Leveraging Big Data to Develop Supply Chain Management Theory: The Case of Panel Data

• DOI: http://doi.org/10.1111/jbl.12188
• Author: Daniel C. Ganster,Stanley E. Griffis,Jason W. Miller
• Published date: 01 September 2018
• Date: 01 September 2018

Leveraging Big Data to Develop Supply Chain Management

Theory: The Case of Panel Data

Jason W. Miller

1

, Daniel C. Ganster

2

, and Stanley E. Griffis

1

1

Michigan State University

2

Colorado State University

Increased data availability is poised to shape both business practice and supply chain management (SCM) research. This article addresses an

issue that can arise when trying to use big data to answer academic research questions. This issue is that distilled data often have a panel

structure whereby repeated measurements are available on one or more variables for a substantial number of subjects. Thus, to fully leverage

the richness of big data for academic research, SCM scholars need an understanding regarding the different types of research questions answer-

able with panel data. In this article, we devise a framework detailing different types of research questions SCM scholars can answer with panel

data. This framework provides a basis to categorize how SCM scholars have examined the services supply chain setting of health care with

public data regarding hospital-level patient satisfaction. We extend prior research by testing a series of three questions not yet examined in this

area by fitting a series of structured latent curve models to seven years of hospital-level patient satisfaction for nearly 4,000 hospitals. The dis-

cussion highlights theoretical and methodological challenges SCM scholars are likely to encounter as they use the panel data in their research.

Keywords: big data; panel data; repeated measures; structured latent curve models

INTRODUCTION

One of the most important changes in the business landscape

is the increased availability of data and consequently the abil-

ity to incorporate these data when making decisions (Lohr

2010; McAfee and Brynjolfsson 2012; Waller and Fawcett

2013a; Schoenherr and Speier-Pero 2015). Although the term

“big data”have various definitions (Arunachalam et al. 2018;

Tonidandel et al. 2018), most converge to the themes put for-

ward by Sanders (2014) that big data refer to very large quan-

tities of data such that, in their raw form, traditional data

handling approaches such as spreadsheets are no longer feasi-

ble. The defining characteristics of big data that make them

unique from data in the past are their volume, velocity, and

variety (Sanders 2016; Arunachalam et al. 2018). Companies’

supply chain operations generate a tremendous amount of such

data in the form of global positioning system coordinates for

delivery vehicles, customer transaction data, and throughput

information (Sanders 2014). Consequently, the topic of big

data has increasingly drawn interest from supply chain man-

agement (SCM) scholars (Waller and Fawcett 2013a,b; Sanders

2014; Schoenherr and Speier-Pero 2015; Arunachalam et al.

2018).

One of the challenges SCM scholars face when trying to transi-

tion to using big data for research is that there is a substantial gap

between traditional methodological training, which tended to focus

on qualitative and survey-based techniques, and the skills neces-

sary to work with big data. In particular, the one challenge many

SCM scholars face is that once big data are distilled

1

into to a for-

mat more suitable for testing theoretical predictions, these data

may have a repeated measures

2

panel structure as in Massimino

et al. (2017). By panel structure, we mean data that consist of a

substantial number of subjects

3

measured on one or more variables

across several occasions (Baltagi 2008). Although panel data are

not big data per se, for SCM scholars unfamiliar with using panel

data, these data have characteristics associated with true big data

Corresponding author:

Jason W. Miller, Department of Supply Chain Management, Eli

Broad College of Business, Michigan State University, 632 Bogue

Street, N370, East Lansing, MI 48824, USA; E-mail:

mill2831@broad.msu.edu

1

The distillation of big data in a form usable for research may

be performed by the authors or via other entities. Two studies

where authors have performed this process themselves are Mas-

simino et al. (2017) and Mukherjee and Sinha (2017). In some

other instances, entities generating the data may compile these

data into a usable form, such as the Federal Motor Carrier Safety

Administration distilling data from millions of motor carrier

inspections into carrier-level aggregates that are disseminated to

the public (The Volpe Center 2016). In the case of the Toxics

Release Inventory data, third parties such as the Environmental

Defense Fund compiled vast quantities of data regarding chemi-

cal emissions to disseminate them in a more user-friendly fashion

(Fung and O’Rourke 2000).

2

We define repeated measures data to mean that two or more

observations of the same metric(s) are taken from the same sub-

ject(s) with the measurement occasions separated by time (Crow-

der and Hand 1990). We avoid describing such data as

“longitudinal”because it is possible for scholars to have a cross-

sectional study design whereby the independent and dependent

variables are measured at different time points. Such a design is

inherently longitudinal in that there is a temporal separation

between the measures, but the data are not of the repeated mea-

sures type.

3

We adopt nomenclature from Cudeck and Harring (2007) and

use “subject”as a generic term to refer to the unit of analysis for

which repeated observations are obtained.

Journal of Business Logistics, 2018, 39(3): 182–202 doi: 10.1111/jbl.12188

© 2018 Council of Supply Chain Management Professionals

in that (1) volume (e.g., number of records) is substantially more

than encountered in most survey studies; (2) velocity (e.g., fre-

quency at which measures are taken) is greater than surveys, which

by and large are cross-sectional; and (3) variety may be greater

than seen with surveys (e.g., more reliance on proxies is neces-

sary). Thus, focusing on how to leverage panel data for research

can provide a first step for scholars interested in developing the

skills to use big data while further strengthening researchers’

applied use of big data given that the distillation process may

transform true big data into a panel structure to answer theoretical

questions. As such, the remainder of this article will focus on

issues pertaining to the use of panel data by SCM scholars with

the understanding that this can directly and indirectly strengthen

the use of big data for research purposes.

Although increased data availability is welcome, having access

to large quantities of data does not, ipso facto, result in scientific

advancement (Davis 2010). Data can only extend scientific

knowledge to the extent they are used to test hypotheses with

sound theoretical foundations (George et al. 2014; Davis 2015;

Barley 2016). To this end, a key impediment for SCM scholars

most fruitfully leveraging panel data is the absence of a clear,

didactic discussion regarding the array of research questions

panel data can address. In particular, we believe the divergent

perspectives regarding how to analyze panel data that exist, on

the one hand, in sociology (Halaby 2004) and economics (Balt-

agi 2008) and, on the other hand, biostatistics (Fitzmaurice et al.

2011) and quantitative psychology (Nesselroade 1991), could

stymie progress.

4

Our concern is that SCM scholars trained in

either research tradition may not appreciate the value, and thus

legitimacy, of answering questions central to the other.

The purpose of this article was to provide an overview for how

SCM scholars can leverage panel data to devise and test theories

that push the boundaries of our existing SCM knowledge, while

still articulating implications for practitioners. To pursue this goal,

we first devise a framework detailing different types of research

questions that can be answered with panel data. Our framework

integrates concepts from econometrics (Baltagi 2008), biostatistics

(Crowder and Hand 1990; Fitzmaurice et al. 2011), and quantita-

tive psychology (Ram et al. 2005; Grimm et al. 2016) into a uni-

fied framework. We use this as a springboard to illustrate how

SCM scholars have examined the services supply chain setting of

health care with public data regarding hospital-level patient satis-

faction. We show that past studies have begun to address some

types of research questions with panel data, but that many other

types of questions remain. Following this review of services sup-

ply chain health care research, we test a series of three questions

not yet examined in this area with panel data by fitting a series of

structured latent curve models (SLCMs) (Browne and Du Toit

1991; Blozis 2004) to seven years of hospital-level patient satisfac-

tion data for nearly 4,000 hospitals in the United States. Keeping

with our theme, the discussion highlights theoretical and method-

ological challenges SCM scholars are likely to encounter as they

use the panel data in their research.

This article contributes to the SCM literature in several

ways. First, it provides an integrated framework explicating the

types of questions that can be asked with panel data by syn-

thesizing different methodological research traditions. Second,

it extends prior services supply chain investigations in the

health care sector by studying new questions concerning hospi-

tal-level improvement in patient satisfaction. This responds to

Ketokivi and McIntosh’s (2017) call for SCM scholars to

focus more emphasis on studying how subjects’performance

changes over time. Third, our discussion section clarifies the

role that the passage of time plays when devising theories

with longitudinal character, a topic that has received limited

attention (Ancona et al. 2001). Fourth, this research highlights

some of the challenges SCM scholars can expect to face when

using panel data.

This article is structured in three sections. The next section

develops a framework for different types of questions address-

able using panel data. The penultimate section discusses SCM

literature in the health care sector, specifically those that have

examined hospital-level patient satisfaction, and extends this

literature by fitting SLCMs to panel data on hospital-level

patient satisfaction to answer unaddressed questions. The final

section describes theoretical and methodological challenges

with using panel data, notes limitations, and suggests directions

for future research.

LEVERAGING PANEL DATA TO BUILD SCM THEORY

We begin by establishing the boundaries for our examination.

First, we assume theoretical constructs

5

serving as outcomes are

continuous, as opposed to categorical. We make this assumption

because categorical constructs must be operationalized and ana-

lyzed using different techniques (Collins and Lanza 2010). Sec-

ond, we assume that the empirical operationalizations

6

of

theoretical constructs within the same study have an acceptable

degree of measurement invariance so that comparing these

4

These divergences concern both research questions studied

and metaphysical issues more than they concern the use of dif-

ferent terminology. An example of the former is that biostatisti-

cians and quantitative psychologists focus much attention on

studying change over time using growth curve models (Fitzmau-

rice et al. 2011; Little 2013), whereas economists give little

attention to these models—Wooldridge’s (2010) text on panel

data does not discuss such models. A metaphysical difference is

that many quantitative psychologists limit their discussion of

parameter bias because they do not believe in “true”models

(Cudeck and Henly 1991, 2003; MacCallum 2003). Economists

and sociologists tend to strongly emphasize issues of parameter

bias resulting from omitted variables and consequently place

greater emphasis on the use of instrumental variables and subject

fixed effects when fitting panel models (Halaby 2004; Wool-

dridge 2010).

5

This does not mean that observed measures in these databases

need to be continuous, as one can operationalize a continuous

construct using multiple binary observed indicators (Newsom

2015).

6

Constructs can be operationalized using either single mea-

sures (i.e., as observed scores) or as multi-item latent variables.

In the case of single measures, testing for measurement invari-

ance is impossible (Little 2013).

Big Data and Panel Data 183