Cultural Effects on the Behavior of Auditors - A Multilevel Structural Equation Model

Cultural Effects on the Behavior of Auditors - A Multilevel

Structural Equation Model

Auke Meijer

March 4, 2010

Abstract

This article models the relationship between auditors’ professional behavior and national cul-tural dimensions included in House et al. (2004) across 49 countries. This study introduces the use of multilevel structural equation models in behavioral science. We establish that differences in auditors’ professional behavior can be explained by cross-national cultural dif-ferences. In particular, we show that institutional collectivism, in-group collectivism, power distance, assertiveness and performance orientation affect auditors’ professional behavior.

1

Introduction

This study examines the effect of culture on auditors’ professional behavior. This paper is the first study considering the impact of culture on behavior that combines a multilevel model and a structural equation model. Therefore, we introduce the use of multilevel structural equation modeling in behavioral science. Cross-national differences in auditors’ professional behavior are analyzed across 49 countries using data from an international accounting firm. Furthermore, we use cultural dimensions from House et al. (2004) since those dimensions form the newest set of dimensions that can be pragmatically applied in management science. This study contributes to a more complete understanding of the impact of culture on auditors’ professional behavior. Given the internalization and globalization of multinational firms, this study provides a deeper knowledge and understanding of how culture affects multinational audit firms.

and auditing firms in particular are affected by culture at the societal level. International auditing organizations face numerous cross-national differences affecting the organization and the audit process, as suggested by Perera (1989), Welton and Davis (1990), Beresford (1990), Wood (1996), Cooper et al. (1998), Gray and Needles (1999), Nobes and Parker (2000), Barret et al. (2005), Radebaugh et al. (2006) and Choi and Meek (2008). These firms face the challenge to provide assurance services at the highest possible level of audit quality throughout the world.

Prior research shows that national cultural differences may explain that auditors’ professional behavior differs between countries. However, the literature suffers from several shortcomings. Firstly, we do not know any study that performs factor analysis while taking into account binary or ordinal data structures. Instead, authors perform factor analysis as if all items would be contin-uous. This yields inconsistent estimates. Secondly, the use of multilevel models is limited. Many authors use basic statistical methods to investigate the relationship between behavior and culture and assume that there is independence between units from the same level of analysis. The result is that standard errors are underestimated, hence, the significance of variables is overestimated. Furthermore, ignoring the hierarchical structure of data yields inconsistent estimates. Some au-thors use multilevel modeling, however, those auau-thors use a linear model where a nonlinear model should be used or they do not establish that random effects are significant.

This paper investigates the relationship between differences in auditors’ professional behavior and cross-national cultural differences as being one of the factors causing these behavioral dif-ferences. Following Rabe-Hesketh et al. (2001), Rabe-Hesketh et al. (2004a) and Skrondal and Rabe-Hesketh (2004), we combine a multilevel model (e.g. Goldstein, 1986) and structural equa-tion model (e.g. J¨oreskog, 1973) in order to explain the determinants of auditors’ professional behavior. Multilevel regression models (multilevel random effects models or hierarchical models) are used when the data structure is hierarchical with units nested in clusters, which in turn may be nested in higher clusters, and so on (Rabe-Hesketh et al., 2004a). Structural equation models, consisting of a factor and structural model, are used when the variables of interest cannot be mea-sured perfectly. Instead, there are sets of items reflecting hypothetical constructs or factors. The structural model, modeling relations between factors and explanatory variables, is the substantive model of interest. This paper uses a multilevel structural equation model since the data structure is hierarchical with teams nested in countries and auditors’ professional behavior is measured by a set of measured items which reflect several constructs of auditors’ professional behavior. Random effects at the country level induce dependence among all teams in the same country.

accountability is easily deferred, so, auditors do not consider themselves to be responsible for some auditing tasks. Contrary to institutional collectivism, in-group collectivism positively affects audi-tors’ team work. We also establish that power distance, assertiveness and performance orientation affect auditors’ professional behavior.

Other studies that analyze auditors’ behavior across countries have relied on cultural dimen-sions developed by Hofstede (1980) and Hofstede and Bond (1988). However, House et al. (2004) discuss many problematic issues associated with those cultural dimensions: the data set is out-dated, individualism and collectivism are not similar dimensions, there is high correlation between power distance and individualism, and there is a lack of validity for the masculinity dimension. Hence, it is important to consider more updated and more methodologically sound measures of culture (Husted, 2000).

The rest of the paper is organized as follows. We first review the literature. The third section describes the data set, consisting of dependent and independent variables. Section 4 lays out the model. The results are discussed in section 5. Section 6 concludes and the appendices contain additional information.

2

Literature Review

Prior research shows that auditors’ professional behavior differs between countries and that those behavioral differences may be caused by national cultural differences. Firstly, prior research sug-gests that behavior in general is affected by, among other factors, societal and cultural systems (Parsons and Shils, 1951; Kluckhohn, 1962). Secondly, some studies focus on differences in profes-sional behavior across national borders in general (Hofstede, 2001; Adler, 2002; House et al., 2004). Thirdly, within the school of behavioral research in accounting, several studies show how auditors’ professional behavior is affected by local environmental and cultural and societal factors within a particular country (Ferris and Larcker, 1983; Soeters and Schreuder, 1988; Birnberg and Shields, 1989; Dillard and Ferris, 1989; Cooper et al., 1998; Donnelly et al., 2003; Almer et al., 2005; Barret et al., 2005). Other authors focus on specific aspects of auditors’ professional behavior in a cross-cultural context (Ferris et al., 1980; Hussein et al., 1986; Gray, 1988; Needles, 1989; Agacer and Doupnik, 1991; Pratt et al., 1993; Cohen et al., 1995; Tsui, 1996; Farrel and Cobbin, 2001; Patel et al., 2002; Chan et al., 2003; Smith and Hume, 2005; Parboteeah et al., 2005; Tsakumis, 2007; Doupnik, 2008; Ehmke et al., 2010; Bik, 2010). In conclusion, prior research has shown that auditors’ professional behavior is driven to a large extent by the auditor’s local environment and culture.

developed by House et al. (2004). Parboteeah et al. (2005) use 21 countries in their analysis while Bik (2010) uses 29 countries in his analysis. Other studies have been undertaken in a relatively small number of countries.

Auditors’ professional behavior cannot be measured perfectly. Instead, there are sets of items reflecting hypothetical constructs, i.e. particular aspects of auditors’ professional behavior. These constructs induce dependence among the items. Factor analysis (e.g. Tabachnik and Fidell, 2007; Bartholomew et al., 2008) is usually used in behavioral studies, it reduces many items to several unobservable behavioral aspects. After identifying auditors’ behavioral aspects, regression methods may be used to examine the relationship between auditors’ professional behavior and culture. It is necessary to perform a multilevel research since auditors’ professional behavior is measured at either the individual or team level, while culture is measured at the country level. Klein and Kozlowski (2000) discuss issues faced by researchers in conducting multilevel research. They first examine procedures used to justify aggregating data from lower level data to higher units of analysis. In fact, data should be homogeneous within countries and heterogeneous between countries1_{. Furthermore, Klein and Kozlowski describe procedures to test relationships between} variables in multilevel models. In particular, multilevel regression models are used when the data structure is hierarchical. However, the use of multilevel regression models is limited in studies considering the impact of culture on behavior. After aggregating data, basic statistical methods are used to investigate the relationship between behavior and culture (e.g. Patel et al., 2002; Chan et al., 2003; Smith and Hume, 2005; Doupnik, 2008; Ehmke et al., 2010; Bik, 2010). Those methods overestimate the significance of variables. Furthermore, the estimates are inconsistent since the nested structure of the data is ignored. Cullen et al. (2004), Hui et al. (2004), Fu et al. (2004) and Parboteeah et al. (2005) use hierarchical modeling. However, Cullen et al. (2004) and Hui et al. (2004) use a linear model where a nonlinear model should be used. Furthermore, Cullen et al. (2004), Hui et al. (2004) and Parboteeah et al. (2005) do not establish that random effects are significant. Finally, Cullen et al. (2004), Hui et al. (2004) and Fu et al. (2004) perform factor analysis ignoring the ordinal structure of the data while Parboteeah et al. (2005) take the average of several items instead of performing factor analysis. This yields inconsistent estimates. Structural equation models combine factor models and regression models. Structural equa-tion models were developed in psychometrics and are used in biostatistics and econometrics as well. Rabe-Hesketh et al. (2004a) are the first authors who combine multilevel and structural equation models without making unrealistic assumptions. Generalized linear latent and mixed models (GLLAMM) unify generalized linear mixed models and structural equation models. An-other approach to specify multilevel structural equations models is the two-stage approach. This approach specifies separate structural equation models for the within and between covariance matrices (Longford and Muth´en, 1992; Poon and Lee, 1992; Longford, 1993; Linda et al., 1993;

1_{We do not discuss these procedures further, yet the requirements for aggregating data are satisfied. Bik (2010)}

Muth´en, 1994; Lee and Shi, 2001; Everson and Millsap, 2004). However, there are some limitations to models specified via separate structural equation models (Rabe-Hesketh et al., 2004a).

3

Data

3.1

Questionnaires

The data come from questionnaires which were executed by an international accounting firm. As part of a yearly process performance improvement project, this firm submitted a questionnaire to a cross-section of their assurance practice throughout its network of affiliated international ac-counting organizations. Access was obtained to the results of the questionnaires covering financial statement audits performed on 2005 financial statements.

In total 1,939 questionnaires were submitted to 117 countries. For each country, there exists several engagement teams whose members jointly respond to 118 questions or items. Therefore, the data have a multilevel structure with 147,584 responses to the 118 items from 1,939 engage-ment teams nested in 117 countries. An audit engageengage-ment team is an accounting team that works on the audit for a particular client2_{. The 118 questions reflect the detailed steps of thirteen main} audit steps. Some main steps have three substeps and thus three questions, where other have twenty or more. Questions are answered yes or no based on review of the detailed audit steps in the audit engagement file. The answers to the questions result in the ultimate performance on each of the thirteen main audit steps, which’ scorings are summaries of questions. An example of a main audit step is to what extent the audit team initially shared their knowledge. This is the first main audit step. Questions corresponding with this audit step include whether or not the audit team initially shared their knowledge of fraud risk and whether or not the audit team held a kick-off meeting. Column one in table 1 describes all the main audit steps. Table 1 also con-tains the number of questions per audit step and the number of observations per audit step. Due to privacy considerations for the international accounting firm, descriptive statistics of questions cannot be included.

The sample of questionnaires submitted per country reflected the size and nature (e.g., in-dustry sectors, listed versus privately owned clients) of the audit practice in each country. The questionnaires were filled-out by the selected number of financial audit engagement teams based on the audit files. Each engagement team has the same probability of being selected. The central project team designed the yearly process performance improvement project and was responsible for the execution of the project. The results of a questionnaire were to be discussed with and cleared for accuracy with the audit engagement partner before submission of the final

question-2_{An audit engagement team consists of an engagement leader (a director or partner), who has the final}

T able 1: Summary of main audit steps Main audit steps: Num b er of ques- tions Num b er of ob-serv a-tions Num b er of

se-lected ques- tions

naire. The main objective of the process was to evaluate the performance of each country as a whole, not the performance of individual audit teams or audit partners. Consequently, the process was designed such that results could not be traced back to individual teams, clients or partners and the outcome of the process had no consequences for monetary rewards. Hence, audit teams and partners had no direct incentive to inflate data. For each questionnaire, we have some team specific information. This will be discussed in the next subsection.

Recall that the questionnaires were designed to evaluate the performance of each country as a whole, not to test for cultural differences. Therefore, the questionnaire contains questions which do not seem to be relevant for this research. Omitting irrelevant questions yields a data set con-taining 79 questions. Furthermore, some questions are only answered for subsamples so that those questions contain a low number of observations. We omitted those questions as well. Finally, some questionnaires contain missing data due to skip pattern. That is why we introduced ordinal items. These variables, containing several ordered categories, do not have missing data. For instance, we introduced four ordinal variables that represent the questions in audit step nine. Those items equal respectively zero, one or two if the audit engagement team validated no controls at all, if the team validated one control, or if the team validated two controls. In the end, we have a data set consisting of 41 items including seven ordinal items. We continue with these 41 items. Some of the answers, especially for small countries, seem not to be plausible since those answers seem to reflect the general accepted answers. We will take this into account by modeling unobserved country specific heterogeneity.

3.2

Explanatory variables

To explain differences in auditors’ professional behavior, we use both country specific and team specific variables. Since the questionnaire data come from an international accounting firm, where the organizational culture is more or less the same, we examine culture at the societal level (Soeters and Schreuder, 1988; Pratt et al., 1993)3_{. According to Hofstede et al. (1993), differences between} countries may be explained by culture at the societal level. Hofstede (1980) defines national culture as ‘the collective programming of the mind which distinguishes the members of one human group from another’. We use cultural dimensions from House et al. (2004) since these authors took prior (cultural) studies, e.g. Kluckhohn and Strodtbeck (1961), Hofstede (1980) and Schwartz (1992), a step further. Furthermore, there has been many critique on conceptual and methodological issues of Hofstede’s model (Parboteeah et al., 2005).

The 10-year research project ‘GLOBE’ (Global Leadership and Organizational Behavior Ef-fectiveness research program), also referred to as House et al. (2004) after the primary researcher, studied cross-cultural interactions based on responses of about 17,000 managers from 951

orga-3_{Choosing teams from several companies could lead to organizational cultural differences affecting teams’}

nizations functioning in 62 societies throughout the world. As a result of this research effort, House et al. present, amongst other, how 62 societies ‘score’ on nine major attributes of culture. The GLOBE dimensions form the newest set of dimensions that can be pragmatically applied in management science. Project GLOBE defines culture as follows:

Culture is defined as ‘shared motives, values, beliefs, identities, and interpretations or meanings of significant events that result from common experiences of members of collectives and are trans-mitted across age generations’ (House et al., 2004).

House et al. have identified nine cultural dimensions in their study, which are included in table 2. Project GLOBE measures the cultural dimensions split in modal practices and modal values at the societal level. Modal practices are measured by the responses to questionnaire items con-cerning common behaviors, institutional practices, proscriptions, and prescriptions. Modal values are expressed in response to questionnaire items concerning judgments of ’What should be . . .’. These questionnaire items are intended as a measure of the respondents’ values concerning the practices reported by the respondents. Cultural practices are considered to be more robust pre-dictors or explanatory factors of actual behavorial differences compared to cultural values. Since the questionnaires measure actual auditors’ behavior, this study will focus on cultural practices as variables reflecting societal culture.

House et al. (2004) ‘scored’ 62 countries on their cultural dimensions, five of which are not in-cluded in the questionnaire data. The other way around, 68 countries inin-cluded in the questionnaire data are not included in House et al. (2004). Four countries included in the questionnaire data are not included on a comparable basis in House et al., since, for those four countries, our country classification does not fit with the country classification of House. For example, observations for Germany were measured by House et al. for Germany East and Germany West. Since we would not have information that indicates if our observations for Germany are in either East or West Germany, and we cannot estimate this either, we could not use the cultural dimensions of House et al. for Germany. Similarly, we cannot use data for Switzerland, South Africa and China. That is why we decided to omit such observations from the dataset. In the end, we have a final set of 49 countries for which both questionnaire data and cultural dimensions are available and included in our analysis. This corresponds to 1,223 audit engagement teams. The reconciliation of the initial questionnaire data to the data included in analysis is summarized in table 3.

Cultural dimensions are country specific and may explain differences between audit engagement teams across countries. There might also be differences between audit engagement teams within the same country. To control for differences between audit engagement teams within the same country, we include the following team specific variables: the industry4 in which the audit client

4_{Four industry classes are distinguished: 1. Consumer and Industrial Products and Services (CIPS); 2. Telecom}

Table 2: House et al. (2004)’s cultural dimensions

Culture construct Definition

Power Distance The degree to which members of an organization or society

expect and agree that power should be stratified and concentrated at higher levels of an organization or government.

Uncertainty avoidance The extent to which members of an organization or society strive to avoid uncertainty by relying on established social norms, rituals, and bureaucratic practices. People in high uncertainty avoidance cultures actively seek to decrease the probability of unpredictable future events that could adversely affect the operation of an

organization or society and remedy the success of such adverse effects.

Humane Orientation The degree to which individuals in organizations or societies encourage and reward individuals for being fair, altruistic, friendly, generous, caring, and kind to others.

Institutional Collectivism The degree to which organizational and societal institutional (Collectivism I) practices encourage and reward collective distribution of

resources and collective action.

In-Group Collectivism The degree to which individuals express pride, loyalty, and (Collectivism II) cohesiveness in their organizations or families.

Assertiveness The degree to which individuals in organizations or societies

are assertive, confrontational, and aggressive in social relationships.

Gender Egalitarianism The degree to which an organization or a society minimizes gender role differences while promoting gender equality. Future Orientation The degree to which individuals in organizations or societies

engage in future-oriented behaviors such as planning, investing in the future, and delaying individual or collective gratification.

Table 3: Reconciliation of the initial questionnaire data to the data included in analysis

Number of countries Number of questionnaires

Questionnaire data 117 1939

Exclude:

countries that are not included in House et al. 68 716

Questionnaire data included 49 1,223

operates, whether or not the client is a public interest entity5_{, the time expressed in hours that the} engagement team put into the audit engagement, the time expressed in hours that the engagement leader put into the audit engagement, the time expressed in hours that the engagement manager put into the audit engagement, and whether or not the audit client is perceived to be of higher risk. We could not include descriptive statistics due to privacy considerations for the international accounting firm.

4

The Model

Recall that the data have a multilevel structure. The multilevel design is unbalanced. The analysis in section 3 led to a data set consisting of 41 items. Selecting countries for which House cultural dimensions are available, we end with 47, 757 responses to 41 items from 1,223 engagement teams in 49 countries. We follow Rabe-Hesketh et al. (2001), Rabe-Hesketh et al. (2004a) and Skrondal and Rabe-Hesketh (2004) in formulating the multilevel structural equation model. They formulate a rather general multilevel type of structural equation model that nests both a random effects model and a general factor model. The multilevel structural equation model is specified using a multilevel regression model. An advantage of using a multilevel regression model is that the multilevel design is allowed to be unbalanced (Rabe-Hesketh et al., 2004a). The framework of the multilevel structural equation model consists of two parts, which are discussed in the next subsections. The first part of the model is a response or factor model. The response model is estimated with confirmatory factor analysis. In confirmatory factor analysis, there exists prior knowledge about the factor structure. So, a hypothesis about the factor structure could be tested. The structural model, which is our main interest, is discussed in the second subsection. It describes the relationship between latent factors, the explanatory variables described in section 3 and other latent variables. The final subsection describes the method of estimation.

5_{A public interest entity is either a firm that is listed on the stock market or an organization with significant}

4.1

Response model

The response or factor model describes how the 41 items are reduced into factors. Recall that questions are answered yes or no based on review of the detailed audit steps in the audit engage-ment file. Furthermore, we introduced some ordinal items. Using indices i for the item (level 1), j for the engagement team (level 2) and c for the country (level 3), yijcis defined as team jc’s answer on item yijc. Item yijc is either a binary or an ordinal variable. Let yijc∗ denote an underlying unobserved continuous random variable for yijc. Suppose that yijc is a binary variable, so, yijc equals either zero or one. y∗_ijcis not observed, however, we do observe the binary or dichotomous variable yijc:

yijc = 1 if yijc∗ ≥ 0, = 0 if y_ijc∗ < 0,

with i = 1, . . . , 41, j = 1, . . . , nc and c = 1, . . . 49. For an ordinal observation yijc, yijc = s if κs−1< y∗ijc≤ κs, where s = 1, . . . , S, are the response categories and κs with κ0= −∞, κ1 = 0 and κS = ∞ are the thresholds to be estimated. A general single level response or factor model is

y_ijc∗ = x0_ijcβ + η(2)_1jcz_1ijc(2) 0λ(2)₁ + . . . + η(2)_M

2jcz (2) M2ijc 0_λ(2) M2+ ijc, (1) where η(2)_m

2jc denotes the m2th latent variable at level 2, m2 = 1, . . . , M2, β denotes the vector

of fixed effects or regression coefficients and xijc is a column vector of explanatory variables associated with the fixed effects. The first element of λ(2)m2 is set to 1 in order to identify the

model. It is assumed that ijc is logistically distributed6. The m2th latent variable η (2) m2jc is

multiplied by a linear combination z(2)_m

2ijc

0_λ(2)

m2 of a column vector of explanatory variables z

(2) m2ijc

and parameter vector λ(2)m2. Let vijc = x

0 ijcβ + PM2 m2=1η (2) m2jcz (2) m2ijc 0_λ(2)

m2 be the linear predictor:

Ey∗_ijc|xijc, z (2) ijc, η (2) jc  = vijc, where z (2) ijc = (z (2) 1ijc, . . . , z (2) M2ijc) 0 _{and η}(2) jc = (η (2) 1jc, . . . , η (2) M2jc) 0_{. The} conditional expectation of the response yijcgiven xijc, z

(2) ijc and η

(2)

jc is linked to the linear predictor vijc via a link function g (·):

ghEyijc|xijc, z (2) ijc, η (2) jc i = vijc,

where it is assumed that yijcis a binary observation. For an ordinal observation yijc, g [P r (yijc≤ s)] = κs− vijc. We use a logit link for the dichotomous and ordinal items yijc7.

6_{Alternatively, }

ijc is assumed to be standard normal distributed, yielding a probit model. It turns out that

results are similar as with the logit model.

7_{Start with y}∗

ijc = vijc+ ijc where vijc = x0ijcβ +

PM2 m2=1η (2) m2jcz (2) m2ijc 0_λ(2)

m2. It follows that, for a

bi-nary observation yijc, P r (yijc= 1) = P r (vijc+ ijc≥ 0) = P r (ijc≥ −vijc) = F (vijc) = e

vijc

1+evijc with

F (·) the cumulative density function of ijc. It is assumed that ijc is logistically distributed. Let g (·) be

a link function. In fact, g (·) is the inverse function of F (·). Using a logit link function, it follows that g (P r (yijc= 1)) = log h _{P r(y} ijc=1) 1−P r(yijc=1) i = vijc.

For an ordinal observation yijc, P r (yijc= s) = P r yijc∗ ≤ κs

− P r y∗

ijc≤ κs−1

= P r (vijc+ ijc≤ κs) −

P r (vijc+ ijc≤ κs−1) = P r (ijc≤ κs− vijc) − P r (ijc≤ κs−1− vijc) = F (κs− vijc) − F (κs−1− vijc) with

F (·) the cumulative density function of ijc. It is assumed that ijc is logistically distributed. Using a logit link

Both the random effects model and the factor model are nested in (1). Modeling random ef-fects, unobserved heterogeneity is taken into account. A random effect could be modeled as a random intercept or a random coefficient. Random intercepts represent heterogeneity between clusters in the overall response and random coefficients represent heterogeneity in the relation-ship between the response and explanatory variables. A random effects model is obtained if x0 ijcβ = x0jcβ, z (2) m2ijc = z (2) m2jc is a scalar and if λ (2)

m2 is a parameter set to one. η

(2)

m2jc denotes

the m2th random effect at the team level8. Moreover, z (2)

m2jc is an element of an explanatory

variable, which may be a scalar equal to one if a random intercept is modeled. This implies that η_m(2) 2jcz (2) m2jc 0_λ(2) m2 = η (2) m2jcz (2) m2jc 0 _{which is either η}(2)

m2jc indicating a team specific random intercept

or η(2)_m

2jcz

(2) m2jc

0 _{indicating a team specific random coefficient. x}

jc includes both team specific and country specific explanatory variables whereas z(2)_jc includes team specific explanatory variables only. xjccontains all variables included in z

(2)

jc . The factor model is obtained if x0ijcβ = βi, so that xijcis a dummy variable indicating a specific item, and if z

(2)

m2ijc is a column vector with elements

set to either zero or one. λ(2)m2is a parameter vector with elements 0 ≤ λ

(2)

im2 ≤ 1. This implies that

η_m(2) 2jcz (2) m2ijc 0_λ(2) m2 is either 0 or η (2) m2jcλ (2)

im2. Factors can be either team or country specific. In (1),

it is assumed that factors are team specific. Each item i loads on factor η_m(2)_2jcwith factor loading z_m(2)

2ijc

0_λ(2) m2 = λ

(2)

im2. Similarly, each factor consists of several items. In the rest of the section we

use a factor model in (1). Therefore, we write βi instead of x0ijcβ.

This research focuses on team specific factors. This implies that scores on aspects of auditors’ behavior may differ both between audit engagement teams and across countries. In order to model for country unobserved heterogeneity, country specific random effects may be included. Country specific random effects induce dependence among all teams in a particular country and could be modeled directly in a multilevel response model:

y∗_ijc = βi+ M2 X m2=1 η_m(2) 2jcz (2)0 m2ijcλ (2) m2+ M3 X m3=1 η_m(3) 3cz (3)0 m3icλ (3) m3+ ijc, = βi+ 3 X l=2 Ml X ml=1 η(l)_m ljcz (l)0 mlijcλ (l) ml+ ijc, (2)

where l = 2, 3 indicates the level, η(l)_m

ljcdenotes the mlth latent variable at level l, ml= 1, . . . , Ml

and it is assumed that there are M2team specific factors and M3country specific random effects9. The random effects are assumed to be mutually independent and independent of the team spe-cific factors. Instead of modeling random effects directly, random effects could be modeled in a

8_{It is more reasonable to assume that random effects are country specific. Country specific random effects are}

included in (2) or in a structural model.

9_{Instead of country specific random effects, country specific factors could be modeled, so that a multilevel factor}

structural model. This will be discussed in the next subsection. Stacking items i yields: y_jc∗ = β + 3 X l=2 Ml X ml=1 η(l)_m ljcZ (l) mljcλ (l) ml+ jc, (3)

where y∗_jc is a column vector with 41 elements y∗_ijc, Z_m(l)

ljc is a matrix of explanatory variables

containing the column vectors z_m(l)

lijc and jc is a column vector with 41 elements ijc.

4.2

Structural model

The M2 latent factors are defined via the response model in (1). Country specific random effects are introduced in (2). These country specific random effects could also be modeled in the structural model. The structural model allows latent variables to be regressed on explanatory variables and on other latent variables. By regressing the latent factors on other latent variables, country specific random effects are introduced.

Recall that we use two types of explanatory variables: country specific cultural dimensions and team specific variables. Cultural dimensions may explain differences between countries. To control for differences between engagement teams within the same country, team specific variables are included. Let wjcbe a column vector containing team specific explanatory variables including the scalar one. Let ujc be a column vector of both country specific cultural dimensions and team specific explanatory variables. Therefore, ujc contains all variables included in wjc. In fact, ujc= u0c, w0jc

0

with ucthe column vector of cultural dimensions. To explain auditors’ professional behavior we will estimate the following structural model:

η(2)_m 2jc = u 0 jcγ + w0jcη (3) c + ζ (2) m2jc, = u0_jcγ + M3 X m3=1 η(3)_m3cwm3jc+ ζ (2) m2jc, (4) η_m(3) 3c = ζ (3) m3c (5)

where γ denotes the vector of fixed effects or regression coefficients, η(3)m3cdenotes the m3th country

specific random effect and ζ_m(2)

2jcdenotes the vector of team specific disturbances. Equations (1), (4)

and (5) form a multilevel structural equation model with team specific factors regressed on country specific random effects10_{. It is assumed that ζ}(2)

m2jc is assumed to be normally and independently

distributed with mean 0 and variance σ2 m2: ζ

(2)

m2jc ∼ N ID(0, σ

2

m2). wm3jc is an element of wjc

and λ(3)m3 is set to one. Since wjc includes a scalar, w

0 jcη

(3)

m3c consists of a country specific random

intercept and country specific random coefficients. Equation (5) indicates that η(3)m3c equals the

disturbance ζm(3)3c. It is assumed that ζ

(3) c ∼ N (0, Σc), with Σc=        σ2 1 0 . . . 0 0 σ22 . . . 0 .. . ... . .. ... 0 0 . . . σ_M2₃        . If we

10_{As discussed above, country specific random effects could also be modeled directly in the response model by using}

(2) with z_m(3)

3ic= wm3jcz

(2)

allow for correlations between the random effects, identification problems arise. It is also assumed that ζm(3)3c is orthogonal to ζ

(2) m2jc.

Model (4) incorporates cluster dependence through two ways. Firstly, regression parameters ηm(3)3c vary across clusters. Secondly, the total residual in (4),

PM3 m3=1ζ (3) m3cwm3jc+ ζ (2) m2jc, is

cor-related within a cluster. In fact, the model allows for heteroscedasticity. Cluster dependence induces the team specific common factor η_mjc(2) to vary between countries. We therefore modeled for unobserved heterogeneity at the country level in a way similar to that in the MIMIC (Multiple-Indicator Multiple-Cause) model (J¨oreskog and Goldberger, 1975). The inclusion of the random intercept makes the model a variance components factor model, which is a generalization of the item response model.

The model is illustrated in a path diagram in figure 1. Latent variables are represented by cir-cles and observed variables are represented by rectangles. The path diagram contains two frames representing the levels. The outer frame, indicated by ‘country c’, contains variables varying be-tween countries. Variables varying bebe-tween teams are located in the inner frame, which is labeled with ‘team jc’. Arrows connecting observed responses yijc and the latent factor η

(2)

m2jcrepresent a

nonlinear relation since we use a logit link function in the response model. Arrows connecting the latent factor η(2)_m2jc and observed variables ujc represent linear relations. The random effects are represented by ηc(3)= (η

(3) 1c , . . . , η

(3) M3c)

0_{. A short arrow pointing at the observed response represents} item level binomial variability.

Substituting the structural model into the factor model in (1), we obtain the reduced form:

yijc∗ = βi+ M2 X m2=1 η_m(2)_2jcz(2) 0 m2ijcλ (2) m2+ ijc, = βi+ M2 X m2=1 " v0jcγ + M3 X m3=1 η(3)m3cwm3jc+ ζ (2) m2jc # z(2) 0 m2ijcλ (2) m2+ ijc, = βi+ M2 X m2=1 " v0_jcγz(2)_m0 2ijcλ (2) m2+ M3 X m3=1 ζ_m(3) 3cwm3jcz (2)0 m2ijcλ (2) m2+ ζ (2) m2jcz (2)0 m2ijcλ (2) m2 # + ijc. (6) The disturbances ζ_m(2) 2jc and ζ (3)

m3c are the only latent variables in the reduced form. Suppose that

item i loads on the m2th factor only. Then z (2)

m2ijc is a column vector with element i equal to one

and zero elsewhere. This implies that y_ijc∗ = βi+ λ (2) im2η (2) m2jc+ ijc, = βi+ v0jcγλ (2) im2+ M3 X m3=1 ζ_m(3)_3cwm3jcλ (2) im2+ ζ (2) jc λ (2) im2+ ijc, with λ(2)_im

country c team jc y2jc yIjc y1jc wjc wjc uc η_m(2) 2jc η(3)c ζ_m(2) 2jc λ2m2 λIm2 1 γ γ

Figure 1: The multilevel structural equation model

4.3

Estimation

The model discussed above can be estimated using maximum likelihood estimation. The likelihood of the observed data is the likelihood marginal to all latent variables yet conditional on explanatory variables. Let θ be the vector of all fixed parameters. So, θ includes the regression coefficients β and γ, the factor loadings λ(2)m2, and the elements of the covariance matrices (σ

2 m2, σ

2

1, . . . , σ2M3).

The number of free parameters in θ will be reduced if restrictions are imposed. Let y(l) be the response vector at level l and let X(l) be the matrix of explanatory variables for all level l units belonging to a particular unit at level l. Moreover, let y and X be respectively the response vector and matrix of explanatory variables for all levels. To perform maximum likelihood estimation, the total likelihood must be obtained. The total likelihood is the product of the contributions of the level three clusters:

l(θ; y, X) = 49 Y

c=1

f(3) y(3)|X(3); θ
. (7)

The contributions do not depend on any latent variable. The total likelihood is constructed recursively. Consider level 1. (6) defines the conditional distribution of the responses given the latent and explanatory variables. The conditional distribution of the responses given the latent and explanatory variables is denoted by f(1)_y

(1)|X(1), ζ (2) jc , ζ (3) c ; θ 

. We observe dichotomous and ordinal responses. For dichotomous responses, f(1)y(1)|X(1), ζ

(2) jc , ζ

(3) c ; θ



equals the probability that yijc equals one11: P r (yijc= 1) = P r (vijc+ ijc≥ 0) = F (vijc) = e

vijc

1+evijc. Hence, the conditional distribution of the responses given the latent and explanatory variables equals the conditional probability that team jc chooses ‘yes’ on item yijc. This is similar with Revelt and Train (1998). Revelt and Train denote the conditional probability that team jc chooses ‘yes’ on item yijc by Lijc(θ). For ordinal responses, f(1)

 y(1)|X(1), ζ (2) jc , ζ (3) c ; θ 

equals the probability that yijc equals s: P r (yijc= s) = P r y∗ijc≤ κs
− P r yijc∗ ≤ κs−1

= P r (vijc+ ijc≤ κs) −

11_{f (y}

P r (vijc+ ijc≤ κs−1) = F (κs− vijc) − F (κs−1− vijc). We are not only interested in team jc’s choice on item yijc, yet we are interested in all choices from a particular team. In Revelt and Train (1998), the conditional probability of team jc’s observed sequence of choices on all items would be the product of the standard logits: Q41

i=1Lijc(θ). The level two conditional distribution of the responses given the explanatory variables and ζc(3) is obtained by integrating the level two disturbances ζ_jc(2) out. Hence,

f(2)y(2)|X(2), ζc(3); θ  = Z ∞ −∞ h(2)ζ_jc(2); θ 41 Y i=1 f(1)y(1)|X(1), ζ (2) jc , ζ (3) c ; θ  dζ_jc(2), (8) where h(2)_ζ(2) jc 

is the multivariate normal density of latent variable ζ_jc(2). The term h(2)_ζ(2) jc ; θ  Q41 i=1f (1)_y (1)|X(1), ζ (2) jc , ζ (3) c ; θ 

is the product of the prior density of ζ_jc(2) and the probability density function of the responses given ζ_jc(2)and ζc(3). In Revelt and Train (1998), the conditional probability of country c’s observed sequence of choices on all teams and items would beQnc

j=1 Q41

i=1Lijc(θ). The level three distribu-tion of the responses marginal to all latent variables yet condidistribu-tional on the explanatory variables is obtained by integrating the level three disturbances ζc(3) out. Hence,

f(3) y(3)|X(3); θ
= Z ∞ −∞ h(3)ζ_c(3); θ nc Y j=1 f(2)y(2)|X(2), ζc(3); θ  dζ_c(3). (9)

The integrals in (8) and (9) do not have closed forms, so, the density function cannot be calculated exactly. Instead, we approximate the integrals. Let v(l) be Ml independent standard normally distributed latent variables at level l with ζ(l) = Clv(l), where Cl is the Cholesky decomposition of Σl. Using Cartesian product quadrature,

f(2)y(3)|X(3), ζc(3); θ  = Z ∞ −∞ h(2)ζ_jc(2); θ 41 Y i=1 f(1)y(1)|X(1), ζ (2) jc , ζ (3) c ; θ  dζ_jc(2), = Z ∞ −∞ φ(v(2)_M 2jc) · · · Z ∞ −∞ φ(v(2)_1jc) 41 Y i=1 f(1)y(1)|X(1), v (2) jc, v (3) c ; θ  dv_1jc(2)· · · dv_M(2)_2jc, ≈ R(2) M2 X r_M2=1 πr_M2· · · R(2)₁ X r1=1 πr1 nc Y j=1 f(1) y(1)|X(1), αr1, . . . , αrM2; θ
,

where φ(·) is the standard normal density, πr_m2 denotes the quadrature weight, αr_m2 denotes the quadrature location, rm2 = 1, 2, . . . , Rm2 with Rm2 the number of repetitions. The multivariate

integral over latent variables ζ_jc(2) is evaluated by integrating over M2 independent standard nor-mally distributed latent variables. The simulation is not over just R repetitions as in Revelt and Train (1998), yet for each latent variable m2 = 1, . . . , M2 we use Rm2 repetitions. In total, the

integrand is evaluated atQM2

m2=1 quadrature points. The quadrature points αrm2 could be

posterior density of latent variables is treated as the weight function. Similarly as above, (9) can be approximated: f(3) y(3)|X(3); θ
= Z ∞ −∞ h(3)ζ_c(3); θ nc Y j=1 f(2)y(2)|X(2), ζc(3); θ  dζ_c(3), = Z ∞ −∞ φ(v_M(3) 3c) · · · Z ∞ −∞ φ(v_1c(3)) nc Y j=1 f(2)y(2)|X(2), v(3)c ; θ  dv(3)_1c · · · dv(3)_M 3c, ≈ R(3) M3 X r_M3=1 πr_M3· · · R(3)₁ X r1=1 πr1 nc Y j=1 f(2) y(2)|X(2), αr1, . . . , αrM3; θ
,

where rm3 = 1, 2, . . . , Rm3 with Rm3 the number of repetitions for latent variable m3= 1, . . . , M3.

Using the approximations for (8) and (9) and substituting the approximation for (9) into (7), the simulated total likelihood is obtained:

ls(θ; y, X) = 49 Y c=1 R(3) M3 X r_M3=1 πr_M3· · · R(3)₁ X r1=1 πr1 nc Y j=1 f(2) y(2)|X(2), αr1, . . . , αrM3; θ
.

Similarly, the simulated log-likelihood function is

lls(θ) = 49 X c=1 log    R(3) M3 X r_M3=1 πr_M3· · · R(3)₁ X r1=1 πr1 nc Y j=1 f(2) y(2)|X(2), αr1, . . . , αrM3; θ
  .

Maximizing the simulated log-likelihood yields estimated parameters θ. The gllamm program uses a Newton-Raphson algorithm in maximizing the log-likelihood.

The results of the estimation method discussed above are discussed in section 5. We also performed a two-step procedure in order to check the results. Since this procedure is only used to check the results obtained with gllamm we do not report these results. The two-step procedure starts with estimating the factor model. If factor scores are obtained, the structural model can be estimated. More details can be found in appendix A-1. The two-step method does not model the factor and structural model jointly. Furthermore, standard errors are underestimated. This implies that the statistical significance of explanatory variables is overestimated. In fact, some cultural dimensions that are not significant using gllamm are significant using the two-step method. Finally, Skrondal and Laake (2001) show that the two-step procedure yields inconsistent estimates. Therefore, the use of gllamm is preferred. Noteworthy is that we used the results of exploratory factor analysis as input for estimating the response model of the multilevel structural equation model.

5

Results

1. Engagement leader led discussion and knowledge sharing in the audit engagement team. 2. Documentation of knowledge about the client and the client’s industry.

3. Documentation of audit work.

4. Compliance by the engagement leader with the required audit steps. 5. Compliance by the engagement team with the required steps on fraud risk. 6. Assessing audit risks and designing the audit plan.

7. Assessment of design effectiveness of internal controls. 8. Testing of operating effectiveness of internal controls.

The numbers corresponding with these factors are used in table 4. The lower part of the ta-ble contains factor loadings and estimates of the team specific variance σ2m2 and country specific

variances σ2_m3. Country specific variances are variances corresponding to random intercepts and random coefficients. Since gllamm is computationally expensive, we did not succeed in estimating random coefficients for all factors. Therefore, we decided to include for each factor only a random intercept. Obviously, this still models unobserved country heterogeneity. Furthermore, there is no other behavioral study modeling random effects in the way discussed in this paper. The items contributing to the factors are described in appendix A-212_{. Some items do not load on any factor} and are omitted in the analysis. The upper part of table 4 is of main interest, it describes the re-lationship between particular aspects of auditors’ professional behavior and explanatory variables. We are not interested in all regression coefficients corresponding to team specific explanatory vari-ables. Hence, we include only the coefficient of the time that the engagement team put into the audit engagement and the coefficient of the dummy variable indicating whether or not the audit client is perceived to be of higher risk.

Table 4 shows that institutional collectivism, in-group-collectivism, power distance, assertiveness and performance orientation are the cultural dimensions that affect auditors’ professional behav-ior. Furthermore, some factors are affected by the time that the engagement team put into the audit. Finally, whether or not a client is perceived to be of higher risk seems to affect auditing. Some auditors’ behavioral aspects are more related to culture than others. Engagement leader led discussion and knowledge sharing in the audit engagement team, the documentation of audit work, and the assessment of design effectiveness of internal controls seem to be the auditors’ behaviors most sensitive to cultural practices. Compliance is hardly driven by culture.

12_{We could not estimate the multilevel structural equation model for factor two. Therefore, we use an ordinal}

Several aspects of auditors’ professional behavior are driven by institutional collectivism. In institutional collectivistic cultures, accountability is easily deferred, hence, individuals may not consider themselves to be responsible for team work that belong to several auditors. In particular, team members do not consider themselves to be responsible for the distribution of knowledge, for documentation, and for identifying and addressing potential risks in the audit. This limits the extent of knowledge sharing in the audit engagement team, the extent of documentation in the audit engagement team, and the identification of potential risks in the audit and the response to those risks. Hence, institutional collectivism negatively affects the performance on three factors: engagement leader led discussion and knowledge sharing in the audit engagement team, documen-tation of audit work, and compliance by the engagement team with the required steps on fraud risk. These effects are enforced by other aspects of institutional collectivism. People from institu-tional collectivistic cultures emphasize maintaining harmony and respecting authority. Therefore, engagement leader led discussion and knowledge sharing are low in institutional collectivistic cul-tures. Moreover, people from institutional collectivistic cultures tend to be relatively less skeptical than people from individualistic countries. Auditors from individualistic countries tend to express their opinions and views and prefer direct and solution-oriented resolution tactics while people from collectivistic countries apply avoidant and compromising resolution tactics. So, auditors from individualistic cultures tend to be more critical in examining fraud risk than auditors from institutional collectivistic cultures. This implies that the engagement team’s compliance with the required steps on fraud risk is low in institutional collectivistic cultures.

In-group collectivism positively affects the audit team’s documentation of knowledge about the client and the client’s industry, the team’s assessment of audit risks and the team’s design of the audit plan, and the team’s testing of operating effectiveness of internal controls. People from high in-group collectivism cultures have high level of trust in people from their own group. However, there is fairly low level of trust in those who do not belong to the same group. This implies that auditors have high level of trust in their team members but not in people that do not belong to the audit engagement team. So, teams members tend to work intensively together and audit teams perform high in team work. Indeed, audit teams perform high in the documentation of knowledge about the client and the client’s industry, in the assessment of audit risks, and in testing of operating effectiveness of internal controls.

high power distance cultures would have less room for independent thought and action. Further-more, it can be dangerous to question authority and to express disagreement. So, auditors in high power distance cultures tend to be less critical to the client’s internal controls.

In addition to institutional collectivism, assertiveness negatively affects the audit engagement team’s documentation of audit work. People in assertive cultures behave more opportunistically and would be more willing to accept a risk than people in low assertive cultures (House et al., 2004). So, auditors in assertive cultures tend to be less inclined to carefully address potential risks.

The audit engagement team’s assessment of design effectiveness of internal controls is not only driven by power distance, it is also negatively affected by risk and performance orientation. These negative effects are in contrast to what is expected. It may indicate that the standard procedure to assess the client’s internal controls is less appropriate to use if the client is perceived to be of higher risk. Moreover, the standard procedure may not be seen as the highest standard, so that performance oriented auditors tend to use another procedure. Risk has also a negative effect on engagement leader’s compliance with the required audit steps. If auditors are faced with risky clients, the engagement leader tend to follow the required audit steps less strictly. This is in contrast to what is expected since it is expected that additional risk induces auditors to work more carefully. Engagement leader’s compliance with the required audit steps is not affected by culture. A Wald test tests the null hypothesis that all cultural dimensions have a regression coefficient equal to zero. We find a Wald statistic of 7.19 with a corresponding p-value of 0.6177. This implies that the null hypothesis cannot be rejected. Hence, engagement leader’s compliance is not driven by culture. Finally, if the client is perceived to be of higher risk, auditors perform high on the documentation of audit work. This is as expected since auditors would work more carefully when they are faced with risky clients.

Engagement leader led discussion and knowledge sharing in the audit engagement team de-pend, in addition to the negative effects induced by institutional collectivism and power distance, positively on total hours. Furthermore, documentation of knowledge about the client and the client’s industry and documentation of audit work seem to be positively affected by total hours. This seems obvious since the audit engagement team performs better if they put high enough effort in the audit engagement file.

of the significance of cultural dimensions.

6

Conclusion

This paper examines the relationship between differences in auditors’ professional behavior and cross-national cultural differences as being one of the factors causing these behavioral differences. This study introduces the use of multilevel structural equation models in behavioral science. We go beyond traditional analysis and combine a multilevel model and a structural equation model. This model yields more precise and robust results than would be obtained using other models. Other methods yield inconsistent estimates. Furthermore, the significance of cultural dimensions is overestimated. We use data from an international accounting firm. Cross-national differences in auditors’ professional behavior are analyzed using 47,757 responses across 49 countries. We use cultural dimensions from House et al. (2004) rather than the outdated cultural dimensions from Hofstede (1980) and Hofstede and Bond (1988). We establish that culture affects auditors’ professional behavior. In particular, we show that several of the measured aspects of auditors’ professional behavior are negatively related to institutional collectivism or positively related to in-group collectivism. In institutional collectivistic cultures, accountability is easily deferred, so, individuals do not consider themselves to be responsible for documentation, for the distribution of knowledge, and for the identification of potential risks. Next to institutional and in-group collectivism, we establish that power distance, assertiveness and performance orientation affect auditors’ professional behavior. Some auditors’ behavioral aspects are more related to culture than others. Engagement leader led discussion and knowledge sharing in the audit engagement team, documentation of audit work, and the assessment of design effectiveness of internal controls seem to be the auditors’ behavioral aspects most sensitive to cultural practices. Compliance seems to be hardly driven by culture.

Table 4: Results multilevel structural equation analysis

1. Discussion and 2. Documentation

knowledge sharing of knowledgea

Parameter Est (SE) Est (SE)

Fixed part: γ1 (power distance) - 2.9689∗ (1.6045) - 1.0976 (0.6707) γ2 (uncertainty avoidance) 1.1436 (1.2756) - 0.0800 (0.5328) γ3 (Humane orientation) - 0.4514 (1.1528) - 0.3685 (0.4410) γ4 (Institutional Collectivism) - 2.2613∗ (1.2490) - 0.7198 (0.5482) γ5 (In-group Collectivism) 1.7505 (0.9994) 0.9364∗∗∗ (0.3554) γ6 (Assertiveness) - 0.8667 (1.6349) - 0.2049 (0.7327) γ7 (Gender Egalitarianism) - 1.2738 (1.2237) - 0.2623 (0.5634) γ8 (Future Orientation) - 0.1580 (1.7285) 0.1111 (0.631) γ9 (Performance Orientation) - 1.3861 (1.6523) - 0.8921 (0.5839)

γ10(Hours total per 1,000) 0.1916∗∗∗ (0.0595) 0.1232∗∗ (0.0551)

γ11(Higher risk) 0.2945 (0.3597) 0.1640 (0.2583)

Random part: team level Factor Loadings: Item 1: 1 Item 2: 1.2425 Item 3: 0.5459 Item 4: 0.3589 Item 5: 0.4058 Item 6: 0.2725

Common factor variance: 10.4541∗∗∗ (2.9194)

Random part: country level

Random intercept variance: 5.7760∗∗∗ (2.1994) 0.3620∗ (0.1911)

Log-likelihood: - 2,600.8843 - 554.1561

* Significant at 10%; ** Significant at 5%; *** Significant at 1%.

a_{There are no estimated factor loadings since we use an ordinal variable for factor two. Furthermore,}

3. Documentation 4. Engagement

of audit work leader’s compliance

Parameter Est (SE) Est (SE)

Fixed part: γ1 (power distance) 0.2274 (0.7521) - 1.6989 (1.2737) γ2 (uncertainty avoidance) - 0.8446 (0.6237) - 1.6246 (1.0735) γ3 (Humane orientation) - 0.3687 (0.5120) - 0.4952 (0.8308) γ4 (Institutional Collectivism) - 1.4770∗∗ (0.6185) 0.9773 (1.0761) γ5 (In-group Collectivism) - 0.3818 (0.4301) - 0.3936 (0.7057) γ6 (Assertiveness) - 1.6978∗∗ (0.8167) - 0.3075 (1.2452) γ7 (Gender Egalitarianism) 0.1772 (0.5801) - 0.8338 (0.9684) γ8 (Future Orientation) 1.1040 (0.7665) 0.2905 (1.2258) γ9 (Performance Orientation) - 0.0067 (0.7332) - 0.8687 (1.1990)

γ10 (Hours total per 1,000) 0.1129∗∗∗ (0.0378) 0.1802∗∗∗ (0.0649)

γ11 (Higher risk) 0.3928∗ (0.2070) - 0.6186∗∗ (0.2777)

Random part: team level Factor Loadings:

Item 1: 1 1

Item 2: 1.0421 0.3086

Item 3: 1.1092 0.9608

Item 4: 0.7827

Common factor variance: 2.0165∗∗∗ (0.6151) 4.5199∗∗∗ (1.2987)

Random part: country level

Random intercept variance: 1.2589∗∗ (0.4989) 3.3881∗∗ (1.3240)

Log-likelihood: - 1,737.3189 - 1,669.6615

5. Engagement 6. Assessing risk and

team’s compliance designing audit plan

Parameter Est (SE) Est (SE)

Fixed part: γ1 (power distance) - 1.2704 (0.9924) - 0.5836 (0.5777) γ2 (uncertainty avoidance) - 0.4684 (0.8211) - 0.1725 (0.4806) γ3 (Humane orientation) 0.1718 (0.6735) 0.2600 (0.4521) γ4 (Institutional Collectivism) - 1.7446∗∗ (0.8535) - 0.4183 (0.6398) γ5 (In-group Collectivism) 0.2789 (0.5508) 0.5439∗ (0.3182) γ6 (Assertiveness) - 0.3132 (1.0241) 0.5499 (0.6041) γ7 (Gender Egalitarianism) - 0.3722 (0.7494) 0.0319 (0.4955) γ8 (Future Orientation) 0.7086 (0.9809) - 0.3286 (0.6018) γ9 (Performance Orientation) - 0.0740 (0.9338) 0.2522 (0.5699)

γ10 (Hours total per 1,000) 0.0133 (0.0188) 0.0012 (0.0090)

γ11 (Higher risk) - 0.1394 (0.2581) - 0.0054 (0.1491)

Random part: team level Factor Loadings: Item 1: 1 1 Item 2: 0.8854 1.4134 Item 3: 0.9323 2.4345 Item 4: 2.7580 Item 5: 3.6162 Item 6: 0.7080

Common factor variance: 4.5656∗∗∗ (1.2015) 2.6344∗∗∗ (0.3869)

Random part: country level

Random intercept variance: 1.9356∗∗∗ (0.7182) 0.6912∗∗∗ (0.2158)

Log-likelihood: - 1,409.8480 - 3,407.2835

7. Assessment of 8. Testing of

design effectiveness operating effectiveness

Parameter Est (SE) Est (SE)

Fixed part: γ1(power distance) - 1.4172∗ (0.0.8063) - 0.7607 (0.5312) γ2(uncertainty avoidance) 0.0837 (0.6803) 0.3532 (0.4575) γ3(Humane orientation) 0.6953 (0.5271) - 0.4267 (0.3651) γ4(Institutional Collectivism) - 0.8570 (0.6341) - 0.3163 (0.4462) γ5(In-group Collectivism) 0.1145 (0.4438) 0.6687∗∗ (0.3200) γ6(Assertiveness) - 0.2992 (0.8872) - 0.5508 (0.5389) γ7(Gender Egalitarianism) - 0.8671 (0.6171) - 0.4550 (0.4120) γ8(Future Orientation) 0.4634 (0.8224) 0.2653 (0.5168) γ9(Performance Orientation) - 1.8423∗∗ (0.8057) - 0.3638 (0.4929)

γ10(Hours total per 1,000) 0.1695∗∗∗ (0.0344) 0.0036 (0.0102)

γ11(Higher risk) - 0.3902∗∗ (0.1938) - 0.0573 (0.1433)

Random part: team level Factor Loadings: Item 1: 1 1 Item 2: 1.0898 1.3375 Item 3: 1.1158 Item 4: 1.3541 Item 5: 1.3909

Common factor variance: 3.8605∗∗∗ (0.6420) 1.1206∗∗∗ (0.4153)

Random part: country level

Random intercept variance: 1.4931∗∗∗ _{(0.4529) 0.5167}∗ _(0.2827)

Log-likelihood: - 2,695.0757 - 1,555.0170

Appendix A-1: Two-step Estimation

With the two-step estimation method, the factor model and the structural model are estimated separately. Firstly, the factor model is estimated so that factor scores are obtained. Next, factor scores are regressed on the explanatory variables discussed in section 3 in the structural model.

Consider the factor model. Both exploratory and confirmatory factor analysis could be per-formed. In exploratory factor analysis, there is no or limited theoretical knowledge about the factor structure (e.g. Bortholomew, 2008). In confirmatory factor analysis, there exists prior knowledge so that a hypothesis about the factor structure could be tested. Noteworthy is that exploratory factor analysis may be used as the input for confirmatory factor analysis. We perform exploratory factor analysis and then check whether the factors correspond with our theoretical knowledge. If necessary, we change the factor structure. Confirmatory factor analysis is performed when we estimate the response model and structural model using gllamm.

Recall that factor analysis is used in order to reduce the 41 items into several factors. The factor model is given in (1). It is assumed that x0_ijcβ = βi and z

(2)

m2ijc is a column vector with elements

set to either zero or one. λ(2)m2 is a parameter vector with elements 0 ≤ λim2 ≤ 1. Each item yijc

loads on factor η_m(2)

2jcwith factor loading z

(2) m2ijc

0_λ(2) m2= λ

(2)

im2. This implies that η

(2) m2jcz (2) m2ijc 0_λ(2) m2 is either 0 or η_m(2) 2jcλ (2)

im2. Factors are team specific. Country specific random effects are taken into

account in the structural model. This implies that (1) becomes

y∗_ijc = βi+ M2 X m2=1 η_m(2) 2jcz (2) m2ijc 0_λ(2) m2+ ijc.

factors, we find eight factors. Moreover, the solution is an interior solution. Hence, we continue with these factors.

Using matrices V and L, the correlation matrix can be rewritten: R = V√L V√L 0

= Λ(2)_Λ(2)0_{, with Λ}(2)_{= V}√_L13_{. Matrix Λ}(2) _{is called the factor loading matrix, i.e. it contains the} factor loadings λ(2)_im

2. Similarly, it is the matrix of correlations between factors and variables. A

high factor loading λ(2)_im

2 indicates that item yijcbelongs to factor m2. However, most of the items

are correlated with several factors. Rotation facilitates the interpretation of factors. It maximizes high correlations between factors and variables whereas low correlations are minimized. Using a rotation matrix Ω, the rotated factor loading matrix is obtained: Λ(2)_rotated = Λ(2)Ω. There are numerous methods of rotation available. The main distinction is made between orthogonal and oblique rotation. In orthogonal rotation, factors are uncorrelated. In oblique rotation, factors may be correlated. We use orthogonal rotation since it is assumed that factors are independent from each other. Moreover, with confirmatory factor analysis, no rotation is used, so, factors are independent from each other. Hence, using orthogonal rotation makes the comparison with the estimated response model using gllamm more intuitive. Finally, the results are robust for the rotation method since we obtained similar results with oblique rotation.

After rotation, the factors could be interpreted and factor scores could be predicted for each audit engagement team. There are two procedures for estimating factor scores. Regression factor scores are obtained by first estimating factor score coefficients: B = R−1Λ(2), where B is the 41 × M2 matrix of factor score coefficients. Next, factor scores can be computed using F = ZB, where F is the 1, 223 × M2 matrix of factor scores and Z is the 1, 223 × 41 matrix of standardized observed variable scores. We use Bartlett factor scores rather than regression factor scores since Bartlett factor scores use Generalized Least Squares rather than Ordinary Least Squares. Furthermore, Bartlett factor scores have a higher correlation with factor scores obtained using gllamm. This makes the comparison with the estimated response model using gllamm more intuitive. Finally, Skrondal and Laake (2001) advise to use Bartlett factor scores.

Using the factor scores, it is possible to estimate (4). The results are similar as the results that are obtained when we estimate the response model and structural model using gllamm, however, standard errors seem to be underestimated using the two-step estimation method. This implies that explanatory variables tend to be more significant with the two-step estimation method than in the multilevel structural equation model. So, the statistical significance of cultural dimensions is overestimated. Furthermore, it is important to include unobserved country specific heterogeneity. Ignoring the random intercept yields lower standard errors so that the results would be even more misleading. It might be suggested that it is possible to use cluster robust standard errors instead of a random intercept with corresponding standard error. However, cluster robust standard errors tend to be inconsistent since the data set includes only 49 countries. Therefore, it is important to model unobserved country heterogeneity by including a country specific random intercept in the

13_{Similarly, L = V}0_{RV . So, the correlation matrix R is diagonalized by post- and premultiplying it by the matrix}

regression model.

Appendix A-2: The Factors with Corresponding Items

This appendix describes the items corresponding to the factors. Each item loads on one factor only. The factors are constructed as follows:

1. Engagement leader led discussion and knowledge sharing in the audit engagement team. • During the planning phase of the audit, did the audit engagement team (including

the audit engagement leader) share their understanding of the audited entity and its environment and internal controls?

• During the planning phase of the audit, did the audit engagement team (including the audit engagement leader) share their knowledge of fraud risks at the audited entity? • During the planning phase of the audit, did the audit engagement team (including the

audit engagement leader) share their knowledge of and discuss the significant risks of material misstatement of the financial statements?

• During the planning phase of the audit, did the audit engagement leader participate in the team’s knwowledge sharing?

• Did the audit engagement team hold an audit planning meeting prior to the start of audit field work?

• Did the audit engagement leader participate in forming the responses to identified (significant) risks of material misstatement of the financial statements?

• Was the audit engagement leader involved in the evaluation of the audit engagement team’s performance?

2. Documentation of knowledge about the client and the client’s industry.

• During the planning phase of the audit, did the audit engagement team document their understanding of the client’s market?

• During the planning phase of the audit, did the audit engagement team document their understanding of the client’s strategy?

• During the planning phase of the audit, did the audit engagement team document their understanding of the client’s value creating activities?

• During the planning phase of the audit, did the audit engagement team document their understanding of the client’s financial performances?

• Did the audit engagement team document identified (significant) risks of material mis-statement of the financial mis-statements?

• Did the audit engagement team make inquiries of management and others within the entity to identify (significant) risks of material misstatement of the financial statements and management’s response thereto (including effectiveness of controls)?

• Is there evidence in the audit engagement file that significant risks were addressed through substantive auditing procedures?

• Did the audit engagement team document the evidence obtained per financial statement line item?

4. Compliance by the engagement leader with the required audit steps.

• Did the audit engagement partner document to be satisfied, through review of the audit documentation and discussion with the audit engagement team, that the audit strategy and the planned scope, timing and extent of the audit to be performed is appropriate prior to the start of the audit field work?

• Did the audit engagement partner document to have reviewed the significant audit matters resulting from the audit prior to issuance of the auditor’s report?

• Did the audit engagement partner document to be satisfied, through review of the audit documentation and discussion with the audit engagement team, that sufficient appropriate audit evidence has been obtained to support the conclusions reached and for the auditor’s report to be issued?

5. Compliance by the engagement team with the required steps on fraud risk.

• Did the audit engagement team discuss the susceptibility of the entity’s financial state-ments to material misstatestate-ments due to fraud?

• Did the audit engagement team enquire management and others within the entity regarding the risks of material misstatement due to fraud or error?

• Did the audit engagement team identify and assess the risks of material misstatement due to fraud, and did the team develop and document their responses to identified risks?

6. Assessing audit risks and designing the audit plan.

• Is there evidence that the audit engagement team evaluated the reliability of the un-derlying data given by the client?

• Is there evidence that the audit engagement team defined a maximum for the difference between the client’s actual financials and the audit engagement team’s expectation? • Is there evidence that the audit engagement team computed the actual difference

be-tween the client’s actual financials and the audit engagement team’s expectation? • Is there evidence that the audit engagement team resolved differences greater than the

maximum difference between the client’s actual financials and the audit engagement team’s expectation?

• Is there evidence that the audit engagement team used the standard (analytical) pro-cedure template?

7. Assessment of design effectiveness of internal controls.

• Did the audit engagement team understand, evaluate and document the client manage-ment’s risk assessment process?

• Did the audit engagement team understand, evaluate and document the control envi-ronment?

• Did the audit engagement team understand, evaluate and document information and communication controls?

• Did the audit engagement team understand, evaluate and document the client manage-ment’s process for monitoring controls?

• Did the audit engagement team understand, evaluate and document control activities at the entity level?

8. Testing of operating effectiveness of internal controls.

• Did the audit engagement team meet the client’s non-financial management and eval-uate audit risks?

• When validating controls, was the reliability of the information used validated?

References

Adler, N. J. (2002): International Dimensions of Organizational Behavior. Cincinnati, South-Western college publishers.

Agacer, G.,and T. Doupnik (1991): “Perceptions of Auditor Independence: A Cross-cultural Study,” The International Journal of Accounting, pp. 220–237.