Towards Decision Support and Goal AchievementIdentifying Ac.docx

Towards Decision Support and Goal Achievement:
Identifying Action-Outcome Relationships From Social
Media
Emre Kıcıman
Microsoft Research
[email protected]
Matthew Richardson
Microsoft Research
[email protected]
ABSTRACT
Every day, people take actions, trying to achieve their per-
sonal, high-order goals. People decide what actions to take
based on their personal experience, knowledge and gut in-
stinct. While this leads to positive outcomes for some peo-
ple, many others do not have the necessary experience, knowl-
edge and instinct to make good decisions. What if, rather
than making decisions based solely on their own personal
experience, people could take advantage of the reported ex-
periences of hundreds of millions of other people?
In this paper, we investigate the feasibility of mining the
relationship between actions and their outcomes from the
aggregated timelines of individuals posting experiential mi-
croblog reports. Our contributions include an architecture
for extracting action-outcome relationships from social me-
dia data, techniques for identifying experiential social media
messages and converting them to event timelines, and an
analysis and evaluation of action-outcome extraction in case

studies.
1. INTRODUCTION
While current structured knowledge bases (e.g., Freebase)
contain a sizeable collection of information about entities,
from celebrities and locations to concepts and common ob-
jects, there is a class of knowledge that has minimal cov-
erage: actions. Simple information about common actions,
such as the effect of eating pasta before running a marathon,
or the consequences of adopting a puppy, are missing. While
some of this information may be found within the free text of
Wikipedia articles, the lack of a structured or semi-structured
representation make it largely unavailable for computational
usage. With computing devices continuing to become more
embedded in our everyday lives, and mediating an increasing
degree of our interactions with both the digital and physical
world, knowledge bases that can enable our computing de-
vices to represent and evaluate actions and their likely out-
comes can help individuals reason about actions and their
Permission to make digital or hard copies of all or part of this
work for personal or
classroom use is granted without fee provided that copies are
not made or distributed
for profit or commercial advantage and that copies bear this
notice and the full citation
on the first page. Copyrights for components of this work
owned by others than the
author(s) must be honored. Abstracting with credit is permitted.
To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires
prior specific permission
and/or a fee. Request permissions from [email protected]
KDD’15, August 10-13, 2015, Sydney, NSW, Australia.
Copyright is held by the owner/author(s). Publication rights
licensed to ACM.

ACM 978-1-4503-3664-2/15/08 ...$15.00.
DOI: http://dx.doi.org/10.1145/2783258.2783310 .
consequences, make better decisions and be more likely to
achieve their individual goals.
In today’s digitally connected world, hundreds of millions
of people regularly and publicly report their goals, actions
and outcomes on social media, including Twitter, Facebook
and other social web sites. Such detailed records of the
events occurring in people’s lives provide an opportunity to
learn the relationships among everyday actions, their out-
comes, and higher-level goals. While there are many data
sources (including web documents, search queries, and a va-
riety of wearable sensors) that potentially capture relation-
ships between actions and outcomes, our initial focus is on
social media data for several reasons. First, status messages
naturally capture the temporal occurrences of events expe-
rienced by individuals, allowing our analysis to exploit tem-
poral relationships among actions and outcomes. Secondly,
status messages capture both the actions that people take
as well as their outcomes across a wide variety of domains.
Finally, social media messages are annotated with persistent
user identifiers that allow us to condition our results on past
actions and other relevant information.
A knowledge base of actions has many potential applica-
tions, such as direct user exploration to aid decisions; review
of recent actions and their likely future impact; and per-
sonalization of automated recommendations based on user’s
medium- and long-term goals. Research in the fields such as
social psychology, medicine and human computer interaction
has shown that information, such as action plans, task and
goal reminders, and reviews can have a significant positive
impact on goal achievement of individuals [35, 17, 23]. Scal-
ing the generation of these aids across an open-ended domain

of actions and goals, tailored appropriately across popula-
tions, and then delivering them at the right time and place
has to date been infeasible. However, with our computing
devices continuing to be integrated more tightly into our ev-
eryday lives, and mediating more of our actions (through dis-
covery, recommendation, purchase, guidance, tracking, etc.),
embedding a knowledge base that can link available and oc-
curring actions with their long-term consequences could en-
able such positive impact on individual outcomes.
This paper describes our efforts to build such a knowledge
base of actions. To realize the full value of the large-scale
longitudinal records of actions and outcomes in social media
archives, there are many potential technical challenges that
must be addressed, from interpreting and aggregating the
natural language text of social media texts, to accounting for
biases inherent in the data. While these are grand issues, we
wonder whether straightforward approaches to these techni-
547
cal challenges might yet produce reasonable and useful, if
limited, representations of actions and their outcomes from
social media. In this paper, we investigate that basic ques-
tion of current feasibility through two case studies analyzing
action and outcome relationships extracted with a general
purpose analysis methodology. Our contributions include:
• An analysis framework to extract action-outcome re-
lationships from social media data (Section 3).
• Algorithmic and implementation details for each ma-
jor component of the framework, including the identi-
fication of experiential social media messages, extrac-

tion of a timeline representation of events from raw
messages, and extraction of precedent and subsequent
actioncome relationships (Section 4).
• Two case studies applying the techniques to Twitter
data: extracting positive and negative outcomes for
decision support, and identifying precedent events for
supporting goal achievement (Section 5).
Addressing many other important and related issues, includ-
ing social media biases, algorithmic scalability, efficacy of
intervention methods, and causal reasoning, is a non-goal of
this paper. These issues are briefly discussed in Section 6.
2. BACKGROUND AND RELATED WORK
2.1 Social Media Content
With the wide-spread adoption of social networking services
over the last 10-15 years, much research has focused on un-
derstanding people’s motivations and participatory behavior
on these sites, both from a qualitative as well as quanti-
tative perspective [27, 14, 19, 21, 13, 32, 26, 36, 6, 42].
Across these studies, common findings are that individuals
are motivated to participate in social networking for a va-
riety of purposes, including communicating and keeping up
with current friends, meeting new people, managing one’s
professional reputation, and learning interesting new things.
This diversity of purpose on social networking sites has
lead to a broad a variety of content being found within social
media messages. Even in this variety, however, status mes-
sages reporting on an individual’s own experiences consti-
tute a significant percentage of content. Naaman, Boase and
Lai categorize tweets and find that such “me now” messages,
describing personal state and current experiences constitute
40% of messages [32]. Ramage, Dumais and Liebling per-

form an large-scale latent dirichlet allocation (LDA) anal-
ysis of Twitter messages at a word-level, and find that on
average, tweets are composed of 11% substance, 5% status,
16% style words, 10% social and 56% other (other includes
non-English words, many numbers, dates and times) [36].
This well-documented behavior of individuals announcing
and discussing a broad range of their current activities and
status in social media is one of the key features of social me-
dia datasets that promises to enable the work in this paper.
2.2 Mining Social Media and Search
Much research has focused on extracting and validating in-
formation and relationships about the off-line world from
social media, search queries and other digital traces of hu-
man activities. In the health domain, social media studies
have looked at the relationships between diseases, medicines,
side-effects, and symptoms [33, 31] as well as disease trans-
mission [40]. Similar studies have been conducted in urban
informatics [8], mental health [9, 16], natural disaster mon-
itoring [11, 41], and other domains. Many of these analy-
ses rely on a co-occurrence analysis: the assumption is that
items that co-occur frequently may share some true relation-
ship. For example, Sadilek et al.’s analysis of disease con-
tagion infers relationships between disease carriers and new
infections based on co-visited locations. Paul and Dredze
studied the relationship between mentioned ailments and
the geographies in which they occur. Becker et al. analyze
social media data to surface information and insights about
real-world events [3].
Studies with similar goals have been applied to search
query logs and other data sources. Richardson uses long-
term query logs to identify topical and temporal relation-
ships about the world [37]; [45] and [44] extract relationships

between drugs and possible consequences (adverse reactions)
from search queries. A closely related body of work frames
the problem of learning about the real-world from social
media, search and other data sets as a prediction problem.
Given a known (historical) signal about the world, the goal
is to predict the current or future signal from current social
media signals. This approach has been applied to prediction
of economic, financial and other signals [4, 7, 15, 2, 1].
Our goals are to extend this prior work by focusing on
extracting action-outcome information from individual-level
timelines at relatively fine granularity. More importantly,
our goal is to explore generalizable techniques that require
minimal information about specific actions, domains and
outcomes.
2.3 Actions and Plans
Recently, there have been several attempts at using crowd
sourcing techniques to create action plans to aid goal achieve-
ment. Law and Zhang use crowdsourced workers to gener-
ate simple plans related to the “high-level missions” driving
search queries, and evaluate the effect of replacing search en-
gine results for the original query with web resources related
to the various steps required by a plan [28]. They find that
organizing web resources in this way is useful for helping
users navigate the space of their problem.
Kuo, Hsu and Shih use crowdsourcing to elicit the common-
sense context that can aid in social media interpretation [25].
Mechanisms such as this, perhaps modified for scalability,
could aid our identification and interpretation of events, ac-
tions and goals in social media. Kokkalis et al. describe a
system to provide individuals with actionable and reusable
plans, to see if plans generated by others are as effective
at improving goal achievement as plans generated by one-
self [23]. They find that, indeed, system-provided plans do

have a positive effect on goal achievement.
We find the effectiveness of these techniques to improve
goal achievement to be promising. We see these techniques
for crowdsourcing action plans as largely complementary to
mining action-outcomes from social media data, and believe
that an existing knowledge base of actions could reduce the
required manual effort to scale out the generation of action
plans for a broader set of scenarios.
3. KNOWLEDGE BASE OF ACTIONS
In this section, we define the problem of extracting action-
outcome relationships. We present details about the implied
548
subproblems and discuss how this framework can be used to
formulate a variety of interesting questions.
3.1 Choice Exploration and Goal Achievement
We consider two major types of questions one might want to
ask: choice exploration, and goal achievement. For the
former, we can help by advising the user what experiences
to expect after taking a particular action (based on other
people who have taken this action). For the latter, we can
convey which actions are most likely to lead to the desired
goal (based on other people who have accomplished the same
goal). Since the social data is open-domain, these two topics
cover a broad range of questions one might have.
One way to measure online users’ desire to answer such
questions is by looking at the queries they submit to a search
engine. Many of these are decision questions beginning with
“should I/you”. The most common ones show their breadth

of topic, including finance, relationships, and health: should
I refinance my mortgage, should I date a co-worker, should
you marry your best friend, should I get a flu shot, should
I file bankruptcy, should I upgrade to windows 8. We also
see many people asking for advice between two options, as
in: should I lease or buy a car, should I file married jointly
or separately, should I eat before or after working out, and
should I call him or wait for him to call me. In both cases,
we would like to provide people with the ability to see what
experiences other people tend to have after taking one of the
actions. For example, among those people who ate before
working out vs. after working out, who was most likely to
lose weight or get a side-ache, and what other unexpected
effects might differ between the two populations?
Similarly, people show a desire for help in achieving goals.
The most common question containing the word “marathon”
is how to train for a marathon. Other common “how to”
questions include how to lose weight, how to draw, how to
get pregnant, and how to speak Spanish. As with decision
support, we could provide people with the ability to see
what actions were more commonly taken among those who
accomplished their goal than those who didn’t.
Though there may be online resources devoted to answer-
ing some of these questions, using social data has many dis-
tinct advantages. First, results are grounded in the real
experiences of users who have taken an action, potentially
leading to more reliable results than simply reading advice
from web pages. Second, a question may be too rare for
someone to have devoted writing advice about, but still have
plenty of social data to answer via data mining. For exam-
ple, someone may ask whether to move to one city vs. an-
other. Web pages may exist to answer such a question for
some city pairs, but surely not for any pair of cities that
may be asked. In contrast, we need only look at social post-

ings from people who have moved to one city vs. the other
and compare their postings to see the relative benefits of
each. Third, an answer may be contextually dependent on
the asker. To the extent that we can infer demographic
information for social media users [24], we can provide an-
swers not just in the abstract, but specifically tailored to the
asker: people similar to you (urban male, age 25-35) have
found that a low-carb diet works best for losing weight.
3.2 Problem Definition
A key advantage of applying our techniques to social data
is that it is fully open-domain. Social data contains experi-
ences about anything that users wanted to post about, and
as a result contains information on an incredibly wide range
of topics. A sampling of the experiential tweets contained
reports on love and relationships, food and alcohol, children,
sleeping, weekends, weather, school, health, and so forth. A
key goal in our problem definition and architecture is to en-
sure that our techniques match the open-domain nature of
the data set and problem domain. Thus, our knowledge base
of actions is simply an architecture for answering questions
based on a large corpus of social data.
We formalize this core problem as follows: Given a corpus
of social media messages and a query defined by two events,
E+ and E−, our goal is to identify the precedent and sub-
sequent relationships of an event E+ that distinguish the
social media timelines containing E+ from timelines com-
paring some event E−. Semantically, E+ and E− can be
thought of as identifying either positive and negative out-
comes or treatment and control classes. A class of events
E+ or E− is specified as, for example, some specific obser-
vation, or a complex matching function.
Depending on the specific query we choose, we can ask

different forms of high-level questions.
Choice Exploration: If we choose a query such that E+
selects a specific action (and E− selects an inverse action or
null action), then the results from our analysis will identify
what is likely to happen after taking the specified action.
Goal Achievement: If instead we choose a query such
that E+ selects the achievement of a specific goal (and E−
selects the non-achievement of that goal), then the prece-
dents identified by our analysis will identify what is likely
done and differentiates between people achieving the goal
and not achieving it.
While this query setup is straightforward, there are sub-
tleties in the selection of query specifiers. For example, if we
which to explore how people achieve some goal E+, we will
find different results if we compare to an E− that captures
timelines of people who attempted but failed to achieve a
goal; versus if we compare to an E−∗ that captures time-
lines of people who never even tried to achieve the goal.
The choice of E− depends on the question that one wants
to answer.
3.3 Architecture
Figure 1 shows the pipeline of data processing steps in our
analysis. We begin with a corpus of social media messages.
These messages consist of the original microblog text posted
by individuals. We expect these messages to include at least
a user identifier and a timestamp, but they may also include
other metadata, such as includes geographic location, au-
thor details (name, brief biographical description, popular-
ity statistics), as well as social network connections.
First, from this corpus of social media data, we extract
a large set of timelines of event occurrences. Each time-

line represents events occurring in a single individual’s life.
Some of these events may be actions explicitly taken by the
individual. Other events may describe outcomes that came
about because of such an action, or background events that
happened due to unrelated causes. These events may be di-
rectly extracted from individual social media messages, or
inferred from the corpus as a whole.
By avoiding an explicit categorization of events as being
actions or outcomes, we greatly simplify the task of generat-
ing timelines for individuals. Leaving this classification and
549
Individual timelinesMessages Query-aligned Timelines
C ardi g an fan ny p ac k Odd
Fu tu re, B an ksy
cre d selv ag e ch il lwa ve ret ro
sel fie s o rg an i c. YOLO
sh ab by c hi c
Th u nd erca ts , lo mo
me di ta tio n
Wi lli ams bu rg pl ai d na rwh al
cru ci fi x M arfa
u1
u2
u1

u2
E
+
E
-
Precedents
Subsequents
E
+
E
+
Figure 1: Steps of our general analysis
interpretation of actions and outcomes outside of the core
data representation and analysis mechanics simplifies our
task, at the cost of potentially requiring additional semantic
understanding at higher-levels. We believe that this is likely
to be a beneficial trade-off as adding additional semantics
when grounded within a specific application context is often
easier than building a general-purpose recognizer up-front.
In the next step, given a query, E+ and E−, we extract
and temporally align a set of timelines that match the cri-
teria E+ and a set of timelines that match the criteria E−.
Representing a query as two distinct events, E+ and E−—
as opposed to comparing a single event class against a back-
ground model of all timelines—provides significant flexibility
to ask a broader range of questions of our collected data.

Finally, from these two sets of event timelines, we extract
the precedent events and subsequent events that distinguish
the E+ and E− timeline subsets from each other.
3.4 Subproblems
There are a number of implied subproblems within the key
tasks of event timeline extraction, subselection of timelines
according to a query, and identification of precedent and
antecdent events, including:
Identification of experiential messages: When extract-
ing a timeline of events experienced by a person, the first
thing we must do is identify experiential messages which re-
port on personal experiences of the author, whether past,
current or (expected) future. Non-experiential messages in-
clude conversational texts, hearsay, pointers to news articles
and current events, among others. We describe our method
for identifying experiential tweets in Section 4.1.
Timestamping event occurrences: While many social
media messages provide in situ reports of an individual’s
experiences, it is not uncommon for authors to also report on
past experiences and anticipated future experiences. For this
reason, it is important to identify the time period referred
to in a message, and timestamp the recognized events. We
describe our approach and findings in Section 4.2.
Recognition and canonicalization of events: A key
step in the generation of a timeline of events is the extrac-
tion of events from the text of social media messages. These
events may be extracted directly from the textual represen-
tation of a message, or inferred from multiple messages. We
discuss the former in Section 4.3 and provide an example of
the latter in our second case study, in Section 5.3.
Identification of precedent and subsequent events
that distinguish the two sets of timelines from each other.
Our framework allows for various implementations, from
correlational to causal analyses. Note that even when calcu-
lated using causal analyses, such as propensity score match-

ing, it is unlikely that the strong assumptions necessary for
inferring causality would hold (i.e., assuming the observabil-
ity of all potential causal factors). Section 4.4 describes our
implementation.
Identification of positive and negative valence of events:
Of course, some outcomes of actions are good and others
bad. In social media, messages describing such outcomes
are often augmented with clear emotional words that sig-
nal the current mood of the author. Detecting these moods
or sentiments and associating them with outcomes can help
with reasoning about their significance. We use a domain-
agnostic affect extractor, described in [10], to extract the
author’s levels of joviality, sadness, fatigue, hostility, etc.
While we do not describe details here, we demonstrate its
application in Section 5.2.
4. ANALYSIS DETAILS
In this section, we present the details of our framework, its
specific application to Twitter data, and how we adapt and
apply existing algorithms to address the challenges of ex-
tracting action-outcome relationships. In addition, we high-
light key descriptive statistics of Twitter social media rele-
vant to our overall tasks, including the percentage of Twitter
messages that are experiential tweets, and the prevalence of
relative time references.
4.1 Experiential Tweets
Social media fulfills a diverse set of roles, including experien-
tial tweets that report on actions and events occurring that
individuals are experiencing first hand, but also includes
the dissemination of information about broader news and
other world events, chit chat with friends, and incitements
to action and advocacy [32, 5, 26, 19]. To extract action-
consequence relationships, we must be able to distinguish
experiential tweets from other social media content.

We tackle this as a straightforward classification task. We
label ≈ 10000 messages using crowdsourced workers, asking
them to specify whether or not a message is a “personal
experience”, defined as
A message where the author is describing or in-
dicating their own personal experience, such as
an action or situation that they are currently in,
have experienced, or are concretely planning to
take in the definite future.
We explicitly instruct workers not to mark messages as per-
sonal experiences if they describe or declare personal de-
sires or intents unless describing a concrete plan or action.
550
Personal Experiences
Just completed a 15.72 km run with @RunKeeper.
Check it out! <URL> #RunKeeper
Just to set the mood I brought some Marvin Gaye
and Chardonnay.
lacrosse is so much fun why didn’t I start earlier
lol
Oh yeah guys we got a new puppy.
@Alice Tell me about it. Knee isn’t hurting today,
but it’s also taped within an inch of its life.
Other (Personal desires and goals)
When i turn 16, i’m driving anywhere and every-
where.
Hope you enjoy England! Wish i could go :(
I wish I could cook

I’ve got real big plans and such bad thoughts
Other (news, 3rd-person, misc.)
New campaign to protect children from second
hand smoke launched... <URL>
Whoa. The kid from Cincinnati just suffered a
horrible injury. Not good.
@Bob I hear you.
@Charlie did you enjoy your night at the club?
Table 1: A sample of experiential and non-
experiential tweets.
Label Count Pct
Personal Experience 2580 26%
Other (Personal Desire/Goal) 755 7.6%
Other (news, 3rd-person, misc.) 6583 66%
Total Tweets 9873 100%
Table 2: Experiential tweet labeling results
To reinforce this, we ask workers to label the non-personal-
experience tweets as either being a personal goal or other.
Table 1 shows example messages for each class of labels.
We train a näıve Bayes classifier on these labeled mes-
sages, using maximum likelihood estimation for the NB pa-
rameters. We tokenize the messages based on whitespace,
removing all non-alphanumeric characters, but not applying
any stemming. We generate a feature t for every pair of
co-occurring tokens in a message.
As shown in Table 4.1, the great majority of tweets labeled
by our workers are found to be non-personal, other tweets.

26% of messages describe personal experiences. The pri-
mary implication for this paper is the confirmation of prior
research that a significant amount of the data in Twitter is
describing the kind of personal experience that is relevant
to our learning of actions and outcomes. To measure the
difficulty of the labeling task, we also collect two additional
labels for each tweet. The inter-annotator agreement, mea-
sured by Fleiss’ kappa, is 0.325, which is regarded as “fair
agreement”. For the remainder of the paper, we ignore the
distinction between desire/goal and other, since we care only
about whether a tweet is a personal experience or not.
4.2 Temporal expressions
Personal experiences are not always reported on social me-
dia as they occur. Often, people will post about an upcom-
1
10
100
1000
10000
100000
1000000
10000000
-180 -150 -120 -90 -60 -30 0 30 60 90 120 150 180
C
o

u
n
t
Days relative to message post date
Figure 2: Distribution of relative time mentions
ing event or experiences in anticipation (“I can’t believe my
marathon is coming up next week”), will reference a recent
event (“We got a cat yesterday”) or a long past event (“I ran
my first marathon ten years ago”). As noted by Ritter et
al., building up a true timeline of event occurrences requires
resolution of the temporal expressions accompanying such
non-concurrent personal experience reports [38].
To do this, we built a simple rule based system, similar to
TempEx [29], that can recognize and resolve basic expres-
sions of relative offsets (“yesterday”, “next weekend”), as well
as references to nearby days and dates (“Tuesday” and “Feb
10th”). Figure 2 presents the distribution of relative times
mentioned in our data sets. We see that most messages, by
default, refer to the current date, and a large number re-
fer to dates within a few days of the current date. As we
look to dates further afield, we see more references to future
events, and also spikes of references at week and month unit
distances.
4.3 Event Extraction
Once we have identified a timeline of messages referring to
the personal experiences of an individual, we wish to break
apart each message into the component representations of
the events (both actions and outcomes) that are being re-
ported. This task is analogous to the task of named entity

recognition [18], and shares many of its challenges, includ-
ing candidate identification (what words in the message refer
to an event of interest), disambiguation (when a candidate
could mean multiple things, which does it mean) and canon-
icalization (can we recognize when two candidates with dif-
ferent forms are referring to the same underlying event).
Given how little information we have about what might
constitute an action or an outcome and because our goal is
an open-domain system, we make a design decision to sim-
ply extract all phrases of the message as potential events,
without attempting to classify them as actions, outcomes,
or neither. An advantage of using phrases instead of single
words is the implicit sense disambiguation provided. For
example, while the word spaghetti often refers to an Ital-
ian noodle dish, it sometimes is used as part of the name
‘spaghetti squash’. Recognizing the phrase as the unit re-
duces the need for additional sense disambiguation.
We maintain an open-domain approach to phrase segmen-
tation and the canonicalization of phrases into events:
Phrase Segmentation: We use a statistical modeling ap-
proach to infer the hidden phrase boundaries in a text. To
efficiently find phrases, we use a phrase unigram language
model, as described in Jin et al. [20]. Briefly, each token in
a phrase unigram language model consists of one or more
551
Cluster name Elements
cat eats bit my ear, bit my nose,
bit my finger,...

woke up at 1 woke up at 3,
woke up at 4,...
sleeping on my bed sleeping on my lap, sleep-
ing on my chest
cheese balls cheese, cheese pizza
loud people people crazy, people suck
Table 3: Example of phrase canonicalization. The
most frequent element is selected as the cluster
name.
white-space separated words. By encoding multiple words
within a single unigram, the phrase language model is able
to capture long distance relationships without requiring high
Markov order statistics and concomitant large models. The
phrase unigram language model itself is trained from a large
corpus of text (in this case, from a complete archive of 16
days of tweets), using an EM process that iteratively seg-
ments a corpus into likely phrases and then retrains a new
phrase unigram language model 1.
Given a phrase unigram language model, identifying phrase
segmentations in a message is a matter of searching for the
most probable combination of component phrase-unigrams.
Below are segmentations of 2 sample messages:
It’s gorgeous outside | so I’m pretty sure | I have no
excuse not | to get this | long run in.
I got a new kitten | and he has blue eyes and | stripes
and | I need a good name | but nothing | that’s normal
Canonicalization: Generally speaking, there are many al-
ternative ways to describe or report on a personal experi-
ence when writing a social media message, leading to the

need to identify and canonicalize phrases with substantially
the same meaning. To do so, we cluster phrases based on
their distributional similarity. Specifically, for each phrase,
we build a distribution of co-occurring (single-word) tokens.
We use agglomerative hierarchical clustering to group to-
gether all phrases that are within a distance threshold d
of each other, where the distance between two phrases is
measured as the cosine similarity between their token dis-
tributions. (We use d = 0.75 in our experiments). Table 3
shows example phrase canonicalizations.
4.4 Precedent and Subsequent Events
There are multiple methods to identify the distinguishing
precedent and subsequent events when comparing timelines
containing an event E+ to those containing an event E−.
In this paper, we report our experiences with two methods:
a simple correlational analysis, and a correlational analysis
with semantic scoping. These two techniques make different
assumptions and are appropriate for different purposes.
Correlational Analysis: Our first technique looks at
simple correlations between a target event and the events
that precede or follow it. Our goal in this analysis is to find
events that are more correlated with occurring before or af-
ter E+ (but not both before and after) than occuring before
1The MSR Phrase Breaker Service is available
for demonstration and programmatic access at
http://weblm.research.microsoft.com/PhraseBreakerDemo.aspx
E
+
t<0 E
+
t>0

E
-
t<0 E
-
t>0
Figure 3: Quadrants of our two sets of timelines
or after E−. As shown graphically in Figure 3, our goal
corresponds to finding events that are more likely to occur
in one quadrant (say, E+ for t > 0) than in its immediately
neighboring quadrants (E− for t > 0 and E+ for t < 0) 2.
More formally, we begin by defining the pair-wise compar-
ison of likelihoods of an event occurrence between a target
quadrant q and a neighboring quadrant u. Let Nq(e) be
the number of occurrences of an event e in a given quad-
rant, |Nq| be the total number of events in a quadrant, and
p̂ q(e) =
Nq(e)
|Nq|
.
Our score, Sq,u(e), is the relative likelihood of an event
occurrence in q as compared to u. We calculate this as:
Sq,u(e) =
p̃q,u(e)
p̂ u(e)
(1)

where p̃q,u(e) is the Laplace-smoothed probability:
p̃q,u(e) =
Nq(e) + p̂ u(e)m
|Nq| + m
(2)
Smoothing the likelihood of p̂ q(e) toward the neighboring
quadrant has the effect of requiring greater evidence of a
difference in likelihood to appear significant. In our exper-
iments, we set m = 104. For an event to be considered
important, we require Sq,u(e) � 1 for both neighboring
quadrants. For example, when considering an event in the
quadrant E+t>0, we will calculate the score for both u = E
+
t<0
and u = E−t>0. The final reported score is the minimum of
the two.
Correlational analysis has the advantage of being straight-
forward and requiring no inputs beyond the definitions of
E− and E+. Because it is not a causal analysis, however,
we expect its results to be better suited for tasks such as pre-
dictions which do not require a causal interpretation. Fur-
thermore, correlational analysis may find relationships that
are difficult to easily explain or interpret, and thus may not
be appropriate for end-user facing applications.
Correlational analysis with semantic scoping: Our
semantic correlation is the same as the correlational analysis
above, with the added restriction that we only consider those
events that are believed to be semantically closely related to
our domain of interest. Let us define E′ to be a set of events

known to be in our domain then we will consider only ei
that co-occurs at least once with E′ in our corpus.
Semantic correlation makes an assumption that if an event
ei is related to our target events E
+ and E−, then at least
one person would have clearly mentioned ei in the recog-
nizable context of our target domain. Our expectation is
2Recall that all of timelines were aligned such that the events
E+ and E− occur at time t = 0
552
that the ranked events ei will be more robust to noise and
confounds. Furthermore, we expect that any events found
to be correlated is more likely to be easily interpretable by
humans, due to the enforced domain proximity. The cost,
however, is that we essentially extend our query model to
require a specification of the domain of interest.
While the outcomes of actions can vary based on context,
our analyses are context-independent. Extending them to
incorporate individual demographics, past actions, location,
seasonality, social and other contextual information is im-
portant future work.
5. CASE STUDIES
In this section, we present two case studies extracting var-
ious forms of action-outcome relationships from social me-
dia data. First, we demonstrate an example of subsequent
event analysis. We evaluate the quality of analysis results
and measure the quality reduction when experiential mes-

sage filtering, phrase clustering, or semantically scoped cor-
relation are removed. Secondly, we demonstrate an example
of precedent event analysis, where we measure the increase
in likelihood of goal achievement given the occurrence of a
precedent event. We ground our first case study in identi-
fying the consequences of pet adoption, and the second in
achieving the goal of running a marathon.
5.1 Data
While we are designing our architecture to process a full,
unfiltered archive of social media data, our first small-scale
implementation demonstrates and evaluate the feasibility of
the techniques through archive subsets. For our first case
study, we create an archive subset of the timelines of English-
language Twitter users who mentioned getting a dog, cat,
puppy or kitten during the period of August 1-15, 2013. This
procedure identified 6232 Twitter users who had mentioned
adopting a pet. We then collected the entire Twitter time-
lines for these users from the period of August 1-September
15, 2013, encompassing a total of 4.6M tweets.
For our second case study, we create an archive subset of
the timelines of English-language Twitter users during the
period of March 1-31, 2014 who mentioned running or train-
ing for a marathon. We then collected 2 month timelines for
each of these users, from February 1-March 31, 2014. In
total, this data set consists of 40,591 users and 21M tweets,
with retweets removed. In addition, we used a random sam-
ple of 260M tweets to provide background statistics.
5.2 Subsequent Events and Choice Exploration
In our first case study, we wish to test the basic compo-
nents of our analysis pipeline to better understand the qual-
ity implications of each analysis stage: Namely, how impor-
tant are the subtasks of identifying experiential tweets and
canonicalizing phrases with similar meaning? How much

perceived benefit is there to restricting precedent and sub-
sequent events to those with a semantic correlation to the
target domain?
To do this, we ground our study in the specific task of au-
tomatically generating a “pros and cons” list to aid people
deciding whether or not to adopt a kitten or cat. A “pros
and cons” list is a simple decision making aid for clearly eval-
uating the benefits (pros) and disadvantages (cons) of taking
some action (in this case, adopting a pet). Writing a pros
and cons list is often recommended to individuals facing a
significant decision to ensure that all potential consequences
are considered and evaluated.
In this case study, we apply our analysis techniques to au-
tomatically extract the subsequent events that follow decla-
rations of pet adoption in social media timelines. More for-
mally, our query consists of an E+ that consists of a boolean
OR search for the following phrases: {“got a pet”, “got a
new pet”} where pet is either “cat” or “kitten”. The set of
E+ timelines consist of all messages written by users who
wrote a tweet matching E+. In this query, our E− is the
null event, capturing all timelines—essentially a background
model of user timelines. The semantic scoping of our cor-
relational analysis consists of limiting our analysis to those
events that co-occurred at least once with the main topic
words “cat” or “kitten”.
Table 4 shows the top entries of the pros/cons list gener-
ated by our system. We split outcome events into pros and
cons by looking at the aggregate affect valence of all men-
tions of these outcomes across all of our E+ set of timelines.
Events with a valence > 0.6 are added to the pros column,
and < 0.4 are added to the cons column. Events are ranked
by their relative likelihood of occurrence, as compared to

their occurrence in E− timelines.
To evaluate the importance of each of the analysis stages,
we regenerate our pros/cons list while disabling aspects of
our pipeline, one at a time. First, we disable experiential
tweet classification, and keep all tweets for analysis. Second,
we disable phrase clustering and treat all distinct phrases
independently. Third, we switch to correlation analysis, in-
stead of semantic correlation.
To evaluate the quality impact of disabling each of these
aspects of the system, we post the items of each of the 4 gen-
erated pros/cons lists for evaluation by crowdsourced work-
ers. For each item, we display to workers the event title,
and 3 messages mentioning the event (Table 4 only shows
1 message due to space limitations). We then ask work-
ers to label, on a scale of 0 to 4 whether or not each item
and messages are useful and relevant to deciding whether or
not to adopt a cat. We use these labels to calculate a dis-
counted cumulative gain (DCG) score for the entire set of
results: DCGp = r1 +
∑p
i=2
ri/log(i), where ri is the label
at rank position i, and DCGp is the accumulated score at
rank position p.
The results provide interesting insights into the role that
each stage of the pipeline plays. Our complete pipeline
achieves the highest DCG score, of 20.7 summed across
both the pros and cons list. Disabled-Experiential filter-
ing is the 2nd best variation with a DCG score of 19.5. The
results are very similar to our complete pipeline, though
there are ranking differences and several results related to

cat videos. Our pipeline without clustering is the third best
variation, achieving a DCG of 16.0 after discounting du-
plicate items. Significant semantic duplication of results is
the biggest drawback to not clustering phrases. Finally, our
fourth variation of regular correlation achieves the worst per-
formance, with a DCG of just 0.38. Most of the items found
by this variation are not clearly related to cats or kittens
at all. While this may be due to the relatively small data
sizes, it is a striking result nonetheless, and emphasizes the
importance of perceived topical relevance and the important
need for an end-user to understand why correlations exist in
results.
553
Pros Cons
Event Example message PosNeg RL Event Example messages
PosNeg RL
1 cat
named
We just got a cat and
named it Versace
0.70 9.3x 1 ran up-
stairs
But I ran upstairs and fell
and now my head hurts
0.20 9.5x
2 I’ve got

a cat
I’ve got a kitten asleep on
my lap, and my heart has
softened.
0.67 7.3x 2 damn
kitten
Had practically no sleep
because the damn kitten
kept going nuts and runniy
round my room
0.22 6.2x
3 Love my
new kit-
ten
I love my new kitten 0.88 7.2x 3 cat is lit-
erally
My cat is literally the devil 0.31 5.9x
4 named
my cat
I named my cat tapenga if
that’s how you spell it
0.63 6.1x 4 cat just
ate
My cat just ate something
off the floor I don’t know

what it was gross
0.24 5.8x
5 love the
fact that
Love the fact that our kit-
ten Marley has a massive
“M” on his forehead
0.64 5.3x 5 cat just
jumped
My cat just jumped on me
and scratched me
0.21 5.7x
Table 4: Top positive and negative events observed to occur
after new cat ownership. PosNeg is the mood
valence (1=good,0=bad). RL is the relative likelihood of the
event occurring, compared to timelines where
a pet adoption did not occur within our observation period.
5.3 Precedent Events and Goal Achievement
In a second analysis, we consider the effect of selected prece-
dent actions on a specific, declared goal. In particular, we
choose to look at the relative importance of various marathon
training actions on the eventual outcome of a marathon race.
5.3.1 Marathon Event Identification
In the first case study we exclusively analyzed events explic-
itly mentioned in social media messages in an open-domain
way, only requiring the user to specify four phrases and
two keywords. Our second case study demonstrates our

pipeline’s ability to incorporate higher-level events, namely,
marathon participation inferred from information mentioned
across multiple social media messages. We infer the date of
a marathon for individuals who have been tweeting about
their training, but do not explicitly tweet about their race
on the day of their run. Secondly, we report on experiments
learning correlations between marathon training actions and
declarations of personal record achievement.
We use official marathon result data from www.marathon-
guide.com to label a small set of 558 Twitter user timelines
with the specific dates on which they ran a marathon by
matching on the person’s name and mentioned race. From
these labeled timelines, we train a classifier to detect marathon
dates. The features for the classifier included tokens used in
tweets during a 3-day sliding window before and after the
official marathon date, and tokens used in tweets that used
temporal expressions to reference a date within 3-days of
the official marathon date. Using these features, we built a
hierarchical classifier by first estimating the likelihood that
any given day was a “immediately-before-marathon” day or
an “immediately-after marathon day”. Then, we learned a
logistic regression classifier over these estimates to find the
most likely actual marathon date. Our final classifier is able
to identify the true marathon date for 83% of a held out set
of 42 test users within an average of 1.3 days of the actual
day. The remaining 17% are not assigned to any marathon
day. We applied this classifier to our entire data set and
identified 1436 individuals with identifiable marathon dates
during the month of March 2014.
Once we have inferred a marathon date for a user, we in-
sert an artificial <inferred marathon event> symbol into
the user’s timeline. Without this additional inference step,
we could certainly rely on explicitly mentioned marathon
phrases, such as “ran a marathon today”. However, implicit

event identification enables us to further recognize individ-
uals who have, for example, mentioned their excitement be-
fore a marathon and their soreness and exhaustion after-
wards.
5.3.2 Measures of Marathon Success
While there are certainly several ways that individuals might
determine the success of their own marathon, we use a sim-
ple definition here: whether the individual declared that
they achieved a personal record (PR) after running the
marathon.
Our query E+ is a boolean AND search for the phrases “PR”
and <inferred marathon event>, where the latter is the
event identifier output by our marathon date inference de-
scribed above. E− is a boolean AND search for <inferred
marathon> and NOT “PR”. Against this, we measure the
correlation between a person tweeting about taking a spe-
cific training action (whether they chose to “taper”, trained
with “long runs”, ate carbs before the race) and reporting
that they achieved a personal record. Table 5 shows the
results. Overall, we found that reporting the action of go-
ing for long runs and tapering (reducing exercise before the
marathon) were most correlated with later reporting a per-
sonal record. Reporting eating carbohydrates (carbs) before
the marathon had a minor effect as well.
Figure 4 shows the temporal dynamics of these precedent
actions. Such a visualization could be useful for understand-
ing when people take actions. For example, we see that peo-
ple eat carbs the day or night before their race; go on long
runs weekly for many weeks before the race; and taper their
exercise 7-10 days before their race.
6. DISCUSSION
There are, of course, several challenges that our presentation
above has so far elided. For example, relying on experiential

social media data to learn outcomes can introduce bias due
554
0.0%
0.1%
0.2%
0.3%
0.4%
0.5%
0.6%
0.7%
-35 -28 -21 -14 -7 0
Carb(s)
0.0%
0.1%
0.2%
0.3%
0.4%
0.5%
0.6%
0.7%
-35 -28 -21 -14 -7 0
Long Run
0.0%
0.1%
0.2%
0.3%

0.4%
0.5%
0.6%
0.7%
-35 -28 -21 -14 -7 0
Taper(-ing)
Days Before Marathon
Figure 4: Temporal dynamics of carbs, long run
and taper mentions. The y-axis is the percentage
of tweets on a given day containing the phrase(s).
Action Increase in PR likelihood
Carbs +9%
Long run +27%
Taper +45%
Table 5: Actions reported by marathon runners on
Twitter and the relative increase in reporting a per-
sonal record.
to population and self-reporting biases [30, 12, 22]. Signifi-
cantly, the absence of an event in our social media timeline
does not necessarily mean that an event did not occur. Un-
derstanding the implications of previous empirical studies
for our inference processes, as well as the implications for
how such biases circumscribe our ability to learn parts of the
semantic space of relationships is important future work.
In our pipeline, we currently ignore much of the semantics
of the language people use, in favor of a simplistic approach
of treating all phrases in experiential tweets as candidate

events in a person’s timeline. Considering additional seman-
tics and even interpreting people’s own statements of causal
inference, is a potentially rich area for future exploration.
An important challenge is that a true action-outcome model
is essentially a model of causal relationships. There is a rich
literature on the inference of causal relationships from purely
observational data [43, 34] though there is debate about the
reliability of causal inference in the absence of randomized,
active intervention [39]. Luckily, at least for some initial
applications of these models, inference of the true causal
relationships seems likely unnecessary and simpler analyses
such as temporal prediction and propensity scored relation-
ships may be sufficient for the extracted results to be useful.
An area left largely unexplored in this paper is the ques-
tion of how information about actions and their outcomes
can best be used to aid people, and the implications of
these application patterns for the action-outcome extraction
pipeline. For example, many decisions involve comparing
multiple choices, rather than the two-sided choice implied
by the query E+, E− in our pipeline. Our pipeline will have
to be adapted to such scenarios—perhaps through all-pairs
comparisons, or multiple comparisons to a single base case.
Perhaps a more immediate consideration is whether or
not the results of a particular algorithm are appropriate for
a particular application or user interaction paradigm. We
saw in our first case study that regular correlational anal-
ysis, when not scoped to a semantic domain, generated re-
sults that were not interpretable and marked as irrelevant
by our labelers. It is quite possible that such correlations
would have worked well if an application called for predictive
power. But in the context of an end-user interface, the hu-
man interpretability of results is paramount. Better under-

standing of how to ensure results are interpretable, through
correct presentation, supporting information and scoping as
necessary, is an important area for further study.
Closely related to this issue is that of actionability. If we
are to recommend actions, as we might be tempted to do
based on the precedent analysis in our second case study,
we must ensure that the actions we are recommending are
feasible. For example, the event most predictive of a suc-
cessful marathon outcome might be the simple declaration
that the author “loves running!”. However, recommending
to a user that they should “love running” to ensure success,
while perhaps insightful, is not necessarily actionable.
7. CONCLUSIONS
As computing devices continue to become more embedded in
our everyday lives, they are mediating an increasing number
of our interactions with the world around us. From helping
people search for the best product to buy, to recommend-
ing a restaurant we are likely to enjoy, computing services
enable users to evaluate options and take action with “one
click”. While such services model many facets of the options
they present, they do not model the higher-level implications
and trade-offs inherent in deciding to take one action instead
of another. For example, a restaurant recommender service
will not know that suggesting a carb-heavy Italian restau-
rant the evening before a person is going to run a marathon
might improve their race outcomes. Today, people reason
about these trade-offs based on their own past experiences
and learnings, combined with their own “gut instinct”. Peo-
ple with a relevant knowledge may do well; but many others
do not. By aggregating the combined experiences of hun-
dreds of millions of people into a knowledge base of ac-tions
and their consequences, we believe that our computing de-
vices may provide significant assistance to augment our own
decision-making abilities.

In this paper, we focused on the question of feasibility:
Can relatively straightforward techniques identify action-
outcome relationships from social media data? As demon-
strated in our initial results, even a relatively small scale of
social media data — weeks as opposed to the years of data
available — allows us to discover rich action-outcome rela-
tionships. As future work, we are continuing to develop more
sophisticated techniques, as well as evaluate with broader
workloads and applications.
8. REFERENCES
[1] H. Achrekar, A. Gandhe, R. Lazarus, S.-H. Yu, and B. Liu.
Predicting flu trends using twitter data. In Intl Workshop
on Cyber-Physical Networking Systems (CPNS). IEEE,
2011.
555
[2] S. Asur and B. A. Huberman. Predicting the future with
social media. In Web Intelligence and Intelligent Agent
Technology (WI-IAT). IEEE, 2010.
[3] H. Becker, M. Naaman, and L. Gravano. Beyond trending
topics: Real-world event identification on twitter. ICWSM,
11, 2011.
[4] J. Bollen, H. Mao, and X. Zeng. Twitter mood predicts the
stock market. Journal of Computational Science, 2(1):1–8,
2011.
[5] D. Boyd, S. Golder, and G. Lotan. Tweet, tweet, retweet:
Conversational aspects of retweeting on twitter. In HICSS.

IEEE, 2010.
[6] G. M. Chen. Tweet this: A uses and gratifications
perspective on how active twitter use gratifies a need to
connect with others. Computers in Human Behavior,
27(2):755–762, 2011.
[7] H. Choi and H. Varian. Predicting the present with google
trends. Economic Record, 88(s1):2–9, 2012.
[8] J. Cranshaw, R. Schwartz, J. I. Hong, and N. M. Sadeh.
The livehoods project: Utilizing social media to understand
the dynamics of a city. In ICWSM, 2012.
[9] M. De Choudhury, S. Counts, and E. Horvitz. Predicting
postpartum changes in emotion and behavior via social
media. In CHI. ACM, 2013.
[10] M. De Choudhury, M. Gamon, and S. Counts. Happy,
nervous or surprised? classification of human affective
states in social media. In ICWSM, 2012.
[11] B. De Longueville, R. S. Smith, and G. Luraschi. Omg,
from here, i can see the flames!: a use case of mining
location based social networks to acquire spatio-temporal
data on forest fires. In Intl. Workshop on Location Based
Social Networks. ACM, 2009.
[12] F. Diaz, M. Gamon, J. Hofman, E. Kiciman, and
D. Rothschild. Online and social media data as a flawed
continuous panel survey. Working Paper
http://research.microsoft.com/flawedsurvey.
[13] J. DiMicco, D. R. Millen, W. Geyer, C. Dugan,
B. Brownholtz, and M. Muller. Motivations for social
networking at work. In CSCW. ACM, 2008.

[14] N. B. Ellison et al. Social network sites: Definition,
history,
and scholarship. Journal of Computer-Mediated
Communication, 13(1):210–230, 2007.
[15] S. Goel, J. M. Hofman, S. Lahaie, D. M. Pennock, and
D. J. Watts. What can search predict. In WWW, 2010.
[16] S. A. Golder and M. W. Macy. Diurnal and seasonal mood
vary with work, sleep, and daylength across diverse
cultures. Science, 333(6051):1878–1881, 2011.
[17] P. M. Gollwitzer and P. Sheeran. Implementation intentions
and goal achievement: A meta-analysis of effects and
processes. Advances in experimental social psychology,
38:69–119, 2006.
[18] S. Guo, M.-W. Chang, and E. Kiciman. To link or not to
link? a study on end-to-end tweet entity linking. In
HLT-NAACL, 2013.
[19] A. Java, X. Song, T. Finin, and B. Tseng. Why we twitter:
understanding microblogging usage and communities. In
Workshop on Web mining and social network analysis.
ACM, 2007.
[20] Y. Jin, E. Kıcıman, K. Wang, and R. Loynd. Entity linking
at the tail: sparse signals, unknown entities, and phrase
models. In WSDM. ACM, 2014.
[21] A. N. Joinson. Looking at, looking up or keeping up with
people?: motives and use of facebook. In CHI. ACM, 2008.
[22] E. Kıcıman. Omg, i have to tweet that! a study of factors
that influence tweet rates. In ICWSM, 2012.

[23] N. Kokkalis, T. Köhn, J. Huebner, M. Lee, F. Schulze, and
S. R. Klemmer. Taskgenies: Automatically providing action
plans helps people complete tasks. ACM Transactions on
Computer-Human Interaction (TOCHI), 20(5):27, 2013.
[24] M. Kosinski, D. Stillwell, and T. Graepel. Private traits
and attributes are predictable from digital records of
human behavior. PNAS, 110(15):5802–5805, 2013.
[25] Y.-L. Kuo, J. Hsu, and F. Shih. Contextual commonsense
knowledge acquisition from social content by
crowd-sourcing explanations. In Proceedings of the Fourth
AAAI Workshop on Human Computation, pages 18–24,
2012.
[26] H. Kwak, C. Lee, H. Park, and S. Moon. What is twitter, a
social network or a news media? In WWW. ACM, 2010.
[27] C. Lampe, N. Ellison, and C. Steinfield. A face (book) in
the crowd: Social searching vs. social browsing. In CSCW.
ACM, 2006.
[28] E. Law and H. Zhang. Towards large-scale collaborative
planning: Answering high-level search queries using human
computation. In AAAI, 2011.
[29] I. Mani and G. Wilson. Robust temporal processing of
news. In ACL. Association for Computational Linguistics,
2000.
[30] A. Mislove, S. Lehmann, Y.-Y. Ahn, J.-P. Onnela, and
J. N. Rosenquist. Understanding the demographics of
twitter users. ICWSM, 11:5th, 2011.

[31] M. Mysĺın, S.-H. Zhu, W. Chapman, and M. Conway.
Using
twitter to examine smoking behavior and perceptions of
emerging tobacco products. Journal of medical Internet
research, 15(8), 2013.
[32] M. Naaman, J. Boase, and C.-H. Lai. Is it really about
me?: message content in social awareness streams. In
CSCW. ACM, 2010.
[33] M. J. Paul and M. Dredze. You are what you tweet:
Analyzing twitter for public health. In ICWSM, 2011.
[34] J. Pearl. Causality: models, reasoning and inference,
volume 29. Cambridge Univ Press, 2000.
[35] A. Prestwich, M. Perugini, and R. Hurling. Can
implementation intentions and text messages promote brisk
walking? a randomized trial. Health Psychology, 29(1):40,
2010.
[36] D. Ramage, S. T. Dumais, and D. J. Liebling.
Characterizing microblogs with topic models. In ICWSM,
2010.
[37] M. Richardson. Learning about the world through
long-term query logs. ACM Transactions on the Web
(TWEB), 2(4):21, 2008.
[38] A. Ritter, Mausam, O. Etzioni, and S. Clark. Open domain
event extraction from twitter. In Proceedings of the 18th
ACM SIGKDD international conference on Knowledge
discovery and data mining, pages 1104–1112. ACM, 2012.
[39] J. M. Robins and L. Wasserman. On the impossibility of
inferring causation from association without background

knowledge. Computation, causation, and discovery, pages
305–321, 1999.
[40] A. Sadilek, H. A. Kautz, and V. Silenzio. Predicting
disease
transmission from geo-tagged micro-blog data. In AAAI,
2012.
[41] T. Sakaki, M. Okazaki, and Y. Matsuo. Earthquake shakes
twitter users: real-time event detection by social sensors. In
Proceedings of the 19th international conference on World
wide web, pages 851–860. ACM, 2010.
[42] T. Spiliotopoulos and I. Oakley. Understanding motivations
for facebook use: Usage metrics, network structure, and
privacy. In CHI. ACM, 2013.
[43] P. Spirtes and C. Glymour. An algorithm for fast recovery
of sparse causal graphs. Social Science Computer Review,
9(1):62–72, 1991.
[44] R. W. White, N. P. Tatonetti, N. H. Shah, R. B. Altman,
and E. Horvitz. Web-scale pharmacovigilance: listening to
signals from the crowd. Journal of the American Medical
Informatics Association, 2013.
[45] E. Yom-Tov and E. Gabrilovich. Postmarket drug
surveillance without trial costs: discovery of adverse drug
reactions through large-scale analysis of web search queries.
Journal of medical Internet research, 15(6), 2013.
556

International Journal of Physical Distribution & Logistics
Management
Toward creating competitive advantage with logistics
information technology
Benjamin T. Hazen, Terry Anthony Byrd,
Article information:
To cite this document:
Benjamin T. Hazen, Terry Anthony Byrd, (2012) "Toward
creating competitive advantage with logistics
information technology", International Journal of Physical
Distribution & Logistics Management, Vol. 42
Issue: 1, pp.8-35, https://doi.org/10.1108/09600031211202454
Permanent link to this document:
https://doi.org/10.1108/09600031211202454
Downloaded on: 06 June 2017, At: 15:25 (PT)
References: this document contains references to 122 other
documents.
To copy this document: [email protected]
The fulltext of this document has been downloaded 4983 times
since 2012*
Users who downloaded this article also downloaded:
(2006),"The impact of information technology on the
competitive advantage of logistics firms
in China", Industrial Management & Data Systems, Vol.
106 Iss 9 pp. 1249-1271 http://
dx.doi.org/10.1108/02635570610712564
(2009),"Role of logistics in enhancing competitive advantage: A
value chain framework for global
supply chains", International Journal of Physical Distribution
& Logistics Management,
Vol. 39 Iss 3 pp. 202-226 <a
href="https://doi.org/10.1108/09600030910951700">https://

doi.org/10.1108/09600030910951700</a>
Access to this document was granted through an Emerald
subscription provided by emerald-srm:485088 []
For Authors
If you would like to write for this, or any other Emerald
publication, then please use our Emerald for
Authors service information about how to choose which
publication to write for and submission guidelines
are available for all. Please visit
www.emeraldinsight.com/authors for more information.
About Emerald www.emeraldinsight.com
Emerald is a global publisher linking research and practice to
the benefit of society. The company
manages a portfolio of more than 290 journals and over 2,350
books and book series volumes, as well as
providing an extensive range of online products and additional
customer resources and services.
Emerald is both COUNTER 4 and TRANSFER compliant. The
organization is a partner of the Committee
on Publication Ethics (COPE) and also works with Portico and
the LOCKSS initiative for digital archive
preservation.
*Related content and download information correct at time of
download.
D
ow
nl
oa

de
d
by
U
ni
ve
rs
ity
L
ib
ra
ry
A
t 1
5:
25
0
6
Ju
ne
2
01
7
(P

T
)
https://doi.org/10.1108/09600031211202454
https://doi.org/10.1108/09600031211202454
Toward creating competitive
advantage with logistics
information technology
Benjamin T. Hazen and Terry Anthony Byrd
Aviation and Supply Chain Management,
Auburn University, Auburn, Alabama, USA
Abstract
Purpose – Successfully implementing and exploiting the right
information technologies is critical to
maintaining competitiveness in today’s supply chain. However,
simply adopting off-the-shelf
technologies may not necessarily induce this competitiveness
unless the organization combines these
technologies with additional complimentary resources. This
study draws on the logistics innovation
literature, resource-advantage theory, and the resource-based
view of the firm with the purpose of
investigating performance outcomes of logistics information
technology (LIT) adoption and the
proposed moderating effect of a complimentary resource. The
paper posits that combining LIT with
positive buyer-supplier relationships may set the stage for
organizations to achieve competitive
advantage.
Design/methodology/approach – A meta-analysis of 48 studies

that report outcomes of EDI or
RFID adoption was performed. Regression was used to
investigate the moderating effect of the
buyer-supplier relationship on the relationship between LIT
adoption and performance outcomes.
Findings – The findings suggest that adoption of LIT promotes
enhanced levels of effectiveness,
efficiency, and resiliency for the adopting firm and that the
quality of the buyer-supplier relationship
moderates the degree of efficiency and resiliency realized via
adoption.
Research limitations/implications – The results of this study
suggest that adoption of a logistics
innovation by itself may not necessarily produce a sustained
competitive advantage. Instead, when
combined with complimentary firm resources, the innovation
may yield a sustained competitive
advantage for the adopting firm.
Originality/value – Logistics innovation needs greater
theoretical development in the literature.
This research extends a foundational logistics innovation model
by incorporating relevant theory to
propose and test an additional dimension of the model.
Keywords Information technology, Innovation, Resource-based
view, Radio frequency identification,
Electronic data interchange, Meta-analysis, Competitive
advantage
Paper type Research paper
1. Introduction
Information technology (IT) has emerged as one of the most

popular categories of
technological innovation being implemented in the supply chain
(Russell and Hoag,
2004). Indeed, IT is purported to be one of the most
managerially-relevant research
topics in extant supply chain management (SCM) literature
(Thomas et al., 2011).
The current issue and full text archive of this journal is
available at
www.emeraldinsight.com/0960-0035.htm
The authors would like to thank Robert Overstreet and Fred
Weigel for their assistance
throughout this research effort. An earlier version of this paper
was presented at the Pacific Asia
Conference for Information Systems in Brisbane, Australia. The
authors would like to thank the
track chairs, anonymous reviewers, and session attendees for
their valuable feedback, which
helped to strengthen this paper.
IJPDLM
42,1
8
International Journal of Physical
Distribution & Logistics Management
Vol. 42 No. 1, 2012
pp. 8-35
q Emerald Group Publishing Limited
0960-0035
DOI 10.1108/09600031211202454

D
ow
nl
oa
de
d
by
U
ni
ve
rs
ity
L
ib
ra
ry
A
t 1
5:
25
0
6
Ju
ne

2
01
7
(P
T
)
Although some firms have reported positive results from
adoption of IT,
implementation can be risky and expensive, especially if the
ramifications and
outcomes of such innovations are not fully understood by the
adopting firm (Heinrich
and Simchi-Levi, 2005). Considering the Council of Supply
Chain Management’s
definition of logistics and Rogers’ (2003) definition of
innovation, we define a logistics
information technology (LIT) innovation as an IT application
that is perceived as new
to the organization of adoption that is used for planning,
implementing, and/or
controlling procedures for the transportation and storage of
goods and services from
the point of origin to the point of consumption. Organizations
looking to adopt LIT are
often interested in understanding how adopting such
technologies will aid in achieving
positive operational and strategic benefits. However,
inconsistent findings in the
literature suggest that additional phenomena may moderate the
relationship between
LIT adoption and positive performance outcomes (Narayanan et

al., 2009).
The resource-based view (RBV) of the firm suggests that capital
resources may be
utilized to create competitive advantage (Barney, 1991).
Although off-the-shelf IT
usually does not directly induce competitive advantage, these
technologies have been
shown to provide capabilities that may lead to enhanced
measures of operational
performance (Kros et al., 2011; Wade and Hulland, 2004). This
operational perspective is
based on the argument that the first-order effects of IT
innovation adoption occur at the
functional/operational level via enhancing various aspects of
efficiency, effectiveness,
and resiliency (Barua et al., 1995; Grant, 1991). However, when
combined with additional
organizational resources, adoption of off-the-shelf information
technologies may
provide the foundation for a firm to realize sustained
competitive advantage (Mata et al.,
1995; Nevo and Wade, 2010; Ray et al., 2005). Thus, the RBV
perspective provides an
adequate context in which to examine the value of LIT
adoption.
As with any innovation, firms generally adopt LIT for the
purpose of realizing
improved measures of performance. However, the logistics
arena may present a unique
set of challenges because of the inherent inter- and intra-
organizational
interdependencies required for the effective transportation and
storage of goods and
services. Thus, adoption of LIT may not automatically translate

into realized
improvements in performance for the adopting firm. For
example, a study of electronic
data interchange (EDI) use in large German and US firms
revealed a variety of
conflicting findings regarding the benefits of adoption (Reekers,
1994). Although EDI
was demonstrated to improve trading partner communication,
data accuracy, and
customer service, other anticipated benefits such as reductions
in inventory and
reductions in paperwork were not demonstrated. One
explanation for these
inconsistencies may be found in lack of inter-organizational
integration and/or forced
adoption by firms with more powerful market position (Reekers,
1994). This assertion
is supported by the work of Riggins and Mukhopadhyay (1994),
whose research
suggests that firms that initiate inter-organizational systems
should take into account
the costs and benefits of the system to their trading partners if
both firms are to reap
maximum benefits of system implementation. These findings
may be explained by
social exchange theory, which views the exchange relationship
between specific actors
as being contingent upon rewarding reactions from others (Blau,
1964). When one firm is
coerced into adopting a collaborative technology that it believes
will only benefit the
other organization, then it may not be motivated toward
successful implementation and
usage. Conversely, firms who have cultivated a positive buyer-
supplier relationship may

Logistics
information
technology
9
D
ow
nl
oa
de
d
by
U
ni
ve
rs
ity
L
ib
ra
ry
A
t 1
5:
25

0
6
Ju
ne
2
01
7
(P
T
)
view the adoption of a given LIT innovation as mutually
beneficial and thus put forth the
effort and resources that are necessary to reap positive rewards.
The first purpose of this study is to integrate existing (and often
contradictory)
research to draw conclusions regarding the nature of the
relationship between LIT
innovation adoption and performance outcomes. Accordingly,
this study asks:
RQ1. Do LIT innovations induce positive performance outcomes
for the adopting
firm?
The second purpose of this study is to investigate whether the
presence of a
complimentary firm resource, specifically a positive trading
relationship, may enhance the

performance outcomes realized by LIT adoption. As such, our
second research question is:
RQ2. How does the buyer-supplier relationship affect the
relationship between LIT
adoption and performance outcomes?
To further develop these questions and adequately articulate the
outcome of our
investigation, the remainder of this manuscript is organized as
follows. First, we review
relevant background literature and develop hypotheses. The
review begins with a
discussion of theories pertinent to logistics innovation
diffusion, to include diffusion of
innovation theory, resource-advantage (R-A) theory, and RBV.
Next, the artifacts used to
characterize LIT, namely EDI and radio frequency identification
(RFID) are introduced.
We then briefly discuss Section 2.3 as cited in the literature,
which leads to development
of our first set of hypotheses. Our conversation then turns to the
proposed moderating
role of buyer-supplier relationships, which leads to our second
set of hypotheses.
Because the purpose of this study is to not only integrate results
of existing studies
(which are often conflicting) to draw meaningful conclusions,
but also to test for
moderation, a meta-analysis method was deemed to be the most
appropriate method to
employ (Glass, 1976; Hunter and Schmidt, 2004). We discuss
the meta-analytic methods
used in this study to illustrate how extant empirical literature is
utilized for analysis. The
results of the research are then presented. Finally, we discuss

the findings, to include
practical and theoretical implications, and end with limitations
and recommendations
for future research.
2. Background literature and hypotheses
2.1 Logistics innovation
Rogers (2003) offers a generalized model of the innovation
diffusion process, which has
been used extensively as the basis of innovation research in the
SCM field. Skipper et al.
(2009) examine the relationship between Rogers’ antecedents of
innovation adoption
(relative advantage, compatibility, ease of use, trialability, and
observability) and a
firm’s adoption of supply chain contingency planning processes,
extending Rogers’
work by proposing two additional antecedents (top management
support and
centralization) from extant management information systems
(MIS) literature (Moore
and Benbasat, 1991; Tornatzky and Klein, 1982). In addition,
the IT implementation
model (Kwon and Zmud, 1987; Zmud and Apple, 1992) has been
used by a variety of
authors to investigate IT diffusion within supply chain settings
(Cooper and Zmud,
1990; Premkumar et al., 1994). As demonstrated above, many
innovation diffusion
studies in the SCM context have focused on IT artifacts (Chen
J.V. et al., 2009;
Germain et al., 1994; Patterson et al., 2003, 2004; Williams,
1994).
IJPDLM
42,1

10
D
ow
nl
oa
de
d
by
U
ni
ve
rs
ity
L
ib
ra
ry
A
t 1
5:
25
0
6

Ju
ne
2
01
7
(P
T
)
Although research in the MIS and SCM fields has expanded
upon diffusion of
innovation theory to develop more discipline-specific
conceptualizations of innovation
diffusion, the literature was previously scarce in offering a
unified model of logistics
innovation. Grawe’s (2009) recent review of logistics
innovation research suggests such
a model of logistics innovation and provides a basis for further
research. As Grawe’s
(2009) review indicates, logistics innovation research has
investigated a wide variety of
antecedents and outcomes. This model is shown in Figure 1.
Grawe (2009) proposes that diffusion of logistics innovation is
positively related to a
firm’s competitive advantage. This proposition is rooted in R-A
theory and based on a
critical survey of logistics innovation literature. As described
by Hunt and Morgan
(1996), the R-A theory of competition posits that organizations
seek competitive

advantage in the marketplace via obtaining a comparative
advantage in resources,
which then leads to superior financial performance. However,
this proposition and
accompanying model do not clearly address how a firm may
create competitive
advantage from the adoption of logistics innovation. As
suggested by Barney (1991) and
the RBV, a firm may only realize competitive advantage when it
implements a value
creating strategy that is not being implemented by current or
potential competitors.
Although adoption of a homogeneous and perfectly mobile
resource (e.g. off-the-shelf
logistics innovations such as EDI, RFID, containerization, etc.)
may induce a short-term
competitive advantage, such adoption likely will not foster
sustained competitive
advantage unless paired with heterogeneous firm resources or
characteristics.
This current research seeks to extend current SCM innovation
diffusion literature,
and specifically the model presented by Grawe (2009), by
investigating one
characteristic that may aid in fostering competitive advantage
for a firm via
the adoption of logistics innovation. By examining the
moderating effect of the
buyer-supplier relationship on the relationship between
adoption of LIT and
performance outcomes, this study proposes the introduction of a
key construct that
may strengthen the existing logistics innovation model. In doing
so, we propose an
additional dimension to Grawe’s (2009) model that may bridge

the gap between logistics
innovation adoption and competitive advantage. Considering
both R-A theory and RBV,
we suggest that complementary firm resources, when combined
with logistics
innovation adoption, may allow a firm to realize competitive
advantage over other firms
who adopt the same logistics innovation yet do not possess
additional complementary
Figure 1.
Logistics innovation
antecedents and outcomes
Source: Grawe (2009, p. 364)
Environmental factors
Organization of labor (–)
Competition
Capital scarcity
Organizational factors
Knowledge
Technology
Relationship network factors
Financial resources
Management resources
Logistics innovation
Logistics innovation diffusion
Competitive advantage
Logistics
information

technology
11
D
ow
nl
oa
de
d
by
U
ni
ve
rs
ity
L
ib
ra
ry
A
t 1
5:
25
0
6

Ju
ne
2
01
7
(P
T
)
firm resources. In this study, we examine the effect of buyer-
supplier relationships as a
potential complementary firm resource. If buyer-supplier
relationships are found to
induce greater levels of effectiveness, efficiency, and/or
resiliency, then it may provide
evidence to warrant additional empirical investigation of the
moderating effect of
additional complimentary resources. This finding may suggest a
slight modification to
Grawe’s (2009) model to account for the moderating effect of
complimentary firm
resources.
The preceding discussion reveals how our study fits into the
current body of SCM
innovation literature. Next, our discussion turns to the specific
LIT innovations that
are investigated in this study.
2.2 LIT artifacts

One purpose of this research is to investigate the effect of LIT
adoption on expected
performance outcomes. To begin, we sought an unbiased method
for choosing the most
appropriate LIT artifacts for the focus of investigation. Ideally,
our study would
investigate the entire population of IT that meet our definition
of LIT. However,
as with any research endeavor, we were required to adopt a
valid sampling technique in
order to study a representative sample of our target population.
Purposive sampling is a
non-random sampling technique in which the researcher uses
judgment in selecting
cases for a specific purpose (Neuman, 2006). This sampling
technique is appropriate to
select unique cases that may be especially informative
(Neuman, 2006). Berelson (1952)
suggests that revealing the focus of attention is one of the
primary uses of content
analysis. As such, we adapted procedures for problem-driven
content analysis
suggested by Krippendorff (2004) to determine which LIT
artifacts may be most
appropriate and especially informative to study. In sum, this
content analysis served as
our purposive sampling technique because we posit that the IT
most common in the
SCM literature will likely offer the best insight into the nature
of the population of LIT.
Krippendorff’s (2004, p. 83) first component of content analysis
involves unitizing,
which is defined as “the systematic distinguishing of segments
of text – images,
voices, and other observables – that are of interest to an

analysis”. In this study, we
sought to determine which LIT are often addressed as the
primary artifact of interest in
the extant SCM literature. Accordingly, the unit of analysis is
an article in a SCM
journal that investigated a specific LIT.
To locate articles that meet the above criteria, the top 20 SCM
journals for research
usefulness as identified by Menachof et al. (2009, p. 151) were
considered. These journals
are shown in Table I in alphabetical order. Of note, since the
purpose of the content
analysis is to determine the most relevant IT artifacts in SCM
literature,
interdisciplinary journal titles such as Harvard Business Review
and Management
Sciencewere not included. As such, 14 of the top 20 journals
identified by Menachof et al.
(2009) were utilized for content analysis. The selected journals
were searched via
ABI/INFORM Complete, Business Source Premier, and
ScienceDirect databases.
A keyword search for “information technology” or “information
system” was conducted
for each journal and the number of results was recorded. Titles
and abstracts were then
reviewed to determine if a specific LIT was addressed as the
primary focus of the article.
The nomenclature of the LIT artifact was then noted and the
total number of articles per
journal that addressed LIT as a primary focus was recorded.
Articles addressing the
general use of IT or loosely defined terms such as EBusiness
were not counted.

IJPDLM
42,1
12
D
ow
nl
oa
de
d
by
U
ni
ve
rs
ity
L
ib
ra
ry
A
t 1
5:
25
0

6
Ju
ne
2
01
7
(P
T
)
At this point in the content analysis, we had compiled a listing
of 28 unique IT
innovation artifacts in the logistics literature. Many of these
technologies did not
receive much attention. For example, IT such as transportation
routing systems and
knowledge management systems only emerged twice. However,
this allowed the top
LIT to be easily identified. EDI and RFID emerged as the two
LIT that are addressed
most often in the SCM literature, accounting for 32 percent of
all articles in which IT is
the primary focus. Because they conform to our strict definition
of LIT and they
together comprise nearly one-third of the SCM literature which
addresses IT, we chose
to adopt EDI and RFID to represent LIT in our study. The
results of this content
analysis are illustrated in Table I.

RFID is a type of automated data collection system that uses
radio waves to identify
objects (Angeles, 2005). Interest in RFID applications in the
supply chain has generated
a rapidly growing body of knowledge in recent years. Some
posit that use mandates
from industry leaders such as Wal-Mart has quickly brought
RFID to the attention of
academicians and practitioners alike (Visich et al., 2007). This
has motivated many
authors to discuss cases of RFID implementation success and
suggest anecdotal or
perceived outcomes of RFID adoption. Academicians are
currently working to develop
the body of empirical literature investigating actual benefits
derived from RFID use
(Visich et al., 2009).
EDI is a technology used to exchange information and data
across organizations
(Germain and Droge, 1995) and may be defined as, “business to
business transfer of
repetitive business processes involving direct routing of
information from one computer
to another without human interference, according to predefined
information formats and
rules” (Holland et al., 1992, p. 539). Unlike RFID, EDI research
has spanned the last two
Journal
Term found in abstract/
citation
IT
innovationsa EDI RFID

European Journal of Operational Research 169 25 1 1
International Journal of Logistics Management 16 1 0 0
International Journal of Logistics: Research and
Applications 17 9 1 4
International Journal of Operations and
Production Management 95 14 1 0
International Journal of Physical Distribution
and Logistics Management 109 18 3 1
Journal of Business Logistics 45 14 3 1
Journal of Operations Management 43 16 1 0
Journal of Purchasing and Supply Management 12 4 3 0
Journal of Supply Chain Management: A Global
Review 29 4 2 0
Operations Research 32 8 0 0
Supply Chain Management Review 39 1 0 0
Supply Chain Management: An International
Journal 41 7 2 2
Transportation Journal 21 4 2 1
Transportation Research: Part E 11 4 0 1
Total 668 125 19 11
Note: aThe specific IT innovation was the primary focus of the
article
Table I.
Results of content
analysis
Logistics
information
technology
13

2
01
7
(P
T
)
decades and is widely viewed as a relatively mature technology
(Narayanan et al., 2009).
However, although the literature is insightful in examining
many phenomena
surrounding EDI (e.g. antecedents to adoption, implementation
techniques, etc.), the
quantitative academic literature investigating actual operational
benefits is not
well assimilated and sometimes inconclusive (Ahmad and
Schroeder, 2001;
Narayanan et al., 2009).
In this study, we combine EDI and RFID into a single unit of
analysis that we label
LIT. No one sample is ever perfectly reflective of the
population. However, because
these technologies are widely used in logistics and meet our
definition of LIT, we
believe that EDI and RFID may be representative of most LIT
artifacts. We chose to
study two LITs in lieu of just one for two specific reasons.
First, research into the
performance outcomes of just one LIT may limit the
genralizability of conclusions
drawn from this study. Although we are still careful to

generalize our results to all LIT,
the study of just one technology would limit our ability to
generalize even further.
Second, the study of more than one LIT will provide more data
for analysis. We
propose that combining these LITs into one unit of analysis may
be appropriate
provided the performance outcomes of each are shown to be
statistically homogenous
(which will be demonstrated later in this manuscript). The
expected benefits of these
LITs are discussed in the following section.
2.3 Expected performance outcomes of LIT adoption
A variety of expected performance outcomes of LIT adoption
are touted in the
literature. As such, many EDI and RFID diffusion studies even
suggest that
anticipation of benefits derived from the implementation of LIT
is a key antecedent to
adoption (Crum et al., 1996; Premkumar, 2003). Benefits
investigated in the literature
range from reduced order cycle times and inventory levels
(Leonard and Davis, 2006) to
reduced labor costs and increased profits (Samad et al., 2010).
Some suggest that this
wide range of benefits related to LIT adoption seems to have
perpetuated many
inconsistencies in construct development and measurement in
the literature
(Narayanan et al., 2009). This problem is exacerbated by the
fact that LIT research
is published in academic journals representing nearly 100
different subject categories
(Irani et al., 2010). Therefore, in order to adequately investigate
our research questions,

these outcomes must be organized in such a way as to allow for
proper analysis. To
this end, we adopt and modify a typology of performance
outcomes proposed in recent
literature (Karimi et al., 2007).
Although each individual technology boasts a unique set of
anticipated benefits, we
suggest that the vast majority of the performance outcomes
(both anticipated and
actual) resulting from the adoption of any LIT may be
categorized into one of three
higher order outcomes. We define a performance outcome as
any result that affects a
business function of the organization, whether in a positive or
negative manner.
Examples of specific performance outcomes in the literature are
offered in Table II.
In this study, we adapt a typology used by Karimi et al. (2007)
to classify performance
outcomes within one of the following three categories:
(1) Efficiency. Encompasses performance outcomes that reduce
cost, reduce cycle
time, or increase productivity.
(2) Effectiveness. Encompasses performance outcomes that
improve decision
making, improve planning, improve resource management, or
improve delivery.
IJPDLM
42,1
14

2
01
7
(P
T
)
(3) Resiliency. Encompasses performance outcomes that build
flexibility into
infrastructure, encourage differentiation of products and
services, or establish
or maintain external linkages to multiple customers and
suppliers.
Of note, we use resiliency in our topology, whereas Karimi et
al. (2007) instead use
flexibility. We use resiliency in our study in lieu of flexibility
because resiliency
accounts for creating both redundancy and flexibility
(Christopher and Peck, 2004;
Sheffi, 2005; Sheffi and Rice, 2005). Thus, our use of resiliency
will allow us to better
categorize those performance outcomes that, although important
in the supply chain
context, do not necessarily fall within the categories of
efficiency or effectiveness.
Table II gives an example of the categorization of performance
outcomes used in this
study. The method for categorizing these outcomes and
enhancing reliability of the
process is described later in Section 3.

This study’s RQ1 investigates the effect of LIT adoption on the
performance
outcomes noted above and reads: do LIT innovations induce
positive performance
outcomes for the adopting firm? The concept of LIT is
operationalized in this study via
EDI and RFID. Performance outcomes are operationalized via
efficiency, effectiveness,
and resiliency and are measured via amalgamation of the
variables investigated in the
literature. This study uses three hypotheses to explore the
relationship posited by our
research question.
Our first hypothesis is concerned with the relationship between
LIT adoption and
business process efficiencies. Efficiency is a measure of
productivity in which what has
been accomplished is measured against what is possible to
accomplish. A technology
may be defined as “a means of uncertainty reduction” (Rogers,
2003, p. 13). Thus, by
definition, any technology should enhance the efficiency of the
process in which it is
applied. However, this has not always been demonstrated in the
literature. Iskandar et al.
(2001) found that employees in firms utilizing EDI perceived no
reduction in the number
of employees required to support operations. These findings are
congruent with that of
Sriram and Banerjee (1994), who found that EDI did not
necessarily reduce employee
workload. Sriram and Banerjee (1994) found that employees
were often still required to
approve routine orders, monitor suppliers, and provide a
signature for EDI orders. This

lack of reduction in labor may be due to the fact that EDI does
not always completely
automate the processes in which it is applied, which results in
the continuance
Performance category
Performance outcome Efficiency Effectiveness Resiliency
Reduce processing costs X
Improve equipment utilization X
Improve planning process X
Improve responsiveness X
Facilitate decision making X
Improve relationship with trading partner X
Decrease number of administrative employees X
Reduce cycle times X
Increase productivity X
Enhance channel cooperation X
Reduce delivery of incorrect product X
Table II.
Example of performance
outcome categorization
Logistics
information
technology
15
D
ow
nl
oa

T
)
of processing work for staff to complete. In addition, EDI may
sometimes merely convert
the type of work that employees are required to carry out. For
example, although EDI
may reduce paperwork processing for some organizations, it
may also increase the
amount of work required at computer terminals.
Although instances are cited above which support the idea that
LIT does not improve
levels of efficiency for an organization, additional research
suggests that LIT does
improve efficiencies. Indeed, consistent with R-A theory, the
literature offers many
examples where use of LIT has been shown to be related to
increased efficiencies. For
instance, contrary to the findings noted above, Wang et al.
(2010) demonstrated via
simulation how LIT may significantly reduce manpower
requirements. Other
simulation studies offer similar findings regarding efficiencies
derived from IT
implementation (Hou and Huang, 2006; Veronneau and Roy,
2009). In addition, Hou and
Huang (2006) demonstrated a variety of operational efficiencies
(e.g. reduced time for
product identification) derived from use of LIT. Similarly,
Bendavid et al.’s (2009) case
study of B-to-B e-commerce applications in the supply chain
suggests that these

technologies may yield significant reductions in transaction
time while also reducing
costs. Because of the results of these and similar studies, we
posit that:
H1. Organizational adoption of LIT increases business process
efficiencies.
Our next hypothesis is concerned with the relationship between
LIT adoption and
business process effectiveness. We define effectiveness as the
degree to which business
objectives are achieved. Thus, measures of effectiveness are
usually concerned with
higher-order organizational outcomes. For instance, efficiency
may be concerned with
reducing order processing costs, whereas effectiveness is
concerned with whether or
not process initiatives affect the bottom line.
The literature offers many examples where LIT is shown to
increase effectiveness,
which also lends support for R-A theory. Srinivasan et al.
(1994) demonstrated the
complimentary effect of LIT on manufacturing supply chains
that utilize just-in-time
( JIT) practices. Their study demonstrated a large reduction in
shipments with
discrepancies when EDI was employed along with JIT. Chow et
al. (2006) found similar
benefits in shipping accuracy via use of RFID. Furthermore,
Clark and Hammond’s
(1997) examination of LIT in the grocery industry found that
EDI adoption led to
increased inventory turns and reduced stock-outs. Hardgrave et
al. (2008) reached

similar conclusions regarding increased effectiveness in his
examination of RFID use
at Wal-Mart.
In contrast, other studies have concluded insignificant
relationships between LIT
adoption and measures of effectiveness. Crum et al.’s (1996)
study concluded that EDI
did not improve decision making for firms in the motor carrier
industry. Further,
Leonard and Davis (2006) realized non-significant results when
investigating the
relationship between adoption of LIT and a variety of
effectiveness measures, to
include increased fill rates and reduced stock-outs. Thus, we
seek to determine if these
contradictory results are an anomaly by investigating whether or
not:
effectiveness.
Our third hypothesis concerns the relationship between LIT
adoption and business
process resiliency. We define resiliency as “the ability to return
to normal performance
levels following supply chain disruption” (Zsidisin and Wagner,
2010, p. 3).
IJPDLM
42,1
16
D
ow

nl
oa
de
d
by
U
ni
ve
rs
ity
L
ib
ra
ry
A
t 1
5:
25
0
6
Ju
ne
2
01

7
(P
T
)
Although resiliency rarely translates into immediate increases in
efficiencies,
effectiveness, or short-term profits, resiliency facilitates an
organization’s preparation
to encounter future, unknown events. This preparedness, then,
often leads to an increase
in (or at least a retention of) efficiency, effectiveness, and
profit in the future.
Both extant research and R-A theory suggest that adoption of IT
facilitates
increased resiliency in the logistics setting. Rogers et al.’s
(1992) study of EDI use
in warehousing suggests that firms using EDI are significantly
more able to
accommodate special or abnormal requests and events than
firms that do not use EDI.
Choe’s (2008) research in the Korean manufacturing industry
corroborates the findings
of Rogers et al. (1992) and suggests that EDI facilitates
increased speed and volume of
new product creation and product changeover, thus increasing
operational resiliency.
Lim and Palvia (2001) posit that this increased resiliency is
achieved primarily via
reduction in paperwork and standardization of procedures.
However, others have

shown that EDI also leads to expansion of a firm’s supplier base
and increased market
channel formalization, which also enhances a firm’s capacity to
adapt to market
conditions (Manabe et al., 2005; Vijayasarathy and Robey,
1997). On the other hand,
conflicting research suggests that EDI does not benefit channel
relationships and
coordination (Johansson and Palsson, 2009; Nakayama, 2003).
Thus, we investigate
whether or not:
resiliency.
2.4 Buyer-supplier relationships
In order to transfer products from the point of origin to the
point of consumption, inter-
and intra-organizational collaboration is inherently a key
component of the supply
chain. As with any collaborative effort, the relationship and
level of integration
between participants may significantly impact the performance
outcomes sought by
each party (Tan et al., 2010). In this study, we define the buyer-
supplier relationship as
the quality of the relationship between buyers and suppliers.
This relationship may be
identified via eight key dimensions:
(1) communication and information sharing;
(2) cooperation;
(3) trust;

(4) commitment;
(5) relationship value;
(6) power imbalance and interdependence;
(7) adaptation; and
(8) conflict (Boeck and Wamba, 2008).
In this study, we consider relationships consisting of the
characteristics of:
(1) communication and information sharing;
(2) cooperation;
(3) trust;
(4) commitment;
Logistics
information
technology
17
D
ow
nl
oa
de
d

by
U
ni
ve
rs
ity
L
ib
ra
ry
A
t 1
5:
25
0
6
Ju
ne
2
01
7
(P
T
)

(5) relationship value; and/or
(6) willingness to adapt as being positive relationships.
Conversely, relationships where these six characteristics are
indicated in a negative
sense or if indications of power imbalance or conflict are
present, we consider to be
negative relationships.
Over the past two decades, a variety of studies have
demonstrated the positive
outcomes derived from buyer-supplier co-operation in the
supply chain. Larson (1994)
found that supply chain relationships consisting of greater
levels of trust, respect,
cooperation, teamwork, unified purpose, and communication
resulted in higher levels of
product quality and lower total costs for both the buyer and
supplier. Additionally, Klein
and Rai’s (2009) study of strategic information flows within the
supply chain suggests
that positive supply chain relationships marked by increased
strategic information
flows between partners yields significant financial and
operational performance
outcomes. In their study, both buyers and suppliers realized
improved management of
assets, reduced operations costs, enhanced productivity,
improved planning, flexibility,
and control of resources.
These positive outcomes derived from positive buyer-supplier
relationships may be

Towards Decision Support and Goal AchievementIdentifying Ac.docx

Towards Decision Support and Goal AchievementIdentifying Ac.docx

Recomendados

Recomendados

Mais conteúdo relacionado

Semelhante a Towards Decision Support and Goal AchievementIdentifying Ac.docx

Semelhante a Towards Decision Support and Goal AchievementIdentifying Ac.docx (20)

Mais de turveycharlyn

Mais de turveycharlyn (20)

Último

Último (20)

Towards Decision Support and Goal AchievementIdentifying Ac.docx