The Impact of Increase in Minimum Wages on Consumer Perceptions of Service:

A Deep Learning Analysis of Online Restaurant Reviews

Abstract

We study the impact of a mandated increase in minimum wages on consumer perceptions of multiple dimensions of service quality in the restaurants industry. When faced with a mandated increase in minimum wages, firms might reduce the number of workers leading to poor consumer service. Alternatively, more motivated workers might improve consumer service. Based on textual data from 533,022 online reviews for 2,870 restaurants, we estimate a deep learning Word2Vec model to extract service quality perceptions. We then exploit a natural experiment in the County of Santa Clara, wherein only the city of San Jose legislated a 25% minimum wage increase in 2013. Using measures of perceived service quality as the dependent variable, we estimate a difference-in-difference model. We find an improvement in perceived service quality of San Jose restaurants in terms of reduced delays, slowness and unfriendliness. This decrease is restricted to independent restaurants and does not extend to chains. This finding is consistent with agency theory based predictions of greater incentives to improve service in independent restaurants. We discuss implications for consumers, restaurants and policy makers.

Key words: Service quality, minimum wages, deep learning methods, text analysis, natural experiments, agency theory.

1. Introduction

Most countries in the world including the United States, China, Japan and Germany have

long legislated minimum hourly wage. The purpose of such legislation is to protect workers

against unduly low pay. It also serves as one of several policy tools to potentially reduce

inequality and poverty (International Labor Organization, 2015). The impact of such legislation

on economic and labor outcomes have been studied by academics and policy makers for decades.

Across several industries, geographies and time periods, minimum wage legislation has been

shown to impact inequality (Autor, Manning and Smith 2016), employment levels (Aaronson

and French, 2007), firm value (Bell and Machin 2018), other wages (Grossman 1983), and non-

wage benefits (Bhorat, Kanbur and Stanwix 2014). Despite this rich body of research, there is

lack of consensus about the directions and magnitudes of these effects. Across academia, media,

policy makers, and the general public, there are strong differences in opinion about the

advantages and drawbacks of such legislation. In contrast to the positive view that minimum

wages reduce poverty and inequality, improve living standards, and boost morale, opponents

opine that setting minimum wages increases unemployment especially of unskilled and

inexperienced workers, and increases the cost of doing business.

Even though minimum wage legislation has been intensely scrutinized and debated, there

is very little academic research or discussion on how such legislation might impact the behavior

or experiences of consumers who interact with minimum wage workers. We investigate the

impact of increases in minimum wage increases by thousands of firms on various dimensions of

the perceived service quality of those firms, as captured by the opinions of the firms’ consumers.

Especially, we view these effects as another angle to consider when measuring the economic

impact of minimum wage.

Our empirical research context is the US restaurants industry. Direct labor costs are about

30% of total operating costs of US restaurants, and about 33% of restaurant workers are paid

within 10% of the minimum wage. Two-thirds of workers in the United States, earning the

minimum wage or less in 2016, were employed in service occupations, mostly in food

preparation and serving related jobs1. For these reasons, several researchers have focused on this

industry to measure minimum wage effects (see Section 2 for details). We measure service

quality by text analysis of 533,022 online reviews for 2,870 unique restaurants posted during

1 https://www.bls.gov/opub/reports/minimum-wage/2016/home.htm

2008-2017 (see Section 3 for details). Traditional approaches to measure service quality, such as

surveys and focus groups, might be expensive, time consuming, and potentially subject to recall

biases and demand effects (Netzer et al. 2012). Unlike such approaches which rely on primary

data collected over a short period of time and for a limited number of firms, our data are

available over many months, and for a large number of firms. Therefore, relative to primary data

collection, our approach is especially useful for studying the impact of temporally distant events

(such as past legislations) by comparing periods before and after such events2.

To analyze text, we use a state-of-the-art unsupervised deep learning model (Mikolov et

al. 2013). Unlike applications where a label is associated with text, and the modeling objective is

label prediction for new text (i.e. supervised learning), unsupervised learning uncovers latent

relationships in the text. As analysts of text, our interests often center on topics (and therefore

words) that occur rarely in the text corpus, which in our context is the set of consumer reviews of

restaurants in our context. Topic models such as the Latent Dirichlet Allocation (Blei et al,

2001), which estimate co-occurrence patterns in a corpus using latent topics, have been

increasingly used in marketing (Tirunillai and Tellis 2014, Puranam et al. 2017). One challenge

with such topic models is that the estimated topics tend to generally focus on most commonly

discussed topics. It is difficult to ensure that the topic model will identify rare topics of interest.

We resolve this issue by using a word of focal interest (i.e. one seed word pertaining to one

dimension of service quality, such as reliability), and automatically identifying other related

words using an appropriate distance measure. Specifically, the Word2Vec model (Mikolov et al.

2013) learns a d-dimension latent vector representation for each word in the vocabulary. These

vector representations are referred to as “embeddings”. Semantic similarity between two words

is captured by the cosine distance between their respective embeddings. We expect that the

semantic relationships between words are domain specific, and consequently we estimate the

word vectors from our specific dataset. The model yields a list of semantically proximate words

for the seed word of substantive interest. For example, timely, attentiveness and speedy might be

semantically proximate to “responsiveness”. This list of words then defines the “topic”

associated with the seed word. In the next step, we measure the pairwise distance of each word in

a consumer review from the “topic” associated with the seed word, and proximate words. The

2 It is plausible that despite large sample sizes, user-generated textual content might also suffer from biases; indeed, no market research technique is perfect. As such we do not propose to replace traditional approaches, but instead to augment them with data that are available for free, in larger quantities, and over longer period of times.

mean of these distances across all words in a review is the distance of the review from the topic,

and serves as our dependent variable of interest. Since the distance measure takes into account

not just the seed word, but other proximate words, reliance of results on our choice of seed words

in minimized. Reviews which are at greater distance from the topic of interest, are semantically

less similar to the topic.

Given our interest in multiple service quality dimensions, we analyze several dependent

variables for each review, each representing the distance of the review from one dimension of

service quality. Seed words for our analysis are rooted in the service quality literature.

SERVQUAL (Parasuraman et al, 1988 and 1991) is a framework that outlines five dimensions of

service quality. SERVPERF (Cronin and Taylor, 1993) operationalizes this framework for

restaurants. Based on this framework, we focus our analysis on the following major dimensions

of service quality of restaurants: tangibles, reliability, responsiveness, assurance and empathy.

Descriptions of these service dimensions and seed words associated with each dimension appear

in Table 1. For robustness, we also simultaneously extract discussions of price from consumer

reviews. This ensures that service quality measurement is not conflated with measures of

satisfaction or dissatisfaction with (potentially changing) prices. Following our approach of

identifying seed words for the different dimensions of service quality, we use price as a seed

word to indicate the price dimension.

====Insert Table 1 Here====

Estimation of causal effects of minimum wage increases is difficult in part because of

lack of exogenous variations in wages (i.e. wage increases might be correlated with service

quality due to unobservables), unavailability of high quality instruments, and infeasibility of

conducting field experiments with meaningful wage increases. To deal with this, we exploit the

following natural experiment. The state of California requires employers to pay full state

minimum wage before tips3 to tipped employees. Exclusions based on the number of employees

of the business are not permitted. California is divided into 58 counties, one of which is the

County of Santa Clara (henceforth CSC). CSC has a population of 1.92 million (2015 Census), is

home to Silicon Valley, and is estimated to have the third highest per capita GDP across all cities

in the world4. CSC is comprised of 8 cities (San Jose, Cupertino, Los Altos, Milpitas, Mountain

View, Palo Alto, Santa Clara and Sunnyvale). Of these cities, San Jose, a city with a population 3 https://www.dol.gov/whd/state/tipped.htm4 https://en.wikipedia.org/wiki/Santa_Clara_County,_California

of over 1 million, increased hourly minimum wage by 20% from $8 to $10 effective March 11,

20135. This is due to an ordinance which aims to raise the minimum wage to $15 over time. The

remaining 7 cities did not witness any change from the statewide rate of $8 per hour in 2013.

Since the scope of the wage increase was limited to San Jose, we analyze data from San

Jose and the remaining 7 cities of CSC separately. That is, the remaining cities serve as a useful

contrast and as a geographically proximate control group (see Allegreto and Reich (2018) who

also use the same treatment group and control group). For this purpose, we estimate a difference-

in-difference model on all reviews of restaurants located in CSC in our dataset which were

posted in a 9-month period before, and 9-month period after the date of the wage increase. This

model employs deep learning based measures of perceived service quality as the dependent

variable. Model estimation is based on a subset of the 533,022 reviews employed for estimating

dimensions of service quality. The 18-month window of analysis leads to a dataset of 24,488

reviews from San Jose before the wage increase, and 27,921 reviews from San Jose after. These

represent reviews of 1,232 restaurants. As controls, we employ reviews from the other 7 cities of

CSC. These include 10,705 reviews before the wage increase, and 12,138 reviews after. These

represent reviews of 528 restaurants in the other 7 cities.

Geographical proximity reduces the possibility of biased estimates due to unobservables

which might affect San Jose differently from the other cities in CSC. To isolate the causal effects

of the wage increase on consumer opinions of service quality in restaurants in San Jose, we

control for differences in characteristics between San Jose and the remaining cities,

characteristics of reviewers who post restaurant reviews, observed review level characteristics,

and for unobserved restaurant level characteristics (via restaurant level fixed effects). In sum, we

combine current methods in text analysis from computer science with causal inference

techniques. Deep learning methods are new to the marketing literature. We are also unaware of

textual analysis applications in the vast and inter-disciplinary literature on minimum wages.

1.1 Research Questions and Agency Theory

Our first research question is whether minimum wage increase leads to an increase or

decrease in consumer opinions of service quality. The direction of this effect is a priori unclear.

A restaurant might react to forced increases in minimum wages by reducing the number of

employees (McBride 2017). Indeed, in 2016, McDonald’s announced the nationwide rollout of

5 http://www.sanjoseca.gov/index.aspx?NID=4125

automated kiosks to replace cashiers, as a way to reduce costs due to minimum wage increases

(Rensi 2016). Several other chain restaurants are currently contemplating replacing minimum

wage workers with robots, explicitly to deal with rising labor costs (Taylor 2018). A reduction in

employee strength might lead to lower service quality. For example, with fewer employees, it

might take more time for an order to be transported from the kitchen to the dining area. A kiosk,

or other substitutes, might not be perceived to be as courteous and responsive as a waiter. This

line of logic suggests a negative effect of minimum wage increases on service quality. On the

other hand, greater wages might lead to greater motivation of employees, and might also increase

the restaurant’s ability to hire more productive or experienced workers. It has been found in at

least one industry study that when employers transitioned to paying a living wage to its

employees, they experienced “significantly lower rates of staff turnover, reputational benefits,

reduction in sick leave, better motivated staff and an increase in productivity6.” A majority of

employers noticed better quality of work, and employees agreed that their work had improved

after the wage increase. This suggests a positive impact on service quality. Therefore, the

direction of impact on service quality is an empirical question.

We do not expect the impact of wage increases on service quality to be the same across

all restaurants. An important aspect of the restaurant industry is the difference in ownership

structures of chain and independent restaurants, and resulting differences on incentives and

managerial abilities (Brickely and Dark 1987, Lafontaine 1992, Lafontaine and Shaw 2005, and

Shepard 1993). Our second research question is whether the impact of minimum wage increase

on consumer opinions of service quality, differs across chain and independent restaurants. Chain

restaurants have two forms of ownership - they are either corporate-owned or owned by a

franchisee7. Independent restaurants are more likely to be owned and/or supervised by the

founder. Per agency theory, corporate-owned chain outlets have the least incentive and ability

to monitor service outcomes since the store manager is a salaried employee (who herself might

also face incentive issues). Franchisee outlets keep residual profits, and therefore are more

motivated to improve service outcomes. They also have greater ability to observe employees if

they are locally based. However, to enable ease of franchising and to maintain brand quality,

many national chains are built on process-driven cultures with little room for local variance in

6 https://www.nefconsulting.com/paying-the-living-wage-benefits-business-as-well-as-employees/7 There are small chain restaurants that are all owned by the founder, but these are far outnumbered by national chains like McDonalds, Burger King, Subway, Chipotle etc.

quality. In contrast, industry reports suggest independents’ success is affected by providing a

better customer experience via service and differentiation (e.g. varied menu) than the more bland

but reliable national chains8. Independents have greater ability to monitor employees (on visible

attributes of service) compared to chains. They also have a higher incentive to monitor for higher

service quality provision as a requirement for higher-paid workers, especially if accompanied by

higher pass-through prices. Given their smaller size and differentiated presence, pricing changes

might be easier implement with higher wages. Based on all these insights, we distinguish

between chain and independents in our analysis of minimum wage effects on perceived service

quality. We expect greater effects on service quality of independents.9

These questions have important implications for several stakeholders including

marketers, consumers and policy makers. A positive effect on service quality would provide an

additional rationale for increasing minimum wages - one that is unrelated to employment or

living standards of employees, but related to consumer satisfaction and welfare. On the other

hand, a negative effect would provide robust evidence consistent with the notion that minimum

wage employees are being replaced with less service-oriented substitutes. A positive effect of

increased wages on service quality should give pause to managers considering replacing

minimum wage employees, and provide a rationale based on business outcomes, for lowly paid

employees to demand wage increases. Finally, in the face of some evidence that minimum wage

increases could lead to price increases (Allegretto and Reich 2018), a positive effect on service

quality could incentivize consumers to remain loyal to restaurants affected by the legislation,

despite having to pay greater prices.

We find that consumers discuss topics associated with delays, slowness and

unfriendliness to a lower extent after the wage increase in San Jose restaurants where workers

had higher minimum wages. This change in topic discussion is consistent with improved service

quality. Consistent with agency theory discussed above, this service quality improvement is

restricted to independent restaurants, and does not extend to chain restaurants in that city. Next,

we discuss the relevant streams of literature and how our work is positioned relative to each

stream. Section 3 presents the textual data. In section 4, we discuss the deep learning model, the

8 https://www.consolidatedfoodservice.com/article/independent-vs-chain-restaurants-an-uphill-battlehttps://www.msn.com/en-us/foodanddrink/restaurantsandnews/mom-and-pop-restaurants-are-beating-out-national-chains-here%E2%80%99s-how/ar-BBBKmhp9 To the extent that some independent restaurants have multiple outlets and are therefore “chains” by our definition, our results can be seen as a lower bound on results of chains versus independents.

difference-in-difference specification, and specific estimation challenges. Section 5 presents the

results from the model and their implications. Section 6 concludes.

2. Literature Review

Our work is related to the literatures on the impact of minimum wages, customer

satisfaction, text mining and service quality. The impact of minimum wage increases on

employment is a politically fraught and an economically complex question. We discuss a small

set of papers to highlight the key issues. There are several theory models (e.g., Aaronson and

French, 2007) on the impact of minimum wage increase on employment and total earnings of

employed workers. These models show the employment effect varies as a function of labor and

product market competition, factor substitutability, and other factors. There is a large stream of

empirical work in this area with conflicting findings, and a fierce debate on reasons for these

conflicting findings. Especially, there are debates on the appropriate control group when

estimating the effect of the “treatment” of minimum wage increases, and controlling for

heterogeneity if pooling units subjected to the regulation, across various geographies. As

mentioned, several studies have analyzed the restaurant industry with its heavy reliance on

minimum-wage workers. Card and Krueger (1994) find that a minimum wage increase leads to

higher employment in limited service fast-food chains. Dube et al. (2010) examine both limited-

service and full-service restaurants (with differing fraction of minimum wage workers and price

elasticity of product demand). They also find positive employment and earnings effects. On the

other hand, several researchers find negative employment effects. Among these are Neumark

and Wascher (2000) and Aaronson et. al. (2008), who look at limited-service restaurants. The

employment effect of minimum wages is not the focus of our study. However, we draw on this

literature for its distinction between types of restaurants and the importance of finding the right

control group to infer causality.

Our work is more closely related to two other papers on the impact of minimum wage

increases on marketing outcomes. Allegretto and Reich (2018) estimate whether minimum wage

increases are passed on to consumers via increased price. They invoke the estimate of Orkent

and Alston (2012) that demand for restaurant products is relatively price inelastic (-0.71).

Therefore, restaurants faced with forced increases of minimum wages might prefer to not reduce

employment, but instead increase prices by a small amount. The alternative of reducing

employment might lead to greater reduction of profits. Using internet-based menus of restaurants

tracked before and after wages increases, they find that prices increased as a result f minimum

wage increases. This result holds even in border areas between higher and lower minimum wage

cities, where competition might have been expected to bid away price differentials. They find

that the price increase magnitude is similar to the previous mark-ups. This implies no negative

employment effects. They also do not find any competitiveness effects, based on their analysis of

border areas.

A recent working paper on the impact of minimum wage increases on a marketing related

outcome is Chakrabarty et al. (2017). They examine whether there is a change in restaurant

hygiene (using hygiene violations data) due to minimum wage increase. Their hypothesis is that

higher labor costs lead to employment cuts and fewer workers, and/or cost-cutting in other ways.

They find an increase in less severe hygiene violations, but not in the most severe violations.

They conclude that consumers face less hygienic conditions when minimum wages increase.

Unlike Allegretto and Reich (2018) and Chakrabarty et al. (2017), we do not estimate price pass-

through or health violations. Our interest is in consumer perceptions, and for that we utilize data

from thousands of consumer reviews. To the extent that service perceptions might be confounded

by price perceptions, we jointly estimate a price topic with service topics in our text analysis.

Our difference-in-difference regression of service topics are robust to the inclusion of price topic

as a control variable. We view our analysis as being broader and more relevant to marketing

research with its focus on service attributes. To the best of our knowledge, ours is the first paper

to study the effect of any wage change by a firm on any kind of consumer behavior or

experience.

The closest stream of research in marketing to our work, is on the relationship between

firm routines and customer satisfaction. Anderson et al. (1997) discuss whether customer

satisfaction and firm costs and productivity are positively or negatively correlated. It is

possible that improving customer satisfaction reduces costs such as those of returns, new

customer acquisition, etc. Alternatively, improving customer satisfaction might increase costs

e.g. of customer support, IT etc. They find that for services in particular, customer satisfaction

comes with increased costs. Note that for our context, an increase in minimum wages will

increase per worker cost. The impact on total cost is unclear since firms might reduce

employment, so whether minimum wage increase leads to greater or lower service quality cannot

be predicted based on this stream of research. In fact, even if total labor costs increase, higher

costs are not a sufficient condition for higher customer satisfaction. So the direction of the effect

of minimum wage increases is an open empirical question.

In related work, researchers have examined the link between employee satisfaction

(especially front-line employees) and business outcomes, including customer satisfaction (e.g.,

Schneider et. al. 1998, Kamakura et. al. 2002, Simon et. al. 2009). The predominant conclusion

is a positive correlation between employee and customer satisfaction. Zablah et al. (2016) find

that while attention has been focused on how satisfied front-line employees can improve

customer satisfaction, there might be larger effects from satisfied customers to satisfied front-line

employees. In our empirical context, it is possible that increases in wages are more than offset

by increased work demands, resulting in lower employee satisfaction. It is also possible that

higher wages result in hiring of more motivated current or new employees (e.g. via the efficiency

wage hypothesis in labor economics, Shapiro and Stiglitz 1984). This might result in more

satisfied workers. Therefore, the link between minimum wage increase and customer satisfaction

(or customer perceptions of quality) is perhaps more complex than the relationship between

employee satisfaction and customer satisfaction. We differ from these papers as follows- first,

we are able to extract via text analysis multiple dimensions of customer perceptions of service

quality, while controlling for price perceptions. Second, our natural experiment allows us to

infer causal effects on customer perceptions of service quality; we do not have any confounds of

reverse causality effect of customer perceptions on higher performance of front-line employees.

Another stream of directly relevant research is text mining methods in marketing. Our

paper is related to the literature on extracting useful information from large masses of text of

reviews (Decker and Trusov 2010, Archak et. al. 2011, Ghose et al. 2012). Lee and Bradlow

(2011) automate extraction of phrases from consumer reviews, which are then rendered into

word vectors which record the frequencies with which words appear in the corresponding phrase.

Phrases are clustered together according to their similarity, measured as the distance between the

word vectors. Importantly, they demonstrate the validity of text-mined measures of product

attributes, in comparison with more traditional measures. Netzer et al. (2012) use a similar

approach, with the difference that they define similarity between products based on their co-

mention in the data. Tirunillai and Tellis (2014) and Puranam et. al. (2017) apply the Latent

Dirichlet Allocation (LDA) model on consumer reviews to infer the latent dimensions of product

quality. Büschken and Allenby (2016) propose an LDA model that makes use of the sentence

structure of reviews, leading to improved prediction of consumer ratings. In this paper, we

propose an alternative method for extracting topics from text (see section 4 for details). Unlike

existing methods in marketing, our Word2Vec model (Mikolov et al, 2013) of deep learning is

well suited for identifying rare topics of interest. Our objective is not to summarize the text

corpus into key attributes, but instead to hone in on a specific multidimensional construct of

interest: service quality. Other approaches using topic modeling yield qualitatively poorer

construct representations in our context (see section 4 for details).

Finally, as discussed earlier, we rely on the literature in marketing on service quality to

identify dimensions of service quality to be investigated.

3. Data

Our primary source of data is Yelp, a California-based multinational corporation which publishes

online reviews of local businesses, and is commonly used by consumers in the United States to

post and view reviews of restaurants. Per Alexa.com, an independent research firm, yelp.com

was among the top 70 most popular websites in the United States in July 2014, the time period of

the data. By the end of 2017, the website had well over 100 million reviews. For the purpose of

estimating word vectors using deep learning methods, we utilize data on all restaurant reviews

posted on this website from the time period January 1st, 2008 to April 3rd, 2017, for all cities in

CSC10. This dataset has 533,022 reviews. It covers 962 chain restaurants (those restaurants with

at least two outlets in California), with 160,866 reviews from 76,774 reviewers with an average

review length of 56.19 words and an average rating of 3.40 (out of 5). In addition, the data

includes 1,908 independent restaurants, with 372,156 reviews, from 136,110 reviewers with an

average review length of 56.94 words and an average rating of 3.65 (out of 5).

Our substantive interest is in estimating the effect of a wage increase in May 2013. As

mentioned, for this purpose, we estimate a difference-in-difference model on all reviews of

restaurants located in CSC in our dataset which were posted in a 9-month period before, and 9-

month period after the date of the wage increase. Our results are robust to shorter and longer time

windows. Very long time windows increase the likelihood that differences in time varying

unobservables across the treatment and control units might affect our results. Also, the effect of 10 Using shorter duration of data for estimating word vectors, or estimating word vectors separately for reviews of chain restaurants and independent restaurants, does not change our results.

wage increase on service quality, and its reflection in consumer reviews, might take a few

months. So very short time windows might not be sufficient for our purpose.

We now discuss the appropriateness of using the other 7 cities of CSC as control units.

As mentioned, these are geographically close to San Jose, and are mostly urban, much like San

Jose. More importantly, wages across San Jose and other cities are similar. For the year 2013,

weekly wages were $361 in San Jose, and $394 in other cities (Allegretto and Reich 2018).

Weekly wages in San Jose and other cities for limited service restaurants averaged $312 and

$319 respectively. Weekly wages in San Jose and other cities for full service restaurants

averaged $400 and $435 respectively. Furthermore, both areas experienced similar

unemployment trends (see Allegretto and Reich 2018 for a more detailed comparison).

We now describe reviews of restaurants in San Jose. Means of ratings (on a 5-point scale)

decreased slightly from 3.60 before the wage increase to 3.57 after. However, since ratings are a

composite metric capturing several constructs, not just service quality, this trend is not

particularly insightful for our research context. The website allows readers of a review to

indicate whether it is “useful”, “funny” and “cool”. On average, a review posted before the wage

increase received 1.15 mentions of being “useful”, 0.58 mentions of being “funny” and 0.57

mentions of being “cool”. The means (per review) of these mentions for reviews posted after the

wage increase are 1.07, 0.50 and 0.51 for mentions of “useful”, “funny” and “cool”. Similar to

ratings, it is unclear if these metrics capture service quality in any discernible way, but do serve

as control variables in our model. Detailed summary statistics appear in Table 2.

====Insert Table 2 Here====

Next we discuss the frequency of occurrence of seed words used to model the five

dimensions of service quality (Table 3). As model free evidence of the positive effect of the

wage increase, occurrence frequency (per review) of the words “knowledge” and “attention”

increased by at least 5% after the wage increase; and occurrence frequency of the words “wait”,

“slow” and “impolite” decreased by at least 5%. However, summary statistics do not paint an

entirely positive picture. Occurrence frequencies of the words “dirty” and “delay” increased after

the wage increase, and those of the words “clean” and “fast” decreased. This indicates that

service quality might have been negatively impacted for at least some dimensions. Differences

across chain and independent restaurants in the occurrence frequencies of seed words provide

interesting model free insights. While the occurrence of the word “mistake” decreased by 11%

after the wage increase among independents, it increased by 9% among chain restaurants.

Occurrence of the word “impolite” decreased by 56% among independents, but only 13% among

chains, suggesting stronger effects of wage increases among independents. However, these

metrics ignore other words which are semantically similar to the seed words, do not account for

CSC-specific time varying unobservables (for which we employ reviews from other CSC cities

as controls), do not control for restaurant characteristics and might be confounded with changing

perceptions of price.

====Insert Table 3 Here====

Summary statistics of reviews for restaurants in CSC cities other than San Jose appear in

Tables 4 and 5. Means of restaurant ratings, review lengths, and indication of whether reviews

are “useful”, “funny” and “cool” are similar across the treatment and control group, both before

and after the wage increase. Changes in the frequency of occurrence of some seed words such as

“clean” and “menu” follow similar trends for San Jose and for other cities. On the other hand,

while the frequency of occurrence of “dirty” increased after the wage increase for San Jose, this

frequency decreased for other CSC restaurants.

====Insert Tables 4 and 5 Here====

4. Model

We describe first our text analysis model and then the differences-in-differences

regression model that employs output from text analysis as the dependent variable.

4.1. Deep Learning Text Analysis

We first discuss the motivation for employing a deep learning model. Service quality is

known to be a multi-dimensional construct. Based on the service quality literature, we are

interested in 5 dimensions of service quality, which are captured by 17 seed words (these appear

in Table 1). Each seed word describes a sub-dimension or topic of substantive interest11. The

objective of our text analysis is to obtain robust quantitative estimates of the extent to which each

review in our data, represents each of the 17 topics of interest. As mentioned earlier, our main

motivation for employing deep learning methods is that descriptors of service quality are not

discussed very frequently in our data. This is typical for online reviews. Puranam et al. (2017)

find that as many as 200 topics are discussed in a set of restaurant reviews from New York City, 11 Adding more seed words, or dropping some of these 17 seed words, does not change the substantive results of the paper.

with the most discussed topic accounting for less than 5% of the overall corpus. In our dataset,

several topics such as “clean”, and its antonym “unclean”, are rarely discussed. Typical topic

models used in marketing are not very well suited to investigate the occurrence of rare topics of

interest. To demonstrate this, we estimated two topic models from the marketing literature:

Latent Dirichlet Allocation and Non Negative Matrix Factorization. Words occurring with high

probability in major topics estimated from each of these models, are presented in Appendixes 1

and 2. We find that none of the estimated topics cleanly captures service quality. Discussion of

service quality, as estimated from these models, is spread across several topics, and is absent in

other topics. This highlights the need for an alternate modelling approach.

Our modelling approach has two major steps. In the first step, we adopt a principled data-

based method to define the 17 topics of interest. In the next step, we estimate the distance of each

review from each topic of interest, such that the model output (and the input for the differences-

in-differences analysis) is a set of 17 parameters for each review, capturing the extent to which

each review represents each topic of service quality. One simple way forward for the first step

could to assume that the seed word is an adequate measure of the topic of interest. In the second

step, we could then estimate the distance of each review from each of the 17 seed words. This

has the obvious drawback that subjectivity in the researcher’s choice of seed words could affect

the results. Our model deals with this issue by automatically identifying words which are

semantically similar to a seed word. For example, for the seed word “fast”, the model could

potentially identify the following semantically similar words or phrases: quick, speedy, relatively

fast, fairly quick, etc. The “topic” of interest is then defined not just by one seed word, but by a

set of semantically close words, estimated from the data. To the extent that the words comprising

a topic are semantically similar, adding or dropping a few words from this set of words which

define a topic, does not affect results, thus ensuring robustness. We also estimate a “price” topic

using “price” as a seed word. Note that seed “price” is general enough to allow semantic

similarity to words like cheap, expensive, and cost (of the meal; not of production). We focus

the rest of the discussion on measurement of service topics since these are the main focus on our

paper. Expectedly, similar logic and steps follow for estimation of the price topic as for any of

the service topics.

For the purpose of estimating topics, the Word2Vec model first learns a d-dimension

latent vector representation for each word in the vocabulary V, i.e. the set of all unique |V| words

across all reviews in the dataset. These vector representations are referred to as “embeddings”.

Semantic similarity between two words is then measured in terms of the cosine distance between

their respective embeddings. We expect that the semantic relationships between words are

domain specific, consequently we estimate the word vectors in our dataset, instead of employing

word lists or dictionaries.

From an estimation perspective, the central idea is to set up a prediction task for each

focal word, taking into account the immediate neighboring words in a “context window”. For

example, in the sentence “The quick brown fox jumped over the lazy dog”, suppose the word

“fox” is the word of focal interest. A context window of 3 would involve the 3 words appearing

immediately before, and the 3 words appearing immediately after the focal word. These are {the,

quick, brown, jumped, over, the}. These context words are also termed outside words. Starting

from the first word in the vocabulary, every word is sequentially treated as the focal word. The

prediction task can be set up in at least two distinct ways. The first is to predict the focal word,

given the context window. The second, is to predict the words in the context window, given a

focal word. The former is referred to as the Continuous Bag of Words model and the latter is

referred to as the skip-gram model. The latter specification has been found to perform better on

semantic accuracy (Mikolov et al. 2013a) and on estimating representations for rare words.

Consequently, we focus on the skip-gram model, and now present the formal model

specification12.

The prediction task is as follows - given the focal word “jumped” in the example

sentence above, predict all words in a window of m = 3 words – i.e. quick, brown, fox, over, the,

and lazy. In the skip-gram model each focal word w is associated with two vectors – vw∧uw .

These are the “input” and “output” vector representations of the word w respectively. The

outside words (words in the context window other than the focal word) in the context window of

size 2m are indexed as o. uo is the embedding of the other word in the context window. In the

above example, suppose the focal word w is fox. Then the word fox has two representations – as

an input vector and as an output vector. One example of a context word is quick, consequently uo

is the vector associated with quick. Another context word is over.

A corpus is a concatenation of all documents in the dataset; in our case, the corpus

comprises all reviews. The length of the corpus is T and each position in the corpus is indexed by 12 Other word vector specifications include the GLoVe model (Pennington et al 2014). Empirical tests on a multitude of semantic applications typically find that the skip-gram model is the gold standard.

t. The objective function is to maximize the average log probability of observing the words

given context, across all T positions in the corpus.

J (θ )= 1T

∑t=1

T

∑-m≤j≤m,j ≠ 0

❑

log p ( w t+j|w t ) (1)

The probability for an outside word o (i . e . wt+ j=o ) in a context for a focal wordf (i . e . wt=f ),

located at position t is evaluated as a logit function of the dot product between the two

embeddings:

p(o|f )=exp ( uo

' v f )

∑w=1

|V|

exp( uw' v f )

(2)

The model has a total of 2 d∨V ∨¿ parameters. This is the dimension of the embedding times the

two vectors estimated for each word. And, while it yields a simple analytical form for

maximization, the summation in the denominator (over all the words in the vocabulary) is

computationally problematic. For example, we have approximately 34.3 million positions in the

corpus, and a vocabulary of 11,121 words.

An alternative model specification incorporates the idea that the model parameters from

(1) should be able to help distinguish the observed focal word (f) and context word (o)

combinations in the data from those combinations generated by a noise distribution. This

approach is called noise contrastive evidence (Gutmann and Hyvarinen 2012). For each focal

word observed in the data, k context words (o) are drawn from a noise distribution (q(o)) and

labeled as 0 (i.e. as coming from noise). This set of k context words are referred to as “negative

samples” as these words are not necessarily observed in the context window of the focal word.

All the f and o combinations observed in the data are labeled as 1. The estimation task is to

predict data from noise. The probability that a particular f and o combination is drawn from

noise is:

(Label=0|f,o)= k × q(o)exp (uo

' v f )+k × q(o)

(3)

Similarly, the probability that a particular observation is observed from data is given by:

(Label=1|f,o)=exp (uo

' v f )exp (uo

' v f )+k × q(o) (4)

The reformulated log-likelihood is:

(θ )= ∑f,oϵData

❑

log p ( label=1|f,o ) +k Eo' ~q ( o) log (1- p (Label=0|f,o' ) ) (5)

This objective function maximizes the probability of observing the data, while minimizing the

probability of word combinations generated from noise. Note the second term still requires a sum

over all |V|. This intensive computation is substituted with a Monte Carlo approximation.

(θ )= ∑f,oϵData

❑

log p (label=1|f,o ) +k ∑fϵData

❑ ∑o' ~q(o)

❑ 1k

log p (Label=0|f,o' ) (6)

Mikolov et al. (2013), assume that the context words are drawn from a uniform distribution of

size |V|. This leads to the following simplified objective function.

(θ )= ∑f,oϵData

❑

σ ( uo' v f ) + ∑

fϵData❑ ∑

o' ~q(o)

❑

σ ( -uo'' v f ) (7)

Here σ (.) is simply the logit transform. The first term in (7) is simply the normalized cosine

distance between the two word vectors. The second term measures a similar distance between the

focal word and words unlikely to be associated with the focal word. The model searches for a

set of parameters (the vector representations of each word) that simultaneously accounts for the

frequency with which all pairs of words co-occur. It has also been shown that, under certain

assumptions, this approach is equivalent to matrix factorization. There are some parallels with

multidimensional scaling (MDS) in that similarity measures are preserved under a d-dimensional

representation. Unlike MDS, the similarity scores are not known, though co-occurrence patterns

are known.

The model is estimated using TensorFlow (an application for deep learning, see Abadi et

al, 2016). It is a model development framework, similar to WinBugs used by Bayesian

researchers. The estimation of the parameters is by maximizing likelihood. Given the large

number of parameters, we make use of the “Adam” optimizer (an adaptive momentum optimizer,

see Kingma and Ba 2014). The algorithm is a first-order gradient-based optimization scheme for

stochastic objective functions.

Although the number of parameters in our model are fewer than in the LDA model (and

is independent of the size of the corpus), we do need to select a few hyperparameters, namely,

the window size (m), the number of negative samples and the dimension for the embeddings (d).

We select a maximal window size of 10, 10 negative samples and 100-dimension embeddings

based on the perplexity of the model and the judged similarities of the word embeddings. Our

results are robust to the choice of estimation choices such as using a context window of size 5

and using embeddings of sizes 50, 100 or 150. Details are available and are not presented for

brevity.

Finally we discuss our choice of seed words. Table 1 lists the seed words selected for

each service dimension. In some instances where more than one possible obvious candidate was

available, multiple seeds were used. For example we have selected both wait and delay as seed

words. In theory, we can also explicitly chose the valence of each attribute and measure both

positive and negative valence for each attribute. For example, we include both fast and slow in

the seed words for reliability. However, incorporating valence in graphical models and

factorization methods is uncommon, complex and requires additional assumptions.

The model described above yields a list of semantically proximate words for each of 17

service seed words (and one for price, included for its potential role as control variable to

separate service quality from price perceptions). This list of words defines the topic associated

with the seed word. As mentioned earlier, in the next step, we measure the average distance of

each review from the seed word and use that as our measure of interest. For each word of a

review, we measure the mean cosine distance of that word from each word of the topic (the seed

word and 19 semantically similar words). The mean of these distances (across the 20 words of

the topic), gives us the distance of that word from the topic. We then compute the mean (across

words of the review) of this mean distance, to obtain the distance of the review from the topic.

We repeat this for each of 17 topics. As a result, for each review we obtain 17 measures,

representing the distance of the review from the 17 topics of interest. These serve as the

dependent variable for the next stage of analysis.

We also calculate similar measures of price perceptions, and run the causal difference-in-

difference analysis below including these price perceptions as control variables. Note this is over

and above factoring out any price perceptions confound in the service perception measurement,

as described in this section. Since the difference-and-difference results are similar in direction

and magnitude with and without the inclusion of the price topic as a control variable, in the

discussion below we discuss results without the inclusion of price topic measure. We have also

examined if there is any impact of wage increase on price topic as a dependent variable. We do

not find any systematic changes in price discussion; this is possibly consistent with the small

1.45% price changes found post-policy by Allegretto and Reich (2018), and the relatively price

inelastic nature of restaurant demand in the United States (Okrent and Alston 2012).

4.2: Difference-in-Differences Regression Model

In an attempt to identify the causal effect of the wage increase on the distances of reviews

from 17 topics of service quality, we implement a “difference-in-differences” methodology. We

calculate the causal effect of a treatment (i.e. the wage increase) on the outcome variable

(distance from a topic of service quality) by comparing the average change in the outcome

variable for the treatment group (CSC restaurants in San Jose) to the average change for the

control group (CSC restaurants outside San Jose). We regress the outcome variable on two main

effects (the effect of belonging to the treatment group on the outcome, and the effect of the

treatment on the outcome), the interaction of these two effects, and several control variables, as

follows:

PerQualir=Const i+β1 SanJoser+ β2 Postr+ β3 SanJoser × SanJoser+β4 Ratin gr +β5Useful .Count r+β9 Cool .Coun t r+β9 Funny .Count r+Restaurant Dummies+β10 reviewlengthr+ϵ i+ϵ r

(8)

The subscripts i indexes the reviewer and r the reviews. The dependent variablePerQualir

represents the cosine distance of review r from one of the 17 topics of service quality. For each

review, there are 17 measures, 1 for each topic, leading to 17 equations. SanJoser is the dummy

variable which accounts for the effect of belonging to the treatment group. It controls for

unobserved factors which might affect perceived service quality measures of San Jose restaurants

and other CSC restaurants differently. Postr is the dummy variable which accounts for the effect

of the treatment. It is possible that at the time of the implementation of the wage increase, there

were unobserved events which affected service quality perceptions of all restaurants in CSC. The

main effect of Postr controls for how service quality perceptions for all restaurants changed after

the wage increase. The coefficient of interest is β3, the coefficient on the interaction of SanJoser

and Postr. As controls, we include the review-specific counts of the words “cool”, “funny” and

“useful”, and the length of the review. It is possible that longer reviews, cooler reviews, funnier

reviews and more useful discuss service quality to a greater or lower extent. Restaurant specific

fixed effects, and reviewer specific intercepts, account for restaurant specific unobservables and

unobserved reviewer characteristics respectively. Error terms are assumed IID and normally

distributed. All parameters are topic-specific. The identifying assumption is that after controlling

for a) all observed review characteristics, b) unobserved reviewer level characteristics (e.g., some

reviewers might write more about service quality than others), c) unobserved differences across

San Jose and other cities in the same county, and d) unobserved differences in service quality

between the time period before and after the wage hike, which affect all restaurants in the

county; any differences in changes in our measures of service quality across San Jose and other

cities, are due to the wage increase.

5. Results

First we present the top 10 closest words to each of the 17 seed words for service and for

price, estimated from the deep learning model (Table 6). Several of the closest words are

antonyms of the associated seed words (e.g. fast and speedy), which provides face validity to our

text model. Other close words are specific to our research context of service in restaurants, which

validates our decision to not use dictionaries for defining topics. For example, the words

“bathroom” and “bathrooms” are estimated to be close to the seed word “clean”, and the words

“refill” and “wave” are estimated to be close to the seed word “attention.” Although these are not

antonyms, bathrooms are likely more salient for the cleanliness of restaurants, and waving is a

common way to attract the attention of the waiter.

====Insert Table 6 Here====

We now present parameter estimates from the difference in difference model (Equation

8), which aims to estimate the causal effect of the wage increase on text based measures of each

of 17 dimensions of perceived service quality. Our first question is whether minimum wage

increase leads to an increase or decrease in consumer opinions of service quality. We find that

perceived service quality of San Jose restaurants changes along three dimensions, as compared to

control units (Table 7). The coefficients of the interaction of SanJoser and Postr are negative for

the topics delay, slow and unfriendly. Reviews of restaurants in San Jose showed a larger

negative change in the representation of these topics after the wage increase, than reviews of

restaurants in other cities of CSC did. This is of course, welcome news for consumers and

restaurants. Delay and slowness are known to measure the reliability of service quality; our

results provide unique evidence that greater wages lead to more reliable service quality. These

results are consistent with the notion that greater wages motive employees to serve better, or that

they increase the restaurant’s ability to hire more productive or experienced workers.

====Insert Table 7 Here====

Interestingly, there is no significant change in the discussion of those topics of service

quality which have positive connotations (such as clean, fast, courteous and friendly). One

possible reason is that workers who provide poor service (e.g. those who are delayed, slow and

unfriendly) are likely to be paid less (and minimum wages) than workers who provide good

service (e.g. clean and fast workers). Since the wage increase does not depend on prior service

quality, but only on existing wages, the regulation disproportionately raised the wages of

workers who were providing poor service (and were paid less prior to wage increase). Workers

providing good service might have been paid above minimum wage prior to the legislated

change, and hence their wages are not affected by the legislation. Consequently, their (good)

service levels might remain unaffected. It is also worth noting that the service dimensions

affected by the wage increase are easily observable to consumers (e.g. fast service is easier to

observe than worker knowledge) or can be more easily attributed to workers (e.g. courtesy can

be more easily ascribed to a restaurant worker than the menu items of that restaurants). This

suggests a potential qualifier for service effects of wage increases: wage increases lead to

positive service perceptions of observable and easily attributable worker service quality. This

nuanced insight is enabled by the multidimensional service quality measures.

Other parameter estimates from this analysis show that ratings and counts of useful, cool

and funny significantly predict the discussion of almost all service topics. Reviews scoring high

on the topics clean, menu and prompt, are also associated with more positive ratings. Greater

discussion impoliteness elicits more “useful” votes, as compared to the discussion of other

topics. Greater discussion of promptness and slowness elicits more “cool” votes. Reviews of

restaurants based on San Jose discuss two service dimensions less than other reviews: delay and

knowledge.

5.1 Impact of Increased Wages on Independent Restaurants and Chain Restaurants

Our second research objective is to investigate whether the effect of wage increase varies

across chain restaurants and independent restaurants. As discussed in section 1.1, agency theory

suggests that the impact of wage increase of service quality should be greater in independent

restaurants, which have greater incentive and ability to monitor service outcomes. To explore

this, we estimated the difference in difference model separately, first on all reviews of

independent restaurants (in San Jose and other cities) and then on all reviews of chain restaurants

(in San Jose and other cities). Results from these analyses appear in Tables 8 and 9.

====Insert Tables 8 and 9 Here====

Separate analysis of independent and chain restaurants provides several unique insights.

The negative effects of wage increase we reported earlier on the discussions of three dimensions

of service quality (delay, slow and unfriendly) are even stronger (i.e., more negative) for

independent restaurants. Furthermore, we find negative effects of wage increase on three other

dimensions of service quality. These are dirty, wait and impolite. In all, wage increase negatively

affects six dimensions of service quality. Negative effects on six negatively valenced dimensions

of service quality suggests that the impact of wage increase is quite positive. Consistent with

earlier results, aspects of service quality that are not easily observable (e.g. knowledgeable) or

are easily attributable to workers (e.g. menu) remain unaffected by wage increases.

Estimation of the difference-in-differences model with data from chain restaurants does

not suggest a causal effect of wage increase on any dimension of service. In other words,

improvement in perceptions of service quality in San Jose restaurants after the wage increase, are

driven by independent restaurants only. These results hold even after controlling for discussion

of price in the regression specification. Comparison of results for the two types of restaurants

provides empirical support to our theory that corporate-owned chain outlets have lower incentive

and ability to monitor service outcomes, than independents. Although franchisee outlets among

chain restaurants might be motivated to improve service outcomes, the imposition of process-

driven service guidelines across most major US chains leaves relatively less scope for

improvement than independent restaurants. It is also possible that independent restaurants are

able to improve service outcomes more easily than chain restaurants due to their smaller size,

and greater operational flexibility.

Another possibility which explains the null effect for chains is that typical consumers of

chain restaurants do not expect service improvements due to wage increases. If this is indeed

true, and if workers in chain restaurants are aware of this, then they do not have incentives to

improve service. Our data suggest that this is unlikely. The frequency of occurrence of the 17

seed words describing service quality in reviews posted before the wage increase, does not vary

much across chains and independents (Tables 3 and 4). Different service expectations across

restaurant types should have led to differences in the levels of discussion of service dimensions

across chains and independents. Our objective is not to rule out all but one explanation; we

expect several mechanisms might be driving this difference in the effect size across chain and

independent restaurants. Our results suggest that agency theory based mechanisms are likely the

larger driver of these effects; mechanisms based on differences in consumers’ service

expectations appear to be less likely to cause our results.

6. Implications and Conclusions

We first discuss the implications of our research for consumers. First, consumers who

value timeliness, speed and friendliness will be well-served by increases in minimum wages

since these service qualities improve in independent restaurants. Second, consumers who are

sensitive to other aspects of service such as the extensiveness of the menu, the likelihood of

mistakes by restaurants workers and the knowledge of restaurant personnel will probably not

notice a change a service levels with increases in minimum wages. Given the small price

increase associated with the wage increase documented in a previous study (Allegretto and Reich

2018), such consumers do not have much to lose by continuing to visit restaurants experiencing

wage hikes. Third, consumers who choose independent restaurants over chain restaurants can

expect greater improvements in service levels, not just in terms of reduced delay, slowness and

unfriendliness, but also in terms of reduced dirtiness, waiting and impoliteness.

Our results bring positive news to restaurants experiencing greater labor costs due to

mandated wage increases, even more so for owners and managers of independent restaurants.

The debate on the higher minimum wage legislation has focused solely on greater labor costs and

prices, without considering the potential demand side upsides. Other than profit impact, our

dimension-specific estimates of service quality are useful for owners and managers. Service

perceptions are more likely to improve along dimensions that are easily observable to consumers,

perhaps suggesting the need to manage other service dimensions. Given the limited impact of

wage increase on service perceptions of chain restaurants, corporate managers and franchises of

such outlets might view these results as a reason to defocus their attention from mandated wage

increases.

Policy makers might view these results as providing directionally positive evidence of

minimum wage increases on consumer welfare. This research could serve as an impetus to

further investigate the consumer welfare implications of labor cost increases in restaurants and

other service industries. They might also note that not all effects of wage increases are positive.

Our results are also of interest to researchers in marketing. Marketing applications of deep

learning methods for text mining are at a nascent stage. As we show in our analysis, these

methods can be useful and are easily implementable. Researchers in customer satisfaction, and

agency theory might also be encouraged to use vast online ratings data and natural experiments

for causal inference.

We conclude with a discussion of areas where this work can be extended. First, we were

unable to obtain robust data on employee satisfaction. Establishing a text-analysis multi-

attribute based link between front-line employee satisfaction and perceived customer service is a

logical next step, especially given the rich literature in this area. Second, we do not have data on

differences in pre-policy wages in chains and independents. Nor do we have information on

franchising within chains in our data location and time period. Extending data on these

dimensions would allow us to test more fine-grained employee and owner incentive effects.

Third, we introduce an unsupervised deep learning application to text analysis in the Marketing

domain. Therefore, we view our study as an exploratory analysis of perceived service quality

using deep learning methods.

Bibliography

Abadi, M., P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. Devin, et al. (2016), “TensorFlow: A System for Large-Scale Machine Learning,” 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI ’16), May 27, 265–84.

Aaronson, D. and E. French (2007), “Product Market Evidence on the Employment Effects of the Minimum Wage,” Journal of Labor Economics, 25(1), pp 167-200

Aaronson, D., E. French and J. McDonald (2008), “The Minimum Wage, Restaurant Prices, and Labor Market Structure,” Journal of Human Resources, 43(3), pp 688-720

Allegretto, S. and M. Reich (2018), “Are Local Minimum Wages Absorbed by Price Increases? Estimates from Internet-based Restaurant Menus,” ILR Review, 71(1), 35-63.

Archak, N., A. Ghose and P. G. Ipeirotis (2011), “Deriving the Pricing Power of Product Features by Mining Consumer Reviews,” Management Science, 57(8), 1485--1509.

Anderson, E., C. Fornell and R. Rust (1997), “Customer Satisfaction, Productivity, and Profitability: Differences Between Goods and Services,” Marketing Science, 16(2), 129-145.

Bhorat, H., R. Kanbur and B. Stanwix (2014), “Estimating the Impact of Minimum Wages on Employment, Wages, and Non-Wage Benefits: The Case of Agriculture in South Africa,” American Journal of Agricultural Economics, 96 (5), pp 1402–1419.

Brian, B. and S. Machin (2018), “Minimum Wages and Firm Value,” Journal of Labor Economics, 36, 1, pp 159-195.

Brickely, J. and F. Dark (1987), “The Choice of Organizational Form: The Case of Franchising”, Journal of Financial Economics, 18, pp 401-420

Büschken, J. and G. M. Allenby. 2016. Sentence-Based Text Analysis for Customer Reviews. Marketing Science. 35(6): 953-975.

Card, D. and A. Krueger (1994), “Minimum Wages and Employment: A Case Study of the Fast-Food Industry in New Jersey and Pennsylvania”, American Economic Review, 84(4), pp 772-793

Chakrabarti, S., S. Devaraj and P. Patel (2017), “Minimum Wage and Restaurant Hygiene Violation: Evidence from Food Establishments in Seattle”, working paper, available at SSRN: https://ssrn.com/abstract=2992783 

Cronin, J. J., and S. A. Taylor (1992), “Measuring Service Quality: A Reexamination and Extension.” Journal of Marketing, 56 (3).

Decker, R. and M. Trusov (2010), “Estimating Aggregate Consumer Preferences from Online Product Reviews” International Journal of Research in Marketing, 27(4), 293--307.

Dube, A., T. W. Lester and M. Reich (2010), “Minimum Wage Effects Across State Borders: Estimates Using Contiguous Counties,” Review of Economics and Statistics, 92 (4), pp 945-964

Ghose, A., P. G. Ipeirotis and B. Li (2012), “Designing Ranking Systems for Hotels on Travel Search Engines by Mining User-generated and Crowdsourced Content, Marketing Science, 31(3), 493--520.

Grossman, J. B. (1983), “The Impact of the Minimum Wage on Other Wages,” The Journal of Human Resources, 18(3), pp 359-378.

Gutmann, M. U. (2012), “Noise-Contrastive Estimation of Unnormalized Statistical Models, with Applications to Natural Image Statistics,” Journal of Machine Learning Research, 13, 307–61.

International Labor Organization (2015), “Minimum Wage Policy Guide,” available at http://www.ilo.org/wcmsp5/groups/public/---ed_protect/---protrav/---travail/documents/genericdocument/wcms_508566.pdf

Kamakura, W., V. Mittal, F. de Rosa and J. A. Mazzon (2002), “Assessing the Service-Profit Chain,” Marketing Science, 21 (3), pp. 294-317

Kingma, D. P. and J. Ba (2014), “Adam: A Method for Stochastic Optimization” Proceedings of the 3rd International Conference on Learning Representations (ICLR), December 22.

Lafontaine, F. (1992), “Agency Theory and Franchising: Some Empirical Results,” RAND Journal of Economics, 23(2), pp 263-283

Lafontaine, F. and K. Shaw (2005), “Targeting Managerial Control: Evidence from Franchising,” RAND Journal of Economics, 36(1), pp 131-150

Le, Q. and T. Mikolov (2014), “Distributed Representations of Sentences and Documents,” International Conference on Machine Learning - ICML 2014, 32, 1188–1196.

Lee, T. Y. and E. T. Bradlow. 2011. Automated Marketing Research Using Online Customer Reviews. Journal of Marketing Research. 48(5), 881--894.

McBride, S. (2017), “Minimum Wage Hikes are Causing Business to Cut Jobs,” Business Insider, March 16, http://www.businessinsider.com/minimum-wage-leads-to-job-losses-2017-3

Mikolov, T., G. Corrado, K. Chen, and J. Dean (2013), “Efficient Estimation of Word Representations in Vector Space,” Proceedings of the International Conference on Learning Representations, ICLR 2013, 1–12.

Mikolov, T., I. Sutskever, K. Chen, G. Corrado and J. Dean (2013), “Distributed Representations of Words and Phrases and Their Compositionality,” In Neural Information Processing Systems, 1, 1–9.

Netzer, O., R. Feldman, J. Goldenberg and M. Fresko (2012), “Mine Your Own Business: Market-Structure Surveillance Through Text Mining,” Marketing Science, 31(3), 521--543.

Neumark, D, and W. Wascher (2000), “A Case Study of the Fast-Food Industry in New Jersey and Pennsylvania: Comment,” American Economic Review, 90(5), pp 1362-1396

Okrent, A. and J. Alston (2012), “The Demand for Disaggregated Food-Away-From-Home and Food-at-Home Products in the United States”, United States Department of Agriculture, Economic Research Service Report No. 139, pp 187-212

Parasuman, A., V. Zeithaml, and L. Berry, L. (1988). A Multiple Item Scale for Measuring Consumer Perceptions of Service Quality. Journal of Retailing, 64(1), 12–37.

Pennington, J., R. Socher, and C. D. Manning (2014), “GloVe: Global Vectors for Word Representation,” Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, 1532–43.

Puranam, D., V. Narayan and V. Kadiyali (2017), “The Effect of Calorie Posting Regulation on Consumer Opinion: A Flexible Latent Dirichlet Allocation Model with Informative Priors”, Marketing Science, 36(5), pp 726-746

Rensi, E. (2016), “Thanks To 'Fight For $15' Minimum Wage, McDonald's Unveils Job-Replacing Self-Service Kiosks Nationwide,” Forbes, November 29, https://www.forbes.com/sites/realspin/2016/11/29/thanks-to-fight-for-15-minimum-wage-mcdonalds-unveils-job-replacing-self-service-kiosks-nationwide/#19fde2214fbc

Schneider, B., S. White and M. Paul (1998), “Linking Service Climate and Customer Perceptions of Service Quality: Test of a Causal Model”, Journal of Applied Psychology, 83(2), pp 150-163

Shapiro, C. and J. Stiglitz (1984), “Equilibrium Unemployment as Worker Discipline Device”, American Economic Review, 74 (3), pp 433-444

Shepard, A. (1993), “Contractual Form, Retail Price, and Asset Characteristics in Gasoline Retailing”, RAND Journal of Economics, 24(1), pp 58-77

Simon, D., M. Gomez, E. McLaughlin and D. Wittink (2009), “Employee Attitudes, Customer Satisfaction, and Sales Performance: Assessing the Linkages in US Grocery Stores”, Managerial and Decision Economics, 30(1), pp 27-41

Taylor, K. (2018), “America’s Fast Food Chains are Contemplating Replacing Minimum Wage Workers with Robots – And It Could Lead to a Crisis,” Business Insider, January 11, http://www.businessinsider.com/minimum-wage-increases-spur-fast-food-chains-to-consider-automation-2018-1

Tirunillai, S. and G. Tellis (2014), “Mining Marketing Meaning from Chatter: Strategic Brand Analysis of Big Data Using Latent Dirichlet Allocation,” Journal of Marketing Research. 51(4), 463-479

Zablah, A, B. Carlson, D. Donovan, J. Maxham, and T. Brown (2016), “A Cross-Lagged Test of the Association Between Customer Satisfaction and Employee Job Satisfaction in a Relational Context,” Journal of Applied Psychology, May, 101(5)

Table 1: Dimensions of Service Quality, and Seed Words

Dimensions Descriptions Seed Words

Tangibles Comfort and cleanliness of the dining area Clean, dirty

Attractive and readable menu Menu

ReliabilitySincere interest in correcting anything that is

wrong Mistake

Serve customers in the time promised Wait, delay, fast, slowServe customer’s food exactly as it was

ordered Ordered

Responsiveness Provide prompt and quick service Prompt, seated

Assurance Consistently courteous with customersCourteous, polite, impolite,

friendly, unfriendlyHave the knowledge to answer customers’

questions such as menu items, their ingredients, and methods of preparations Knowledge

Empathy Give customers personal attention Attention

Table 2: Summary Statistics of Reviews for San Jose Restaurants

  Before Wage Increase After Wage Increase Total   Indepen-

dentChain All Indepen-

dentChain All

Number of reviews 16,787 7,701 24,488 19,077 8,844 27,921 52,409Number of unique reviewers

9,988 5,184 12,860 11,798 6,272 15,404 25,009

Number of restaurants reviewed

698 483 1,181 709 486 1,195 1,232

Rating 3.67 (1.23)

3.45 (1.31)

3.60 (1.26)

3.66 (1.26)

3.38 (1.36)

3.57 (1.30)

3.59 (1.28)

Review length 59.84 (47.51)

61.05 (48.83)

60.22 (47.93)

56.55 (46.06)

57.05 (45.90)

56.71 (46.01)

58.35 (46.95)

Number of mentions: "Useful"

1.182 (2.83)

1.07 (2.14)

1.15 (2.63)

1.11 (2.34)

0.96 (2.09)

1.07 (2.26)

1.10 (2.44)

Number of mentions: "Funny"

0.58 (2.31)

0.58 (1.78)

0.58 (2.16)

0.50 (1.69)

0.52 (1.65)

0.51 (1.67)

0.54 (1.92)

Number of mentions: "Cool"

0.59 (2.33)

0.53 (1.60)

0.57 (2.13)

0.51 (1.73)

0.48 (1.63)

0.50 (1.70)

0.53 (1.91)

Table 3: Frequency of Occurrence of Seed Words in Reviews of San Jose Restaurants

  Before Wage Increase After Wage Increase Total  Independ. Chain All Independ. Chain All

Occurrence frequency (per review) for seed words for TANGIBLESClean 0.0508

(0.2401)0.0551

(0.2585)0.0521

(0.2460)0.0453

(0.2311)0.0475

(0.2320)0.0460

(0.2314)0.0489

(0.2384)Dirty 0.0116

(0.1225)0.0171

(0.1449)0.0133

(0.1300)0.0132

(0.1325)0.0185

(0.1609)0.0148

(0.1421)0.0142

(0.1366)Menu 0.1608

(0.4769)0.1424

(0.4585)0.1551

(0.4713)0.1449

(0.4525)0.1244

(0.4142)0.1384

(0.4408)0.1462

(0.4554)Occurrence frequency (per review) for seed words for RELIABILITY

Wait 0.1458 (0.4453)

0.2274 (0.6123)

0.1714 (0.5502)

0.1352 (0.4381)

0.2165 (0.5919)

0.1609 (0.4935)

0.1659 (0.4990)

Delay 0.0004 (0.0189)

0.0005 (0.0279)

0.0004 (0.0221)

0.0007 (0.0280)

0.0002 (0.0150)

0.0005 (0.0247)

0.0005 (0.0235)

Fast 0.0561 (0.2540)

0.1160 (0.4064)

0.0749 (0.3113)

0.0536 (0.2462)

0.0871 (0.3398)

0.0642 (0.2797)

0.0692 (0.2949)

Slow 0.0283 (0.1809)

0.0306 (0.1875)

0.0290 (0.1830)

0.0258 (0.1752)

0.0292 (0.1831)

0.0268 (0.1777)

0.0278 (0.1802)

Ordered 0.2718 (0.6371)

0.2623 (0.6030)

0.2688 (0.6266)

0.2595 (0.6073)

0.2398 (0.5745)

0.2532 (0.5972)

0.2605 (0.6111)

Mistake 0.0043 (0.0743)

0.0062 (0.0881)

0.0049 (0.0789)

0.0038 (0.0730)

0.0067 (0.0900)

0.0048 (0.0788)

0.0049 (0.0788)

Occurrence frequency (per review) for seed word for RESPONSIVENESSPrompt 0.0038

(0.0611)0.0038

(0.0613)0.0038

(0.0612)0.0035

(0.0592)0.0036

(0.0600)0.0035

(0.0594)0.0036

(0.0603)Occurrence frequency (per review) for seed words for ASSURANCE

Courteous 0.0027 (0.0545)

0.0029 (0.0545)

0.0028 (0.0545)

0.0025 (0.0511)

0.0035 (0.0591)

0.0028 (0.0537)

0.0028 (0.0541)

Polite 0.0073 (0.0883)

0.0054 (0.0753)

0.0067 (0.0845)

0.0063 (0.0797)

0.0047 (0.0703)

0.0058 (0.0768)

0.0062 (0.0805)

Impolite 0.0001 (0.0109)

0.0002 (0.0161)

0.0001 (0.0127)

5.2419 (0.0072)

0.0002 (0.0150)

0.0001 (0.0103)

0.0001 (0.0115)

Friendly 0.1060 (0.3255)

0.1023 (0.3156)

0.1048 (0.3224)

0.0972 (0.3103)

0.0907 (0.3093)

0.0952 (0.3100)

0.0997 (0.3159)

Unfriendly 0.0015 (0.0499)

0.0024 (0.0496)

0.0018 (0.0498)

0.0014 (0.0396)

0.0023 (0.0486)

0.0017 (0.0427)

0.0017 (0.0461)

Knowledge 0.0010 (0.0327)

0.0006 (0.0254)

0.0009 (0.0306)

0.0011 (0.0339)

0.0010 (0.0318)

0.0011 (0.0333)

0.0010 (0.0320)

Occurrence frequency (per review) for seed word for EMPATHYAttention 0.0066

(0.0835)0.0075

(0.0964)0.0069

(0.0878)0.0093

(0.1017)0.0099

(0.1090)0.0095

(0.1040)0.0083

(0.0968)

Table 4: Summary Statistics of Reviews for Restaurants in CSC Cities other than San Jose

  Before Wage Increase After Wage Increase Total   Indepen-

dentChain All Indepen-

dentChain All

Number of reviews 7,991 2,714 10,705 9,176 2,962 12,138 22,843Number of unique reviewers 5,606 2,234 7,007 6,615 2,453 8,147 13,718Number of restaurants reviewed 327 182 509 332 181 513 528Rating 3.56

(1.25)3.40

(1.33)3.52

(1.27)3.55

(1.27)3.33

(1.38)3.50

(1.30)3.51

(1.29)Review length 59.28

(48.13)58.36

(46.85)59.04

(47.80)57.75

(47.84)54.30

(41.98)56.91

(46.50)57.91

(47.12)Number of mentions: "Useful"

1.15 (2.41)

1.00 (2.25)

1.11 (2.37)

1.18 (2.58)

0.87 (1.87)

1.11 (2.43)

1.11 (2.40)

Number of mentions: "Funny"

0.55 (1.79)

0.52 (1.80)

0.54 (1.79)

0.51 (1.80)

0.44 (1.57)

0.50 (1.75)

0.52 (1.77)

Number of mentions: "Cool"

0.53 (1.71)

0.47 (1.64)

0.52 (1.70)

0.49 (1.85)

0.38 (1.44)

0.46 (1.76)

0.49 (1.73)

Table 5: Frequency of Occurrence of Seed Words in Reviews of CSC Restaurants not in San Jose

  Before Wage Increase After Wage Increase Total  Independ. Chain All Independ. Chain All

Occurrence frequency (per review) for seed words for TANGIBLESClean 0.0561

(0.2480)0.0459

(0.2374)0.0535

(0.2454)0.0501

(0.2327)0.0554

(0.2512)0.0514

(0.2374)0.0524

(0.2412)Dirty 0.0116

(0.1304)0.0165

(0.1513)0.0128

(0.1360)0.0086

(0.1096)0.0221

(0.1725)0.0119

(0.1280)0.0123

(0.1318)Menu 0.1660

(0.4773)0.1464

(0.4579)0.1610

(0.4725)0.1455

(0.4581)0.1491

(0.4758)0.1464

(0.4625)0.1533

(0.4672)Occurrence frequency (per review) for seed words for RELIABILITY

Wait 0.1357 (0.4274)

0.1791 (0.4905)

0.1467 (0.4446)

0.1434 (0.4398)

0.1592 (0.4522)

0.1473 (0.4429)

0.1470 (0.4437)

Delay 0.0002 (0.0158)

0.0007 (0.0271)

0.0003 (0.0193)

0.0006 (0.0255)

0.0010 (0.0317)

0.0007 (0.0272)

0.0005 (0.0238)

Fast 0.0593 (0.2570)

0.0732 (0.2962)

0.0628 (0.2675)

0.0555 (0.2460)

0.0759 (0.3211)

0.0605 (0.2665)

0.0616 (0.2670)

Slow 0.0285 (0.1876)

0.0412 (0.2511)

0.0317 (0.2057)

0.0300 (0.1907)

0.0440 (0.2298)

0.0334 (0.2010)

0.0326 (0.2032)

Ordered 0.2781 (0.6259)

0.2792 (0.6284)

0.2784 (0.6265)

0.2756 (0.6117)

0.2473 (0.5946)

0.2686 (0.6077)

0.2732 (0.6166)

Mistake 0.0040 (0.0688)

0.0069 (0.0876)

0.0047 (0.0740)

0.0041 (0.0807)

0.0036 (0.0606)

0.0040 (0.0763)

0.0043 (0.0752)

Occurrence frequency (per review) for seed word for RESPONSIVENESSPrompt 0.0065

(0.0804)0.0055

(0.0740)0.0062

(0.0788)0.0052

(0.0721)0.0043

(0.0659)0.0050

(0.0706)0.0055

(0.0746)Occurrence frequency (per review) for seed words for ASSURANCE

Courteous 0.0043 (0.0660)

0.0040 (0.0634)

0.0042 (0.0654)

0.0034 (0.0589)

0.0013 (0.0366)

0.0029 (0.0543)

0.0035 (0.0597)

Polite 0.0061 (0.0796)

0.0073 (0.0854)

0.0064 (0.0811)

0.0088 (0.0980)

0.0087 (0.0930)

0.0088 (0.0968)

0.0076 (0.0898)

Impolite 0.0003 (0.0193)

0.000 (0.000)

0.0002 (0.0167)

0.0003 (0.0233)

0.0003 (0.0183)

0.0003 (0.0222)

0.0003 (0.0198)

Friendly 0.1077 (0.3304)

0.0960 (0.3093)

0.1047 (0.3252)

0.1016 (0.3187)

0.0924 (0.2999)

0.0994 (0.3142)

0.1019 (0.3194)

Unfriendly 0.0007 (0.0316)

0.0022 (0.0469)

0.0011 (0.0361)

0.0013 (0.0361)

0.0030 (0.0607)

0.0017 (0.0434)

0.0014 (0.0402)

Knowledge 0.0008 (0.0335)

0.0018 (0.0428)

0.0011 (0.0361)

0.0011 (0.0346)

0.0010 (0.0317)

0.0011 (0.0339)

0.0011 (0.0349)

Occurrence frequency (per review) for seed word for EMPATHYAttention 0.0075

(0.0946)0.0095

(0.0973)0.0080

(0.0953)0.0069

(0.0870)0.0090

(0.0983)0.0074

(0.0899)0.0077

(0.0925)

Table 6: Top Words (or Phrases) Semantically Proximate to Seed Words

Service Quality Dimensions

Seed Words

Top 10 Closest Words ( in terms of cosine distance)

Tangibles Clean well_maintained spotless nicely_decorated bathroom_clean clean_modern clean_bathrooms bathrooms_clean tidy clean_bright clean_spacious

Dirty filthy unsanitary tables_sticky wiped grimy unclean dirty_floors wipe dirty_sticky sanitary

Menu items menus extensive_menu listed menu_extensive descriptions extensive specials item selections

Reliability Mistake made_mistake error wrong screwed corrected messed argued correct incorrectly insisted

Wait waiting waits minute_wait long waited busy hour min ended_waiting arrive

Delay delayed delays arrive ended_waiting apologetic apologies backed takes_longer unacceptable receiving

Fast quick speedy fast_efficient efficient fairly_quick relatively_fast inexpensive quickly quick_efficient relatively_quick

Slow extremely_slow understaffed spotty slower inattentive incredibly_slow busy staffed forgetful disorganized

Ordered orders ordering ordered ready pick ask placing_order point correct placed_order

Responsiveness Prompt timely attentive speedy courteous friendly_prompt prompt_friendly friendly_attentive efficient fast_efficient prompt_courteous

Assurance Courteous polite friendly prompt attentive helpful friendly_helpful friendly_efficient friendly_courteous cheerful efficient

Polite courteous friendly helpful cheerful attentive respectful cordial smiling pleasant welcoming

Impolite acted rude extremely_rude unfriendly rude_unfriendly unwelcoming condescending rudeness ignored_us ignored

Friendly friendly_helpful courteous polite welcoming cheerful nice friendly_attentive friendly_welcoming helpful friendly_efficient

Unfriendly rude extremely_rude grumpy unwelcoming rude_unfriendly indifferent inattentive surly attitude unhappy

Knowledge knowledgeable knowledgable suggestions informative recommendations educated genuine asked_questions ben answer_questions

Empathy Attention flag someone_attention water_refill ignored flag_someone wave_someone refills_water flag_server refill_water wave

Price Price prices pricing expensive pricey quality value bit_pricey priced cost little_pricey

Table 7: Results of Difference in Difference Analysis (All Restaurants)

Service topic Post X SanJose

SanJose

Post Rating UsefulCount

CoolCount

FunnyCount

Rev. Length

Clean -0.024 (0.052)

1.012 (0.962)

-0.040 (0.044)

0.804 (0.011)

-0.049 (0.010)

0.043 (0.016)

-0.092 (0.014)

-0.021 (0.000)

Dirty -0.054 (0.036)

-0.559 (0.667)

0.049 (0.031)

-0.738 (0.007)

-0.023 (0.007)

-0.033 (0.011)

0.028 (0.010)

0.000 (0.000)

Menu -0.017 (0.042)

-0.842 (0.774)

-0.018 (0.036)

0.519 (0.009)

-0.039 (0.008)

0.076 (0.013)

-0.126 (0.011)

-0.007 (0.000)

Wait -0.104 (0.052)

-1.721 (0.973)

0.142 (0.045)

0.098 (0.011)

-0.028 (0.010)

0.031 (0.017)

-0.116 (0.014)

-0.008 (0.000)

Delay -0.092 (0.043)

-1.639 (0.800)

0.172 (0.037)

-0.780 (0.009)

0.024 (0.008)

0.038 (0.014)

-0.132 (0.012)

0.011 (0.000)

Fast -0.053 (0.054)

0.040 (0.993)

0.004 (0.046)

0.708 (0.011)

-0.045 (0.010)

0.036 (0.017)

-0.086 (0.014)

-0.023 (0.000)

Slow -0.129 (0.047)

-1.289 (0.875)

0.151 (0.040)

-0.424 (0.010)

-0.069 (0.009)

0.092 (0.015)

-0.128 (0.013)

-0.010 (0.000)

Ordered -0.047 (0.058)

-1.655 (1.072)

0.075 (0.049)

-0.129 (0.012)

-0.066 (0.011)

0.079 (0.018)

-0.119 (0.015)

0.000 (0.000)

Mistake -0.063 (0.044)

-1.226 (0.812)

0.097 (0.037)

-1.166 (0.009)

0.026 (0.009)

-0.046 (0.014)

-0.004 (0.012)

0.014 (0.000)

Prompt -0.031 (0.052)

-0.564 (0.969)

0.072 (0.044)

0.841 (0.011)

-0.037 (0.010)

0.111 (0.017)

-0.197 (0.014)

-0.018 (0.000)

Courteous -0.028 (0.051)

-0.689 (0.956)

0.085 (0.044)

0.881 (0.010)

0.018 (0.010)

0.010 (0.016)

-0.132 (0.014)

-0.018 (0.000)

Polite -0.054 (0.050)

-0.669 (0.940)

0.126 (0.043)

0.443 (0.010)

0.019 (0.010)

-0.040 (0.016)

-0.074 (0.013)

-0.013 (0.000)

Impolite -0.061 (0.034)

-0.880 (0.619)

0.120 (0.028)

-0.847 (0.007)

0.046 (0.007)

-0.094 (0.011)

0.036 (0.009)

0.002 (0.000)

Friendly -0.034 (0.062)

-0.581 (1.148)

0.031 (0.053)

1.314 (0.013)

-0.028 (0.012)

0.021 (0.020)

-0.126 (0.016)

-0.029 (0.000)

Unfriendly -0.084 (0.040)

-0.578 (0.735)

0.113 (0.034)

-0.871 (0.008)

0.011 (0.008)

-0.053 (0.013)

-0.017 (0.011)

-0.005 (0.000)

Knowledge 0.007 (0.035)

-2.374 (0.651)

0.018 (0.030)

0.298 (0.007)

0.022 (0.007)

-0.003 (0.011)

-0.028 (0.009)

-0.002 (0.000)

Attention -0.058 (0.039)

-1.106 (0.729)

0.157 (0.033)

-0.802 (0.008)

0.007 (0.008)

-0.018 (0.012)

-0.051 (0.011)

0.007 (0.000)

Note: parameters in bold are significant at the 5% level

Table 8: Results of Difference in Difference Analysis (Independent Restaurants)

Service topic

Post X SanJose

SanJose Post Rating UsefulCount

CoolCount

FunnyCount

Rev. Length

Clean -0.054 (0.060)

0.035 (0.718)

-0.030 (0.050)

0.802 (0.013)

-0.064 (0.012)

0.050 (0.019)

-0.081 (0.016)

-0.021 (0.001)

Dirty -0.091 (0.042)

-1.542 (0.495)

0.077 (0.035)

-0.749 (0.009)

-0.028 (0.008)

-0.041 (0.013)

0.040 (0.011)

0.001 (0.000)

Menu -0.029 (0.048)

-1.722 (0.576)

-0.028 (0.040)

0.508 (0.010)

-0.041 (0.009)

0.082 (0.015)

-0.125 (0.013)

-0.008 (0.000)

Wait -0.143 (0.060)

-1.958 (0.715)

0.144 (0.050)

0.090 (0.013)

-0.034 (0.012)

0.036 (0.019)

-0.108 (0.016)

-0.009 (0.000)

Delay -0.101 (0.050)

-0.480 (0.594)

0.163 (0.042)

-0.787 (0.010)

0.019 (0.010)

0.032 (0.016)

-0.111 (0.014)

0.011 (0.000)

Fast -0.089 (0.062)

-2.044 (0.741)

0.020 (0.052)

0.688 (0.013)

-0.053 (0.012)

0.040 (0.019)

-0.077 (0.017)

-0.024 (0.000)

Slow -0.168 (0.055)

-1.248 (0.652)

0.162 (0.046)

-0.455 (0.011)

-0.078 (0.011)

0.088 (0.017)

-0.112 (0.015)

-0.010 (0.000)

Ordered -0.069 (0.067)

-4.137 (0.805)

0.093 (0.057)

-0.155 (0.014)

-0.070 (0.013)

0.091 (0.021)

-0.123 (0.018)

0.000 (0.000)

Mistake -0.074 (0.051)

-1.795 (0.608)

0.114 (0.043)

-1.180 (0.011)

0.031 (0.010)

-0.057 (0.016)

0.005 (0.014)

0.014 (0.000)

Prompt -0.042 (0.061)

0.718 (0.724)

0.052 (0.051)

0.832 (0.013)

-0.050 (0.012)

0.116 (0.019)

-0.183 (0.016)

-0.019 (0.000)

Courteous -0.035 (0.060)

0.063 (0.712)

0.067 (0.050)

0.894 (0.012)

0.005 (0.011)

0.014 (0.019)

-0.115 (0.016)

-0.019 (0.000)

Polite -0.064 (0.058)

-0.206 (0.698)

0.113 (0.049)

0.448 (0.012)

0.004 (0.011)

-0.034 (0.018)

-0.058 (0.016)

-0.014 (0.000)

Impolite -0.080 (0.039)

0.556 (0.459)

0.129 (0.032)

-0.840 (0.008)

0.041 (0.007)

-0.101 (0.012)

0.050 (0.010)

0.002 (0.000)

Friendly -0.059 (0.072)

-0.657 (0.860)

0.024 (0.060)

1.318 (0.015)

-0.045 (0.014)

0.034 (0.022)

-0.116 (0.019)

-0.030 (0.000)

Unfriendly -0.109 (0.046)

-0.204 (0.546)

0.118 (0.038)

-0.884 (0.010)

0.002 (0.009)

-0.055 (0.014)

0.000 (0.012)

-0.005 (0.000)

Knowledge

0.031 (0.041)

-0.244 (0.486)

-0.017 (0.034)

0.308 (0.009)

0.026 (0.008)

-0.001 (0.013)

-0.031 (0.011)

-0.002 (0.000)

Attention -0.084 (0.045)

-0.996 (0.539)

0.148 (0.038)

-0.810 (0.009)

-0.001 (0.009)

-0.026 (0.014)

-0.028 (0.012)

0.007 (0.000)

Note: parameters in bold are significant at the 5% level

Table 9: Results of Difference in Difference Analysis (Chain Restaurants)

Service topic

Post X SanJose

SanJose Post Rating UsefulCount

CoolCount

FunnyCount

Rev. Length

Clean 0.044 (0.103)

-0.846 (0.755)

-0.071 (0.090)

0.823 (0.019)

-0.004 (0.020)

0.030 (0.033)

-0.126 (0.027)

-0.020 (0.001)

Dirty 0.033 (0.073)

-1.826 (0.531)

-0.021 (0.063)

-0.715 (0.014)

-0.013 (0.014)

-0.018 (0.023)

0.003 (0.019)

-0.001 (0.000)

Menu 0.032 (0.084)

-1.375 (0.616)

-0.023 (0.073)

0.549 (0.016)

-0.028 (0.017)

0.091 (0.027)

-0.146 (0.022)

-0.006 (0.000)

Wait 0.017 (0.108)

-0.407 (0.788)

0.107 (0.094)

0.129 (0.020)

-0.007 (0.021)

0.020 (0.035)

-0.147 (0.028)

-0.009 (0.001)

Delay -0.069 (0.088)

-1.765 (0.640)

0.186 (0.076)

-0.777 (0.016)

0.034 (0.017)

0.061 (0.028)

-0.188 (0.023)

0.010 (0.000)

Fast 0.056 (0.107)

-2.953 (0.783)

-0.050 (0.093)

0.766 (0.020)

-0.021 (0.021)

0.033 (0.034)

-0.116 (0.028)

-0.023 (0.001)

Slow -0.022 (0.095)

-0.878 (0.691)

0.106 (0.082)

-0.352 (0.018)

-0.048 (0.019)

0.115 (0.030)

-0.172 (0.024)

-0.011 (0.000)

Ordered 0.016 (0.114)

-3.707 (0.829)

0.004 (0.099)

-0.073 (0.021)

-0.053 (0.023)

0.089 (0.037)

-0.135 (0.029)

0.001 (0.001)

Mistake -0.031 (0.087)

-3.222 (0.640)

0.043 (0.076)

-1.145 (0.016)

0.010 (0.017)

-0.021 (0.028)

-0.025 (0.022)

0.014 (0.000)

Prompt 0.001 (0.104)

0.454 (0.759)

0.112 (0.090)

0.862 (0.019)

-0.004 (0.021)

0.118 (0.033)

-0.242 (0.027)

-0.018 (0.001)

Courteous -0.008 (0.103)

-0.393 (0.752)

0.124 (0.090)

0.850 (0.019)

0.051 (0.020)

-0.011 (0.033)

-0.174 (0.026)

-0.018 (0.001)

Polite -0.024 (0.102)

-0.848 (0.743)

0.147 (0.089)

0.436 (0.019)

0.055 (0.020)

-0.076 (0.033)

-0.110 (0.026)

-0.013 (0.001)

Impolite -0.014 (0.068)

-1.075 (0.497)

0.105 (0.059)

-0.872 (0.013)

0.052 (0.013)

-0.104 (0.022)

0.013 (0.017)

0.001 (0.000)

Friendly 0.042 (0.122)

-0.863 (0.888)

0.027 (0.106)

1.324 (0.023)

0.022 (0.024)

-0.019 (0.039)

-0.158 (0.031)

-0.028 (0.001)

Unfriendly -0.028 (0.080)

-1.456 (0.582)

0.111 (0.069)

-0.847 (0.015)

0.033 (0.016)

-0.070 (0.026)

-0.046 (0.021)

-0.006 (0.000)

Knowledge -0.073 (0.070)

1.400 (0.513)

0.109 (0.061)

0.269 (0.013)

0.012 (0.014)

-0.009 (0.023)

-0.018 (0.018)

-0.001 (0.000)

Attention -0.006 (0.081)

0.048 (0.590)

0.169 (0.07)

-0.786 (0.015)

0.025 (0.016)

-0.007 (0.026)

-0.103 (0.021)

0.007 (0.000)

Note: parameters in bold are significant at the 5% level

Appendix A1: Topics Estimated with Latent Dirichlet Allocation

Topic0: food place ordered didn bad got wasn really tasted taste better even came good back much one go eat never

Topic10: good sauce place fries shrimp wait wings really got spicy ordered order seafood also pretty go hot came much try

Topic1: pizza good place cheese great toppings one pizzas crust really love sauce slice order go also ordered best large fresh

Topic11: always place food love ve go great service good come time wait order usually never best favorite one times eat

Topic2: place one don food people re go ve restaurant now know time location going ll think want see way even

Topic12: food good place chicken great indian delicious also try restaurant really best dishes fresh go one spicy tried rice definitely

Topic3: good ordered dinner came salad steak restaurant nice menu dish meal great dessert also us service really one pasta $

Topic13: good burrito food tacos mexican place salsa carne_asada super taco meat one chicken go really sauce burritos also best great

Topic4: pho place good vietnamese noodles broth soup dim_sum restaurant bowl beef meat also pretty one really rice places order service

Topic14: us food table came service wait time minutes order got one server seated asked didn ordered took waitress waiter people

Topic5: great food service place friendly staff amazing delicious us back definitely restaurant nice excellent love best awesome experience good made

Topic15: food good service place pretty price great restaurant really better nice decent quality stars go bad ok prices much average

Topic6: sandwich sandwiches bread good place salad fresh one also lunch love go chicken $ cheese really meat great ve delicious

Topic16: ramen good place bowl broth really noodles tea also got pretty one spicy soup try pork little flavor small $

Topic7: sushi good place rolls roll fish fresh really also rice pretty sashimi one great $ definitely japanese service salmon go

Topic17: breakfast good place coffee food great really ordered got came service try also menu nice wait definitely delicious pretty pancakes

Topic8: meat good chicken bbq sauce side place ribs really beef rice food also korean got pretty ordered came spicy plate

Topic18: burger good place fries drinks bar great burgers happy_hour drink really pretty beer $ nice food also go one menu

Topic9: chicken good food dish dishes thai soup spicy rice ordered chinese beef restaurant sauce fried_rice also curry really place noodles

Topic19: order said food asked back one told go didn got us never even time ordered went don service manager customer_service

Appendix A2: Topics Estimated with Non-Negative Matrix Factorization

Topic0: don go order one know re want eat people much

Topic 10: chicken fried wings salad curry teriyaki butter rice ordered dry

Topic1: pizza crust toppings pizzas cheese slice thin_crust delivery slices sauce

Topic 11: place try nice recommend clean go definitely small awesome looking

Topic2: great service friendly staff awesome atmosphere definitely delicious amazing drinks

Topic 12: ramen broth noodles pork bowl egg salty santouka soup Japanese

Topic3: sauce rice spicy ordered dish dishes also beef soup meat

Topic 13: food service restaurant fast quality chinese excellent friendly indian price

Topic4: sushi rolls roll fresh fish sashimi quality japanese tempura salmon

Topic 14: burrito super tacos carne_asada mexican salsa burritos taco orange_sauce meat

Topic5: good pretty service price overall prices though decent little portions

Topic 15: time wait first long come back next every lunch try

Topic 6: pho broth vietnamese noodles bowl meat beef places spring_rolls soup

Topic 16: sandwich sandwiches bread cheese lunch fresh meat bbq steak salad

Topic 7: us came table got asked minutes ordered didn server order

Topic 17: love delicious amazing favorite yummy also absolutely coming awesome everything

Topic 8: burger fries burgers cheese sweet_potato bacon garlic bun five_guys toppings

Topic 18: best ve one tried far san_jose bay_area amazing places favorite

Topic 9: always friendly staff favorite fresh never order usually come go

Topic 19: really nice enjoyed liked also super lot got didn enjoy