Demystifying variance in performance: A longitudinal multilevel perspective

• Published date: 01 June 2017
• DOI: http://doi.org/10.1002/smj.2555
• Author: Guangrui Guo
• Date: 01 June 2017

Strategic Management Journal

Strat. Mgmt. J.,38: 1327–1342 (2017)

Published online EarlyView 1 November2016 in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/smj.2555

Received 23 January 2015;Final revisionreceived 20 June 2016

DEMYSTIFYING VARIANCE IN PERFORMANCE: A

LONGITUDINAL MULTILEVEL PERSPECTIVE

GUANGRUI GUO1,2,3,*

1Odette School of Business, University of Windsor, Windsor, Ontario, Canada

2School of Business Administration, University of Miami, Coral Gables, Florida,

U.S.A.

3School of Economics and Management, Tsinghua University, Beijing, China

Research summary: This study employs longitudinal multilevel modeling to re-examine the rela-

tive importance of business unit, corporation, industry, and year effects on business unit perfor-

mance. Total variance in performance is partitioned into stable variance and dynamic variance.

Sourcesof these two parts of variance are explored. Empirical results indicate that (1) stable effects

of corporation-industry interaction are substantially important, but were unequally confounded

with stable effects of business unit, corporation, and industry in results of previous studies; (2)

stable effects of corporation, industry,and corporation-industry interaction, taken together, are of

similar relative magnitude to stable effects of business unit; and (3) random and nonlinear year

effects are very important in explaining dynamic variance. These ndings extend our theoretical

and empirical understanding of performance variability.

Managerial summary: Whether stable or changing, businessunits themselves, corporate-parents,

and industries inuence business unit operations. This article investigates the relative effects of

these factors on business unit performance. Although the traditionalwisdom is that business unit is

critical, this research nds that corporate-parent, industry, and interactions between these, taken

together, are as inuential as business unit. Specically, interactions between corporate-parent

and industry are important for over-time average business unit performance, indicating that a

given corporate-parent unevenlyinuences its business units in different industries and that a par-

ticular industry unevenly inuences business units within itself from different corporate-parents.

This study also demonstrates that changes in business unit, corporate-parent, and industry are

important drivers of over-time volatility of business unit performance and that effects of these

changes differ. Copyright © 2016 John Wiley & Sons, Ltd.

INTRODUCTION

Understanding sources of variation in performance

is a dening issue of strategic management. One

prominent stream of literature has examined the

relative importance of business unit (BU), corpora-

tion, industry, and year effects on performance by

decomposing variance in performance (e.g., Brush,

Bromiley, and Hendrickx, 1999; Chang and Singh,

2000; Hough, 2006; Karniouchina et al., 2013;

Keywords: stable variance and dynamic variance in per-

formance; corporation-industry interaction effects; random

and nonlinear year effects; over-timechanges and dynamic

variance; longitudinal multilevel modeling

*Correspondence to: Guangrui Guo, Odette School of Business,

401 Sunset Avenue, Windsor, Ontario N9B 3P4, Canada.

E-mail: grguo@uwindsor.ca

Copyright © 2016 John Wiley & Sons, Ltd.

McGahan and Porter, 1997, 2002; Misangyi et al.,

2006; Roquebert, Phillips, and Westfall, 1996;

Rumelt, 1991; Schmalensee, 1985; Short et al.,

2006; Wernerfelt and Montgomery, 1988).1Histor-

ically, results of studies in this research stream have

been conicting (Bowman and Helfat, 2001; Brush

and Bromiley, 1997; Hough, 2006; Misangyi etal.,

2006). However, most recent studies have found

that overall, BU effects were the most important,

while effects of corporation, industry, and year

were usually statistically signicant but much

1Some studies have extended this stream of research by adding

business group effects and country or institution effects (Bhat-

tacharjee and Majumdar, 2011; Chang and Hong, 2002; Diaz

Hermelo, and Vassolo, 2012; Khanna and Rivkin, 2001, 2006;

Majumdar and Bhattacharjee, 2014; Makino, Isobe, and Chan,

2004; McGahan and Victer, 2010; Tong et al., 2008).

1328 G. Guo

smaller (Hough, 2006; Karniouchina et al., 2013;

McGahan and Porter, 2002; Misangyi et al., 2006).

Empirical ndings from this line of research have

sometimes been used to argue the explanatory

power of different theories of rm performance,

such as the resource-based view (RBV) (Barney,

1991; Wernerfelt, 1984), the industry organization

(IO) perspective (Porter, 1979, 1980, 1981), and

theories of corporate strategy (Chandler, 1962;

Eisenhardt and Piezunka, 2011; Porter, 1987;

Rumelt, 1974). Arguably, these recent ndings

indicate that BU-specic resources are most impor-

tant for performance, while corporate strategy,

industry structure, and time are much less impor-

tant. These ndings also imply that the foregoing

effects are both independent (lack of interaction

effects) and stable (marginal year effects).

This study notes that largely due to the inability

of statistical approaches employed, previous stud-

ies have not adequately captured interaction effects,

random year effects, or nonlinear year effects.

In particular, effects of interactions may be con-

founded with BU, corporation, and industry effects,

and year effects may be dramatically underesti-

mated, potentially leading to biased interpretation

of the relevant importance of these components.

Drawing on longitudinal multilevel modeling (Fitz-

maurice, Laird, and Ware, 2012; Hoffman, 2015),

this study endeavors both to extract important sta-

ble effects of corporation-industry interaction from

previously confounded effects and to capture ran-

dom as well as nonlinear year effects.

The following section reviews alternative

approaches utilized in this stream of literature.

Then, the article introduces the data and examines

sources of stable variance by tting two uncon-

ditional means multilevel models. Additionally,

an unconditional longitudinal multilevel model

is estimated in order to explore and analyze year

effects. The last section presents conclusions,

theoretical implications, limitations, and avenues

for future empirical research.

ALTERNATIVE APPROACHES

TO PERFORMANCE VARIANCE

DECOMPOSITION

Generally, three parametric approaches2have been

utilized in this research stream:3(1) components

2Ruei and Wiggins (2003) employed a nonparametric approach

to stratifying performance into categories ranging from inferior

of variance (COV) (aka variance component anal-

ysis), (2) analysis of variance (ANOVA), and (3)

multilevel modeling (MLM) (aka hierarchical lin-

ear modeling or mixed effects modeling).4Each

approach is reviewed belowin terms of its strengths,

limitations, and potential to produce biased results.

Components of variance (COV)5

Some early studies in this stream of literature

used COV (Chang and Singh, 2000; McGahan

and Porter, 1997; Roquebert et al., 1996; Rumelt,

1991; Schmalensee, 1985). The strength of COV

is that it provides explicit variance decomposition

for calculating the relative importance of differ-

ent components. However, this approach requires

the assumptions that (1) effects of different com-

ponents are independent (Garson, 2012; McGahan

performance to modal performance to superior performance and

then estimated an ordinal regression model.

3In a parallel research stream (McGahan, 2009), studies have

examined the persistence of differenteffects on performance (e.g.,

Bou and Satorra, 2007, 2010; Diaz Hermelo and Vassolo, 2012;

Diaz Hermelo, Hetiennot, and Vassolo,2014; Furman and McGa-

han, 2002; Gschwandtner, 2012; McGahan and Porter, 1999,

2003). These studies used a variety of autoregressive models,

which are usually referred to as transitional models or dynamic

models for panel data (DMPD). Both multilevel modeling (MLM)

and DMPD are sophisticated approaches to analyzing longitudi-

nal data (Rabe-Hesketh and Skrondal, 2012). However, the two

have different viewpointso n correlation amongobservations. For

MLM, observations are correlated because they are from the same

subject and share the same underlying processes. Thus, MLM cap-

tures the correlation by introducing random intercepts, random

slopes, and appropriate variance-covariance structures (Hoffman,

2015). For DMPD, on the other hand, observations are corre-

lated because the past inuences the present. Accordingly,DMPD

directly models relationship between current outcome and previ-

ous outcomes by using the latter as predictors. Choice between

these largely depends on research questions and academic disci-

plines. Using DMPD, Bou et al. (2007, 2010), and Diaz Hermelo,

Etiennot, and Vassolo(2014) examined the persistence of perfor-

mance as well as decomposing variance in performance or abnor-

mal performance into rm level, sector (industry) level,and coun-

try level. These three studies also captured country-industry (sec-

tor) interaction effects. The present study, however, decomposes

performance variance into traditional components: BU, corpora-

tion, industry, and year effects. In particular, this study captures

not only effects of corporation-industry interaction but also ran-

dom and nonlinear year effects on performance. MLM and DMPD

are complementary approaches to analyzing longitudinal or panel

data. These capture the dynamic nature of longitudinal data from

different perspectives.

4Brush et al. (1999) used two-stage least squares (2SLS) as an

alternative method to estimate the relativeimportance of different

components. However,2SLS can model only subsamples in which

all corporations must have the same number of businesssegments.

See Hough (2006) for a detailed discussion of 2SLS.

5For detailed discussions of COV, see Brush and Bromiley,

(1997), Garson (2012), and McGahan et al. (1997).

Copyright © 2016 John Wiley & Sons, Ltd. Strat. Mgmt. J.,38: 1327–1342 (2017)

DOI: 10.1002/smj