*3.3. Common Method Bias*

Common method bias (CMB) is recognized to be a validity-threatening problem that can jeopardize the study's results. The main sources of CMB result from the following (Podsakoff et al. 2003): (a) the same respondent at a single point rate for both the independent and dependent variables; and (b) the characteristics of the measurement items. To mitigate CMB problems, the respondents were assured of anonymity, i.e., the answers provided would only be used for research/investigation purposes and that personal information, regarding the company or the respondent, was not asked (Podsakoff et al. 2003; Tourangeau et al. 2000). Moreover, as referred to above, the measurement items were pilot tested with academicians for clarity to reduce their characteristic effects (Podsakoff et al. 2003). Then, even after implementing these preventative procedures to mitigate CMB, Harman's (1967) single-factor analysis was used to assess the potential of CMB after data collection: it was possible to conclude that the single factor of the data that emerged explained less than 50% of the variance, suggesting that CMB was not an issue.

#### *3.4. Data Analysis*

The statistical analysis of the data obtained was conducted through the partial least squares method of structural equation modeling (PLS-SEM), using the SmartPLS 3.0 software; this is because PLS-SEM is considered a robust data analysis tool when the sample size is relatively small, which is the case, although the sample exceeds the minimum

size of 200 responses (Hair et al. 2016). On the other hand, PLS-SEM is the most robust methodology to use because it is robust when working with non-normal data and when the theoretical framework is at an early stage of development, which involves testing and validating an exploratory model (Henseler and Chin 2010). Moreover, despite the differences between covariance-based structural equation modeling (CB-SEM) and PLS-SME (Dash and Paul 2021), as this research sought to explore the concepts and relationships between them in emerging countries, in which the context may play an important role, PLS-SEM was chosen. Furthermore, although both methods are similar, PLS-SEM seems to present higher item loadings, better construct reliabilities and a higher validity than CB-SEM (Dash and Paul 2021).

Data were analyzed to test the psychometric properties of the scales used, namely the reliability, validity, and uni-dimensionality of the constructs, using specific statistical tests (Hair et al. 2019). The internal consistency was examined using Cronbach's alpha coefficients.

The model was tested by assessing the reliability, convergent validity, and discriminant validity of the items and constructs. As PLS-SEM employs bootstrapping to test the significance of relationships, it works well when dealing with non-normal variables, which is usually the case when multiplying two normally distributed variables (Bollen and Stine 2014; Efron 1988). PLS performs well in mediation analysis (Hair et al. 2016). Moreover, most of the proposed solutions to implement mediation analysis are fulfilled (Aguinis et al. 2017).

The coefficient of determination (R2) of the endogenous variables was used to explain the percentage of the variability of the dependent variable, explained by the independent variable(s). R<sup>2</sup> is a good indicator of how well observed outcomes are replicated by the model (Hair et al. 2010). The effect size (f2), which assesses how exogenous constructs help explain an endogenous construct in terms of R2, was used to explain the relative importance of the differentiation strategy and positional advantage. Finally, the standardized root mean squared residual (SRMR) was used to test the model fit (Cho et al. 2022).

#### **4. Results and Discussion**

Tables 2–4 present the loadings of the items obtained through bootstrapping with 5000 replications. Items EP3, SD1, SD5, APRD3, ASERV1 and ASERV3 were removed because they presented factor loadings below the minimum threshold value required. All other items presented loadings equal to or higher than the minimum recommended threshold of 0.7 (Götz et al. 2010).

Before analyzing the hypotheses presented in Figure 1, the industrial sector, firm size (number of employees) and firm age were tested as control variables. The results are shown in Table 5. As such, it is possible to conclude that there are no statistically significant differences for the variables analyzed. Therefore, the resources of larger firms are not very different from the resources of smaller firms, older firms do not possess better resources or capabilities than smaller firms, and the different sectors do not perform differently.

**Table 5.** Analysis of control variables.


Table 6 presents the analysis of the internal consistency and reliability of the scales of the three constructs used, based on Cronbach's alpha, Rho-A, and the composite reliability (CR). Cronbach's alpha coefficients were 0.924 for the differentiation strategy, 0.938 for the export performance, and 0.830 for the positional advantage, all above the recommended value of 0.70 (Hair et al. 2016). The Rho-A coefficients were 0.925 for the differentiation strategy, 0.947 for the export performance, and 0.880 for the positional advantage, all above the recommended value of 0.70 (Dijkstra and Henseler 2015). The CR indicators ranged from 0.880 and 0.938, and are above the recommended value of 0.7 (Hair et al. 2016). Finally, the convergent validity is assured, as the values of AVE are larger than the recommended value of 0.5 (Fornell and Larcker 1981; Götz et al. 2010); this means that all items converge when measuring the underlying constructs under assessment.

**Table 6.** Internal consistency, reliability and convergent validity.


Source: Own elaboration.

Discriminant validity seeks to demonstrate that a certain construct explains the variance of its own indicators better than the variance of other latent constructs (Henseler et al. 2015). Therefore, the discriminant validity was assessed using two perspectives: the Fornell–Larcker and the heterotrait–monotrait (HTMT) criteria. The results are presented in Table 7. While the Fornell–Larcker method compares the square root of the AVE with the correlation of all latent constructs (Hair et al. 2019), the HTMT criterion indicates the need to compare a predefined threshold that ranges from 0.85 (Kline 2011) to 0.9 (Gold et al. 2001) to the actual result for each construct. When using the HTMT criteria, if the result is lower than this threshold—0.85 (Kline 2011) or 0.9 (Gold et al. 2001)—it is possible to claim that there is discriminant validity among the constructs (Henseler et al. 2015). Table 7 shows that the Fornell–Larcker and the HTMT criteria show discriminant validity.

**Table 7.** Discriminant Validity.


Note: diagonal elements (in **bold**) are the square root of AVE; the correlations between the variables are presented below the diagonal. They are used to assess the Fornell–Larcker criterium; HTMT scores are presented above the diagonal in *italics*.

#### *4.1. Hypothesis Testing*

The model shown in Figure 1 was tested based on the sign, magnitude, and statistical significance of the parameters of the relationships tested, as supported by Götz et al. (2010). The coefficient of determination (R2) of the endogenous variables was also analyzed.

Linear regression coefficients were used to test hypothesis H1, H2 and H3. A hierarchical regression analysis (Aguinis and Gottfredson 2010; Arnold 1982; Sharma et al. 1981) was used to test H4. The variable 'differentiation strategy' was used as the independent variable. Finally, the effect of the variable 'positional advantage' was analyzed with the inclusion of the two relationships to be tested: differentiation strategy x positional advantage.

When testing for the mediation effects (Hair et al. 2016), firstly, the significant direct effect between the independent and dependent variable should be established when the mediating variable is excluded. Secondly, the indirect effect of the mediating variable should be significant when the mediating variable is included in the model. Finally, the relationship between the independent and dependent variables should be significantly reduced when the mediator is added. These three steps were carried out in this study using PLS-SEM.

Table 8 and Figure 2 present the results of the effects of the variables and the confirmation of the hypotheses. It is possible to conclude that all the structural relationships tested

have positive signs and parameters, which is in accordance with the assumptions made. According to the results, the differentiation strategy has a positive effect on export performance, i.e., the results obtained support and validate hypothesis H1 (β = 0.462; *p* < 0.001); therefore, the differentiation strategy positively affects the export performance. The model tested the relationship between the differentiation strategy and the positional advantage, indicating that the differentiation strategy has a positive effect on the positional advantage, validating hypothesis H2 (β = 0.622; *p* < 0.001). The results confirm H3 (β = 0.221; *p* < 0.001), since the positional advantage has a positive impact on the export performance. Finally, the data confirm and validate hypothesis H4, whereby the positional advantage has a positive mediating effect on the relationship between the differentiation strategy and the export performance (β = 0.138; *p* < 0.001).



Note: DiffStr: differentiation strategy; ExpPer: export performance; PosAvd: positional advantage; CILL: confidence interval lower limit with corrected bias; CIUL: confidence interval upper limit with corrected bias. Source: Own elaboration.

**Figure 2.** Moderating effect of positional advantage.

Figure 2 shows that the determination coefficient (R2) of the export performance is 0.390, i.e., both the differentiation strategy and the positional advantage explain 39% of the variance in the export performance. Moreover, the differentiation strategy alone explains 38.7% of the variability in the positional advantage of exporting Mozambican firms, giving a clear importance to the differentiation strategies the Mozambican firms manage to deploy.

When assessing the effect size, as shown in Table 8, it is possible to claim that, based on Cohen (1988), the differentiation strategy has a medium effect (f2 = 0.215) on the export performance, the positional advantage has a weak effect (f<sup>2</sup> = 0.049) on the expert performance, and the differentiation strategy has a strong effect (f<sup>2</sup> = 0.631) on the positional advantage. These results corroborate the explanatory power of the differentiation strategy when explaining the R2 obtained for the positional advantage and export performance.

Table 8 shows that the direct effect of differentiation strategy on export performance (β = 0.462) is larger than the indirect effect (β = 0.138). To complement this analysis, and to assess the strength of the mediation effect, Zhao et al.'s (2010) three-factor approach was used, in which it is possible to assess how the indirect effect absorbs the direct effect.

Given the values presented in Table 8 and Figure 2, the proportion of the indirect effect versus the total effect is 0.23 − (0.138)/(0.138 + 0.462) = 0.23 –, which indicates that the positional advantage partially mediates the relationship between the differentiation strategy and the export performance. These results also corroborate the results obtained when assessing R<sup>2</sup> and f2: competitive strategy plays a more important role in explaining the export performance than the positional advantage does.

The results for the sample of 250 firms show that the cut-off value for the standardized root mean squared residual (SRMR) is 0.044, which is lower than the threshold value of the SRMR = 0.06 for samples with N > 500, and lower than the recommended value of the SRMR = 0.091 for samples with 200 respondents (Cho et al. 2022). As such, the model fit is assured.
