Worlds Colliding

Jesse Michael Fagan

Lecturer of Data Analytics

University of Exeter Business School

http://www.jessefagan.com/

Luxury and Standard

Standard, Inc.

(target)

Luxury, Inc.

(acquirer)

Standard with about 1.5B in revenue, and Luxury, with just over 1.5B in revenue, together they made

Luxury-Standard, Inc.

June, T1

March

June, T2

Merger

Formally

Ratified

Merger Process

Email data collection

Networks are computed using 30 days of communications in June T1, and June T2.

Topics were extracted from communications across all time points.

Data Sources

Survey of Job Attitudes

(3 year longitudinal, N > 1,400)

HR Performance / Demographic Data

Email Network

(sampled at 4 time points over 15 months, N > 2,000)

Email Content

(4M cleaned English, internal messages; attachment info; subject headers; just all the stuff)

April

2013

Immediately after the ratification of the merger.

Colors indicate the function or department.

The first function to start merging was Information Technology.

Luxury

Standard

April

2013

Immediately after the ratification of the merger.

Colors indicate the function or department.

The first function to start merging was Information Technology.

Luxury

Standard

August 2013

June

2014

Luxury

Standard

Standard

Luxury

Time 1

Time 2

11

12

Researchers have proposed a multitude of mechanisms to explain this advantage. And while brokers may benefit from learning, autonomy, or control, researchers frequently propose some form of non-redundant, diverse, heterogenous, or novel information is responsible for providing a vision advantage. (Rodan, 2010; Burt, 2008; Aral & van Alstyne, 2011)

Broker

I’m going to be talking about the bva

Brokers have an advantage. Lots of discussion of mechanisms. But this comes down to you have access to these information things. What I’m going to do is break apart this vision advantage and

Understand this vision advantage.

Predicting Manager Performance

Junior executives were rated on their promotability by the executives in the firm.

An strong predictor was the amount structural holes in their network.

Those who had dense networks were rated as exit (they should be removed).

Those who had lots of structural holes were rated as promote

15

Constraint

Constraint relates network density and size.

An increase in constraint tends to increase an increase in closure or a decrease in network size.

A decrease in constraint tends to indicate an increase in structural holes.

http://yannatry.blogspot.com/2010/05/hampton-court-palace.html

Better ideas.

Improved creativity.

Early recognition of opportunities.

(Brass, 1995; Burt, 2004, 2005; Perry-Smith, 2014; Arenius & De Clercq, 2005; Ozgen & Baron, 2007; Rodan, 2010; Brass, 1984)

Vision

Socially distant information

Using locally rare

information

Diverse Information

Recombining diverse and varied information

e.g. Fleming & Sorenson 2004, Laursen 2012

e.g. Granovetter 1973, Rosenkopf & Nerkar 2001

Splitting

the Vision

Advantage

Is it just that you need diverse information, or is it more important that you actually reach out and find socially distant information.

Is it just access or is it actually adopting the information into use?

Topic Models

Topics are an indication of common patterns of language use in email messages.

Each email is a mixture topics.

Each topic is a mixture of terms.

(Blei, Carin, & Dunson, 2010; Blei et al., 2003)

Quantify and measure information. To do that I use probilistic topic modeling. This is an example email. It’s a mixture of mainly three different topics. And each topic is a mixture of different terms. Maybe this topic is about blah blah. So every email a person sends is a mixture of topics. I can use these results to determine which topics a person has access to, and what topics that have adopted into use by sending messages that are weighted on a topic.

Luxury

Standard

Standard

Luxury

V81

(truck, trailer, pick, load, driver,...)

Both networks are from Time 1. The gold nodes are in the top quintile for discussing the topics described in the titles.

These topics are very predictive of legacy membership.

V44

(price, pricing, catalog, book, prices,,...)

Information Variance

Information variance is the average cosine distance between the topic vectors of each email message and the mean topic vector for the individual. (c.f. Aral and Van Alstyne, 2011)

This is computed for both sent and received messages.

The final information variance score is then multiplied by the total volume of messages the individual sends or receives.

Socially Distant Information

Because information is unevenly distributed throughout the network, the greater the number of steps from, the more likely you are to encounter information that is entirely new to you.

The “value” of a topic is related to the average distance that topic is from the ego. The distances are normalized from [0,1].

Luxury

Standard

Time 1

Time 2

Red node is one of the top nodes for adopting socially distant information. One of the biggest contributors was topic V40. The gold nodes are in the top quintile for discussing V40.

Access to Diverse Information and Salary

There is a stark difference in the benefits of increases in access to information variance in the stable context compared to the turbulent context.

Benefiting from Information

Those at luxury benefitted from having diverse information.

And that benefit was amplified if the information also came from socially distant sources.

Those at Standard benefited more from Luxury information or high ranking information.

What does it take to succeed following a merger?

Luxury

Standard

Standard

Luxury

Time 1

Time 2

An Integrating Topic Space

Merging ideas, but not ties

People moved through the topic space with ease..

..but had less success with forming new ties with people.

Luxury and High-Rank at Standard

There was no interaction effect at Standard between high ranking and luxury information.

Thus it doesn’t matter where the high-ranking information comes from.

High ranking Luxury information doesn’t have an added multiplier over high ranking Standard information.

Knowledge is power

Knowledge is a competitive advantage.

The mechanism tying network position to individual advantage is frequently the knowledge that flows through position (e.g. jobs, opportunities)

(e.g. Tsai 2001, Burt 1992 / 2004)

Rank Topic Change

Luxury Topic Change

Δ Salary

Friends in High places

Contacts in high places lead to career sponsorship or references.

(Seibert 2001)

Social networks provide resources.

(Lin 1999)

Δ Salary

Δ Network Rank Difference

Δ Inter-Legacy Ties

Connecting by Knowing

Adopting the cultural patterns and knowledge of powerful people will make it easier to form a relationship with powerful people and increase one’s social capital (Weber & Camerer, 2003)

In a study of the merger of a Finnish and a Swedish bank (Vaara, Tienari, Piekkari, & Säntti, 2005), the newly merged organization was forced to make the choice of the official language of the organization - Swedish or Finnish.

Merging organizations have two separate languages, and the use of language is an early indicator if someone belongs to a group or does not (Milroy & Milroy, 1992).

Does info gives access to networks?

Rank Topic Change

Luxury Topic Change

Δ Salary

Δ Network Rank Difference

Δ Inter-Legacy Ties

Does a network give access to info?

Rank Topic Change

Luxury Topic Change

Δ Salary

Δ Network Rank Difference

Δ Inter-Legacy Ties

Topic Models

Topics are an indication of common patterns of language use in email messages.

Each email is a mixture topics.

Each topic is a mixture of terms.

(Blei, Carin, & Dunson, 2010; Blei et al., 2003)

Quantify and measure information. To do that I use probilistic topic modeling. This is an example email. It’s a mixture of mainly three different topics. And each topic is a mixture of different terms. Maybe this topic is about blah blah. So every email a person sends is a mixture of topics. I can use these results to determine which topics a person has access to, and what topics that have adopted into use by sending messages that are weighted on a topic.

Predicting Legacy Membership

Used a penalized regression model (LASSO) to predict the legacy membership of an individual based on the mean topic scores they use.

Model has an accuracy of 96%, a specificity of 0.94 and sensitivity of 0.94.

Measuring Luxury Topic Change

Topic change is measured by first looking at the per-topic difference from 2013 to 2014. Then each difference is weighted using the coefficients from a LASSO model. The weighted differences are then summed to create the final measure.

Luxury Topics

V139

V186

V189

V145

head

lunch

base

ship

covers

bring

PRODUCT

shipped

PRODUCT

hot

foundation

shipping

PRODUCT

house

foundations

order

seat

eat

touch

tracking

PRODUCT

buy

bases

shipment

foot

nice

PRODUCT

ordered

comfort

food

adjustable

ups

tup

lol

flat

arrive

Standard Topics

V2

V161

V181

V17

BRAND

plant

truck

PRODUCT

PRODUCT

NAME

trailer

NAME

hybrid

plants

pick

set

level

PLANT NAME

load

firm

BRAND

PLANT NAME

driver

PRODUCT

gel

PLANT NAME

trailers

sets

COMPANY

PLANT NAME

drop

top

BRAND

PLANT NAME

trucks

frame

latex

PLANT NAME

deliver

end

Predicting Rank

Used a penalized regression model (LASSO) to predict the rank of the individual based on the mean topic scores they use.

Model has an R2 of 0.82

Low Rank Topics

V149

V176

V3

V118

order

told

love

system

bom

called

lol

show

subassembly

asked

forget

enter

released

back

gsm

entered

release

left

baby

showing

hold

yesterday

doc

oracle

entered

spoke

header

set

cancel

gave

miss

shows

purchase

talked

follow

manually

High Rank Topics

V70

V126

V62

V170

forward

agreement

team

understand

move

legal

review

feel

working

contract

discuss

agree

moving

NAME

meeting

clear

asked

signed

plan

situation

work

company

discussion

decision

assist

agreements

discussed

making

complete

sign

agenda

things

moved

terms

thoughts

comfortable

Inter-Legacy Ties

Inter-legacy ties is the count of ties in an ego’s network that connect to a member of the other legacy organization.

These values are standardized within year.

Network Rank Difference

Network rank difference is computed at each time point as the mean difference between the ego and the alters.

Analyzing

Change on Change

Latent difference score models provide a reliable way of estimating “true change”, accounting for

measurement errors at each

time point, and controlling for

Interindividual differences

in change.

The hypothesis is tested

on the path

β between

△Constraint

and △Salary

Results

Rank Topic Change

Luxury Topic Change

Δ Salary

Δ Network Rank Difference

Δ Inter-Legacy Ties

Results - Ties give access to info

AIC = 6939

CFI = 0.987

Chi-Square = 73.17

Df = 18

Rank Topic Change

Luxury Topic Change

Δ Salary

Δ Network Rank Difference

Δ Inter-Legacy Ties

Results - Info gives access to ties

Rank Topic Change

Luxury Topic Change

Δ Salary

Δ Network Rank Difference

Δ Inter-Legacy Ties

Results - Info gives access to ties

AIC = 6935

CFI = 0.988

Chi-Square = 69.83

Df = 18

Rank Topic Change

Luxury Topic Change

Δ Salary

Δ Network Rank Difference

Δ Inter-Legacy Ties

What do the

linguistic and social features

of email communication

reveal about workplace attitudes?

Job Insecurity

Job Satisfaction

Intention to Turnover

Organizational Identification

Affective Commitment

Continuance Commitment

Distributive Justice

Procedural Justice

Research is revealing the subtle ways in which personality traits, emotions, moods, and are reflected in language

(Cohen, Minor, Baillie, & Dahir, 2008; Duriau, Reger, & Pfarrer, 2007; Pearl & Steyvers, 2010; Pennebaker & King, 1999; Platanou et al., 2017; Yarkoni & Westfall, 2017)

Rather than using machines to predict behavior and attitudes..

This study showed everyone that we could use machine learning to change people’s behavior and attitudes.

Marketers have known we change change behavior and attitudes with text.

Recent Facebook studies have shown how efficient, targetted, and effective these methods have become.

Kramer, A. D. I., Guillory, J. E., & Hancock, J. T. 2014. Experimental evidence of massive-scale emotional contagion through social networks. Proceedings of the National Academy of Sciences of the United States of America, 111(24): 8788–8790.

Bond, RM, Fariss, CJ, Jones, JJ, Kramer, ADI, Marlow, C, Settle, JE, Fowler, JH (2012) A 61-million-person experiment in social influence and political mobilization. Nature 489: 295–298.

https://www.theverge.com/2018/3/27/17165150/facebook-face-recognition-how-to-turn-off-disable

Linguistic Signature of Job Attitudes

Linguistic Properties of Emails Authored

Mental State

Linguistic Properties of Emails Received

Personality &

Organizational Situation

Responses to Survey on Job Attitudes

LIWC Operator Manual

https://s3-us-west-2.amazonaws.com/downloads.liwc.net/LIWC2015_OperatorManual.pdf

LIWC 2015 Work category

Talk a bit about

The Development and Psychometric Properties of LIWC2015

https://repositories.lib.utexas.edu/bitstream/handle/2152/31333/LIWC2015_LanguageManual.pdf

LIWC 2015 Categories

Luxury

Standard

Standard

Luxury

Time 1

Time 2

Predicting Legacy Membership on Linguistic Style

Using an elastic net model with α = 0.2.

Reference

Prediction 0 1

0 430 61

1 44 332

Accuracy : 0.8789

Reference

Prediction 0 1

0 396 40

1 56 388

Accuracy : 0.8909

Luxury

Standard

Standard

Luxury

Time 1

Time 2

Data comes from Time 2 -->

Trained eight xgboost models optimized using Bayesian optimization to predict high scores for each attitude.

Job Attitude: (Test-AUC)

Intent to Stay: 0.77

Job Satisfaction: 0.75

Job Insecurity: 0.82

Org. Identification: 0.85

Procedural Justice: 0.81

Distributive Justice: 0.73

Affective Commitment: 0.77

Continuance Commitment: 0.91

Trained eight xgboost models optimized using Bayesian optimization to predict high scores for each attitude.

Each model had a good or very good fit using the area under the curve on a test sample.

The outcome was binarized as a 6 or 7 on 7-point scale for each attitude, binary outcome

Used the mean, minimum, maximum, and percent of messages greater than zero of the LIWC features from the time the survey was conducted..

Model performance

for predicting

Job Satisfaction

Job Attitude: (Test-AUC)

Intent to Stay: 0.77

Job Satisfaction: 0.75

Job Insecurity: 0.82

Org. Identification: 0.85

Procedural Justice: 0.81

Distributive Justice: 0.73

Affective Commitment: 0.77

Continuance Commitment: 0.91

Model performance for predicting

Job Satisfaction

Job Attitude: (Test-AUC)

Intent to Stay: 0.77

Job Satisfaction: 0.75

Job Insecurity: 0.82

Org. Identification: 0.85

Procedural Justice: 0.81

Distributive Justice: 0.73

Affective Commitment: 0.77

Continuance Commitment: 0.91

Intent to Stay

Words like:

fuck, damn, shit

Intent to Stay

Words like:

mate, talk, they

Distributive Justice

Words like:

few, many, much

Distributive Justice

Words like:

feel, touch

Affective Commitment

Words like:

cause, know, ought

Affective Commitment

Words like:

ally, win, superior, take

Dyadic Changes

Get base prediction using each job attitude model.

Dyadic Changes

Base prediction

Distrib. Justice

Org. Ident

Intent to Stay

Dyadic Changes

Base prediction

Distrib. Justice

Org. Ident

Intent to Stay

Dyadic Changes

Base prediction

Distrib. Justice

Org. Ident

Intent to Stay

Dyadic Changes

Base prediction

Distrib. Justice

Org. Ident

Intent to Stay

Dyadic Changes

Base prediction

Distrib. Justice

Org. Ident

Intent to Stay

Dyadic Changes

Base prediction

Distrib. Justice

Org. Ident

Intent to Stay

Dyadic Changes

Base prediction

Distrib. Justice

Org. Ident

Intent to Stay

Dyadic Changes

Base prediction

Distrib. Justice

Org. Ident

Intent to Stay

Each point is an individual. This is the mean difference in their predicted attitude from removing received or sent dyads.

This is aggregated by job attitude. It shows the variance of the mean of the differences.

Generally it shows how influential individual dyads are on the prediction of each job attitude.

There seems to be no dyadic influence on job satisfaction. While OrgID, procedural justice, and distributive justice, seem to be influenced by others.

This is the variance of the

Since the “diff” variable was incredibly skewed. m

A positive score here means that removing messages sent to a higher rank tends to increase the predicted job attitude measure.

Removing messages received from older colleagues increased the predicted job attitude score.

People within Standard are better off if they aren’t sending messages to other members of Standard. Except for continuance commitment.

All effects here seem to be within the Male to Male dyads.

Thanks!

Jesse Fagan

@jessemfagan

jessefagan.com/worlds-colliding/

Summary

There are strong signals of job attitudes in email content.

The signals for some attitudes are concentrated in one or a few relationships, others are more dispersed.

Demographic characteristics influence the likelihood of changes (rank, age, legacy membership)

Conclusions

Interactions

Haven’t looked at interactions yet.

Information without people doesn’t go far Adopting the information without connecting with people doesn’t seem to do much for individual success.

Enculturation Learning the language first seems to help more when making social connections, instead of the other way around, but the difference is not profound.

Turbulent and Stable Contexts

One organization always dominates during a merger. The identities, routines, and practices of Luxury replaced those of Standard. Those at Luxury experienced more continuity and stability than those at Standard.

In turbulent settings, members are likely to withdraw to protect their gains, thus socially distant information is likely to be even more rare, and more valuable.

In a turbulent context, where the landscape of information is changing quickly and unpredictably, the information in the local network is less likely to be useful to the individual than socially distant information.

In a stable context, the benefits of the of adopting local information is likely to outweigh the benefits of spending time and resources to find single bits of valuable information that may be a long distance away.

This slide might be too wordy. Bring this down to a few sentences.

In my setting, I study two organizations going through a merger. One organization is dominant and controls the character of the merger, while the other organization endures most of the change.(Cowen, 2012; Fried, Tiegs, Naughton, & Ashforth, 1996)

Luxury, Inc. bought Standard Inc. and became Luxury-Standard, Inc. Both are consumer products firms in the same industry.

There were more from Standard (N = 357) and fewer from the Luxury (N = 250).

Final sample is 607 corporate professionals and 4M email English-language email messages.

June, T1

March

June, T2

Merger

Formally

Ratified

Merger Process

Email data collection

Networks are computed using 30 days of communications in June T1, and June T2.

Topics were extracted from communications across all time points.

Luxury

Standard

Standard

Luxury

Time 1

Time 2

notes

Start with a history and background of this thing.

Abstract

Worlds Colliding: Examining the social networks and linguistic patterns of a merging organization through email

During a merger the acquiring organization is often a dominant force. It overwhelms the target organization and replaces its norms, routines, and formal structures. I will present the results from an ongoing analysis of a massively rich dataset of emails, longitudinal surveys, individual performance, and ethnography that paints a detailed picture of an unfolding organizational merger. Topics are extracted from the email content and then estimated to belong to either the dominant organization, or to higher ranking members of the organization. I measure how much of this “powerful” knowledge is then integrated into conversations, and how employees change their networks to connect to powerful people. I show how job attitudes, such as job satisfaction, job insecurity, and organizational commitment, are transmitted among between individuals through email communication. And I show how diverse information fails to provide benefits when the organizational context is rapidly changing. I would encourage those attending to ask questions, and press for new potential directions for this research.

[Q-Step Talk] Worlds Colliding - Google Slides