**3. Data**

We consider a comprehensive database from the Australian service sector (as classified by Australian Bureau of Statistics) for the period 2009 to 2019. The data was collected from Datanalysis database-a database that publishes financial data of companies in di fferent Australian sectors. Although our initial sample was much larger than what we have included in the study, due to matching inconsistency in variable definition and the availability of all variables of all companies, we have truncated the data to 91 companies that have same data set for the entire time period. These companies are all listed in the Australian Stock Exchange.

Table 1 shows that a total of nine sectors are considered to conduct research for the period 2009 to 2019. Based on the availability of the entire data set with chosen variables, some sectors had the most samples and others had only a few companies. The percentage column shows the degree of weight from each sector of our sample. Based on the literature surveyed, we consider several variables shown in Table 2 to investigate our research question.

To avoid spurious regression estimates in our empirical analysis, variables under consideration should ideally be stationary. To confirm this, we used the panel unit root test of Levin et al. (2002). Table 3 shows that the unit-roots hypothesis is rejected by all variables at the 1% level of significance. Following (Canarella and Miller 2018; Köksal and Orman 2015; Khan et al. 2018; M'ng et al. 2017), we also checked for stationarity using a unit root test and observed that all variables were stationary with respect to the dependent variables (Return on Equity (ROE), Operating Margin (op\_margin), Return on Asset (ROA) and Return on Invested Capital (ROIC)), confirmed by the tests for heteroscedasticity and autocorrelation diagnostics.

The panel regressions were run for four dependent variables (return on equity, return on assets, return on invested capital, and operating margin), two treatment variables (leverage and long-term debt to total assets ratio), and five control variables (size, liquidity, revenue growth for three years, tangibility and depreciation tax shield). A series of regressions were run for these variables and diagnostic tests were conducted to confirm the appropriateness of fixed or random e ffects panel regressions models.

For each of the dependent variables, outputs for two models are presented, after eliminating the inappropriate models using Hausman tests. The Breausch Pagan test was employed to confirm the outputs of the Hausman test for this purpose. Earlier studies in capital structure used the Hausman test to identify the appropriate panel data model from two available models: fixed e ffect model and random e ffect model (Dalci 2018; Mayuri and Kengatharan 2019; Sivalingam and Kengatharan 2018; Suntraruk and Liu 2017). Breausch Pagan Lagrange Multiplier tests were used for confirming the appropriateness of the random e ffects model (see for example, Dalci 2018; Ghasemi et al. 2018; Khan et al. 2018). The tables below present the outputs of these models.


**Table 1.** Table shows the various Australian service sector companies considered for this research.

> (Source: Authors' compilation).



**Table 3.** Unit root tests results.

Tax shield Depreciation/Total assets (Fitim et al. 2019; Shalini and Biswas 2019; Yazdanfar and

Operating revenue (size) Log of Operating revenue (Fitim et al. 2019; Shalini and Biswas 2019) Revenue growth (3-year) % of revenue growth (3 yearly average, given) (Chadha and Sharma 2015; Chakrabarti and Chakrabarti 2019)

Öhman 2015)

