1 of 164

Behavioral Research in the Wild

On Amir

Rady School of Management

UC San Diego

2 of 164

Recent research…

Consumer Decision Making

Behavioral Change

Methods

Behavior in the wild

3 of 164

Consumer Decision Making

Stuck in a Rut: The Behavioral Entrenchment Effect (with Allie Lieberman and Ziv Carmon) –2^nd round - OBHDP

A Threshold Value Theory of Sequential Choice (with Coby Morvinski and Evan Weingarten) – 2^nd round JMR

Weingarten, et al. (2020). Experts Outperform Technology in Creative Markets, She Ji: The Journal of Design, Economics, and Innovation.

Kristen E. Duke & On Amir (2019). Guilt Dynamics: Consequences of Temporally Separating Decisions and Actions, JCR, Ferber Award.

How Product Unavailability Leads to Choice of Lesser Alternatives—The Role of Hedonic Aggregation (With Jean Zhang and Gal Smitizsky)

Reflecting on the Reflection Effect: The Role of Probability Representation in Risky Choice (with Kristen Duke and Daniel Mochon)

Kristen Duke, Kelly Goldsmith, & On Amir (2018). Is the Preference for Certainty Always so Certain? JACR

Coby Morvinski & On Amir (2018). Liking Goes with Liking: An Intuitive Congruence between Preference and Prominence, JEP:LMC.

Daniella Kupor, Wendy Liu, & On Amir (2017). The Effect of an Interruption on Risk Decisions, JCR

Coby Morvinski, On Amir, & Eitan Muller (2017). “Ten Million Readers Can’t Be Wrong!,” or Can They? On the Role of Information About Adoption Stock in New Product Trial, Marketing Science.

4 of 164

Behavioral change

Aharon, et al.(2022). Improving Cardiac Rehabilitation Patient Adherence via Personalized Interventions, PLOS-ONE.

Alicea Lieberman, Juliana Schroeder, & On Amir (2022). A Voice Inside My Head: The Psychological and Behavioral Consequences of Auditory Technologies, OBHDP.

Coby Morvinski, Silvia Saccardo, & On Amir (2022). Mis-Nudging Morality, Management Science, 1-11.

Alicea Lieberman, Andera Morales, & On Amir (2021). Tangential Immersion: Increasing Consumer Persistence, JCR.

Elanor Williams, Allie Lieberman, & On Amir (2021). Perspective Neglect: Inadequate Perspective Taking Limits Coordination, JDM.

Shuval, et al. (2017). Physical Activity Counseling in Primary Care: Insights from Public Health & Behavioral Economics, CA: A Cancer Journal for Clinicians.

Alicea Lieberman , Kristen Duke, & On Amir (2019). How Incentive Framing Can Harness the Power of Norms, OBHDP.

Goal Proximity, Social Information, and Giving: When Norms Backfire (with Coby Morvinski and Matt Lupoli) 3^rd round, PLOS-One

The Impact of Linguistic Structure on Judgment Confidence and Source Attitude (with Giulia Maimone and Uma Karmarkar) – Under review at JMR

Skeptical but Still Indebted: Understanding How Reciprocity is Immune to Skepticism (with Freeman Wu and Andrea Morales)

5 of 164

Methods

What Methods Succeed at the Journal of Consumer Research? (with Andrea Morales and Evan Weingarten) – 2^nd round JCR

Weingarten, et al. (2022). What Makes People Happy? Decoupling the Experiential-Material Continuum, JCP

Thomadsen, et al. (2018). How Context Affects Choice, Customer Needs and Solutions.

Andrea Morales, On Amir, & Leonard Lee (2017). A Tutorial in Consumer Research: Keeping it Real in Experimental Research – Understanding When, Where, and How to Enhance Realism and Measure Consumer Behavior, JCR.

6 of 164

In the wild…

Kristen Duke & On Amir (2022). The Importance of Selling Formats: When Integrating Purchase and Quantity Decisions Increases Sales, Marketing Science

Dominance in the Wild (with Ariel Fridman and Karsten Hansen) – 2^nd round JCR

Consumer Coping: A Stage Model of Pandemic Response (with Jean Zhang) – 3^rd round JACR

Price Contrast in the Wild (with Ariel Fridman and Karsten Hansen)

Consideration Sets in the Wild (with Ariel Fridman, Karsten Hansen, and Wendy Liu)

An unexpected bias: High levels of achievement change the biases women face at work (with Jean Zhang and Elizabeth Campbell)

Communication Style as an Impediment to Trade (with Gal Smitizsky and Kaidi Wu)

Tipping in a Digital Services Marketplace (with Seung Hyun Kim and Kenneth C. Wilbur)

7 of 164

Topics

The context of search results assortment

Dominance

Price contrast

Communication styles and global trade

The dark side of discrimination awareness

(Tipping online)

8 of 164

Dominance in the wild

(with Ariel Fridman and Karsten Hansen, JCR 2023)

9 of 164

(weak) Dominance

But… Frederick et al. (2014), Yang and Lynn (2014)

Dominating

Dominated

In this project, I’ll be focusing on the dominance effect – one of the most fascinating phenomena in the consumer behavior literature.

When I say the ‘dominance effect’, I’m referring to the asymmetric dominance effect, also referred to as the attraction effect or decoy effect, first demonstrated by Huber, Payne, and Puto (1982), and there’s been LOTS of subsequent work.

Here’s an example of dominance: I was recently searching for collectible quarters on Amazon, and I saw this assortment of options. I focused in on this first one on the left – it’s a bicentennial quarter from 1976. And when I looked more closely at the other options on Amazon, I noticed that there was this option as well. I took a closer look and realized they’re the exact same coin, but it’s more expensive, lower average rating, and lower number of ratings. So the option on the right was dominated or inferior, compared to the one of the left. Of course, I would never purchase the option on the right, but it’s presence in the assortment does make the left one a dominating option. The dominance effect occurs if I am more likely to purchase the option on the left (as opposed to the other options like the middle two), when it is in the presence of an alternative that it is inferior to it (or a dominated option) like the one on the right.

The dominance effect is really a fundamental behavioral phenomena, and part of what I think makes it so interesting is that it intersects research on perception, attention, cognition, emotion, and more specific theories of similarity, range, and others. And obviously decision making.

10 of 164

Research Questions

Is the dominance effect real?

(beyond the lab and 2-3 item choice sets; preference uncertainty)

Are dominance relationships prevalent?

(why would there be dominated items?)

How does the effect work?

Perceptual vs. Cognitive

In this work, I'll start by using a very large-scale scale field experiment that ask: is this effect real?

I will also shed light on how prevalent dominance relationships are – they can actually be quite common in the real-world.

And I’ll also use the rich field data I have to shed light on the psychology driving this effect – telling you when to expect to find it in the field.

EXTRA:

My last hypothesis, which I’ll only get in to a little here, and in a longer talk I’d love to talk more about it, is that the effect is consistent with a perceptual mechanism underlying it, where the dominating alternative simply “seems better” in the presence of a dominated gig, and the findings help to rule out some alternative mechanisms that have been proposed in the literature.

I want to point out that this hypothesis is not always considered by researchers in the field: for example, Frederick et al. 2014 attempted to use more realistic stimuli, but in using products people knew and had likely purchased before and experienced (like apples and oranges), they also likely had low preference uncertainty – in other words they knew their preferences.

11 of 164

12 of 164

13 of 164

14 of 164

Let’s focus on this particular gig on the top left. It appeared in 669 assortments in the period of data I have. (660 of those assortments were unique.)

So if you wanted to test for the real effect of dominance, and could design the perfect experiment, what would that look like? You would need to vary the assortment around this offer, such that in one condition it is dominating, in another condition it is dominated, and a control condition when it is neither; moreover, you would want to allow people not to chose any of the offers they see – this would be the cleanest test, maybe even cleaner than most lab study designs out there, which use a forced-choice design.

Well? This is exactly what I have in the Fiverr data – this gig appears in the 3 conditions that the natural experiment creates.

EXTRA:

because it would allow you to cleanly identify and measure both sides of the dominance effect – the effect becoming dominating, and the effect of becoming dominated

15 of 164

Dominance Statuses

Dominating

(Weakly) Dominated

93% of assortments have at least one dominance relationship

Let’s take a closer look at the attributes of the focal gig here, and I’ll explain how I use them to operationalize dominance in this setting.

It’s average rating is 4.9 out of 5 stars, it has over 1,000 ratings (top censored at 1,000), and it’s price is $10. I’ll focus on these numeric attributes for defining dominance, though I’ll not ignore the all of the other non-numeric characteristics, in fact they’ll be very relevant later on; I’ll come back to that.

I’ll define this gig as dominating (or superior in a sense) when there exists another gig present in the assortment that has a lower average rating and a lower number of ratings, offered at the same price (and I relax the same-price restriction in a robustness check). When the gig on the right is present in the assortment, I’ll label the focal gig as dominating.

I’ll use an analogous definition to define this gig as dominated (inferior in a sense): when there exists another gig present in the assortment that has a higher average rating and a higher number of ratings – though in this example they’re actually equal, so this will be a case of weak dominance – offered at the same price. And I conservatively allow for cases of weak dominance in my analysis.

In real-life situations, it's also possible for a gig to be both dominated and dominating. In this example, the focal gig would be both dominating and dominated when the two gigs on the right are present in the assortment.

Lastly, there can be other gigs in the assortment – even priced at $10 – that have no dominance relationship with the focal gig. In this example, the gig on the right has a greater average rating, but lower number of ratings, so there is no dominance here.

Though the dominance effect is sometimes criticized for being a stylized effect that marketers cooked up in lab experiments and is rarely present in the real world, I actually find that it can be quite common. In fact, in the real-world Fiverr dataset I use, 93% of assortments have at least one dominance relationship, underscoring the need to better understand the relevance of the dominance effect outside the lab.

EXTRA:

One of the things we might worry about, particular when doing experimental work, is how relevant it is within the scope of this marketplace. I’ll show that dominance relationships can actually be quite prevalent naturally in the market.

16 of 164

No Dominance

64% of assortments

Dominated

6% of assortments

Dominating

29% of assortments

0.7%

0%

1.0%

40% increase

So how does the purchase likelihood of this example focal gig change, when it is in different dominance conditions?

Out of the 669 assortments of 12 gigs where appeared, it had no dominance in 64% of them – this is essentially the control condition – it was dominated in 6% of them, and was dominating in 29%.

In the ‘no dominance’ condition, this example gig was purchased 0.7% of the time. And when it was dominated, it was never purchased. What about when it was dominating? It was purchased 1% of the time! This represents (I think) a dramatic increase in purchase probability, about a 40% increase in conversation rate for this gig, simply by changing one of the other 11 options on the screen, and remember there’s also a no-purchase option as well. While this gig is just one example, in my analysis I will essentially repeat this across hundreds of thousands of gigs, and find a similar pattern of results.

EXTRA:

I’ll note that nearly all experiments that have studied dominance did not include a no-purchase option, which is generally present in the real world.

17 of 164

Experienceability of Categories – preference uncertainty

Experienceable

Non-Experienceable

Some product categories, like Articles and Blog Posts from the example earlier or Translation here on the screen, are non-experienceable: there is no way for sellers to provide any way for consumers to view or sample or experience their work, both on the browse page and the gig’s page. What’s kind of funny is this sellers put the logo to show where they went to school in their icon, presumably as a way to directly signal their quality to consumers. And since the sellers on Fiverr are not big brand names, and repeat purchases on the platform are rare, it’s particularly difficult for customers to make direct inferences about a particular gig’s quality and see if it matches their taste.

By contrast, there are other categories that are experienceable, like logo design on the right, where sellers can showcase various logos they’ve made (and you can click through them right on the browse page). Or for example in the voice over category, you can listen to a sample of the speaker’s voice.

So due to the standardized template that Fiverr uses for sellers in each product category, some product categories are experienceable while others are not, allowing me to test whether experience moderates the effect.

18 of 164

Pre-test: Experienceability and Preference Uncertainty

DV: How certain are you that the choice you selected is the right one for you?

(0 = not at all certain, 10 = completely certain)

We actually pre-tested this claim that customers, on average, face greater preference uncertainty in non-experienceable categories. On Qualtrics, I replicated the Fiverr browse page for the largest 6 product categories (3 experienceable and 3 non-experienable). It actually required quite a bit of time and I had to learn more HTML work to get it took look just right on Qualtrics. This is what it looked like in the articles & blog post category. Participants were asked which gig they would purchase, and how certain they were that the one they selected is the right one for them. I found that participants reported greater preference uncertainty in the non-experienceable product categories where they were not able to experience prior work from the seller.

Why would the ability to experience products reduce preference uncertainty?

Experience has been shown to be very engaging and can amplify differences among alternatives making the options more distinct and the choice easier.

19 of 164

Pre-test: Experienceability and Preference Uncertainty

N = 347

DV: How certain are you that the choice you selected is the right one for you?

(0 = not at all certain, 10 = completely certain)

P = .036

We actually pre-tested this claim that customers, on average, face greater preference uncertainty in non-experienceable categories. On Qualtrics, I replicated the Fiverr browse page for the largest 6 product categories (3 experienceable and 3 non-experienable). It actually required quite a bit of time and I had to learn more HTML work to get it took look just right on Qualtrics. This is what it looked like in the articles & blog post category. Participants were asked which gig they would purchase, and how certain they were that the one they selected is the right one for them. I found that participants reported greater preference uncertainty in the non-experienceable product categories where they were not able to experience prior work from the seller.

Why would the ability to experience products reduce preference uncertainty?

Experience has been shown to be very engaging and can amplify differences among alternatives making the options more distinct and the choice easier.

20 of 164

Data

Large dataset: 51.6 million observations, 4.3 million assortments

assortment of gigs
clicks and purchases
numeric attributes of the gigs
seller and customer characteristics

Timespan: October 1, 2018 to October 2, 2019
Variation: ~500,000 unique gigs; price range: $5-$20,000, median: $25
SMBs: ~80% of buyers are small and medium-sized businesses

21 of 164

22 of 164

23 of 164

Baseline Regression Equation

Purchase ~

Dominance Status x Experienceable +

Fixed Effects:

Gig ID

Country

Date

Position

Controls: Median price, SD price, N gigs at same price, buyer type, etc.

The empirical strategy works as follows: I observe the same gig in various dominance states or conditions (e.g., dominated, dominating, both, or neither), and I rely on that within-gig variation (by including gig id fixed effects in bold here) to identify the dominance effect. This is the approach that allows me to treat this as a perfect experiment. Because by holding a gig constant, but changing the assortment which randomly places the gig in the different dominance conditions, it allows me to estimate the effect of changes in dominance status on the purchase likelihood of a gig (the variable in bold now). And I’ve interacted it with the moderator, experienceable, to test if the dominance effect is moderated by experienceability.

And I ran this as a logit regression, where whether is a gig is purchased is the outcome. I also included controls like the median gig price in the assortment, the SD of the prices, the number of gigs offered at same price, buyer type, etc. And actually the results are the same whether or not the controls are included.

24 of 164

Baseline Regression Equation

Controls: Median price, SD price, N gigs at same price, buyer type, etc.

Purchase ~

Dominance Status x Experienceable +

Fixed Effects:

Gig ID

Country

Date

Position

The empirical strategy works as follows: I observe the same gig in various dominance states or conditions (e.g., dominated, dominating, both, or neither), and I rely on that within-gig variation (by including gig id fixed effects in bold here) to identify the dominance effect. This is the approach that allows me to treat this as a perfect experiment. Because by holding a gig constant, but changing the assortment which randomly places the gig in the different dominance conditions, it allows me to estimate the effect of changes in dominance status on the purchase likelihood of a gig (the variable in bold now). And I’ve interacted it with the moderator, experienceable, to test if the dominance effect is moderated by experienceability.

And I ran this as a logit regression, where whether a gig is purchased is the outcome. I also included controls like the median gig price in the assortment, the SD of the prices, the number of gigs offered at same price, buyer type, etc. And actually the results are the same whether or not the controls are included.

25 of 164

Neither dominated nor dominating

N = 51.6 million

Customers = 1.2 million

I’ll display the results in plots like these, where the y-axis captures the odds ratios.

Here’s how to interpret it: we’re going to compare the odds of purchasing a gig when it is dominated, dominating, or both, relative to the odds of purchasing it when it is not dominated or dominating (captured in the dashed horizontal line). A value greater than 1 indicates an increase in the odds of purchase, and a value less than one represents a decrease in the odds of purchase, again relative to the control condition when a gig’s status is not dominated or dominating.

I’ll start by describing the results for non-experienceable categories. What happens to the odds of purchasing a gig when it’s status changes to dominating? Well, if those who question whether dominance can be found in the wild are correct, then it should not not be statistically different from the dashed horizontal line.

26 of 164

Dominance Effect

N = 51.6 million

Customers = 1.2 million

27 of 164

N = 51.6 million

Customers = 1.2 million

28 of 164

Magnitude of Dominance

>>

≳

29 of 164

Magnitude of Dominance

>>

≳

Perceptual

Simple, automatic processing

(Simonson 1989; Pocheptsova et al. 2009)

Similarity

Similarity drives cognitive assessments – pairwise comparisons

(Hamilton et al. 2007)

30 of 164

Magnitude of Dominance

Sometimes, by random chance, a dominated and dominating gig will be very similar in attribute space (in average rating and number of ratings) and other times they’ll be further apart. I tested whether increasing the degree or magnitude by which the dominating option is better than the dominated option moderates the effect.

I’m interested in this because different theories that make different predictions for the direction of the effect here. Some argue that the effect should be strongest when the dominated and dominating option are most similar – because they’re similar they’re mostly likely to be directly compared, while others predict the opposite – that the effect is strongest when they’re farther apart in attribute space.

The x-axis here captures the difference in the number of ratings between the dominated and dominating gig, and the y-axis captures the difference in average rating.

EXTRA:

For example, a perceptual mechanism suggest that dominated alternative makes the dominating one look (or seem) better, where a stronger contrast make the dominating gig appear even better.

The original authors found null effects: Huber et al. 1982, Huber and Puto 1983. More recent experimental work (Padamwar et al. 2021) has found that range extensions amplified the effect.

31 of 164

Magnitude of Dominance

32 of 164

Magnitude of Dominance

33 of 164

34 of 164

Visual Distance & Count of Dominated Gigs

35 of 164

Robustness Checks

Linear probability models

Category-level models

Dominance on price

Clicks

Non-WEIRD populations

Lab experiment

36 of 164

Research Questions

Is the dominance effect real?

YES

Are dominance relationships prevalent?

YES

How does the effect work?

Perceptual vs. Cognitive

In this work, I'll start by using a very large-scale scale field experiment that ask: is this effect real?

I will also shed light on how prevalent dominance relationships are – they can actually be quite common in the real-world.

And I’ll also use the rich field data I have to shed light on the psychology driving this effect – telling you when to expect to find it in the field.

EXTRA:

My last hypothesis, which I’ll only get in to a little here, and in a longer talk I’d love to talk more about it, is that the effect is consistent with a perceptual mechanism underlying it, where the dominating alternative simply “seems better” in the presence of a dominated gig, and the findings help to rule out some alternative mechanisms that have been proposed in the literature.

I want to point out that this hypothesis is not always considered by researchers in the field: for example, Frederick et al. 2014 attempted to use more realistic stimuli, but in using products people knew and had likely purchased before and experienced (like apples and oranges), they also likely had low preference uncertainty – in other words they knew their preferences.

37 of 164

Price contrast in the wild

(with Ariel Fridman and Karsten Hansen)

38 of 164

39 of 164

40 of 164

95th percentile

41 of 164

42 of 164

Research Questions

Contrast or Assimilation?

Past research predicts contrast, but positive or negative?

Do the behavioral patterns shed light on the underlying psychology?

Trust/confidence vs. reference point shift?

(Affective/Holistic)

(Range-Frequency)

43 of 164

$5

$10

$20

$10

$20

$25

$30

$X

224 times

44 of 164

$5

$10

$20

$10

$20

$25

$30

-

$150

147 times

2.7%

$5

$10

$20

$10

$20

$25

$30

$170

-

$875

77 times

0%

45 of 164

Data

Large dataset: 51.6m observations, 4.3m assortments, 1.2m relevant assortments

assortment of gigs
clicks and purchases
numeric attributes of the gigs
seller and customer characteristics

Timespan: October 1, 2018 to October 2, 2019
Variation: ~500,000 unique gigs; price range: $5-$20,000, median: $25
SMBs: ~80% of buyers are small and medium-sized businesses

46 of 164

47 of 164

48 of 164

49 of 164

Baseline Regression Equation

Purchase from assortment ~

Extreme Condition x Experienceable +

Fixed Effects:

Price Distribution (prices 1-11) x Category

Date

The empirical strategy works as follows: I observe the same gig in various dominance states or conditions (e.g., dominated, dominating, both, or neither), and I rely on that within-gig variation (by including gig id fixed effects in bold here) to identify the dominance effect. This is the approach that allows me to treat this as a perfect experiment. Because by holding a gig constant, but changing the assortment which randomly places the gig in the different dominance conditions, it allows me to estimate the effect of changes in dominance status on the purchase likelihood of a gig (the variable in bold now). And I’ve interacted it with the moderator, experienceable, to test if the dominance effect is moderated by experienceability.

And I ran this as a logit regression, where whether is a gig is purchased is the outcome. I also included controls like the median gig price in the assortment, the SD of the prices, the number of gigs offered at same price, buyer type, etc. And actually the results are the same whether or not the controls are included.

50 of 164

All data

Purchase from assortment ~

Extreme Condition x Experienceable +

Fixed Effects:

Price Distribution (prices 1-11) x Category

Date

51 of 164

All data

52 of 164

Clean data

53 of 164

Buyer experience

54 of 164

Buyer experience

55 of 164

Buyer experience

56 of 164

Extremeness

57 of 164

Extremeness

58 of 164

Extremeness

*** (P < .001)

P = .35

59 of 164

Different specifications

Purchase from assortment ~

Extreme Condition x Experienceable +

Median Price x Category

IQR Price x Category

Fixed Effects:

Date

60 of 164

Robustness Checks

Linear probability models

Excluding largest categories (Logo Design and Translation)

Entire dataset (median, IQR controls)

Outcome: purchase on prices 1-11

Non-WEIRD populations

61 of 164

What about average selling price?

62 of 164

*Excludes one outlier at $332 in non-experienceable

63 of 164

ASP�~-4%

64 of 164

LTV: Carryover or persistent effect

65 of 164

Data: all

Outcome:

any future purchase

SEs: Clustered by user

66 of 164

Data: all

SEs: Clustered by user

Outcome:

any future purchase

67 of 164

Data: all

SEs: Clustered by user

Outcome:

any future purchase

*** (P < .001)

P = .19

68 of 164

Data: future purchase

SEs: Clustered by user

Outcome:

log future value

69 of 164

Research Questions

Contrast or Assimilation?

Past research predicts contrast, but positive or negative?

As uncertainty increases – negative!

Do the behavioral patterns shed light on the underlying psychology?

Trust/confidence* vs. reference point shift?

* Heterogeneity

(Affective/Holistic)

(Range-Frequency)

70 of 164

Robustness – 2022 Data

Large dataset: 37m assortments; 3.6m relevant assortments

assortment of gigs
clicks and purchases
numeric attributes of the gigs
seller and customer characteristics
more than twice the number of sub-categories
New exogenous list of experiencable and non-experiencable categories

Timespan: Jan 1, 2022 to June 31, 2022
Variation: price range: $5-$20,000
SMBs: ~80% of buyers are small and medium-sized businesses

71 of 164

All assortments

72 of 164

Only first 8

~500,000 assortments

73 of 164

“The wild” opens a window to cultural differences and other factors relevant for diversity, equity and inclusion

74 of 164

Communication style

(with Gal Smitizsky)

75 of 164

Communication style theories

Each country gets a score between 0 and 100 for different cultural factors, based on large amount of lab research

Some examples of cultural factors :

Communication style (direct / indirect)

Evaluation habits (direct / indirect)

Decision habits (consensual / top-down)

Trust habits (task / relationship based)

76 of 164

Direct communication

Good communication is precise, simple and clear. Messages are expressed and understood at face value. Repetition is appreciated if it helps clarify the communication.

Responsibility for understanding: The speaker.

Indirect communication

Good communication is sophisticated, nuanced and layered. Messages are both spoken and read between the lines. Messages are often implied but not plainly expressed.

Responsibility for understanding: The listener.

Indirect

Direct

77 of 164

Indirect

Good communication is sophisticated, nuanced and layered

Direct

Good communication is precise, simple and clear.

78 of 164

Sufficient variability

US

Canada

UK

Australia

94

89

67

94

France

Italy

Germany

Portugal

Brazil

Spain

Mexico

Netherlands

35

39

75

44

40

84

43

37

79 of 164

80 of 164

Cancellation rate

Buyer comm. score

AVG cancellation rate

81 of 164

Relational mobility – no significant effect

82 of 164

Buyer comm. score

Conversion rate

Conversion from conversation

83 of 164

Cancellations after Conversion from conversation

84 of 164

��We find support for the lab results even when aggregated by country: �Asking questions facilitates trade�(non-verbal cues absent in digital commerce)��Platforms should accommodate cultural differences

85 of 164

Dark side of discrimination awareness

(with Jean Zhang and Elizabeth Campbell)

86 of 164

Discrimination awareness

Gender discrimination is real

People are aware (especially women)

This drives expectations of successful women up!

Is this a good thing?

87 of 164

Expectancy disconfirmation theory - Revision requests

Buyers can reject a delivery and ask for a revision

The theory suggests this pattern would follow expectations

Carries real economic costs

Do gender-based expectations lead to differential revision requests?

88 of 164

Variable	(1) Logit, FE	(2) Logit, FE + Price
Male Buyer	-0.043	-0.041
Male Buyer	(0.041)	(0.022)
Male Seller	0.048	0.064*
Male Seller	(0.039)	(0.028)
Pro	0.810**	0.687***
Pro	(0.256)	(0.197)
Extra Revisions Purchased	0.444**	0.430***
Extra Revisions Purchased	(0.166)	(0.067)
Rating Level	0.010***	0.010***
Rating Level	(0.001)	(0.002)
Number of Past Rated Orders	0.000	0.000
Number of Past Rated Orders	(0.000)	(0.000)
Male Buyer x Male Seller	-0.029	-0.042
Male Buyer x Male Seller	(0.048)	(0.026)
Male Buyer x Pro	-0.469*	-0.543*
Male Buyer x Pro	(0.238)	(0.236)
Male Seller x Pro	-1.147***	-1.115***
Male Seller x Pro	(0.327)	(0.290)
Male Buyer x Male Seller x Pro	0.768*	0.878**
Male Buyer x Male Seller x Pro	(0.317)	(0.306)

Log-Likelihood	-236,361.1	-235,576.5
Adj. Psuedo R²	0.077783	0.080045
BIC	496,046.2	495,919.1
Squared Cor.	0.102208	0.105509
Observations	397,435	397,435

89 of 164

Variable	(3) Logit, FE	(4) Logit, FE+Price
Male Buyer	-0.075***	-0.075***
Male Buyer	(0.021)	(0.026)
Male Seller	0.215***	0.213***
Male Seller	(0.019)	(0.013)
Female-Male Sub-Category Ratio	-0.304***	-0.344***
Female-Male Sub-Category Ratio	(0.027)	(0.023)
Extra Revisions Purchased	0.556***	0.536***
Extra Revisions Purchased	(0.031)	(0.029)
Rating Level	0.001***	0.001***
Rating Level	(0.000)	(0.000)
Male Buyer x Male Seller	-0.089***	-0.110***
Male Buyer x Male Seller	(0.018)	(0.024)
Male Buyer x Female-Male Sub-Category Ratio	0.019	0.025
Male Buyer x Female-Male Sub-Category Ratio	(0.021)	(0.027)
Male Seller x Female-Male Sub-Category Ratio	-0.482***	-0.432***
Male Seller x Female-Male Sub-Category Ratio	(0.017)	(0.010)
Male Buyer x Male Seller x Female-Male Sub-Category Ratio	0.195***	0.120***
Male Buyer x Male Seller x Female-Male Sub-Category Ratio	(0.009)	(0.011)

Log-Likelihood	-256,964.8	-255,507.4
Adj. Psuedo R2	0.017241	0.021991
BIC	534,524.2	533,080.5
Squared Cor.	0.028943	0.035466
Observations	404,949	404,949

90 of 164

Expectancy disconfirmation theory - Revision requests

Buyers can reject a delivery and ask for a revision

The theory suggests this pattern would follow expectations

Carries real economic costs

Do gender-based expectations lead to differential revision requests?

YES

91 of 164

This is where the field started!

Collaboration across research methods

92 of 164

Thank you!

93 of 164

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Tipping in a Digital Services Marketplace

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

University of California, San Diego

Tipping in a Digital Services Marketplace

94 of 164

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Motivating Question

Why do People Tip?

Why do People Tip in the Gig Economy?

Tipping in a Digital Services Marketplace

95 of 164

Motivation

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Tips are voluntary, post-transaction payments that can motivate service effort, enable price discrimination and help retain talented workers (Lynn 2008)

16% of Americans surveyed have earned money from online gig platforms, 31% of whom said it was their main job within the past year (Pew Research 2021)

Digital services usually feature no face-to-face interaction or observational learning, so tipping practices are opaque.

Digital market design affects consumption and production behaviors, but design effects can be hard to discern, even in thick digital data

Tipping in a Digital Services Marketplace

96 of 164

Research Questions

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Report key descriptives about online tipping behavior

^▶Describe 341K tips among 4.1MM transactions

Does digital marketplace design affect tipping? Why and how?

^▶Field experiment manipulated the standard tip suggestion: Reciprocity vs Norms

New buyers vs. repeat buyers
What can we learn about the mechanism?

Tipping in a Digital Services Marketplace

97 of 164

Empirical Context: �Online Marketplace for Freelancer Digital Services

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Search Buyer explores service listings

Order Buyer contacts seller(s) about service attributes, price, delivery date, reqs. Buyer pays platform upon ordering

Delivery Seller delivers the work via platform. Buyer confirms receipt, platform pays seller

Rating and Tip Platform asks buyer to rate, then asks buyer to tip

Tipping in a Digital Services Marketplace

98 of 164

99 of 164

100 of 164

Pre-treatment Data

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

4.1MM transactions between 1.3MM buyers and 171K sellers

^▶$152MM in total spending, inc. $3.9MM in 341K tips

Period: 01/01/2019 ∼ 06/09/2019 Buyer, seller & transaction characteristics Tipping measures:

^▶Tipping Rate: the proportion of transactions that were tipped

^▶Average Tip Percentage: the average % of price among tips observed

Tipping in a Digital Services Marketplace

101 of 164

Tipping Rate and Tip Percentages

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

(a) Tipping Rate (b) Tip Percentages

Buyers tipped sellers in 8.3% of transactions overall

Among tipped transactions, the average tip observed is 51.8% of the order price

Tipping in a Digital Services Marketplace

102 of 164

Tipping and Other Factors

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Tipping covaries with Price

^▶High ATP is mainly driven by the low prices

(a) Tipping Rate

(b) Average Tip Percentage

Tipping in a Digital Services Marketplace

103 of 164

Tipping and Buyer & Seller Region

Observed Fact #1 - Online tipping behavior is associated with regional tipping culture

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

(a) Tipping Rate

(b) Average Tip Percentage

North Americans tipped in 11.5% of all transactions ATP varies much less across buyer regions

Tipping in a Digital Services Marketplace

104 of 164

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Tipping and Buyer & Seller Region

Observed Fact #1 - Online tipping behavior is associated with regional tipping culture

Little or no “Home-region Bias” or “Tourist Effect”

Tipping in a Digital Services Marketplace

105 of 164

Tipping and Transaction Quality

Observed Fact #2 - Tipping covaries with the quality of the transaction

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

(a) Tipping Rate (b) Average Tip Percentage

TR was four times higher among 5-star than 3-star transactions ATP did not change as much with buyer rating

Tipping in a Digital Services Marketplace

106 of 164

Tipping and Transaction Quality

Observed Fact #2 - Tipping covaries with the quality of the transaction

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

(a) Tipping Rate (b) Average Tip Percentage

TR increases with sellers’ average rating, but ATP did not vary

ATP did not vary across sellers’ average rating

Tipping in a Digital Services Marketplace

107 of 164

Tipping and Buyers’ Characteristics

Observed Fact #3 - Buyer’s previous tipping behavior and previous platform behavior both predict tipping

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

(a) Tipping Rate (b) Average Tip Percentage

Buyers’ tipping history predicts tipping

ATP varies much less with buyer prior tips than tipping rate

Tipping in a Digital Services Marketplace

108 of 164

Tipping and Buyers’ Characteristics

Observed Fact #3 - Buyer’s previous tipping behavior and previous platform behavior both predict tipping

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

(a) Tipping Rate (b) Average Tip Percentage

TR generally increases with buyer tenure up to a peak at 3 years

ATP is nearly flat among newer buyers but then rises with buyer tenure among older cohorts

Tipping in a Digital Services Marketplace

109 of 164

Tipping and Other Factors

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Tipping covaries with Price

^▶5% of $5 transactions are tipped, rising to 11% of $30-$40 transactions, and then falling back to 9% of $150+ transactions

^▶ATP is nearly constant at the 20% default for prices above $25

Tipping covaries with Seller Tip Mention

^▶Tipping Rate increased with seller tip mentions from 8.1% to 13%

^▶ATP did not change

Tipping covaries with the type of service provided

^▶Tipping Rate was 10.6% in the Graphics Design category, the first-most frequent category

^▶Digital Marketing was the second-most frequent category had a Tipping Rate of just 2.6%

^▶ATP did not change across the service category

Tipping in a Digital Services Marketplace

110 of 164

Field Experiment Design - Message

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Period: 06/10/2019-07/07/2019

Randomly assign 45,497 repeat buyers and 7,880 new buyers Treatment revealed after buyer rating, then held constant

Control Group: “Would You Like to Leave a Tip to (Sellername)?”

Implicit Reciprocity: “Show your appreciation to your seller by giving a tip” Reciprocity: “Leave (Sellername) a tip to show your appreciation for a job well done” Norm: “It’s customary to leave a tip for the seller’s service”

Tipping in a Digital Services Marketplace

111 of 164

Field Experiment Design - Default Tip Scale

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Platform simultaneously changed the default tips in the Reciprocity and Norms conditions New default tip scale contained three options rather than two, with custom tip

Tipping in a Digital Services Marketplace

112 of 164

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Treatment Examples

Tipping in a Digital Services Marketplace

113 of 164

Pre-test: Testing the anticipated theoretical constructs of messages

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Method

Participants from Amazon Mechanical Turk were randomly assigned to one of eight conditions (N = 1000)

4 x Message (control [status-quo], implicit reciprocity, explicit reciprocity, and norms) by 2 x Seller name (present vs. not present)

Six reasons presented followed Cialdini’s six “weapons of influence” (Reciprocity/ Social Proof/ Commitment/ Authority/ Liking/ Scarcity)

All participants were asked to mark all explanations that fit how they interpreted each tip message.

Tipping in a Digital Services Marketplace

114 of 164

Pre-test: Testing the anticipated theoretical constructs of messages

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Control (Status Quo) message motivates multiple interpretations, and no distinct one emerges Participants perceived the three treatment messages as intended

Tipping in a Digital Services Marketplace

115 of 164

Pre-test: Testing the anticipated theoretical constructs of messages

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Prevalent perception of all messages does not change regardless of the inclusion of the seller’s

name

Experiment Design

Tipping in a Digital Services Marketplace

116 of 164

Does the language of the request for a tip make a difference?

New buyers vs. repeat buyers

117 of 164

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Nonparametric Treatment Effect on New Buyers

Tipping in a Digital Services Marketplace

118 of 164

Specification for New Buyers

Fixed Effect Regression (pooled)

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Tipping in a Digital Services Marketplace

119 of 164

Specification for New Buyers

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Fixed Effects Regression

Estimate treatment effects separately

All test-period transactions
New buyers’ first transactions
New buyers’ second transactions
New buyers’ third transactions

Tipping in a Digital Services Marketplace

120 of 164

New Buyers Tipping Regression

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Pool (1)

First (2)

Second (3)

Third (4)

Implicit Reciprocity

−0.00001

(0.006)

0.006

(0.006)

0.027^∗∗∗(0.006)

0.257^∗∗∗(0.023)

(0.009)

0.003

(0.009)

Reciprocity

Norms

−0.009

0.010 −0.017

(0.011)

0.006

(0.011)

0.069^∗∗∗(0.012)

0.010

(0.012)

0.013

(0.012)

0.016

(0.011)

0.361^∗∗∗(0.037)

(0.009)

0.318^∗∗∗(0.030)

0.003 −0.030

(0.041)

I previous tip

I previous tip:Implicit Reciprocity

−0.021

(0.033)

(0.053)

I previous tip:Reciprocity

−0.033

−0.036

−0.032

(0.032)

0.055

(0.031)

(0.041)

0.111^∗∗∗(0.039)

(0.052)

0.025

(0.050)

I previous tip:Norms

Buyer Country Seller Country # of Buyers

Observations _R2

Adjusted R²

Y Y

7,880

26,436

0.172

0.162

Y Y

7,880

0.102

0.070

Y Y

7,258

0.264

0.235

Y Y

3,545

0.321

0.270

Note:

^∗∗p<0.05; ^∗∗∗p<0.01

Tipping in a Digital Services Marketplace

121 of 164

Limitation

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Message treatment effect might be confounded with scale treatment effect

To address this concern, we use post-treatment period data

The platform back to show the Control messages with the new default tip scale for a few weeks

	Control	Treatment
Test	Control Message + Three Default	Treatment Message + Four Default
Post Test	Control Message + Four Default	Control Message + Four Default

Table: Platform Design across Time

To identify the effect of the new default tip scale and the (reverse-)message effect, we use difference-in-differences with buyer fixed effects, seller fixed effects, and transaction characteristics

Tipping in a Digital Services Marketplace

122 of 164

Specification for New Buyers

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Tipping in a Digital Services Marketplace

123 of 164

Specification for New Buyers

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

β₁captures the effect of the new default tip scale among Status-Quo message

β^Mⁱcaptures the (reverse-)treatment effect by comparing the change in Tipping Rate

2

across time within treatment groups

Assume the additive specification (treatment effect = message + default tip scale)

Tipping in a Digital Services Marketplace

124 of 164

Difference-in-Differences Result

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Tipping Rate

Post

0.026^∗∗(0.012)

Implicit Reciprocity:Post

−0.015

(0.016)

Reciprocity:Post

−0.055^∗∗∗

(0.017)

Norms:Post

−0.041^∗∗

(0.017)

Observations

# of Buyers _R2

Adjusted R²

20,023

2,928

0.786

0.458

Note:

^∗∗p<0.05; ^∗∗∗p<0.01

Effect of the new default tip scale among Status-Quo message was significantly positive to new buyers

Treated new buyers during test period (i.e., Reciprocity, Norm) were less likely to pay tip after platform gave

Status-Quo message

Both ‘message’ and ‘default tip scale’ design could change online tipping behavior of new buyers

Tipping in a Digital Services Marketplace

125 of 164

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Nonparametric Treatment Effect on Repeat Buyers

Tipping in a Digital Services Marketplace

126 of 164

Difference-in-Differences

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Tipping in a Digital Services Marketplace

127 of 164

Difference-in-Differences

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Estimate treatment effects separately

All test-period transactions
Repeat buyers’ first transactions
Repeat buyers’ second transactions
Repeat buyers’ third transactions

Tipping in a Digital Services Marketplace

128 of 164

Repeat Buyer Tipping Regression

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Pool (1)

First (2)

Second (3)

Third (4)

Implicit Reciprocity:POST

−0.001

0.002

(0.004)

0.006

(0.004)

0.020^∗∗∗(0.004)

0.002

(0.003)

−0.006

(0.002)

0.001

(0.002)

0.004

(0.002)

(0.004)

0.002

(0.004)

0.005

(0.004)

0.002

(0.003)

(0.005)

Reciprocity:POST

−0.004

(0.005)

Norms:POST

−0.005

(0.005)

0.0002

(0.004)

POST

−0.0002

(0.002)

Buyer FE	Y	Y	Y	Y
Seller FE	Y	Y	Y	Y
Category FE	Y	Y	Y	Y
# of Buyers	40,823	40,823	34,871	22,226
Observations	773,979	640,714	584,083	466,511
_R2	0.544	0.562	0.562	0.569
Adjusted R²	0.454	0.462	0.461	0.470
Residual Std. Error	0.211 (df = 646783)	0.210 (df = 521970)	0.204 (df = 474938)	0.195 (df = 378992)
Note:				^∗∗p<0.05; ^∗∗∗p<0.01

Tipping in a Digital Services Marketplace

129 of 164

Limitation

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Like New buyers case, message treatment effect might be confounded with scale treatment effect

To address this concern, we use post-treatment period data

	Control	Treatment
Test	Control Message + Three Default	Treatment Message + Four Default
Post Test	Control Message + Four Default	Control Message + Four Default

Table: Platform Design across Time

To identify the effect of the new default tip scale and the (reverse-)message effect, we use difference-in-differences with buyer FE, seller FE, and transaction characteristics

Tipping in a Digital Services Marketplace

130 of 164

Specification for Repeat Buyers

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Tipping in a Digital Services Marketplace

131 of 164

Difference-in-Differences Result

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Tipping Rate

POST

−0.003

(0.002)

Implicit Reciprocity:POST

−0.001

(0.003)

Reciprocity:POST

−0.003

(0.003)

Norms:POST

−0.003

(0.003)

Buyer FE Seller FE Observations

# of Buyers _R2

Adjusted R²

Y Y

236,072

26,892

0.660

0.504

Note:

^∗∗p<0.05; ^∗∗∗p<0.01

Effect of the new default tip scale among Status-Quo message was negative but not significant

Tipping Rate was directionally negative for treated repeat buyers during the test period compared to untreated repeat buyers

The change in platform design is less likely to change market behaviors of repeat buyers

Tipping in a Digital Services Marketplace

132 of 164

Potential Moderators

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Study why and how the treatment effect increased tipping by interacting treatment effects with potential moderators related to observed facts

Regional tipping culture: North America vs. other regions (I NorthAmerica)

^▶Observed fact #1 showed that buyers from countries with established tipping culture are more likely to tip than those without

^▶This fact indicates that regional tipping norms could motivate online tipping

^▶Test whether the regional tipping culture could moderate the treatment effect

5-star satisfaction rating (I 5star)

^▶Observed Fact #2 shows that a positive correlation between tipping and service quality

^▶It implies that buyers are more likely to reward sellers for the high-quality service

^▶Test whether the high service quality could moderate the treatment effect

Tipping in a Digital Services Marketplace

133 of 164

How treatment effects vary with the two moderators for new buyers

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Regional tipping culture: North America vs. other regions

^▶The interaction between North American dummy and norm treatments are not significant, but positive

5-star satisfaction rating

^▶The interaction between satisfaction and norm treatment is positive and significant

^▶Satisfaction moderates the effect of norm treatment

Result

Tipping in a Digital Services Marketplace

134 of 164

How treatment effects vary with the two moderators for repeat buyers

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Regional tipping culture: North America vs. other regions

^▶The interaction between North American dummy and norm treatments are not significant, but positive

^▶The interaction between North American dummy and both Reciprocity treatments are not significant, and negative

5-star satisfaction rating

^▶Two-way interaction terms with Norm treatment is not significant, but positive

^▶Two-way interaction terms with both Reciprocity treatments are not significant, and negative

Result

Tipping in a Digital Services Marketplace

135 of 164

Takeaways

Seung Hyun Kim, On Amir, Kenneth C. Wilbur

Online tipping behavior - descriptives

^▶Tipping likelihood is more situational than tip percentages

^▶Buyer factors drive tipping; Buyer Region, Buyer Satisfaction and Buyer previous tipping/platform behavior

Does digital marketplace design affect tipping? Why and how?

^▶Yes. An injunctive Norm message increased both new buyers’ and repeat buyers’ tipping

^▶Analysis with post-treatment data shows that both the message’ and the ‘default tip scale’ design could change online tipping behavior of new buyers

^▶The Norms treatment effect increased with buyer satisfaction

^▶Other treatment messages (e.g., [Implicit] Reciprocity) did not change buyer behaviors

Tipping in a Digital Services Marketplace

136 of 164

Thank you!

137 of 164

Lab Experiment

Field Evidence

Discussion

Lab Evidence

Questions

Motivation

138 of 164

Methods

Choice set of songs

Recruited students from UCSD Behavioral Lab (N = 395 after exclusions)

Conditions: non-experienceable, experienceable

Data, analysis code, research materials: https://osf.io/8dasb/?view_only=973dbd5f9d204316ad0fcc7ffbe8e2ea

Field Evidence

Discussion

Lab Evidence

Questions

Motivation

I recruited students to the behavioral lab, presented them with a choice set of songs, and asked them to indicate which they would purchase. The songs were actually created by a musician for this experiment. We mimicked the setup of Fiverr where we gave participants information about the price, ratings, and number of ratings. Those 3 attributes were all the information they had in the non-experienceable condition, and the in the experienceable condition they were also able to listen to a 30-second sample. Pre-registration and open science is very important to me – this experiment was pre-registered, and all the materials and data are posted on OSF.

In the interest of time, I’ll skip over the details of how we ran the experiment, but similar to Fiverr I experimentally manipulated whether songs were dominating.

139 of 164

Song Attributes

A’ makes A dominating, B’ makes B dominating

	Song A’	Song A	Song B	Song B’
price	95 cents	90 cents	70 cents	80 cents
average rating (out of 5)	4.8	4.9	4.7	4.6
number of ratings	96	119	135	82

Randomized

Experienceable

Field Evidence

Discussion

Lab Evidence

Questions

Motivation

140 of 164

Song Attributes

A’ makes A dominating, B’ makes B dominating

	Song A’	Song A	Song B	Song B’
price	95 cents	90 cents	70 cents	80 cents
average rating (out of 5)	4.8	4.9	4.7	4.6
number of ratings	96	119	135	82

Randomized

Experienceable

Field Evidence

Discussion

Lab Evidence

Questions

Motivation

141 of 164

Song Attributes

A’ makes A dominating, B’ makes B dominating

	Song A’	Song A	Song B	Song B’
price	95 cents	90 cents	70 cents	80 cents
average rating (out of 5)	4.8	4.9	4.7	4.6
number of ratings	96	119	135	82

Randomized

Experienceable

Field Evidence

Discussion

Lab Evidence

Questions

Motivation

142 of 164

Results: Choice of Dominating Song

Field Evidence

Discussion

Lab Evidence

Questions

Motivation

P < .001

143 of 164

Discussion

Real and prevalent

Experienceability

Perceptual mechanism

Naturalistic setting

Large assortment

Actionable

Field Evidence

Lab Evidence

Discussion

Questions

Motivation

This work is among the first to find evidence for dominance effects in a real and consequential setting – and that dominance relationships can be prevalent in such settings.

Converging evidence from field and lab experiments showed that the ability to sample or experience alternatives is a moderator for the effect

I only briefly discussed it in this short talk, but the results are consistent with a perceptual mechanism underlying the effect

I found the effect in naturalistic conditions where customers had the option not to purchase any items in the assortment. This setting also had a relatively large assortment of 12 alternatives, instead of the typical 3.

These findings have broad implications for marketers, choice architects, user interface designers, and policymakers. When bringing products to customers, especially new ones that consumers haven’t experienced before, sellers can increase its purchase share of a particular target by introducing a dominated option (or several) to the choice set.

144 of 164

Clean data

145 of 164

Other Real-World Evidence

Three papers tested for dominance effects in field settings:

Retracted: Li, Sun, and Chen (2019)

Cannot isolate the dominance effect from other context effects

Cannot observe choice set, only market-level data

Lab-in-the-field evidence, small sample

146 of 164

Control Variables

Variable Name	Description
Set Characteristics
Median price	Median price of gigs in set
SD price	Standard deviation of the prices of the gigs in set
Minimum price	Minimum price of gigs in set
Maximum price	Maximum price of gigs in set
N gigs at same price	Number of gigs at the same price in the set

Variable Name	Description
Buyer Characteristics
Total Orders Lifetime	Count of orders placed by the buyer
Total Orders Cancellation	Count of orders canceled by the buyer
Buyer Type	Type of buyer: guest, registered not converted (rnc), first time buyer (ftb), second time buyer (stb), repeat buyer, unknown
Operating System	Operating system of buyer: Windows, Mac, Linux, Chrome OS, iOS, Anrdoid, other, unknown
RFM Segmentation	Recency, Frequency and Monetary segmentation of buyer: A, B, C, D, E, OTB, unknown

147 of 164

Count of Dominated Gigs &

Visual Distance

Lab Evidence

Discussion

Questions

Field Evidence

Motivation

148 of 164

Dominance on:

average rating, number of ratings

average rating, number of ratings, price

Lab Evidence

Discussion

Questions

Field Evidence

Motivation

149 of 164

Logit

Linear Probability

Lab Evidence

Discussion

Questions

Field Evidence

Motivation

150 of 164

Category-level Models

Top 6 Categories account for:

57% of searches
57% of clicks
68% of purchases

Lab Evidence

Discussion

Questions

Field Evidence

Motivation

151 of 164

Purchase

Click

Lab Evidence

Discussion

Questions

Field Evidence

Motivation

152 of 164

Advanced

Developing

Lab Evidence

Discussion

Questions

Field Evidence

Motivation

153 of 164

Lab Evidence

Discussion

Questions

Field Evidence

Motivation

154 of 164

Position

Matters

155 of 164

Clicks per Set

156 of 164

Conceptual Assortment

157 of 164

Shaded box: Dominated Options

158 of 164

Ideal Experiment

92% of sets contain a dominating gig

159 of 164

Dominated Relationship Only

160 of 164

Dominated Relationship Only

161 of 164

Preference Uncertainty Experienceability Assumption:

A Manipulation Check

When customers face greater preference uncertainty: more clicks before making a purchase

Clicks would be used to develop and learn their own preferences

In sets with no dominance:

Experienceable: 23.7 clicks for every purchase
Non-experienceable: 35.1 clicks for every purchase (p < .001)

162 of 164

Mixed Effects Models

Mirrored results from fixed-effects models

163 of 164

Choices Shares

Dominating option very rarely chosen in non-experienceable

Dominance Effect

164 of 164

Future Work

What about other context effects? E.g. Contrast Effect

Formation of consideration sets and demand

How much does the unstructured data for each gig matter?