PRINCIPLES AND APPLICATIONS

OF MULTILEVEL MODELING
IN HUMAN RESOURCE
MANAGEMENT RESEARCH
JIE SHEN

Multilevel modeling is important for human resource management (HRM)

research in that it often analyzes and interprets hierarchal data residing at more
than one level of analysis. However, HRM research in general lags behind other
disciplines, such as education, health, marketing, and psychology in the use of
a multilevel analytical strategy. This article integrates the most recent literature
into the theoretical and applied basics of multilevel modeling applicable to HRM
research. A range of multilevel modeling issues have been discussed and they
include statistical logic underpinning multilevel modeling, level conceptualiza-
tion of variables, data aggregation, hypothesis tests, reporting mediation paths,
and cross-level interactions. An empirical example concerning complex cross-
level mediated moderation is presented that will suffice to illustrate the princi-
ples and the procedures for implementing a multilevel analytical strategy in HRM
research. © 2015 Wiley Periodicals, Inc.

Keywords: human resource management (HRM), multilevel modeling (MLM),

multilevel structural equation modeling (MSEM), research methods

H
uman resource management (HRM) (Ostroff & Bowen, 2000). Employee attitudinal and
research often involves hierarchal data behavioral responses to HRM policies and prac-
from more than one level of analysis. tices may be similar in the same organization and
Individual employees are nested in teams different in others due to contextual effects (Bliese
or departments that are entrenched & Hanges, 2004). Ignoring the inherent depen-
within organizations. In turn, organizations are dence of hierarchal data would result in deflated
nested in industries embedded in larger environ- standard errors and inflated values of model fit
ments, such as geographic regions, nations or or correlations (Rowe & Hill, 1998). This type of
economic or political blocks. HRM, as a subset of dependence in data structure is likely to lead to
organizational policies, is a higher-level variable, gross errors of prediction if using nonmultilevel
as individuals within the same organization/unit modeling statistical approaches such as ordinary
share the same HRM policies and practices, but least squares (OLS) regression, designed to analyze
individuals in different organizations/units do not the same level of data (Snijders & Bosker, 2012).

Correspondence to: Jie Shen, School of Management, University of South Australia Business School, City West
Campus, Adelaide, South Australia, SA5001, Phone: +61 430355248, E-mail: [email protected]

Human Resource Management

© 2015 Wiley Periodicals, Inc.
Published online in Wiley Online Library (wileyonlinelibrary.com).
DOI:10.1002/hrm.21666
2 HUMAN RESOURCE MANAGEMENT

Multilevel regression models account for vari- & Chuang, 2004; Messersmith, Patel, & Lepak,
ance among variables at different levels, han- 2011; Sun, Aryee, & Law, 2007; Takeuchi, Chen, &
dling sources of errors more rigorously than OLS, Lepak, 2009). By exploring cross-level relationships
although parameter estimates are not substantially between HRM/HPWS and a range of employee out-
different (Rowe & Hill, 1998). Multilevel regression comes, these studies contribute to HRM literature
models have been variously named as hierarchical in bridging macro and micro perspectives. The
linear models (HLM) (Raudenbush & Bryk, 2002), use of multilevel modeling in education, health,
mixed effects or mixed models (Littell, Milliken, psychology, organizational behavior, and market-
Stroup, & Wolfinger, 1996), random coefficient ing research began in the late 1970s (Mathieu &
models (Kreft & de Leeuw, 1998), and variance Chen, 2011; Raudenbush & Bryk, 2002). Despite
component models (Longford, 1986). Despite the the growing interest from HRM researchers, HRM
different nomenclature, these models all simulta- research overall lags behind in the use of multilevel
neously test relationships within a certain level modeling. Currently, the majority of HRM research
and between or across hierarchical levels, allowing does not adopt multilevel theoretical perspectives,
researchers to disentangle effects of between- and but is conducted at the same level of analysis. This
within-group variance on the dependent variable indicates a great need for further multilevel HRM
while using individual independent variables at research, metaphorically described by Kulik (2012)
the individual level, and group inde- as “picking high-hanging fruits” and “climbing the
pendent variables at the group level higher mountain” (p. 447).
The perceived (Hofmann, 1997). A reason for the inadequate implementation
Multilevel modeling (MLM) also of multilevel modeling in HRM research is a lack
complexity and
offers advantages over alternative of guiding theoretical frameworks linking HRM
difficulties involved techniques including disaggrega- practices and other variables across multiple lev-
tion and aggregation approaches els (Ostroff & Bowen, 2000). Ostroff and Bowen’s
in multilevel design, in dealing with hierarchal data (2000) meso paradigm integrates individual-level,
(Osborne, 2000). The disaggregation organization-level, and cross-level effects, con-
data collection and
approach reduces upper-level vari- tributing to filling this important literature gap
analysis, and result ables to a lower level. Individuals and calling for more multilevel HRM research
in the same unit are assigned the (Hofmann et  al., 2000). Another general rea-
interpretation and same mean unit-related scores, son for such deficiency is that multilevel models
which ignore between-group varia- have only recently become available, and com-
presentation may
tions. Consequently, shared vari- puter software has become technically feasible
have hindered HRM ance is no longer accounted for, and for only a couple of decades. Multilevel analytical
the assumption of independence techniques primarily emerged from educational
researchers from of errors is violated, resulting in research and have only been recently introduced
inflated correlations. The aggrega- to HRM research. The perceived complexity and
conducting multilevel
tion approach raises the lower-level difficulties involved in multilevel design, data col-
research. variables to the higher hierarchal lection and analysis, and result interpretation and
level, thereby ignoring individual presentation may have hindered HRM researchers
differences. This approach results from conducting multilevel research. Therefore,
in research findings focused on higher hierar- it is helpful if theories and procedures for mul-
chal-level predictability, misrepresenting the tilevel modeling are explained in a systematic
relationships between variables (Hofmann, 1997; manner using empirical HRM examples that HRM
Raudenbush & Bryk, 2002). Multilevel modeling researchers new to multilevel modeling can easily
corrects these violations (Osborne, 2000). follow without frustration.
It is important for HRM researchers to integrate The main purposes of this study are to distill
macro and micro levels of analysis by simultane- the literature with reference to statistical logic and
ously taking into account organizational effects the most recent debates and development regard-
and individual effects (Hofmann, Griffin, & Gavin, ing multilevel modeling, and to demonstrate the
2000; Ostroff & Bowen, 2000; Wei, Han, & Hsu, main procedures for conducting multilevel HRM
2010; Whitener, 2001). Recently, a growing num- research. The procedures covered in this study
ber of empirical studies have conceptualized and include level conceptualization of variables, data
measured general HRM practices or high-perfor- handling (e.g., collection, preparation, and aggre-
mance work systems (HPWSs) or high-performance gation), preconditions for conducting multilevel
HRM at higher levels (Aryee, Walumbwa, Seidu, & analysis, hypothesis testing (direct effect, media-
Otaye, 2012; Den Hartog, Boon, Verburg, & Croon, tion, and moderation), and reporting results in
2013; Jensen, Patel, & Messersmith, 2013; Liao statistical or graphical forms.

Human Resource Management DOI: 10.1002/hrm

 MULTILEVEL MODELING IN HRM RESEARCH 3

Published pedagogical multilevel modeling where γ00 (gamma subzero zero) = Level 2 intercept
manuals often use the simplest model contain- term, which is the mean DV when controlling for
ing Level 1 direct effect, Level 2 direct effect, and Level 2 predictor variable, γ01 = Level 2 slope term,
cross-level interaction as examples. However, which is the mean effect of Level 2 predictor vari-
HRM models are commonly complex, including able on DV scores, Zj = Level 2 predictor variable
mediation, mediated moderation, or moderated of unit j, and u0j = residual error when modeling
mediation. To appropriately guide HRM research- variation in intercepts.
ers, this article helpfully presents an empirical Equation 3 represents the main effect of Level
example of cross-level mediated moderation to 2 predictor variable Z on the between-unit vari-
illustrate the procedure for performing complex ance in Level 1 dependent variable after control-
multilevel analysis in HRM research. As such, this ling for Level 1 predictor variable. For the slopes:
article equips HRM researchers with the most up-
to-date multilevel modeling knowledge, contrib- Bij = γ10 + γ11Zj + uij (3)
uting to the promotion of further multilevel HRM
research. where γ10 = Level 2 intercept term, which is the
mean DV when controlling for Level-2 predictor
Statistical Logic Underpinning Multilevel variable; γ11  =  Level 2 slope term, which is the
Modeling mean effect of Level 2 predictor variable on DV
Essentially, multilevel modeling investigates scores; Zj  =  Level 2 predictor variable of unit j;
simultaneously within-unit and between-unit and uij (i.e., t00, tau subzero zero) = residual error at
relationships by estimating within-unit and Level 2 when modeling variation in slopes.
between-unit models separately (Osborne, 2000). Equation 3 represents a cross-level interaction
For the sake of simplicity and clarity, the dis- in which Z moderates the relationship between
cussions throughout this paper focus on two- X and Y. Integrating Equations 2 and 3 results in
level models. The Level 1 model illustrates the a mixed-effect model, as shown in Equation 4,
relationships between individual level predic- which highlights the interaction term γ10Xij Zj as a
tor variables and the individual level outcome cross-level moderation.
variable. The outcome of estimating the Level 1
model is intercepts and slopes that vary between Yij = γ00 + γ01Zj + γ10Xij + γ10Xij Zj + u0j + uijXij + eij (4)
units because each unit has its own intercept and
slope. The Level 2 model indicates how Level 2 Multilevel analyses generate both fixed effects
predictor variables predict the varying intercepts at Level 1, as shown in the first four terms, with
and slopes resulting from the analysis of Level 1 gammas γ in Equation 4 nested within Level 2,
model (Hofmann, 1997; Hofmann et  al., 2000). and random effects shown in the last three terms
As Level 1 regression parameters are regressed that vary across units (Hofmann et al., 2000). u0j
onto Level 2 variables in Level 2 analysis, a mul- and uij are residual error terms.
tilevel modeling approach is actually a regression To test the hypothesized relationships indi-
of regressions (Arnold, 1992). The Level 1 model cated in Equation 4, the following conditions
that does not contain Level 2 predictor variables must be met:
is shown in Equation 1:
• There is systematic within- and between-
Yij = βoj + βij Xij + eij (1) group variance in the DV. Multilevel analysis
partitions within- and between-group vari-
where Yij  =  value of Level 1 dependent variable ance. Systematic within-and between-group
(DV) for individual i in unit j, βoj = the intercept variance in the DV indicates the DV is influ-
(average DV) for unit j, βij = regression coefficient enced by both individual factors and group
slope associated with Level 1 predictor variable level factors, satisfying the need for examin-
for unit j, Xij = value of Level 1 predictor variable ing the effect of group level factors.
for individual i in unit j, and eij (i.e., rij or s2) = the • Mean Level 1 slopes in the DV across organi-
residual error at Level 1 nested with Level 2. The zations are significantly different from zero.
Level 2 model takes intercepts and slopes for units This indicates significant variance in the DV
as dependent variables and uses Level 2 predictor is, at least partially, due to Level 1 predictor
variable as a covariate at Level 2. For the inter- variable.
cepts, the equation is shown in Equation 2 and • There is significant variance in Level 1 inter-
for the slopes in Equation 3. For the intercepts: cepts (the average score of the DV across orga-
nizations). This indicates significant variance
βoj = γ00 + γ01Zj + u0j (2) in the DV is due to Level 2 predictor variable.

Human Resource Management DOI: 10.1002/hrm

4 HUMAN RESOURCE MANAGEMENT

• There is significant variance in Level 1 slopes, 2013), organizational performance, such as service
indicating that while significant variance in performance and customer satisfaction (Liao &
the DV is caused by Level 1 predictor variable, Chuang, 2004), and financial performance (Den
the DV also differs in groups by Level 2 predic- Hartog et al., 2013) have been conceptualized and
tor; showing Level 1 predictor variable and measured at higher levels, such as unit, branch,
Level 2 predictor variable interactively influ- department, and organization. Employee work
ence the DV. attitudes and behaviors traditionally deemed as
individual-level constructs can also be aggregated
to higher levels. For instance, job satisfaction,
Variable Level Conceptualization affective commitment, psychological empower-
According to Rousseau (1985, p. 4), the “[l]evel ment, and organizational citizenship behavior
of measurement refers to the entities from which have been aggregated to the departmental level
the data are drawn or are attached (e.g., rat- in Messersmith et  al. (2011) and organizational
ers, individuals, organizations etc.).” Lower- level in Sun et al. (2007). Whether to conceptual-
level entities do not have to be individuals; they ize and measure a variable at a higher level should
can be groups, departments, organizations or be determined by research design and theoretical
regions. Repeated measurements and methodological rationales.
of individuals may also be exam-
Traditional multilevel ined (Luke, 2004). When research Empirical Example
modeling can be is conducted at an interpersonal The illustrative research project is designed to
or intrapersonal level of analysis, explore the effect of organizational high com-
used only to explore the person is the higher-level unit. mitment HRM practices on individual employee
Multilevel regression analysis can knowledge sharing behavior and the underly-
the effects of higher- also be applied to longitudinal data, ing mechanisms to answer “why” and “when”
level variables on where levels are defined by the mea- questions. Several studies have explored the
surement occasions nested within relationship between HRM and knowledge shar-
Level 1 variables or individuals (Snijders, 1996; Willms ing (Collins & Smith, 2006). However, previous
& Raudenbush, 1989). With the research was conducted at the same level of anal-
the effects of the traditional MLM approach, mostly ysis, not taking into account the interaction of
same-level predictor due to the constraints of computer organizational contextual influence and individ-
software, the outcome variable is ual effects. Moreover, the mechanism via which
variables on Level always situated at the lowest level high-commitment HRM influences employee
of the hierarchy (Castro, 2002; knowledge sharing is unclear. The example proj-
1 outcome, rather Hofmann, 1997; Hofmann et  al., ect addresses these limitations. High-commitment
than the effects of 2000; Luke, 2004). In other words, HRM practices, such as developmental appraisals,
traditional multilevel modeling can comprehensive training and development, and
lower-level variables be used only to explore the effects competitive and equitable pay, create “conditions
of higher-level variables on Level 1 that encourage employees to identify with the
on higher-level variables or the effects of the same- goals of the organization and work hard to accom-
outcomes. level predictor variables on Level 1 plish those goals” (Whitener, 2001, p. 517).
outcome, rather than the effects of In their meso paradigm, Ostroff and Bowen
lower-level variables on higher-level (2000) noted cross-level effects of organization-
outcomes. However, with multilevel structural level HRM practices on individual work attitudes
equation modeling (MSEM), the outcome variable and behaviors. Building on this meso paradigm,
can be situated at higher levels. For example, if a this article hypothesizes a positive cross-level
Level 1 predictor variable influences a Level 2 out- effect of high-commitment HRM on employee
come variable via a Level 1 mediating variable, the knowledge sharing. High-commitment HRM is
model is called a 1-1-2 design. If a Level 2 predic- indicative of personified organizational support
tor variable influences a Level 2 outcome variable for employees, who will reciprocate organizational
via a Level 1 mediating variable, the model is a support with more positive discretional work
2-1-2 design (Preacher, Zyphur, & Zhang, 2010). behaviors, such as knowledge sharing. Hence,
In HRM research, in addition to HRM prac- perceived organizational support (POS) would
tices, a wide range of variables such as concern for mediate the relationship between high-commit-
employee climate (Takeuchi et  al., 2009), service ment HRM and knowledge sharing. As knowledge
climate (Liao & Chuang, 2004) and empowerment sharing is cooperative organizational citizenship
climate (Aryee et al., 2012; Den Hartog et al., 2013), behavior, the effect of HRM and POS on knowl-
organizational communication (Den Hartog et al., edge sharing is influenced by the organizational

Human Resource Management DOI: 10.1002/hrm

 MULTILEVEL MODELING IN HRM RESEARCH 5

cooperative climate. That is, organizational coop- Sample and Measures

erative climate moderates the relationships of
A sample of 738 employees from 30 firms in the
high-commitment HRM and POS with knowledge
People’s Republic of China was used. The data
sharing. Specifically, this article tests the following
for the study were collected between July and
hypotheses:
October 2012 through the network of Chinese
MBA students who also held managerial positions
Hypothesis 1: High-commitment HRM will be posi- in the participating firms. These firms cover man-
tively related to employee knowledge-sharing behavior. ufacturing, retail, finance, electronics, distillers,
communication, food processors, chemicals, and
Hypothesis 2: High-commitment HRM will influence hotels—representing various industry sectors. The
employee knowledge-sharing behavior through the mean number of employees is 532, with a stan-
mediation of perceived organizational support. dard deviation (SD) of 741. The multisourced data
were used to avoid common method variance
Hypothesis 3a: Organizational cooperative climate will (Podsakoff, MacKenzie, Lee, & Podsakoff, 2003).
moderate the relationship between high-commitment Specifically, employees answered questions about
HRM and employee knowledge-sharing behavior. high-commitment HRM, POS, and cooperative cli-
mate. Supervisors rated their subordinates’ knowl-
Hypothesis 3b: Organizational cooperative climate edge-sharing behavior. Each firm was required to
will moderate the relationship between perceived randomly select 50 employees from one depart-
organizational support and employee knowledge-shar- ment or production/service unit to participate in
ing behavior. the study. The total number of received matched
completed surveys was 761, accounting for a 51
In this illustrative research, employee knowl- percent individual-level response rate. The num-
edge sharing is the individual-level dependent ber of usable surveys is 738.
variable. High-commitment HRM is the organiza- The measure for high-commitment HRM prac-
tion-level predictor influencing employee knowl- tices was adapted from the five items developed
edge sharing via the mediation of POS at the by Snell and Dean (1992). A sample item is “My
individual level. Organizational cooperative cli- organization provides comprehensive, adequate
mate is conceptualized as the organizational-level training to employees.” POS was measured using
variable moderating the direct effect of high-com- the eight-item scale adapted from Eisenberger,
mitment HRM and the second stage of the media- Huntington, Hutchison, and Sowa (1986). A
tion of POS. Figure 1 depicts the conceptual model sample item is “My organization cares about my
indicating the multilevel nature of the research opinions.” Cooperative climate was measured
design: Hypothesis 1 proposes a cross-level main using a five-item scale adapted from Chatman and
effect; Hypothesis 2 is a meso-mediation relation- Flynn (2001). A sample item is “There is harmony
ship with a 2-1-1 design; Hypotheses 3a and 3b within my team.” Knowledge-sharing behavior
are cross-level moderations. Jointly, the model is was measured using seven items adapted from
a meso-mediated moderation, indicating the need Bock and Kim (2002). A sample item is “He or she
to use a multilevel modeling strategy. This exam- always provides constructive ideas to help col-
ple therefore will sufficiently demonstrate how to leagues improve performance.” Likert-type scales
conduct complex multilevel HRM research. ranging from 1 = strongly disagree to 5 = strongly

High-Commitment H1
HRM
Organizational H3a
Level
Cooperative
Climate

H3b
Perceived
Individual Knowledge-Sharing
Organizational
Level Behavior
Support

H2

FIGURE 1. The Conceptual Model and Hypotheses

Human Resource Management DOI: 10.1002/hrm

6 HUMAN RESOURCE MANAGEMENT

agree were used for all study variables. The scales are not allowed at Level 2 or above. For this rea-
have Cronbach alpha values between .84 and .89. son, groups with missing data at Level 2 or above
Gender, age, education, hierarchal position, and should be removed.
tenure are controlled at the individual level, and
organizational size (log of employee numbers) Data Aggregation
and response rate at the organizational level. Some computer programs such as HLM require
Data Issues separate datasets for performing multilevel analy-
ses. In the case of a two-level analysis, two datas-
Data Requirements ets are required—one containing individual-level
variables and the other Level 2 variables. The
A sufficient sample size for multilevel modeling group ID links two datasets. Statistical Package
is important for accurate estimation in terms of for the Social Sciences (SPSS) has the function to
regression coefficients and variance. aggregate data: “If global measures of constructs
Statistical power for multilevel mod- are not available, data must be gathered from mul-
Statistical power
els depends on the number of both tiple employees within a firm, and from multiple
for multilevel individual observations at the indi- firms” (Ostroff & Bowen, 2000, p. 253).
vidual level and groups at the unit Although HRM practices should be conceptu-
models depends level, as Level 1 sample size influ- alized at the organizational level, it is preferable if
ences the statistical power to detect the data are collected from individual employees
on the number of
Level 1 direct effects and Level 2 due to the importance of employee perceptions
both individual sample size is relevant to the statis- of HRM practices (Ostroff & Bowen, 2000). Liao,
tical power to detect Level 2 direct Toya, Lepak, and Hong (2009) noted significant
observations at effects (Raudenbush & Liu, 2000). differences between managerial perceptions and
The requirement for both large indi- employee perceptions of HRM practices in the
the individual
vidual sample size and group sample same organization. In the illustrative project, the
level and groups size is regarded as a disadvantage data for high-commitment HRM and cooperative
of multilevel modeling (Kreft & de climate were collected from individual employ-
at the unit level, Leeuw, 1998). However, “the group- ees. When higher-level variables are composites
level sample size is generally more of lower-level variables, researchers should justify
as Level 1 sample
important than the total sample the aggregation of the data for lower-level vari-
size influences the size or observations per group, with ables (Mathieu & Chen, 2011). The following two
a large individual level sample size preconditions should be met in order to aggregate
statistical power partially compensating for a small data as Level 2 variables (Rousseau, 1985).
number of groups” (Maas & Hox,
to detect Level 1
2006, p. 87). Condition 1: A high level of interrater agreement for
direct effects and Specifically, according to Kreft Level 2 variables within the Level 2 units.
and de Leeuw (1998), a sample of
Level 2 sample size 20 is recommended as the small- Interrater agreement or homogeneity means
est acceptable number for groups, the reliability of unit-level variables takes into
is relevant to the
and five for individual observa- account differences within units relative to dif-
statistical power to tions per group. To have sufficient ferences between units. Rwg has been developed
power (e.g., .90) to detect cross-level by Bliese (2000) and James, Demaree, and Wolf
detect Level 2 direct interactions, a sample of 30 units
(1984) to assess the level of interrater agreement.
containing 30 individuals each is Rwg can be calculated using the following equation
effects.
desirable (Hofmann et  al., 2000). (LeBreton & Senter, 2008, p. 819):
Notably, in general, the larger the
group-level sample, the stronger the power of Rwg(J)= J[1–(s̄2Xj)/ σ2E] / J[1–(s̄2Xj)/σ2E)+s̄2Xj)/σ2E ] (5)
predictability (Hofmann, 1997). Hence, it would
be helpful to increase sample size, the group- where J = the number of items ranging from j = 1
level sample size in particular, especially when to J, X = an observed score, typically measured on
researchers are interested in effects of higher-level an interval scale of measurement, s̄2Xj = the mean
predictor variables.1 The illustrative sample con- observed variance on X, σ2E = the variance expected
tains 30 firms and the mean number of responses when there is a complete lack of agreement among
at Level 1 is 24.6; meeting the recommended the raters and s̄2Xj)/σ2E  =  the proportion of error
requirements for sample size for multilevel analy- variance caused by random responding.
sis. As for missing data, multilevel modeling han- Readers are referred to LeBreton and Senter
dles missing data only at Level 1. Missing data (2008, pp. 841–845) for the syntax for calculating

Human Resource Management DOI: 10.1002/hrm

 MULTILEVEL MODELING IN HRM RESEARCH 7

rwg using SPSS. According to James et al. (1984), an variations in Level 2 variables, justifying the
rwg greater than .70 is acceptable and the higher aggregation of high-commitment HRM and coop-
the value of rwg, the stronger the within-group erative climate as Level 2 constructs.
agreement of the construct. Rwg is assessed in one
Data Centering
group at a time. If rwg is below .70 for some groups,
researchers should determine how many groups Centering of predictor variables reduces unneces-
and variables have low rwg values and why. It is sary multicollinearity and improves the interpret-
common to report either the range or average for ability of lower-order coefficients in multilevel
each group for each variable (LeBreton & Senter, analysis (Hofmann & Gavin, 1998; Kreft & de
2008). With the illustrative sample, rwgs range from Leeuw, 1998; Mathieu, Aguinis, Culpepper, &
.74 to .91 for high commitment HRM, and from Chen, 2012). Group mean centering centers pre-
.75 to .94 for cooperative climate, respectively. dictors on group means. In this case, the intercept
is the average outcome for each group; allowing
Condition 2: Systematic between-group variations in interpretation of parameter estimates as person-
Level-2 variables. level effects within each group. In the grand mean
centered model, predictors are centered on overall
Generally, researchers should adopt the follow- means. This represents the group mean value for a
ing three procedures to investigate between-group person with a (grand) average on every predictor. It
variations in Level 2 variables. Firstly, one-way is recommended that Level 1 predictors normally
ANOVA (analysis of variance) is performed to are group mean centered to more accurately esti-
examine between-group variations. ANOVA at this mate intercepts. Level 2 predictors
stage can be performed on SPSS. Second, intraclass are suggested to be grand mean cen-
tered. This results in Level 2 inter- Multilevel analysis
correlation (ICC(1)) is calculated to check the pro-
portion of variance due to team variability using cept being equal to the mean score
involves testing
the following equation: of the outcome variable (Mathieu
et  al., 2012). Dummy variables four hierarchical
ICC(1) = [MSB – MSW] / can be centered, although dummy
[MSB + (n – 1) × MSW] (6) variables do not change the inter- models: null model,
pretation of the intercepts when
random intercepts
where MSB is the between-group mean square, group mean-centering is employed
MSW is the mean square within group and n is (Hofmann & Gavin, 1998; Kreft & model, intercepts-as-
the average number of members within groups. de Leeuw, 1998).
MSB and MSW can be obtained by conducting When interaction terms are cre- outcome model, and
one-way ANOVA on SPSS in which Level 2 vari- ated using the variables at the same
slopes-as-outcomes
able is the DV and the group ID is IV. Bliese (2000) level, researchers need to create
suggests that ICC(1) values different from zero are interaction terms separately (e.g., on model.
desirable, with values close to .20 indicating high SPSS) and normally use grand mean
scores for group-level analysis. centering for both focal predictors
Thirdly, reliability of the mean (ICC(2)) is cal- and moderators (Mathieu et  al., 2012). Preacher
culated to examine the extent to which teams can et al. (2010) suggested that if more than a couple of
be used to reliably differentiate in terms of indi- the variables have ICCs below .05, the estimation
viduals’ ratings. ICC(2) can be calculated using of the indirect effect is likely to be unstable with
Equation 7. ICC(2) values greater than .60 are a potentially large bias. Under this circumstance,
regarded as desirable (Glick, 1985). group mean should be used to obtain stable indi-
rect effect. Accordingly, in the illustrative research,
ICC(2) = (MSR – MSW) / MSR (7) individual-level control variables are not centered,
Level 1 variables are group mean centered, and
With the illustrative sample, one-way ANOVA Level 2 variables are grand mean centered.
analyses demonstrate significant variations in
high commitment: SRHRM (F(29) = 3.41, p < .001)
Hypothesis Testing
and in cooperative climate (F(29) = 2.73, p < .01) Multilevel analysis involves testing four hierar-
among the 30 participating companies. ICC(1) chical models: null model, random intercepts
and ICC(2) are .18 and .64 for high-commitment model, intercepts-as-outcome model, and slopes-
HRM, and .16 and .71 for cooperative climate, as-outcomes model. A null model is also labeled
respectively. These results show a high level of an unconstrained or intercept-only model, which
interrater agreement for Level 2 variables within includes no explanatory variables. A null model
the Level 2 units and systematic between-group allows intercepts to vary and assumes that slopes

Human Resource Management DOI: 10.1002/hrm

8 HUMAN RESOURCE MANAGEMENT

are fixed across higher-level units. It is used to esti- groups, and units are treated as interchangeable
mate whether the residual variance in the individ- throughout the article when referring to Level 2
ual-level model by Level 2 units is significantly units.
different from zero, and to confirm whether mul-
tilevel modeling is necessary. A random intercepts
Null Model
model is also labeled random coefficient regression In the example, the null model is tested using the
model. This model tests the relationship between following equation:
Level 1 predictor variable and the same-level out-
come variable. An intercepts-as-outcomes model Knowledge sharing behavior = γ00 + u0j + rij (8)
is also labeled a fixed slope model or means-as-out-
comes model. It is used to examine the direct effect where the fixed part of the model γ00 = grand
of the higher-level predictor variable on the lower mean knowledge sharing (i.e., the average score
level criterion variable. A slopes-as-outcomes of employee knowledge sharing across all firms),
model is also labeled random slopes model, which the random part of the model u0j  =  individual
examines whether cross-level interactions signifi- (within-group) variance in knowledge sharing,
cantly account for some variance in Level 1 slopes. and rij  =  between-group variance in knowledge
It is prudent to note that to test a conven- sharing. As there are no explanatory variables in
tional simple multilevel model containing Level the model, the variance is used in the calculation
1 direct effect, Level 2 direct effect and cross-level of the ratio of the between-group variance to the
interaction one normally tests these four hierar- total variance, termed ICC for the criterion vari-
chical models in sequence as shown earlier by able. In the illustrative example, the result of the
entering the control variables, the predictor vari- analysis using HLM 7 is statistically significant
ables, and interaction terms at different stages. [u0, χ2(29) = 18. 81, p < .001]. Hence, additional
However, some research does not need to test all variance exists by organizational level predictors,
four models. For example, one does not need to justifying the application of multilevel modeling.
test the random intercepts model if the researcher ICC is calculated using Equation 9:
is interested only in direct cross-level effects.
ICC = τ00 / (τ00 + σ2(rij)) (9)
Moreover, the four models do not always have
to be tested sequentially and can sometimes be ICC = .07941 / (.07941 + .22995) = .26, indi-
tested simultaneously. For instance, if the research cating 26% of variance in Level 1 criterion variable
interest is cross-level mediation, the random inter- resides at a higher level of analysis (i.e., between
cepts model and intercepts-as-outcome model firms). This result further suggests the imple-
should be estimated simultaneously (Zheng, mentation of a multilevel data-analytic strategy
Zyphur, & Preacher, 2009). A slopes-as-outcomes (Snijders & Bosker, 2012).
model is often tested together with a random
intercepts model. When tested together, the com- Intercepts-as-Outcomes Model
bined  model is labeled a random intercepts and An intercept term is generally interpreted as the
slopes model. expected average of the dependent variable given
In the illustrative research, our first research that all predictors in a model are equal to zero.
objective is to explore Level 2 direct effect. Hence, In the example, the variance in the average score
we first estimate a null model, which is fol- of employee knowledge sharing across organiza-
lowed by an intercepts-as-outcomes model. Due tions may be due to organizational factors such
to the fact that our second research objective is as high-commitment HRM. This model is run to
to explore cross-level mediation, our next step is estimate the following equation (Equation 10)
to simultaneously estimate a random intercepts that also includes the control variables to confirm
model and an intercepts-as-outcome model by Hypothesis 1:
adopting the MSEM approch. Finally, we estimaite
a slopes-as-outcomes model to test cross-level Knowledge sharing behavior = γ00 + γ01 (HRM)
moderation and an intercepts-as-outcomes model + γ10 (gender) + γ20 (age) + γ30 (position) + γ40
to test Level-1 moderation. For the purpose of the (education) + γ50 (tenure) + γ60 (POS) + u0 + rij (10)
illustrative analysis, it is assumed that all assump-
tions of multilevel modeling are adequately met. where γ00 = Level 2 intercept, γ01 = Level 2 slope
Specifically, there is no multicollinearity, normal (Hypothesis 1), γ10–γ60  =  mean (pooled) slopes,
distribution of error terms at every level of the u0 = residual intercept variance, and rij = Level 1
model, homogeneity of variance and indepen- residual variance.
dence of observations (see Hofmann et al., 2000, The analysis using HLM 7 shows high com-
for the details of assumptions of MLM). Firms, mitment HRM is significantly and positively

Human Resource Management DOI: 10.1002/hrm

 MULTILEVEL MODELING IN HRM RESEARCH 9

related to employee knowledge sharing behavior Nevertheless, this procedure has been criti-
(γ01 = .35, t = 4.1, p < .001) holding POS constant. cized by authors such as Zheng et al. (2009) and
Every SD increase in high-commitment HRM at Preacher et  al. (2010) for bias of indirect effects,
the organization-level results in .35 SD increase not taking into account measurement error, pro-
in employee knowledge sharing behavior. Hence, viding goodness-of-fit indices and the inability to
Hypothesis 1 receives support. model effects involving higher level dependent
variables. To address these limitations, Preacher
Random Intercepts Model et al. (2010) and Zheng et al. (2009) suggest using
MSEM, which has a range of advantages such as
As this model tests the relationship between Level allowing for separate estimation of between-group
1 predictor variable and the outcome variable, and within-group relationships, simultaneous
with a simple multilevel model without cross-level estimation first stage and second stage media-
mediation, it requires testing prior to testing the tion, and treating variables as latent. Preacher has
intercepts-as-outcomes model. In the illustrative provided syntax for performing MSEM with the
research, Hypothesis 2 proposes cross-level media- bootstrapping procedure to test cross-level media-
tion, which involves testing a random intercepts tion using Mplus software.2 According to Preacher
model, and an intercepts-as-outcomes model et  al. (2010), cross-level indirect effects using
(Equation 10). The random intercepts model is MSEM approach are more accurate than using tra-
tested using Equation 11: ditional MLM approach. The procedure for using
the MSEM approach to test Hypothesis 2 is dem-
Knowledge sharing behavior = γ00 + γ10 (gender)
+ γ20 (age) + γ30 (position) + γ40 (education) onstrated later.
+ γ50 (tenure) + γ60 (POS) + u0 + rij (11) A partial mediation model was fitted with
the direct relationship between high commit-
where γ00 = the mean of intercepts across groups, ment HRM and knowledge sharing,
γ10–γ60  =  means of the slopes across groups, as well as the indirect relationship
u0 = variance in intercepts, and rij = Level 1 resid- via POS and compared with a full It is important to note
ual variance. A simple multilevel model does not mediation model only with the that there are yet
involve the procedure for estimating indirect indirect relationship: “Any media-
effect, as demonstrated next. tion effect in a model at least one of no standard cutoffs
X, M, or Y is assessed at Level-2 must
occur strictly at the between-group for fit indices for
Test for Cross-Level Mediation
level” (Preacher et al., 2010, p. 210). multilevel modeling.
Hypothesis 2 in the illustrative research is cross- Hence, both models are assessed at
level mediation that requires testing for cross-level the between-group level.
indirect effect. Traditional MLM tests cross-level It is important to note that there are yet no
indirect effect adopting Baron and Kenny’s (1986) standard cutoffs for fit indices for multilevel mod-
approach developed for testing single-level media- eling. Hence, this article refers to the fit indices
tion. For example, Freedman and Schatzkin (1992) for single level SEM models. The partial media-
have suggested that the meso-mediation effect can tion model fits well into the data: χ2(9) = 17.52,
be calculated using the formula: “γ01 without controlling for M the comparative fit index (CFI) = 1.00, the Tucker
(mediator) – γ01 controlling for M” or “γ01 without controlling for M (mediator) Lewis (TLI) = 1.00, the root mean square error of
× γ01 controlling for M.” Recently, some authors, includ- approximation (RMSEA)  =  .05, the standardized
ing Kenny (Kenny, Kashy, & Bolger, 1998), have root mean square residual (SRMR) within = .01, and
recognized that the existence of mediation does the SRMR between  =  .04. This partial mediation
not require the existence of a significant direct fits significantly better [Δχ2(1) = 6.22, p < .05] than
relationship between the independent variable the full mediation model [χ2(8) = 23.74, CFI = .87,
and the dependent variable. TLI = .86, RMSEA = .06, SRMR within = .03, and
Accordingly, to test Hypothesis 2, calculating SRMR between  =  .18]. Hence, the partial media-
the cross-level indirect effect that can be obtained tion model is deemed as the final model.
using the coefficient between high commitment Checking coefficients at the within-organiza-
HRM and POS (γ01, in Equation 12), the coeffi- tion level, POS is significantly related to knowl-
cient between POS and knowledge sharing (γ60, in edge sharing behavior (γ  =  .22, p < .01). At the
Equation 11) is needed: between-organization level, high-commitment
HRM is significantly related to POS (γ  =  .77, p <
POS = γ00 + γ01 (HRM) + γ10 (gender) + .001) and knowledge sharing (γ = .40, p < .001). POS
γ20 (age) + γ30 (position) + γ40 (education) + is related to knowledge sharing (γ = .32, p < .001).
γ50 (tenure) + u0 + rij (12) The indirect effect at the between-organization

Human Resource Management DOI: 10.1002/hrm

10 HUMAN RESOURCE MANAGEMENT

level from high commitment via POS to knowl- Step 1—testing whether cross-level moderation
edge sharing is significant (.25, p < .01). The boot- exists; and Step 2—testing significance of slopes
strapping test shows 95 percent of the confidence of interactions.
interval of the indirect effect is [.39, .10] and does Step 1: Testing whether cross-level moderation
not contain zero. exists. Although Edwards and Lambert’s (2007)
The bootstrapping procedure (Preacher & procedure was initially developed for single level
Hayes, 2008) and the PRODCLIN procedure mediated moderation, it is equally applicable to
(MacKinnon, Fritz, Williams, & Lockwood, 2007) multilevel mediated moderation and has been
are currently regarded as effective in rigorously used previously by Liu, Liao, & Loi (2012) and Tse,
testing the significance of indirect effects. It is Dasborough, and Ashkanasy (2008). This proce-
argued that the Sobel test does not work well in dure was adopted by bringing moderations on the
small samples and is not recommended for use if direct effect of the independent variable and in
researchers have the assess to raw data (Preacher the second stage of mediation into the equation,
& Hayes, 2008). The bootstrapping approach does as demonstrated in Equation 15:
not impose distributional assumptions resulting
in greater control of Type I error rates and higher Knowledge sharing behavior = γ00 + γ01 (HRM)
power; it provides a better alternative with a small + γ02 (cooperative climate) + γ03 (HRM ×
sample size (MacKinnon et al., 2007; Preacher & cooperative climate) + γ10 (gender) + γ20
Hayes, 2008). (age) + γ30 (position) + γ40 (education) + γ50
Pseudo R2 for the whole model (tenure) + γ60 (POS) + γ61 (cooperative
and for the variance attributed to climate × POS) + u0 + u5 (POS) + rij (15)
Researchers must high commitment HRM can be cal-
culated. Pseudo R2 for the whole where g00  =  Level 2 intercept, g01, g02, g03, and
take the following model is estimated using Equation g61  =  Level 2 slopes, g10, g20, g30, g40, g50 and
13 and for variance by high- g60 = Level 2 intercepts, u0 = residual intercept vari-
two steps to test ance, u5 = residual slope variance and rij = Level 1
commitment HRM Equation 14,
interactions with respectively. residual variance. The results show that the inter-
Pseudo R2 for the model: action of high commitment HRM and cooperative
multilevel modeling: climate is significantly related to knowledge shar-
(σ2null – σ2intercepts-as-outcomes) / σ2null (13) ing: γ03 = .09, t = 4.1, p < .05, and the interaction
Step 1—testing of POS and cooperative climate is significantly
whether cross-level Pseudo R2 for HRM: related to knowledge sharing: γ61  =  .07, t  =  4.7,
p < .05. Notably, although γ02, the main effect of
moderation exists; (τ00 random intercept – τ00 intercepts-as-outcomes) / cooperative climate on knowledge sharing is non-
τ00 random intercept (14) significant, the significance of the interaction
and Step 2—testing terms still requires testing.
Results show that the pseudo R2 Following the procedure suggested by Bliese
significance of slopes for the model is (.23 – .18) / .23 = .22 (2002), Equation 16 was used to calculate a pseudo
of interactions. and for high-commitment HRM is R2 for the whole model (.23 – .13) / .23 = .43 and
(.014 – .012)/.014  =  .14, indicating Equation 17 for the moderator variable, coopera-
the whole mediation model and tive climate (.012 – .009) / .012 = .25.
high-commitment HRM explain 22 percent and Pseudo R2 for the model:
14 percent of variance in Level 1 criterion vari-
able, respectively. Combined, Hypothesis 2 is (σ2 null – σ2 intercepts-as-outcomes) / σ2 null (16)
confirmed.
Pseudo R2 for cooperative climate:
Slopes-as-Outcomes Model
(τ00 intercepts as outcomes – τ00 slopes as outcomes) /
Interactions can occur between two predictors at τ00 intercepts as outcomes (17)
Level 1, at Level 2, or cross-level. Strictly speak-
ing, only cross-level interaction is a slopes-as- Step 2: Testing significance of slopes of inter-
outcomes model. The interaction between two actions. Preacher, Curran and Bauer (2006) have
predictors within Level 2 is a means-as-outcomes created an online calculator for simple slope
model. In the illustrative research, Hypothesis 3a tests.3 Hypothesis 3a is the same as Preacher’s
is a slopes-as-outcomes model while Hypothesis Case 2 and Hypothesis 3b is the same as Preacher’s
3b is a means-as-outcomes model. Case 3. The values for fixed coefficients and the
Researchers must take the following two steps asymptotic covariance matrix of fixed regression
to test interactions with multilevel modeling: estimates are needed to calculate simple slopes.

Human Resource Management DOI: 10.1002/hrm

 MULTILEVEL MODELING IN HRM RESEARCH 11

The values of the moderator, as suggested by between-group effects. However, Mplus does not
Aiken and West (1991), may be one SD below tolerate errors in data or in syntax.
the mean and one SD above the mean when the The HLM program is commonly utilized
moderator is a continuous variable, or zero and for conducting two- or three-level analysis. The
one if it is dichotomous. advantages of HLM are the requirement of fewer
Using Preacher’s calculator, the following assumptions to be met than other programs
results were obtained: γ01 = .19, p < .01 when coop- (Raudenbush & Bryk, 2002), accommodating a
erative climate is low and γ01 = .51, p < .001 when lack of sphericity, missing data, small and/or dis-
cooperative climate is high for high commitment crepant group sample sizes, and heterogeneity
HRM; γ60 = .16, p < .05 when cooperative climate of variance across repeated measures (Osborne,
is low and γ60  =  .47, p < .001 when cooperative 2000). More importantly, for novice multilevel
climate is high for POS at the between-organiza- researchers, it is user-friendly soft-
tion level. At the individual level, γ60 = .13, p < .05 ware. The major disadvantage is
This study aims to
when cooperative climate is low and γ60  =  .37, p that it is unsuitable for performing
< .001 when cooperative climate is high for POS. MSEM. The procedure for conduct- serve as a guide for
Taken together, when cooperative climate is low, ing simple multilevel analysis with-
the relationships of high commitment HRM and out cross-level mediation using the a general readership
POS with knowledge sharing are weaker and when HLM software including program
of HRM researchers
high, the relationships are stronger. Therefore, setup and hypothesis tests is dem-
Hypotheses 3a and 3b are supported. onstrated in the Appendix. in understanding
The results of the moderated path analyses are
shown in Table I. Some statistical packages such as
Concluding Remarks multilevel modeling
SAS, SPSS, R, and HLM, have graphing features for Multilevel modeling is important
concepts and
plotting cross-level interactions. Online utility for to HRM research as it takes into
plotting HLM two- and three-way interactions are account the interactions of contex- procedures for
also available.4 tual factors of higher levels of units
and individual factors, so that pre- conducting multilevel
Statistical Packages dicting accuracy increases. Despite
research.
Several major statistical packages, such as HLM, recent growing interest in the use
R, SPSS, MPlus, LISREL, and MLWiN are available of multilevel analytical strategies,
for performing multilevel analysis. Each package HRM research lags behind other disciplines such
has advantages and disadvantages in terms of as education, marketing, and psychology, in this
operation and output production. MPlus is gain- regard. This study aims to serve as a guide for a
ing popularity as it provides desired output for general readership of HRM researchers in under-
ICC and effects at different stages of mediation, standing multilevel modeling concepts and proce-
including direct, first stage mediation, second dures for conducting multilevel research.
stage mediation, indirect and total effects. It is also This study discusses, from a theoretical per-
more suitable for performing MSEM, which pro- spective, the importance of multilevel modeling
duces model fit indices, and separates within- and in HRM research and the advantages of multilevel

TABLE I Results of the Moderated Path Analyses

HCHRM (X) POS (M) Knowledge Sharing (Y)
First Stage Second Direct Effect Indirect Effect Total Effect
Cooperative Climate PMX Stage PYM PYX PMX*PYM PYX+ PMX*PYM
Low Between- β = .77, β = .16, β = .19, β = .12, β = .31,
levels of level p < .001 p < .05 p < .01 p < .05 p < .01
cooperative
Within- β = .13,
climate
level p < .05
High Between- β = .77, β = .47, β = .51, β = .36, β = .87,
levels of level p < .001 p < .001 p < .001 p < .001 p < .001
cooperative
Within- β = .37,
climate
level p < .01
Note: HCHRM = high-commitment HRM; high levels of cooperative climate = 1 SD above the mean; low levels of cooperative climate = 1 SD
below the mean.

Human Resource Management DOI: 10.1002/hrm

12 HUMAN RESOURCE MANAGEMENT

modeling over classical analytical methods. It and assisting HRM researchers in, conducting
distills the literature on statistical logic underpin- multilevel HRM research.
ning multilevel modeling, variable level concep-
tualization, and data aggregation and analysis. Notes
An empirical HRM research example is offered 1. Mathieu et  al. (2012) created a computer program
to demonstrate the procedures for meeting pre- that allows researchers to estimate the power of
conditions for the implementation of multilevel cross-level interactions to determine the sample size
analytical strategies and hypothesis test. A brief of multilevel variables prior to data collection. This
assessment of statistical packages for performing program is available online at: http://mypage.iu.edu
multilevel analysis is also provided. /haguinis/~-crossless.html.
As HLM software is an invaluable, user-friendly 2. The software is available at http://www.quantpsy.org
computer program for novice users, the proce- /selig_preacher_mplus_syntax.htm.
dures for the HLM program setup and hypothesis
tests when conducting simple multilevel analysis 3. Available at http://www.quantpsy.org/interact/hlm2
without cross-level mediation are described in the .htm.
appendix. In summary, this article provides thor- 4. Such as from http://www.people.ku.edu/~preacher
ough, actionable knowledge aimed at promoting, /interact/shacham/index.hlm.

JIE SHEN is associate professor of HRM at the University of South Australia. Previously, he
held the positions of full professor of HRM at Curtin University and associate professor of
HRM at Monash University. He is an editorial board member of three international journals
and the guest editor of the special issue “Multilevel HRM Research” for the International
Journal of HRM. Since 2004, he has published 7 research books and 55 refereed journal
articles, including articles in the Journal of Management and Human Resource Management.

References Bock, G. W., & Kim, Y. G. (2002). Breaking the myths of

rewards: An exploratory study of attitudes about
Aiken, L. S., & West, S. G. (1991). Multiple regression: Testing knowledge sharing. Information Resources Management
and interpreting interactions. Newbury Park, CA: Sage. Journal, 15(2), 14–21.
Arnold, C. I. (1992). An introduction to hierarchical linear Castro, S. L. (2002). Data analytic methods for the analysis
models. Measurement and Evaluation in Counselling and of multilevel questions: A comparison of intraclass cor-
Development, 25, 58–90. relation coefficients, rwg(j), hierarchical linear modeling,
Aryee, S., Walumbwa, F. O., Seidu, E. Y. M., & Otaye, L. E. within- and between-analysis, and random group resam-
(2012). Impact of high performance work systems on pling. Leadership Quarterly, 13, 69–93.
individual and branch-level performance: Test of a multi- Chatman, J. A., & Flynn, F. J. (2001). The influence of demo-
level model of intermediate linkages. Journal of Applied graphic heterogeneity on the mergence and conse-
Psychology, 97, 287–300. quence of cooperative norms in work teams. Academy of
Baron, R. M., & Kenny, D. A. (1986).The moderator-mediator Management Journal, 44(5), 956–974.
variable distinction in social psychological research: Con- Collins, C. J., & Smith, K. G. (2006). Knowledge exchange
ceptual, strategic and statistical considerations. Journal and combination: The role of human resource practices
of Personality and Social Psychology, 51, 1173–1182. in the performance of high-technology firms. Academy of
Bliese, P. D. (2000). Within-group agreement, non-independ- Management Journal, 49, 544–560.
ence, and reliability: Implications for data aggregation Den Hartog, D. N., Boon, C., Verburg, R. M., & Croon, M. A.
and analyses. In K. J. Klein & S. W. J. Kozlowski (Eds.), (2013). HRM, communication, satisfaction, and perceived
Multilevel theory, research, and methods in organiza- performance: A cross-level test. Journal of Management,
tions: Foundations, extensions, and new directions 39, 1637–1665.
(pp. 349–381). San Francisco, CA: Jossey-Bass. Edwards, J. R., & Lambert, L. S. (2007). Methods for inte-
Bliese, P. D. (2002). Multilevel random coefficient modeling grating moderation and mediation: A general analytical
in organizational research. In F. Drasgow & N. Schmitt framework using moderated path analysis. Psychological
(Eds.), Measuring and analyzing behavior in organiza- Methods, 12, 1–22.
tions (pp. 401–445). San Francisco, CA: Jossey-Bass. Eisenberger, R., Huntington, R., Hutchison, S., & Sowa,
Bliese, P. D., & Hanges, P. J. (2004). Being both too liberal D. (1986). Perceived organizational support. Journal of
and too conservative: The perils of treating grouped Applied Psychology, 71(3), 500–507.
data as though they were independent. Organizational Freedman, L. S., & Schatzkin, A. (1992). Sample size for
Research Methods, 7, 400–417. studying intermediate endpoints within intervention

Human Resource Management DOI: 10.1002/hrm

 MULTILEVEL MODELING IN HRM RESEARCH 13

trials of observational studies. American Journal of MacKinnon, D. P., Fritz, M. S., Williams, J., & Lockwood, C. M.
Epidemiology, 136, 1148–1159. (2007). Distribution of the product confidence limits
Glick, W. H. (1985). Conceptualizing and measuring organi- for the indirect effect: Program PRODCLIN. Behavior
zational and psychological climate: Pitfalls in multilevel Research Methods, 39, 384–389.
research. Academy of Management Review, 10(2), Mathieu, J. E., & Chen, G. (2011). The etiology of the
601–616. multilevel paradigm in management research. Journal of
Hofmann, D., Griffin, M., & Gavin, M. (2000). The application Management, 37, 610–641.
of hierarchical linear modeling to organizational research. Mathieu, J. E., Aguinis, H., Culpepper, S. A., & Chen, G.
In K. Klein & S. Kozlowski (Eds.), Multilevel theory, (2012). Understanding and estimating the power to detect
research, and methods in organizations: Foundations, cross-level interaction effects in multilevel modeling.
extensions, and new directions (pp. 467–511). San Journal of Applied Psychology, 97, 951–966.
Francisco, CA: Jossey-Bass. Messersmith, J. G., Patel, P. C., & Lepak, D. (2011). Unlock-
Hofmann, D. A. (1997). An overview of the logic and ration- ing the black box: Exploring the link between high-
ale of hierarchical linear models. Journal of Manage- performance work systems and performance. Journal of
ment, 23, 723–744. Applied Psychology, 96(6), 1105–1118.
Hofmann, D. A., & Gavin, M. B. (1998). Centering decisions Osborne, J. W. (2000). Advantages of hierarchical lin-
in hierarchical linear models: Implications for research ear modeling. Practical Assessment, Research, and
in organizations. Journal of Management, 24(5), Evaluation, 7, 1–3.
623–641. Ostroff, C., & Bowen, D. E. (2000). Moving HR to a higher
James, L. R., Demaree, R. G., & Wolf, G. (1984). Estimating level: HR practices and organizational effectiveness. In
within-group inter-rater reliability with and without K. L. Klein & S. W. J. Kozlowski (Eds.), Multilevel theory,
response bias. Journal of Applied Psychology, 69, 86−98. research, and methods in organizations: Foundations,
Jensen J. M., Patel, P. C., & Messersmith, J. G. (2013). extensions, and new directions (pp. 211–266). San
Higher-performance work systems and job control: Francisco, CA: Jossey-Bass.
Consequences for anxiety, role overload, and turnover Podsakoff, P. M., MacKenzie, S. B., Lee, J. Y., & Podsakoff,
intentions. Journal of Management, 39, 1699–1724. N. P. (2003). Common method biases in behavioral
Kenny, D. A., Kashy, D. A., & Bolger, N. (1998). Data analy- research: A critical review of the literature and
sis in social psychology. In D. Gilbert, S. T. Fiske, & G. recommended remedies. Journal of Applied Psychology,
Lindzey (Eds.), Handbook of social psychology (4th ed., 88, 879–903.
Vol. 1, pp. 233–265). New York, NY: McGraw-Hill. Preacher, K. J., Curran, P. J., & Bauer, D. J. (2006). Compu-
Kreft, I. G. G., & de Leeuw, J. (1998). Introducing multilevel tational tools for probing interaction effects in multiple
modeling. Newbury Park, CA: Sage. linear regression, multilevel modeling, and latent curve
Kulik, C. (2012). Climbing the higher mountain: The chal- analysis. Journal of Educational and Behavioral Statis-
lenges of multilevel, multisource, and longitudinal tics, 31, 437–448.
research designs. Management and Organization Review, Preacher, K. J., & Hayes, A. F. (2008). Asymptotic and resa-
7(3), 447–460. mpling strategies for assessing and comparing indirect
LeBreton, J. M., & Senter, J. L. (2008). Answers to 20 ques- effects in multiple mediator models. Behavior Research
tions about interrater reliability and interrater agreement. Methods, 40, 879–891.
Organizational Research Methods, 11, 815–852. Preacher, K. J., Zyphur, M. J., & Zhang, Z. (2010). A general
Liao, H., & Chuang, A. (2004). A multilevel investigation of multilevel SEM framework for assessing multilevel medi-
factors influencing employee service performance and ation. Psychological Methods, 15, 209–233.
customer outcomes. Academy of Management Journal, Raudenbush, S. W., & Bryk, A. S. (2002). Hierarchical linear
47, 41–58. models: Applications and data analysis methods (2nd
Liao, H., Toya, K., Lepak, D. P., & Hong, Y. (2009). Do they see ed.). Newbury Park, CA: Sage.
eye to eye? Management and employee perspectives of Raudenbush, S. W., & Liu, X. (2000). Statistical power and
high-performance work systems and influence processes optimal design for multisite randomized trials. Psycho-
on service quality. Journal of Applied Psychology, 94, logical Methods, 5, 199–213.
371–391. Rousseau, D. M. (1985). Issues of level in organizational
Littell, R. C., Milliken, G. A., Stroup, W. W., & Wolfinger, R. D. research: Multi-level and cross-level perspectives.
(1996). SAS systems for mixed models. Cary, NC: SAS Research in Organizational Behavior, 7, 1-37.
Inst. Rowe, K. J., & Hill, P. W. (1998). Modeling educational effec-
Liu, D., Liao, H., & Loi, R. (2012). The dark side of leadership: tiveness in classrooms: The use of multi-level structural
A three-level investigation of the cascading effect of equations to model students’ progress. Educational
abusive supervision on employee creativity. Academy of Research and Evaluation: An International Journal on
Management Journal, 55(5), 1187–1212. Theory and Practice, 4(4), 307–347.
Longford, N. T. (1986). VARCL—Interactive software for vari- Snell, S. A., & Dean, J. W. Jr. (1992). Integrated manufactur-
ance component analysis: Applications for survey data. ing and human resource management: A human capital
Professional Statistician, 5, 28–32. perspective. Academy of Management Journal, 35,
Luke, D. A. (2004). Multilevel modeling (3rd ed.). Thousand 467–504.
Oaks, CA: Sage. Snijders, T. A. B. (1996). Stochastic actor-oriented models for
Maas, C. J. M., & Hox, J. J. (2006). Sufficient sample sizes for network change. Journal of Mathematical Sociology, 21,
multilevel modelling. Methodology, 1(3), 86–92. 149–172.

Human Resource Management DOI: 10.1002/hrm

14 HUMAN RESOURCE MANAGEMENT

Snijders, T. A. B., & Bosker, R. J. (2012). Multilevel analysis: Wei, Y. C., Han, T. S., & Hsu, I. C. (2010). High-performance HR
An introduction to basic and advanced multilevel practices and OCB: A cross-level investigation of a causal
modelling (2nd ed). Los Angeles, CA: Sage. path. International Journal of Human Resource Manage-
Sun, L. Y., Aryee, S., & Law, K. S. (2007). Higher- ment, 21(10), 1631–1648.
performance human resource practices, citizenship Whitener, E. M. (2001). Do “high commitment” human
behavior, and organizational performance: A relational resource practices affect employee commitment? A
perspective. Academy of Management Journal, 50, cross-level analysis using hierarchical liner modeling.
558–577. Journal of Management, 27, 515–535.
Takeuchi, R., Chen, G., & Lepak, D. P. (2009). Through the Willms, J. D., & Raudenbush, S. W. (1989). A longitudinal
looking glass of a social system: Cross-level effects of hierarchical linear model for estimating school effects
high-performance work systems on employees’ attitudes. and their stability. Journal of Educational Measurement,
Personnel Psychology, 62, 1–29. 26, 209–232.
Tse, H. M., Dasborough, M., & Ashkanasy, N. M. (2008). A Zheng, Z., Zyphur, M. J., & Preacher, K. J. (2009). Testing
multilevel analysis of team climate and interpersonal multilevel mediation using hierarchical linear mod-
exchange relationships at work. Leadership Quarterly, 19, els: Problems and solutions. Organizational Research
195–211. Methods, 12(4), 695–719.

Human Resource Management DOI: 10.1002/hrm

 MULTILEVEL MODELING IN HRM RESEARCH 15

APPENDIX The Procedure for Conducting Multilevel Analysis Using HLM Software

HLM Set-Up
After launching the HLM program, click File-Make new MDM file (if using an existing MDM file, click
Create a new model using an existing MDM file) -Stat package input-HLM2 (for three level models,
click HLM3). A dialogue box will open, in which researchers are required to:
1. Create a MDM file using the .mdm suffix and save by clicking Save mdm file.
2. Specify nesting of input data by choosing persons within groups.
3. Load Level-1 and Level-2 files and choose variables. Firstly, choose Organization as ID in both files
as it links two levels and secondly, choose Anything else if using HLM version 6 and SPSS/Windows
if using HLM version 7.
4. Provide missing data information by choosing Yes or No. If there is missing data at Level-1,
researchers need to specify whether they wish to delete missing data when creating a mdm file or
running analyses. Missing data are not allowed at Level-2 and Level-3.
It is now possible to create a mdm file. Researchers should click Make MDM and check the statistics
including variables, n, mean and SD by clicking Check Stats. If satisfied with the statistics, click Done.
After this process is complete, researchers can perform the required analyses.
Hypothesis Testing
To run the null model, click Level 1 and enter Level-1 dependent variable as Outcome variable and click
Run analysis. To check the results, go to File and click View Output. To test the intercepts-as-outcomes
model, click Level 1, enter Level-1 outcome variable as Outcome variable, and enter control variables
as uncentered and Level-1 independent variable as a group centered. Then, click Level 2, enter Level-2
control variables and the Level-2 independent variable as a grand centered.
To test the slopes-as-outcomes model, click the error term for Level-1 variable, after the error term
appears, click Level-2 variable. This will enter an interaction of Level-1 and Level-2 predictors. After all
variables are entered, click Run analysis. To check the results, go to File and click View Output.

Human Resource Management DOI: 10.1002/hrm