4.1 Sample description
To test our hypotheses, we conducted a survey among German companies using the DAFNE database by Bureau van Dyk. To reach potential companies that currently deal with topics of digitization, we restricted our sample selection to companies that employ at least 100 employees and have annual revenue of at least 30 million euros. This results in approximately 17,000 potential addressees, of which we have randomly chosen 2000 companies. We sent a paper-based questionnaire with a signed cover letter to the management accounting (“Controlling”) departments of the companies. Additionally, we provided a short link within the cover letter for participants who wanted to answer the survey online.
The survey period ran from mid-June to mid-July 2018. In total, we received 159 questionnaires, which equates to a response rate of approximately 8%. 107 out of 159 respondents answered the survey online. However, several questionnaires exhibit missing entries of certain items. While we cannot rule out that some data is missing systematically, we decided to exclude observations with missing data instead of extrapolating missing values (e.g., Byrne
2001). Therefore, the final number of observations for our hypothesis tests is reduced to 115. Table
1 provides an industry breakdown of our sample. While metal and mechanical engineering businesses have a slight edge in our sample, the breakdown presents a rather balanced and broad set of companies.
Table 1Industry breakdown
01 Consumer goods | 6 (5.2) |
02 Chemicals, pharmaceuticals, and health care | 14 (12.2) |
03 Metal producers and machinery | 19 (16.5) |
04 Other product manufacturers | 11 (9.6) |
05 Construction | 10 (8.7) |
06 Real estate | 6 (5.2) |
07 Transportation | 3 (2.6) |
08 Financial | 2 (1.7) |
09 Media, telecommunication, and IT | 7 (6.1) |
10 Utilities | 10 (8.7) |
11 Retailers, servicing, and maintenance | 11 (9.6) |
12 Automotive | 6 (5.2) |
13 Miscellaneous | 10 (8.7) |
Total | 115 (100.0) |
Table
2 presents sample characteristics. Specifically, our questionnaire targeted employees who were working for the finance or (management) accounting (“Controlling”) department of their company, favorably in an executive position. We achieved this goal because all respondents indicated that they work in finance or controlling. In addition, almost 75% of these respondents stated to have an executive function. The respondents’ average working experience within their current company is 12.7 years (median: 10.5 years). On average, survey participants self-assessed their proficiency in information technologies with 4.6 (median: 5), based on a 6-point Likert scale. Besides, only 6 respondents stated a 3, indicating more of a beginner level while about 95% would rate themselves between an advanced or expert level.
1 More than half of our respondents state that their companies’ total annual revenues lie between 50 and 1000 million euros and more than a fourth report revenues above 1000 million euros. Finally, most of our respondents state that their company belongs to a group that has several consolidated companies, which is an indicator of the importance of budgets within these organizations.
Table 2Sample characteristics (n = 115)
Respondents’ function |
CEO/CFO | 5 | (4.3) |
Head of finance/accounting | 81 | (70.4) |
Management accountant, financial analyst, etc. | 29 | (25.2) |
Company-specific tenure (n = 113) |
Less than 5 years | 20 | (17.7) |
Between 5 and 9 years | 29 | (25.7) |
Between 10 and 20 years | 42 | (37.2) |
More than 20 years | 22 | (19.5) |
IT proficiency |
1 (Beginner) | 0 | |
2 | 0 | |
3 | 6 | (5.2) |
4 | 46 | (40.0) |
5 | 54 | (47.0) |
6 (Expert) | 9 | (7.8) |
Total annual revenue (in million euros) |
Less than 50 | 15 | (13.0) |
Between 50 and 1000 | 68 | (59.1) |
Between 1000 and 5000 | 24 | (20.9) |
More than 5000 | 8 | (7.0) |
Consolidated companies |
1 | 14 | (12.2) |
Between 2 and 5 | 37 | (32.2) |
Between 6 and 20 | 38 | (33.0) |
Between 21 and 50 | 17 | (14.8) |
Between 51 and 100 | 3 | (2.6) |
More than 100 | 6 | (5.2) |
To test for a potential (non-)response bias (see, e.g., Armstrong and Overton
1977), we divide our sample into three parts, depending on when the survey has been answered. Performing a Kruskal–Wallis test, all survey constructs show no significant differences between early, middle and late respondents (
p > 0.180, two-tailed). Hence, we infer that a potential non-response bias is unlikely to affect our inferences.
4.2 Variable measurement and control variables
Our questionnaire consists of three main areas. First, we gather information about the budgeting process and related aspects, such as satisfaction with the process. Second, we gather information about current software use, data governance, and deployed analytical techniques. Third, we pose questions about the respondents’ companies and personal attributes.
Based on a thorough literature review, we found that there are no existing instruments that will adequately capture the construct of using analytical methods in the budgeting process. We, therefore, use self-developed items because of at least two reasons. First, we intended to create items, which are particularly suited for German companies because of their specific style of budgeting (Kloock and Schiller
1997; Guenther
2013). Second, by using self-developed items we follow the call of Sivabalan et al. (
2009) for alternative ways to measure budgeting variables. Although we are aware that the budgeting process can differ vastly (e.g., annual budgets vs. rolling forecasts), the general reasons for budgeting do not vary between its concrete types. Irrespective of the frequency of setting the budget, companies regularly pursue a specific purpose with budgets. In line with that reasoning, Sivabalan et al. (
2009) find that there is no difference in the importance of the motives of planning/control and evaluation between rolling forecasts and annual budgets. In addition, they indicate that it is hard to separate annual budgets from rolling forecasts because companies often use rolling forecasts in parallel to annual budgets.
While developing new scales is always considered a sensitive process (Schmitt and Klimoski
1991), the frequently listed difficulties in the development process are mainly due to the aspiration to create behavioral measures (McCoach et al.
2013). In contrast, our scale captures methods or techniques that are deployed. Thus, the use of questions regarding a firm’s deployment of an analytical method is less prone to social desirability reporting (Edwards
1957) because it does not induce the respondent to self-report a favorable behavior. Besides, the wording of the items is likely to be less of an issue here since the items consist only of the analytical methods (e.g., “online analytical processing”). Nevertheless, we follow the typical steps to validate our measure. After an extensive item generation period that comprised extant literature, we pre-tested the content of our scale by using expert judges (DeVellis
2017). For this purpose, we have conducted several interviews with practitioners and adjusted our items accordingly. Lastly, we use a confirmatory factor analysis for data reduction and refining constructs (Ford et al.
1986) and goodness-of-fit indices to assess the model’s fit (MacKenzie et al.
1991). This will be further discussed in the measurement model section.
In accordance with Elbashir et al. (
2011), our final measurement instrument consists of key components that capture the dimensions of business analytics, data infrastructure sophistication, as well as the importance of the planning and evaluation functions of budgeting. We use multiple items to measure each of our hypothesized variables since the use of multiple items for measuring theoretical constructs is considered robust (Gorsuch
1997). Unless stated otherwise, we deployed 6-point Likert-scales to prevent respondents from simply choosing the midpoint.
Use of business analytics in the budgeting process (ANALYT) Due to the multifaceted nature and the correspondingly various definitional approaches of business analytics, we created a large number of items that more or less accurately reflect our desired construct. In total, we ask survey participants about the extent to which their company (1) has or uses automated data integration and harmonization, (2) deploys technologies such as OLAP, big data analytics, text mining, natural language generation, regression analyses, time-series analyses, or classification, and we ask participants to rate the (3) overall extent to which methods from descriptive analytics, diagnostic analytics, predictive analytics, and prescriptive analytics are used. Lastly, we survey how they would assess the company’s level of automated processes during the budgeting process. However, the large number of items is also associated with a certain amount of dispersion, which means that not all items meet the validity criteria in order to be included in the final latent variable. As indicated, the measurement model section describes how we derive the final model.
Data infrastructure sophistication (DATA_INFRA) With this construct, we also aim to cover a broad field with our used items. We capture the extent to which the respondent’s company has standardized master data, data governance as well as the perceived data quality in terms of precision, completeness, timeliness, consistency, unambiguousness, and accessibility.
Importance of the planning function of budgeting (PLAN) To measure one of the two major functional areas, we ask respondents to indicate the importance of different functions of budgeting, anchored by 1 (extremely unimportant) and 6 (extremely important). More specifically, respondents assessed how important their companies consider coordination, codification, resource allocation, forecasts, alignment with the company’s objectives, and assignment of decision-making and spending authority as a function of budgeting.
Importance of the evaluation function of budgeting (EVAL) To measure the second functional area, we ask respondents to indicate how important their company considers motivating target attainment, performance evaluation, deviation analyses, and compensation as a function of budgeting.
Satisfaction with the budgeting process (BUDG_SAT) To assess employees’ satisfaction with the process of budget preparation, we break the satisfaction with the process down into single process components by examining respondents’ satisfaction with the fulfillment of budgeting functions as well as the process’ duration, use of resources, and costs.
Control variables In addition to our main variables, we include control variables that could have an impact on the budgeting process or satisfaction with it, distort the response to the survey, or have an impact on business analytics.
With regard to the budgeting process, size and complexity appear as important attributes (Brynjolfsson and McElheran
2016). Hence, we take the company’s annual revenue (
REV) and the number of consolidated companies (
CONSOL) into account. By using a dummy variable, we split the sample into almost equal parts and divide large companies from smaller ones. We code
REV as 1 if the respondent marked an annual revenue of at least one billion euros, 0 otherwise. Regarding
CONSOL, we also perform a sample split and code it as 1 when the respondents indicated at least 6–20 consolidated companies on the scale, resulting in a sample split of 51 companies with comparably low consolidation efforts and 64 highly consolidated companies.
Moreover, we add two aspects that are potentially associated with the satisfaction of the budgeting process. First, we include the approximate duration of the budget preparation process in weeks (
BUDG_DURA), as time consumption is a major point of criticism (Hansen et al.
2003). Second, we add a factor measuring the perceived participation in the company’s budget preparation process (
BUDG_PART) because a process design allowing for higher levels of participation is likely to increase employees’ satisfaction with the process (Derfuß
2016). The corresponding factor consists of three items, capturing the extent of employees’ input when setting the budgets.
Further, the respondent’s position (EXEC) within the company may alter his or her view on satisfaction with the budgeting process. We code EXEC as 1 if the respondent indicated that he or she has an executive position either in top management or in the finance or accounting department.
Finally, we include attributes that could influence (the evaluation of) the use of business analytics in the budgeting process. In accordance with Brynjolfsson and McElheran (
2016), we control for the respondents’ knowledge and expertise regarding IT (
IT_EXP), which may potentially lead to biased responses in terms of the state of the use of business analytics. Moreover, we distinguish between group subsidiaries (
SUBSID) and group headquarters since this will likely affect the budgeting process and, hence, the degree of use of digitization. We code
SUBSID as 1 if the respondent stated that he or she works for a subsidiary, 0 otherwise.
4.3 Measurement model
We evaluate our theoretical model with confirmatory factor analysis (Gerbing and Anderson
1988). We assess the model fit by using a two-step modeling approach that first evaluates the measurement model to assure its fit and then examines the full structural model (Schumacker and Lomax
2016). Since the majority of the constructs in this study were new constructs, factor analysis has been performed initially to clarify the elements of the constructs. In this way, we also check the reliability, convergent validity, and discriminant validity (Chin
1998; Fornell and Bookstein
1982; Tenenhaus et al.
2005) of our measures. The scales resulting from the analysis were then used in the final confirmatory factor analysis. This approach is similar to that in Fullerton and Wempe (
2009).
We use average variance extracted (AVE) and composite reliability (CR) to assess the general reliability (Chin
1998; Fornell and Larcker
1981). With respect to AVE, only
BUDG_SAT (0.64) and
EVAL (0.55) exceed the general threshold of 0.5. However, Fornell and Larcker (
1981) point out that if AVE is less than 0.5, but CR is higher than 0.6, the convergent validity of the construct is still adequate. All latent variables exceed the threshold of 0.6 (Bagozzi and Yi
1988), with
PLAN having the worst CR with 0.73.
2 These results indicate good consistency of all constructs, even when using another suggested threshold of 0.70 (Nunnally
1978).
Regarding convergent validity, we check the factor loadings of the measures on their respective constructs (Chin
1998; Tenenhaus et al.
2005). We use the common threshold of 0.5 as a cut-off value (Hair et al.
1995,
2010; Hulland
1999) and retain all items whose reliability exceeds this value. In contrast, this means that we delete 9 out of 38 items due to insufficient factor loadings. Table
3 provides an overview of the final items, their corresponding factor loadings, means, and standard deviations.
Table 3Confirmatory factor analysis and measurement model estimates (convergent validity)
Satisfaction with the budgeting process (BUDG_SAT) (CR = 0.839, AVE = 0.641) |
budg_sat1 | Satisfaction with the process’ duration | 0.843 | 3.347 | 1.379 |
budg_sat2 | Satisfaction with the process’ commitment of resources | 0.910 | 3.339 | 1.206 |
budg_sat3 | Satisfaction with the process’ costs | 0.619 | 3.922 | 1.200 |
Use of business analytics in the budgeting process (ANALYT) (CR = 0.828, AVE = 0.415) |
analyt1 | Data automation | 0.867 | 3.261 | 1.185 |
analyt2 | OLAP | 0.574 | 3.035 | 1.955 |
analyt3 | Degree of automation | 0.707 | 2.617 | 0.960 |
analyt4 | Descriptive analytics | 0.508 | 2.200 | 1.540 |
analyt5 | Diagnostic analytics | 0.559 | 2.296 | 1.561 |
analy6 | Predictive analytics | 0.605 | 1.739 | 1.185 |
analyt7 | Prescriptive analytics | 0.623 | 1.600 | 1.041 |
Data infrastructure sophistication (DATA_INFRA) (CR = 0.856, AVE = 0.432) |
data_infra1 | Standardized master data | 0.66 | 3.774 | 1.178 |
data_infra2 | Level of data governance | 0.60 | 3.287 | 1.283 |
data_infra3 | Precision | 0.81 | 4.287 | 1.049 |
data_infra4 | Completeness | 0.76 | 4.296 | 1.043 |
data_infra5 | Timeliness | 0.70 | 4.357 | 0.993 |
data_infra6 | Consistency | 0.82 | 4.052 | 1.083 |
data_infra7 | Unambiguousness | 0.77 | 4.139 | 1.050 |
data_infra8 | Accessibility | 0.55 | 4.200 | 1.149 |
Planning function (PLAN) (CR = 0.727, AVE = 0.347) |
plan1 | Coordination | 0.62 | 4.652 | 1.101 |
plan2 | Codification | 0.77 | 2.896 | 1.385 |
plan3 | Resource allocation | 0.77 | 4.183 | 1.240 |
plan4 | Alignment with the company’s objectives | 0.61 | 5.035 | 0.999 |
plan5 | Assignment of decision-making and spending rights | 0.77 | 3.922 | 1.285 |
Evaluation function (EVAL) (CR = 0.781, AVE = 0.550) |
eval1 | Motivating target attainment | 0.602 | 4.217 | 1.168 |
eval2 | Performance evaluation | 0.901 | 3.965 | 1.344 |
eval3 | Compensation | 0.691 | 3.817 | 1.418 |
With regard to the means, Table
3 also shows that respondents indicate a predominantly average level of satisfaction with the duration (3.347) and commitment of resources (3.339) of the budgeting process. Satisfaction with the costs of the process is significantly higher than the scale’s midpoint of 3.50 (
t = 3.77,
p < 0.001; one-tailed). Regarding the use of business analytics (
ANALYT), the items of OLAP, descriptive analytics, and diagnostic analytics have the highest standard deviations of the sample, indicating considerable differences between companies. Moreover, all items of
ANALYT have mean values that are significantly lower than the scales midpoint (all
p ≤ 0.016, one-tailed). In addition, the frequency of use decreases when we cross the line from reporting and analyzing historical data to proactively forecasting the future (Halper
2013; Huikku et al.
2017). Specifically, sample firms use descriptive analytics significantly more frequently than predictive analytics (
t = 4.09,
p < 0.001, one-tailed). This marks an interesting IT-related productivity paradox while digitization is a trending topic with great opportunities, but companies currently struggle to implement corresponding solutions (Brynjolfsson et al.
2019).
To ensure discriminant validity, we compare the square root of the AVE of each latent variable with the correlation of the latent constructs in the model. Hence, Table
4 presents the correlations between the different constructs in the lower left, off-diagonal elements of the matrix, and the square root of the AVE for each of the constructs along the diagonal (marked in bold). As can be seen from Table
4, every square root of each construct’s AVE has a higher value than the correlations with other latent constructs. Therefore, the Fornell–Larcker criterion is met so that we can assume adequate discriminant validity (Fornell and Larcker
1981; Hulland
1999).
Table 4Correlation matrix and discriminant validity
(1) BUDG_SAT | 0.8004 | | | | |
(2) ANALYT | 0.2566 | 0.6444 | | | |
(3) DATA_INFRA | 0.1250 | 0.3593 | 0.6572 | | |
(4) PLAN | − 0.1654 | 0.2686 | − 0.1314 | 0.5893 | |
(5) EVAL | 0.0406 | 0.0474 | − 0.1983 | − 0.3196 | 0.7418 |