Categorization and coordination

Key Question Being Explored in This Insight:

Vessela Daskalova provides a methodology to determine whether decision-makers may benefit from employing fewer categories, even without exogenous costs.

View the page of: Vessela Daskalova

Institution: University College Dublin

Explore more insights in:

Behavioral economics Economics Social Sciences

Media Literacy Toolkit

Faculti's Media Literacy Toolkit helps viewers critically engage with academic insights by analyzing the research context, identifying perspectives, and encouraging thoughtful evaluation.

Critical Questions to Consider ▼

What assumptions does the research make?
Are there alternative perspectives not explored?
What are the limitations of the research method?

Bias and Perspective Awareness ▼

This research comes from University College Dublin . Reflect on how the institution's academic focus and research partnerships may shape the questions being explored.

Further Reading ▼

Read the Study

Vessela Daskalova provides a methodology to determine whether decision-makers may benefit from employing fewer categories, even without exogenous costs.

Image courtesy of interviewee. January 26, 2024

Log-in or Sign-up to Faculti
Currently viewing this subject insight as a guest. You have insight(s) remaining for this month. Login to view 8000+ figures on the platform.

Generating Workspace...

Export as PDF

The paper titled "Categorization and Coordination," co-authored by the speaker and Nick Green from Queen Mary University of London, explores how individuals utilize mental categories to make predictions based on past experiences. It examines the trade-offs between coarse and fine categorization and investigates when each may be beneficial. Examples include employers categorizing applicants and equity researchers categorizing firms.

The authors argue that categorization is a decision-making process influenced by previous encounters, and individuals can employ either coarse or fine categories, impacting their predictions based on the category averages. The study emphasizes that while fine categorization may minimize bias in deterministic environments, coarse categorization can be advantageous in stochastic environments, where noise and sample size play significant roles.

Key findings reveal that coarse categorization may help in minimizing prediction errors when coordinating with others, especially in noisy contexts. The paper discusses three scenarios: individual prediction, coordination of predictions, and a combination of both, employing a model that incorporates game theory and statistical analysis. Through a bias-variance decomposition, the authors identify factors influencing the effectiveness of categorization and highlight the importance of the environment's characteristics.

Ultimately, the paper suggests that while coarse categorization has been linked to biases, it can also arise rationally in certain contexts and may provide valuable insights into decision-making biases in various fields.

Language:

Export as PDF

X-TIMESTAMP-MAP:LOCAL=00:00:00.000,MPEGTS=900000

00:00.265 --> 00:03.645
The paper that I'm going to talk about is called,

00:03.645 --> 00:05.285
categorization and Coordination

00:05.905 --> 00:07.725
and it's joint work,

00:07.755 --> 00:10.685
with my co-author Nick Green from Queen Mary

00:10.735 --> 00:11.805
University of London.

00:12.705 --> 00:16.685
So, what we do in this paper is we are interested in the

00:16.705 --> 00:20.045
use of mental categories to make predictions.

00:20.985 --> 00:23.245
And so you could think of many examples

00:23.825 --> 00:25.925
of people using mental categories.

00:26.385 --> 00:31.085
So it could be employers categorizing applicants based

00:31.305 --> 00:34.125
on their field of study where they went to university,

00:34.355 --> 00:36.805
perhaps their gender, race and so on.

00:37.465 --> 00:41.005
Or you could think of equity researchers,

00:41.325 --> 00:45.885
categorizing firms, based on the sector

00:46.035 --> 00:49.405
that they operate in based on their size,

00:49.625 --> 00:51.165
and other characteristics.

00:51.485 --> 00:54.045
Likewise, customers categorizing hotels.

00:54.185 --> 00:55.925
So there are multiple examples

00:56.145 --> 00:58.445
of people categorizing out there,

00:58.745 --> 01:01.365
and that's why we are interested in categorization.

01:02.105 --> 01:04.405
But, well, what does it mean to categorize?

01:04.985 --> 01:07.805
So, categorization is a form

01:08.185 --> 01:11.045
of decision making on the pa on the basis

01:11.105 --> 01:12.285
of past experiences.

01:12.865 --> 01:15.445
So the idea is,

01:15.665 --> 01:18.605
an individual will have accumulated a set

01:18.605 --> 01:22.445
of past experiences and they store them into categories.

01:23.305 --> 01:26.405
And, when encountering a new situation

01:26.835 --> 01:30.205
that resembles this past experiences based on its observable

01:30.205 --> 01:33.205
characteristics, they make a prediction based on the

01:33.605 --> 01:35.285
category average for this category.

01:35.985 --> 01:38.405
So this is what categorization is.

01:38.905 --> 01:43.165
Um, what we look at this at in this paper is the idea

01:43.355 --> 01:47.565
that in any situation there will be many different possible

01:47.565 --> 01:48.565
ways to categorize.

01:49.105 --> 01:53.525
Um, and what we focus on in particular is the idea

01:53.915 --> 01:58.765
that individuals may be using a few course categories.

01:59.225 --> 02:02.965
So basically putting their past experiences in

02:02.965 --> 02:04.245
just a few categories.

02:04.245 --> 02:06.405
And this will mean that they will be,

02:07.205 --> 02:10.245
possibly many experiences in each one of these categories

02:10.985 --> 02:13.925
or individuals could sort their past experiences

02:14.435 --> 02:16.325
into many fine categories.

02:16.865 --> 02:19.285
And this means that there would be,

02:20.045 --> 02:22.685
possibly fewer experiences in each category.

02:23.025 --> 02:25.605
So the way the individual categorizes matters,

02:25.605 --> 02:27.645
because they're using the category average

02:27.745 --> 02:28.805
to make predictions.

02:29.305 --> 02:32.085
And that's why we are interested in the question,

02:32.795 --> 02:36.845
when is course categorization actually good for individuals?

02:36.865 --> 02:39.245
Can it be ever good beneficial for them

02:39.425 --> 02:40.805
to categorize coursely?

02:40.825 --> 02:43.165
Can they make better predictions if they

02:43.165 --> 02:44.405
categorize coursely?

02:44.945 --> 02:46.725
Uh, this is the key question of the paper.

02:47.305 --> 02:49.205
And, we,

02:49.725 --> 02:52.165
answered this question using a simple model.

02:52.515 --> 02:54.925
Combining game theory and statistics.

02:55.515 --> 02:58.965
There's abundant literature in psychology showing

02:59.195 --> 03:02.405
that people use categories in various contexts.

03:03.145 --> 03:06.805
And more recently there's a literature in economics

03:07.275 --> 03:10.085
that links course categorization

03:10.705 --> 03:13.565
to various biases in decision making.

03:14.225 --> 03:16.245
And so, this, they range

03:16.755 --> 03:21.285
from discrimination against minorities to over

03:21.545 --> 03:23.845
and, underpricing of assets

03:23.865 --> 03:27.085
and financial markets to customers reacting

03:27.085 --> 03:28.685
to advertising in various ways.

03:28.705 --> 03:30.285
So in many different contexts,

03:30.505 --> 03:32.885
course categorization has been linked

03:32.905 --> 03:34.645
to biases in decision making.

03:35.425 --> 03:38.645
And so from, inspired by this, we, we are kind

03:38.645 --> 03:39.845
of interested in the question

03:39.845 --> 03:43.325
where if course categorization link is linked

03:43.465 --> 03:44.685
or can lead to biases

03:44.905 --> 03:48.565
and decision making, why do people categorize coursely?

03:48.905 --> 03:52.245
Uh, is it one could think that it's due to some sort

03:52.245 --> 03:53.805
of cognitive limitations?

03:53.825 --> 03:56.165
So when an innate limitation on the number

03:56.165 --> 03:58.605
of categories we can use,

03:58.785 --> 04:00.885
or, basically it could be

04:00.885 --> 04:03.965
that there are computational costs of using more categories,

04:04.305 --> 04:07.205
even if we could have many available in our brains.

04:07.745 --> 04:10.325
But so what we show in this paper is

04:10.325 --> 04:14.805
that it's not necessarily the case, that,

04:15.145 --> 04:19.205
that basically course categorization has to be due

04:19.425 --> 04:20.765
to cognitive limitations.

04:20.825 --> 04:23.165
So course categorization can arise

04:23.805 --> 04:25.485
rationally under some environments.

04:25.555 --> 04:28.645
This is the point of probably the main message of the paper.

04:28.705 --> 04:29.845
So this is the main message.

04:30.425 --> 04:34.085
And, in showing this, we are inspired

04:34.385 --> 04:37.965
by literature from different fields, from cognitive science,

04:38.505 --> 04:40.485
from, machine learning,

04:40.795 --> 04:43.085
from econometrics in particular,

04:43.085 --> 04:46.485
there's the literature on the bias variance,

04:46.505 --> 04:49.045
the composition of expected prediction error,

04:49.265 --> 04:50.805
and this what we draw on.

04:51.305 --> 04:52.525
But the way,

04:52.875 --> 04:57.165
what our paper focuses on is taking this kind of intuition

04:57.515 --> 05:00.165
that, there can be this different components

05:00.305 --> 05:03.965
of expected prediction error to a strategic context, so

05:03.965 --> 05:08.125
to a, a game theoretical context where individuals want

05:08.125 --> 05:10.485
to coordinate their predictions with each other.

05:10.825 --> 05:13.645
And that's how it differs from a lot of this literature,

05:13.895 --> 05:16.765
which is focusing on individual decision making.

05:17.425 --> 05:19.325
We have a model of an individual

05:19.425 --> 05:23.285
who categorizes their past experiences, in order

05:23.285 --> 05:26.165
to make predictions about a new situation

05:26.165 --> 05:27.365
that they will encounter.

05:28.065 --> 05:30.885
Um, so the individual will observe

05:31.435 --> 05:34.565
some observable characteristics of this past experiences

05:34.985 --> 05:38.285
and places the past experiences in categories based

05:38.305 --> 05:39.445
on the observables.

05:40.025 --> 05:42.645
Um, and then what the individual is trying

05:42.645 --> 05:46.005
to do is they're trying to predict some unobserved value,

05:46.425 --> 05:49.285
of the situation, let's say the, suitability

05:49.345 --> 05:51.165
of an applicant for a particular job.

05:51.225 --> 05:54.645
And they're trying to predict this based on observable

05:54.645 --> 05:57.245
characteristics such as where, where the applicant went

05:57.245 --> 06:01.445
to university, or, their race and gender and so on.

06:01.985 --> 06:06.165
Um, and so the individual places this past experience into a

06:06.365 --> 06:08.565
category and makes a prediction based

06:08.705 --> 06:09.925
on the category average.

06:10.585 --> 06:12.525
Um, and as I mentioned before,

06:12.995 --> 06:16.165
because there are many different possible ways

06:16.225 --> 06:17.405
to categorize, and

06:17.405 --> 06:20.485
because the individual is making a prediction in the

06:20.685 --> 06:22.925
category average, this will,

06:23.115 --> 06:24.645
this will influence their prediction.

06:24.645 --> 06:26.685
So whether the individual categorizes coarsely

06:26.685 --> 06:29.805
or whether the individual categorizes finally has an impact

06:29.825 --> 06:31.565
on the prediction that the individual makes.

06:32.145 --> 06:35.285
The key focus of the paper is to look at the question,

06:35.595 --> 06:38.685
when is it good for this individual to categorize coarsely?

06:39.185 --> 06:40.845
And, in particular,

06:41.105 --> 06:43.245
we take an X under perspective on this.

06:43.505 --> 06:45.965
So in this model, we kind of think

06:46.025 --> 06:49.605
of this category averages as sets of estimators

06:49.605 --> 06:52.045
that the individual uses to make predictions.

06:52.625 --> 06:54.285
And we look,

06:54.465 --> 06:56.925
and we assume the objective of the individual is

06:56.925 --> 06:59.165
to minimize their expected prediction error

06:59.305 --> 07:00.485
on the next object.

07:00.985 --> 07:02.605
Um, and the question is

07:03.035 --> 07:05.925
what determines which categories are issues?

07:06.225 --> 07:08.965
The minimize expected prediction error, if it's course

07:08.985 --> 07:10.725
or fine categorizations,

07:10.985 --> 07:13.165
we do some comparative statics on that.

07:13.945 --> 07:16.245
So the three types of cases,

07:16.275 --> 07:20.765
that we consider in this research is individual prediction,

07:20.965 --> 07:24.725
coordination of predictions and individual prediction

07:24.745 --> 07:26.565
and coordination of predictions together.

07:27.105 --> 07:30.805
So what do I mean by that? Um, individual predictions.

07:31.285 --> 07:33.725
A prediction means that we have an individual

07:34.105 --> 07:38.485
who is interested in predicting the true value

07:38.865 --> 07:41.685
of the next situation that they encounter correctly.

07:42.185 --> 07:45.325
Um, and, so basically their objective is

07:45.325 --> 07:47.725
to minimize expected prediction error in

07:47.725 --> 07:48.765
individual prediction.

07:48.765 --> 07:51.205
This is what we are assuming this is the first case.

07:52.025 --> 07:54.645
Now the second benchmark case that we consider

07:55.225 --> 08:00.205
is two individuals who are matched to play a type

08:00.205 --> 08:01.925
of coordination game with each other.

08:02.265 --> 08:05.205
And their objective is to try to

08:05.765 --> 08:08.405
minimize their expected prediction error,

08:08.635 --> 08:09.725
from the other person.

08:09.905 --> 08:13.125
So they no longer in this case, in this particular case,

08:13.125 --> 08:15.445
they no longer care about predicting the true value

08:15.465 --> 08:16.685
of the object correctly,

08:16.915 --> 08:18.925
they just care about coordinating their

08:18.925 --> 08:20.125
prediction with the other person.

08:20.585 --> 08:24.645
And so if they only care about coordinating their prediction

08:24.645 --> 08:27.005
with the other person, what we,

08:27.005 --> 08:28.925
what we assume is their objective is

08:28.925 --> 08:30.845
to minimize their expected prediction error

08:31.155 --> 08:32.285
from coordination.

08:32.745 --> 08:36.285
Um, and, and this is the second case we consider.

08:36.585 --> 08:39.605
And then the third case we consider is naturally in many

08:39.605 --> 08:42.565
situations, they will care both about correct individual

08:42.565 --> 08:45.405
prediction and coordination of predictions.

08:45.665 --> 08:49.085
And so basically they placed some weight on individual

08:49.085 --> 08:52.045
prediction and some weight on coordination of predictions.

08:52.625 --> 08:55.645
Now, the key thing is, in all of these three cases,

08:56.235 --> 09:00.285
what we do is we do a bias variance decomposition

09:00.345 --> 09:01.805
of expected prediction error.

09:02.065 --> 09:03.925
And this will give us a trade off

09:04.105 --> 09:05.605
of when it's good to categorize.

09:05.605 --> 09:08.405
Finally, versus when it's good to categorize coursely.

09:08.755 --> 09:11.325
Basically there will be a couple of key factors

09:11.635 --> 09:16.205
that determine, when course categorization is beneficial

09:16.385 --> 09:19.845
or when fine categorization is beneficial for an individual

09:19.905 --> 09:21.205
who wants to make predictions

09:21.225 --> 09:23.005
or coordinate predictions with others.

09:23.425 --> 09:25.205
But let me start by distinguishing the factors.

09:25.745 --> 09:27.245
So the factors,

09:27.305 --> 09:30.685
or among the factors we consider are the sample size

09:30.905 --> 09:33.965
of past experiences that the individual has accumulated,

09:34.825 --> 09:38.325
and, the noise of the environment.

09:38.325 --> 09:41.965
So whether the environment is deterministic or stochastic.

09:41.985 --> 09:43.205
So this what type

09:43.205 --> 09:46.165
of environment the individual is operating in is

09:46.165 --> 09:47.205
very important here.

09:48.065 --> 09:50.845
So what is additional realistic environment?

09:50.955 --> 09:52.525
What is a stochastic environment?

09:53.125 --> 09:57.445
A deterministic environment is an environment in which,

09:58.865 --> 10:02.485
the observable characteristics fully determine the

10:02.485 --> 10:04.285
unobserved value of the object.

10:04.745 --> 10:09.725
So basically, whether where I went to school,

10:09.915 --> 10:13.565
what, a type of, what field of study I had,

10:13.705 --> 10:17.245
my gender, my race fully determined, my ability

10:17.385 --> 10:19.485
or my suitability for a particular job.

10:19.865 --> 10:22.445
Uh, and everyone who has the same

10:22.445 --> 10:24.845
characteristics would have the same ability

10:24.865 --> 10:26.125
or suitability for a job.

10:26.705 --> 10:28.845
Now, of course, this is not realistic, right?

10:28.865 --> 10:32.405
So this is just kind of a benchmark case in which everyone

10:32.425 --> 10:34.525
who looks the same has the same value,

10:34.625 --> 10:36.365
has the same mobility, suitability,

10:36.805 --> 10:38.205
whatever it is we are trying to predict.

10:38.745 --> 10:40.685
Uh, the more interesting case

10:40.705 --> 10:43.005
and the more realistic case is really a

10:43.005 --> 10:44.165
stochastic environment.

10:44.885 --> 10:48.165
A stochastic environment is an environment in which

10:48.435 --> 10:51.045
observables do not,

10:51.415 --> 10:56.285
fully determine on individual's, unobserved value

10:56.345 --> 10:57.925
or an individual's suitability.

10:58.465 --> 11:00.725
Um, and so these are the environments

11:00.725 --> 11:03.805
that we very much focus on stochastic environments.

11:04.545 --> 11:06.285
And now back to my three cases.

11:06.945 --> 11:11.245
So an individual prediction, what happens is

11:11.595 --> 11:13.925
that if an individual is just trying

11:13.925 --> 11:16.165
to predict the true value of the object correctly,

11:16.185 --> 11:19.045
and they're operating in a deterministic environment,

11:19.425 --> 11:23.045
so basically observe both fully determined value, then

11:23.745 --> 11:26.765
the best thing the individual could do is

11:26.945 --> 11:29.805
to use the finest possible categorization.

11:30.065 --> 11:32.725
So have a separate category for each object type.

11:32.745 --> 11:35.765
So splitting their experiences into many categories.

11:36.665 --> 11:41.645
Now, this is if the environment is deterministic, if

11:42.415 --> 11:44.885
noise over sample size increases.

11:44.885 --> 11:47.365
So if the environment becomes stpa stochastic

11:47.365 --> 11:49.845
and the individual has a limited number of experiences

11:49.905 --> 11:52.965
or the environment is changing, then

11:53.155 --> 11:55.845
basically the finest possible categorization

11:56.385 --> 11:58.125
may no longer be the best.

11:58.625 --> 12:01.565
So it may become rational for an individual

12:01.625 --> 12:03.725
to use a coarser categorization

12:03.835 --> 12:04.965
because this type

12:04.965 --> 12:09.045
of categorization might now minimize expected prediction error.

12:09.585 --> 12:11.085
Now, what is the reason for this?

12:11.555 --> 12:14.405
Well, the reason is this bias variance trade off.

12:14.705 --> 12:18.085
So, in a deterministic environment, there's no variance.

12:18.105 --> 12:21.405
So that's why the finest possible categorization was best

12:21.555 --> 12:23.565
because it minimizes the bias.

12:24.065 --> 12:26.885
But in a stochastic environment, the noisiness

12:26.885 --> 12:29.005
of the environment starts to play a role.

12:29.425 --> 12:32.365
And what the individual can do in order to

12:33.085 --> 12:37.325
minimize the effect of this noisy past experiences on their

12:37.625 --> 12:39.165
new prediction is

12:39.385 --> 12:42.805
to lump their experiences in one big category.

12:43.105 --> 12:46.085
So this means that there will be a higher sample size per

12:46.325 --> 12:48.165
category, more experiences per category,

12:48.385 --> 12:51.925
and this is what helps to decrease the variance component,

12:52.305 --> 12:54.445
of the expected prediction error.

12:54.625 --> 12:56.405
So there's this trade off between bias

12:56.465 --> 12:57.525
and variance components.

12:57.955 --> 13:00.045
This is the case for individual prediction.

13:00.465 --> 13:03.805
Now, the focus of this paper is really on the case

13:03.945 --> 13:06.085
of coordination and predictions.

13:06.705 --> 13:11.205
Uh, and so there two individuals are, matched,

13:11.265 --> 13:14.325
to play a one-shot simultaneous move game

13:14.325 --> 13:17.325
where each individual chooses what categorization

13:17.385 --> 13:19.085
to use in order to make a prediction.

13:19.505 --> 13:21.805
And we analyze this type of setting.

13:21.915 --> 13:24.965
They're trying to minimize their expected prediction error

13:25.195 --> 13:28.165
from each other, from just coordinating their predictions,

13:28.465 --> 13:31.885
and we can, derive, equilibrium existence,

13:31.885 --> 13:33.485
equilibrium efficiency and so on.

13:33.485 --> 13:35.485
But I'll just focus on the main results here.

13:35.945 --> 13:38.005
And so the main results are

13:38.315 --> 13:40.765
that if the environment it is deterministic

13:40.765 --> 13:43.285
and the individual just cares about coordinating

13:43.285 --> 13:46.485
with the other person, then the best thing that they can do,

13:46.945 --> 13:51.125
is use the same categorization as the other, just kind

13:51.125 --> 13:53.765
of like use the same model for making predictions.

13:54.185 --> 13:58.405
Uh, but it does not really matter which exact model they're

13:58.405 --> 14:01.645
using as long as they're using the other in a deterministic

14:01.645 --> 14:04.245
environment if they want to coordinate predictions.

14:05.465 --> 14:08.485
Now, this is basically saying

14:08.485 --> 14:10.525
that it doesn't matter if they use a finer

14:10.525 --> 14:13.405
or a cosar categorization in a deterministic environment,

14:13.865 --> 14:15.965
in a stochastic environment,

14:15.965 --> 14:17.725
however, it matters greatly.

14:18.065 --> 14:22.365
So as noise increases relative to sample size,

14:23.255 --> 14:27.725
finer, categorization profiles, may no longer

14:28.345 --> 14:29.485
be equilibrium.

14:29.745 --> 14:33.365
Uh, and also for any kind of stochastic environment,

14:34.015 --> 14:37.405
forer categorization profiles will help,

14:37.405 --> 14:39.405
will be more efficient, will,

14:39.405 --> 14:43.445
will always help agents always lead to lower

14:44.005 --> 14:45.205
expected prediction errors.

14:45.205 --> 14:47.445
So courses, symmetric categorization profiles.

14:47.705 --> 14:51.045
So this is kind of saying, well, if the individual wants

14:51.145 --> 14:53.965
to minimize their expected prediction error from

14:53.965 --> 14:57.005
coordination from the other person's prediction,

14:57.355 --> 15:01.685
then they're better off categorizing coarsely if there is,

15:02.025 --> 15:03.805
any noise in the environment.

15:04.185 --> 15:05.885
Um, and, as I said,

15:05.885 --> 15:08.405
it's the various equilibrium efficiency,

15:08.405 --> 15:11.605
equilibrium existence arguments, but this is the key point.

15:11.905 --> 15:15.365
What's the intuition of why, agents

15:16.015 --> 15:19.245
would be better off categorizing coursely if they want

15:19.245 --> 15:21.005
to coordinate their prediction with each other?

15:21.875 --> 15:25.525
Well, this has to do with the variance component

15:25.525 --> 15:27.005
of the expected prediction error.

15:27.385 --> 15:30.045
As I said before, this expected prediction error can be

15:30.045 --> 15:32.565
decomposed into bias, into variance component.

15:33.025 --> 15:36.925
Now, course categorization, leads to, leads to a bias.

15:37.385 --> 15:40.525
Uh, but what it helps to do is it helps

15:40.625 --> 15:42.525
to minimize the variance component

15:42.715 --> 15:46.805
because if individuals lump their experiences

15:47.355 --> 15:48.605
into fewer categories,

15:48.735 --> 15:51.485
there are more experiences per category.

15:51.865 --> 15:53.965
And what this means is

15:54.315 --> 15:56.885
that the idiosyncratic noise

15:57.105 --> 16:01.805
of each individual's past experience will be minimized,

16:01.945 --> 16:05.125
or, by using a courser categorization.

16:05.665 --> 16:09.845
So that's the intuition briefly of why it may be better off

16:09.845 --> 16:12.885
to categorize coursely if individuals want to,

16:12.885 --> 16:14.725
coordinate their predictions with each other.

16:15.145 --> 16:16.285
And this whole intuition

16:16.285 --> 16:18.365
that I've developed here carries over

16:18.425 --> 16:21.485
to the individual prediction and coordination case as well.

16:21.865 --> 16:25.605
The key message of this paper, again, is

16:25.635 --> 16:29.525
that course categorization can be beneficial,

16:29.955 --> 16:32.405
both in an individual prediction environment.

16:32.785 --> 16:35.485
Uh, but what we focus on mostly is this environment

16:35.485 --> 16:37.805
where individuals want to coordinate with others.

16:38.305 --> 16:41.525
So course categorization, can be beneficial

16:41.665 --> 16:42.965
for individuals who want

16:42.965 --> 16:44.645
to coordinate their predictions with others.

16:45.105 --> 16:48.045
And this is of course, under some assumptions,

16:48.045 --> 16:50.565
they sample independently from each other and so on.

16:51.265 --> 16:53.565
So now why does this matter?

16:54.065 --> 16:57.365
Uh, why does it matter that course categorization,

16:57.365 --> 17:00.045
can be good in stochastic environments

17:00.215 --> 17:02.765
where individuals have limited past experiences

17:03.295 --> 17:05.565
where there are many such environments out there.

17:06.105 --> 17:10.605
And the previous literature has linked various biases in

17:10.805 --> 17:14.445
decision making, to, course categorization.

17:14.825 --> 17:16.845
So we think this is kind of a step

17:16.905 --> 17:20.405
to understanding why course categorization may

17:20.405 --> 17:21.565
arise, nevertheless.

17:21.945 --> 17:24.645
So why do people still categorize coursely,

17:24.645 --> 17:26.805
even if this leads to some sort of biases,

17:27.365 --> 17:28.685
mistakes in decision making?

17:29.225 --> 17:32.645
Um, and also, the findings

17:32.645 --> 17:34.245
of the paper are really quite nuanced.

17:34.245 --> 17:36.005
So it kind of says,

17:36.195 --> 17:39.725
that in some environments we may expect,

17:39.795 --> 17:41.605
only minor categorization

17:41.665 --> 17:43.925
to be present In some environments,

17:44.025 --> 17:47.045
we may expect worse categorization to be present.

17:47.305 --> 17:48.725
And in some environments we may

17:48.725 --> 17:51.445
Expect categorizations at different levels, of course.

17:51.905 --> 17:54.845
So this is consistent, with what we observe,

17:55.105 --> 17:56.165
out there really.

17:56.585 --> 17:58.765
Um, so the key point is,

17:58.875 --> 18:01.365
that worst categorization can be beneficial,

18:01.665 --> 18:04.165
and this is shown using a bias variance straight off.

Quick Answers

Benefits of Coarse Categorization in Decision Making

00:00:00 - The paper that I'm going to talk about is called, categorization and Coordination and it's joint work, with my co-author Nick Green from Queen Mary University of London. So, what we do in this paper is we are interested in the use of mental categories to make predictions. And so you could think of many examples of people using mental categories. So it could be employers categorizing applicants based
00:00:34 - perhaps their gender, race and so on. Or you could think of equity researchers, categorizing firms, based on the sector that they operate in based on their size, and other characteristics. Likewise, customers categorizing hotels. So there are multiple examples of people categorizing out there, and that's why we are interested in categorization.
00:01:04 - So, categorization is a form of decision making on the pa on the basis of past experiences. So the idea is, an individual will have accumulated a set of past experiences and they store them into categories. And, when encountering a new situation that resembles this past experiences based on its observable characteristics, they make a prediction based on the
00:01:35 - So this is what categorization is. what we look at this at in this paper is the idea that in any situation there will be many different possible ways to categorize. and what we focus on in particular is the idea that individuals may be using a few course categories. So basically putting their past experiences in just a few categories.
00:02:07 - possibly many experiences in each one of these categories or individuals could sort their past experiences into many fine categories. And this means that there would be, possibly fewer experiences in each category. So the way the individual categorizes matters, because they're using the category average to make predictions. And that's why we are interested in the question, when is course categorization actually good for individuals?
00:02:39 - to categorize coursely? Can they make better predictions if they categorize coursely? this is the key question of the paper. And, we, answered this question using a simple model. Combining game theory and statistics. There's abundant literature in psychology showing that people use categories in various contexts. And more recently there's a literature in economics

Cognitive Limitations and Course Categorization Biases

00:03:10 - to various biases in decision making. And so, this, they range from discrimination against minorities to over and, underpricing of assets and financial markets to customers reacting to advertising in various ways. So in many different contexts, course categorization has been linked to biases in decision making. And so from, inspired by this, we, we are kind
00:03:39 - where if course categorization link is linked or can lead to biases and decision making, why do people categorize coursely? is it one could think that it's due to some sort of cognitive limitations? So when an innate limitation on the number of categories we can use, or, basically it could be that there are computational costs of using more categories, even if we could have many available in our brains. But so what we show in this paper is
00:04:15 - that basically course categorization has to be due to cognitive limitations. So course categorization can arise rationally under some environments. This is the point of probably the main message of the paper. So this is the main message. And, in showing this, we are inspired by literature from different fields, from cognitive science, from, machine learning,
00:04:43 - there's the literature on the bias variance, the composition of expected prediction error, and this what we draw on. But the way, what our paper focuses on is taking this kind of intuition that, there can be this different components of expected prediction error to a strategic context, so to a, a game theoretical context where individuals want to coordinate their predictions with each other.
00:05:13 - which is focusing on individual decision making. We have a model of an individual who categorizes their past experiences, in order to make predictions about a new situation that they will encounter. so the individual will observe some observable characteristics of this past experiences and places the past experiences in categories based on the observables. and then what the individual is trying
00:05:46 - of the situation, let's say the, suitability of an applicant for a particular job. And they're trying to predict this based on observable characteristics such as where, where the applicant went to university, or, their race and gender and so on. and so the individual places this past experience into a category and makes a prediction based on the category average. and as I mentioned before,

Understanding Coarse vs. Fine Categorization

00:06:16 - to categorize, and because the individual is making a prediction in the category average, this will, this will influence their prediction. So whether the individual categorizes coarsely or whether the individual categorizes finally has an impact on the prediction that the individual makes. The key focus of the paper is to look at the question, when is it good for this individual to categorize coarsely? And, in particular, we take an X under perspective on this.
00:06:46 - of this category averages as sets of estimators that the individual uses to make predictions. And we look, and we assume the objective of the individual is to minimize their expected prediction error on the next object. and the question is what determines which categories are issues? The minimize expected prediction error, if it's course or fine categorizations, we do some comparative statics on that.
00:07:16 - that we consider in this research is individual prediction, coordination of predictions and individual prediction and coordination of predictions together. So what do I mean by that? individual predictions. A prediction means that we have an individual who is interested in predicting the true value of the next situation that they encounter correctly. and, so basically their objective is
00:07:47 - individual prediction. This is what we are assuming this is the first case. Now the second benchmark case that we consider is two individuals who are matched to play a type of coordination game with each other. And their objective is to try to minimize their expected prediction error, from the other person. So they no longer in this case, in this particular case, they no longer care about predicting the true value
00:08:16 - they just care about coordinating their prediction with the other person. And so if they only care about coordinating their prediction with the other person, what we, what we assume is their objective is to minimize their expected prediction error from coordination. and, and this is the second case we consider. And then the third case we consider is naturally in many situations, they will care both about correct individual prediction and coordination of predictions.
00:08:49 - prediction and some weight on coordination of predictions. Now, the key thing is, in all of these three cases, what we do is we do a bias variance decomposition of expected prediction error. And this will give us a trade off of when it's good to categorize. Finally, versus when it's good to categorize coursely. Basically there will be a couple of key factors that determine, when course categorization is beneficial

Deterministic vs. Stochastic Environments in Predictions

00:09:19 - who wants to make predictions or coordinate predictions with others. But let me start by distinguishing the factors. So the factors, or among the factors we consider are the sample size of past experiences that the individual has accumulated, and, the noise of the environment. So whether the environment is deterministic or stochastic. So this what type of environment the individual is operating in is very important here.
00:09:50 - What is a stochastic environment? A deterministic environment is an environment in which, the observable characteristics fully determine the unobserved value of the object. So basically, whether where I went to school, what, a type of, what field of study I had, my gender, my race fully determined, my ability or my suitability for a particular job.
00:10:22 - characteristics would have the same ability or suitability for a job. Now, of course, this is not realistic, right? So this is just kind of a benchmark case in which everyone who looks the same has the same value, has the same mobility, suitability, whatever it is we are trying to predict. the more interesting case and the more realistic case is really a stochastic environment. A stochastic environment is an environment in which observables do not,
00:10:56 - or an individual's suitability. and so these are the environments that we very much focus on stochastic environments. And now back to my three cases. So an individual prediction, what happens is that if an individual is just trying to predict the true value of the object correctly, and they're operating in a deterministic environment, so basically observe both fully determined value, then
00:11:26 - to use the finest possible categorization. So have a separate category for each object type. So splitting their experiences into many categories. Now, this is if the environment is deterministic, if noise over sample size increases. So if the environment becomes stpa stochastic and the individual has a limited number of experiences or the environment is changing, then basically the finest possible categorization
00:11:58 - So it may become rational for an individual to use a coarser categorization because this type of categorization might now minimize expected prediction error. Now, what is the reason for this? Well, the reason is this bias variance trade off. So, in a deterministic environment, there's no variance. So that's why the finest possible categorization was best because it minimizes the bias. But in a stochastic environment, the noisiness

Trade-Offs in Categorization for Prediction Accuracy

00:12:29 - And what the individual can do in order to minimize the effect of this noisy past experiences on their new prediction is to lump their experiences in one big category. So this means that there will be a higher sample size per category, more experiences per category, and this is what helps to decrease the variance component, of the expected prediction error. So there's this trade off between bias and variance components.
00:13:00 - Now, the focus of this paper is really on the case of coordination and predictions. and so there two individuals are, matched, to play a one-shot simultaneous move game where each individual chooses what categorization to use in order to make a prediction. And we analyze this type of setting. They're trying to minimize their expected prediction error from each other, from just coordinating their predictions,
00:13:31 - equilibrium efficiency and so on. But I'll just focus on the main results here. And so the main results are that if the environment it is deterministic and the individual just cares about coordinating with the other person, then the best thing that they can do, is use the same categorization as the other, just kind of like use the same model for making predictions. but it does not really matter which exact model they're using as long as they're using the other in a deterministic
00:14:05 - Now, this is basically saying that it doesn't matter if they use a finer or a cosar categorization in a deterministic environment, in a stochastic environment, however, it matters greatly. So as noise increases relative to sample size, finer, categorization profiles, may no longer be equilibrium. and also for any kind of stochastic environment,
00:14:37 - will be more efficient, will, will always help agents always lead to lower expected prediction errors. So courses, symmetric categorization profiles. So this is kind of saying, well, if the individual wants to minimize their expected prediction error from coordination from the other person's prediction, then they're better off categorizing coarsely if there is, any noise in the environment.
00:15:05 - it's the various equilibrium efficiency, equilibrium existence arguments, but this is the key point. What's the intuition of why, agents would be better off categorizing coursely if they want to coordinate their prediction with each other? Well, this has to do with the variance component of the expected prediction error. As I said before, this expected prediction error can be decomposed into bias, into variance component. Now, course categorization, leads to, leads to a bias.

Benefits of Coarse Categorization in Decision Making

00:15:40 - to minimize the variance component because if individuals lump their experiences into fewer categories, there are more experiences per category. And what this means is that the idiosyncratic noise of each individual's past experience will be minimized, or, by using a courser categorization. So that's the intuition briefly of why it may be better off
00:16:12 - coordinate their predictions with each other. And this whole intuition that I've developed here carries over to the individual prediction and coordination case as well. The key message of this paper, again, is that course categorization can be beneficial, both in an individual prediction environment. but what we focus on mostly is this environment where individuals want to coordinate with others. So course categorization, can be beneficial
00:16:42 - to coordinate their predictions with others. And this is of course, under some assumptions, they sample independently from each other and so on. So now why does this matter? why does it matter that course categorization, can be good in stochastic environments where individuals have limited past experiences where there are many such environments out there. And the previous literature has linked various biases in decision making, to, course categorization.
00:17:16 - to understanding why course categorization may arise, nevertheless. So why do people still categorize coursely, even if this leads to some sort of biases, mistakes in decision making? and also, the findings of the paper are really quite nuanced. So it kind of says, that in some environments we may expect, only minor categorization to be present In some environments, we may expect worse categorization to be present.
00:18:01 - Expect categorizations at different levels, of course. So this is consistent, with what we observe, out there really. so the key point is, that worst categorization can be beneficial, and this is shown using a bias variance straight off.

Faculti AI can make mistakes

Research

Policy

Analysis

Data

Originals

Explainers

Podcasts