots will increase the average quality of the talks you do accept and will help keep the conference small.
  • No presentations during meals – We don’t talk when others have the stage, but we understand why other people do. Conference are rare opportunities to connect with far-away friends and collaborators. It’s kind of cruel not to let people catch up during meals. Presidents deserve a separate session to get their ideas across. They should not have to compete with the clanking of forks and people talking. In banquet halls, half the audience can’t see the speaker anyway.
  • Better yet, more meals on your own – Letting people choose where they eat keeps conference costs down and allows each attendee to spend according to his or her preferences. It’s also nice to see more of a city than just a convention hotel.
  • Avoid social events that people can’t walk out of – H/T Eric Johnson. Attendees face a lot of constraints. Some are dead tired from international travel. Others need to juggle catching up with family and various groups of people in the city. The social event shouldn’t keep the attendees captive.
  • Centrality – Keep in mind that the location of the conference matters when people decide to submit. Don’t forget the Europeans.

Those are our preferences. We’re open for suggestions!

By the way, registration is now open for the 2014 SJDM conference, which will be held November 21-24 in Long Beach. Early-registration (through Nov. 10) is $225 for members, $260for non-members, and $100 for students. Information on the conference can be found at www.sjdm.org — to register, visit www.sjdm.org/join.html.

Our favorite conference, the Society for Judgment and Decision Making (SJDM) Conference, is coming up soon. Our other favorite conference Behavioral Decision in Research Management (BDRM), just went down splendidly in London over the summer. We often hear people say that the SJDM and BDRM conferences are better than other psychology, decision science, consumer behavior, marketing, policy, and behavioral economics conferences. Perhaps this is due to homophily–SJDM is our favorite conference and we tend to hang out with similar others. But perhaps it is due to characteristics. We don’t have data for BDRM, but we can talk about the JDM conference. Here’s what we observe.

  • JDM is small - We have put together some data on JDM attendance over the years (below). Until last year, the conference had fewer than 500 attendees. Pretty sure BDRM has always been similarly small. The Membership panel shows the number of people in the society. Attendance picked up ahead of membership in 2003.  Perhaps this has something to do with Kahneman’s Nobel Prize in 2002? Correlation is not causation. We simply remember that around 2005, every grad student in Marketing started saying that they were “doing JDM”.

multi.s
Click to Enlarge

 

  • JDM has few parallel sessions – We went from two tracks to three in 1999. I will fight to keep it that way.
  • JDM is selective – JDM routinely has acceptance rates of less than 30%, despite being a rather specialized and self-selected group of researchers. It also tends to favor people who were not accepted in the previous year. And it has a limit on the number of submissions on which someone can list themselves as a presenter. This keeps people from dominating the program.
  • JDM has short talks – 20 minutes is the norm.
  • JDM usually has separate Presidential and Plenary talks – Dan Ariely’s talk was the first we remember in which the President didn’t have to compete with a meal.
  • JDM has meals on your own – This makes JDM an incredible value for the money. Over the last 5 years, full price registration was $200 on average, and student registration was $98! (Exclamation, not factorial.)
  • JDM social events tend to be those people can walk out of  – And they tend to be pretty simple. Music, dancing, drinks, done.
  • JDM and location – Because JDM follows the Psychonomics Conference around, it tends to be located in major hubs that are easy to get to. The decision as to whether to break with Psychonomics comes up from time to time, but we always decide to stay. I think it’s a good move. It gives attendees the chance to attend the decision-making sessions at Psychonomics, and it saves the society the decision-making costs of figuring out where to have the conference. Yep, real decision experts consider the decision costs. Psychonomics has a policy of moving like a pendulum across the USA: East, Central, West, Central, East, …. The location does seem to affect attendance. See below and the graph at the top of this post.  We think that the high attendance in Canada was because it was easier for many Europeans to get to. The estimated changes in probabilities of attending aren’t that great, so it’s probably not too bad to keep following Psychonomics around.

 

attendance.c.s

TO DO YOU OWN ANALYSES IN R