Modelling structural breaks in the tourism-led growth hypothesis

ABSTRACT

Structural breaks represent periods of turmoil that may influence how tourism affects economic growth. Current research on the tourism-led growth hypothesis (TLGH) measures the effect of structural breaks using dummy variables in regression models. However, the drawback of this approach is that there could be multiple structural breaks which result in an overfitting problem and reduce degrees of freedom in small samples. It also becomes difficult to isolate the effect of individual breaks when multiple structural breaks occur within the same year. We thus highlight the role of the Fourier ARDL model in addressing these shortcomings. We use three Pacific Island Countries: Fiji, Tonga, and Vanuatu as case studies to evaluate the efficacy of the Fourier ARDL model. Contrary to earlier research, our results indicate that tourism does not always lead to economic growth. Appropriate modelling of structural breaks also influences the outcome of asymmetric effects. These findings imply that future research should pay close attention to the effects of structural breaks in the TLGH.

KEYWORDS:

• Tourism-led growth hypothesis
• Fourier approximation
• Fourier autoregressive distributed lag model
• tourism demand
• Pacific Island countries

1. Introduction

The tourism-led growth hypothesis (TLGH) is a key area of research within the tourism economics literature (Song & Wu, 2022). Tourism supports income by providing employment and generating exports for the destination economy (Song & Wu, 2022). However, Nunkoo et al. (2020) and Fonseca and Sánchez-Rivero (2020) argue that TLGH research is subject to publication bias. Most research is inclined to report positive and statistically significant estimates of the effect of tourism on growth. This compromises the development of theories and policies related to the tourism sector. Shahzad et al. (2017), argue that estimates of the effect of tourism on economic growth are sensitive to the length and quality of the data, the functional form of the model, and features like structural breaks. Events like the COVID-19 pandemic and major recessions influence the effect of tourism on economic growth (Gunter & Smeral, 2016; Gössling et al., 2021). Nunkoo et al. (2020) reiterate this by calling for more research on the TLGH using updated methods that address these methodological shortcomings.

The objective of this study is to model the effects of structural breaks within the TLGH. Excluding breaks subjects the model to the omitted variables bias. Including breaks, however, raises three empirical issues. First, if multiple structural breaks occur within the same year, isolating the effect of each break, particularly with annual data, is difficult. Second, is the overfitting problem (Banerjee et al., 2017). Multiple break dummies lead to an overfitting problem that diminishes degrees of freedom in finite samples. Third, is determining whether structural breaks have a sharp immediate or gradual effect on the dependent variable. Banerjee et al. (2017), argue that structural break dummies do not capture the gradual effect of structural breaks. Break tests also exhibit low power if the number of breaks is greater than the number allowed by the test (Enders & Li, 2015). The Fourier approximation in autoregressive distributed lag model is able to take into account the above issues (Banerjee et al., 2017). It avoids the challenging procedure of estimating break dates and is suitable for an unknown number and form of breaks.

We position our study within the literature that considers both linear and nonlinear effects of tourism on economic growth. Our starting point of reference is Song and Wu (2022) who argue that the use of the Solow growth model for research on the TLGH raises two key issues. These are whether tourism is consistent as a factor input and whether it leads to long-run growth or merely drives economic fluctuations. We argue that tourism as a source of demand-led growth influences aggregate productivity through learning by doing (Fazzari et al., 2020). Being demand-led raises the issue of cyclical growth patterns and asymmetric effects (Neftci, 1984). Convex aggregate supply curves mean that positive tourism shocks increase output by a smaller amount than negative tourism shocks decrease output (Karras & Stokes, 1999). The opposite pattern of asymmetric effects of tourism may arise if productivity grows faster in expansions (c.f. Bhaduri, 2006; Hein & Tarassow, 2010). The issue of the asymmetric effect of tourism on growth is thus an open empirical question (Balsalobre-Lorente et al., 2021).

We use three Pacific Island Countries, namely, Fiji, Tonga, and Vanuatu as case studies. We select these countries because they are highly reliant on tourism for growth and development in the Pacific region (Kumar & Patel, 2022). We then follow Banerjee et al. (2017) and integrate the trigonometric Fourier terms into the linear (ARDL) and nonlinear autoregressive distributed lag models (NARDL) developed by Pesaran et al. (2001) and Shin et al. (2014), respectively. This is important because the ARDL and NARDL models may not automatically accommodate structural breaks (Pesaran et al., 2001; Cho et al., 2023). We confirm that tourism does not have asymmetric growth effects in Fiji. While the outcome of linearity is consistent with Kumar and Patel (2022), our estimate of effect of tourism on economic growth differs substantially. We further find that tourism does not have any significant effect on economic growth in Tonga. One percent negative tourism shocks reduce growth by about 0.1 percent in Vanuatu. Positive tourism shocks are insignificant. The findings suggest the contrasting outcome that tourism does not always increase economic growth. This reflects the view that tourism is potentially a sub-optimal growth strategy (Higgins-Desbiolles, 2020).

2. Brief literature on tourism, structural breaks, and economic growth

This literature review briefly discusses the current stance of research on tourism and economic growth and then highlights the key role of structural breaks. The tourism-led growth hypothesis describes how tourism increases economic growth. First, tourism encourages infrastructural development and private sector competition which facilitates human capital accumulation (Brida et al., 2016). Second, tourism has spill-over effects on other industries which integrates tourism into the local economy value chain (Cernat & Gourdon, 2012). Third, tourism demand eases foreign exchange constraints and allows the host economy to import capital goods for production (McKinnon, 1964). Finally, tourism allows the host country to benefit from economies of scale and scope which lower average total costs for hotels (Weng & Wang, 2004). Overall, Nunkoo et al. (2020) confirm via a meta-analysis of 545 estimates drawn from 113 studies that tourism increases economic growth even after correcting for publication biases.

Song and Wu (2022) however raise concerns that research on the TLGH is not based on economic theory. They argue that the Solow growth model can include the growth effects of tourism. Yet, there are two overarching issues with this approach. First, tourism demand is incompatible as a factor input like labour and capital. Second, it is unclear whether tourism drives the long-run growth trajectory or instead drives business cycles. A demand-led growth framework may resolve these issues (Fazzari et al., 2020). With this framework, tourism could increase aggregate productivity through learning by doing (LBD). This is an economic concept that argues that productivity gains arise due to practice, minor innovations, and self-perfection (Arrow, 1962; Sheshinski, 1967). LBD leads to an increase in supply even when factor inputs like capital and labour do not increase. This is true for the labour-intensive tourism sector (Liu & Wu, 2019). Tourism sector LBD could thus lead to economy-wide productivity gains because increasing tourism demand has spill-over effects in other sectors and spurs aggregate productivity (Banerjee et al., 2015).

Arguing that tourism is demand-led raises the issue of cyclicality and asymmetric effects. Following the seminal work of Neftci (1984), a large body of research subsequently documented asymmetries in economic variables across the business cycle. Asymmetric effects could arise due to convex short-run aggregate supply curves (Karras & Stokes, 1999). Convexity means that any increase in aggregate demand would result in incrementally smaller contributions to GDP. Negative tourism shocks thus reduce growth by a larger amount than positive shocks increase growth. Balsalobre-Lorente et al. (2021) empirically confirm this outcome. It also means that tourism also has a stronger effect on GDP in recessions. Liu et al. (2022) empirically confirm this outcome. In contrast, Bhaduri (2006) and Hein and Tarassow (2010) argue that productivity grows faster in expansions to prevent real wage growth from labour shortages and increased demand. Tourism could thus have a stronger effect on GDP in expansions. Positive tourism shocks may increase growth by a larger amount than negative tourism shocks reduce growth. Kumar et al. (2020), confirm this pattern of asymmetric effects using the Cook Islands as a case study. Overall, the pattern of asymmetric effects is an empirical matter.

Underpinning the TLGH research is the assumption that the economy does not experience structural changes. Mérida and Golpe (2016) show that ignoring/miss-specifying structural breaks can bias the outcome of causality and the effect of tourism on economic growth. Current research identifies data-specific structural breaks using statistical tests and measures their effects using dummy variables (e.g. Narayan & Popp, 2010). Ignoring structural breaks altogether could result in omitted variables and raise the risk of endogeneity. Banerjee et al. (2017), underscore that including structural breaks raises three empirical issues. The first issue is isolating the effect of individual breaks if multiple structural breaks occur within the same year. The second issue is the potential overfitting problem. Including multiple dummy variables for structural breaks diminish degrees of freedom in finite samples. Third, structural break dummy variables may not capture the gradual effect of structural breaks (Banerjee et al., 2017). The final issue with break tests in general is that the tests themselves may exhibit low power if the number of breaks is greater than the maximum number allowed by the test procedure (Enders & Li, 2015). We employ the Fourier ARDL model to address the above issues.

3. Research design

We use the extended Solow growth model with tourism as a shift variable (c.f. Kumar et al., 2017; Kumar & Patel, 2022). Total factor productivity is assumed to comprise two components. One part of technology evolves independently and is described by the Hicks-neutral technical progress. The second part assumes that technology is a function of tourism demand per worker.

Thus, we specify¹: (1) $y_t=A_0{e}^{gt}tu{r}_t^{\vartheta }\times k_t^{\alpha }$(1) where $y_t$ is real GDP per worker, $A_0$ is the initial level of technology, $g$ is the exogenous growth rate of technology, $tu{r}_t$ is visitor arrivals per worker, $k_t$ is capital stock per worker, $\vartheta >0$ is the elasticity of tourism, and $\alpha$ is the capital share with constant returns to scale.

The general nonlinear Fourier ARDL model is specified below: (2) $\mathrm{\Delta }\ln y_t=\mu +{\text{β}}_1\sin \left(2\pi \rho t/T\right)+{\text{β}}_2\cos \left(2\pi \rho t/T\right)+{\text{β}}_3\ln y_{t-1}+{\text{β}}_4^{+}\ln tu{r}_{t-1}^{+}+{\text{β}}_4^{-}\ln tu{r}_{t-1}^{-}+{\text{β}}_5\ln k_{t-1}+\mathrm{s}\mathrm{h}\mathrm{o}\mathrm{r}\mathrm{t}\mathrm{r}\mathrm{u}\mathrm{n}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{m}\mathrm{s}+u_t$(2)

where $\mu$ is the intercept, $-1<{\text{β}}_3<0$ is the adjustment coefficient, $\ln tu{r}_t^{+}$ and $\ln tu{r}_t^{-}$ are the positive and negative partial sum decompositions defined below, and $u_t$ is the error term.

The trigonometric terms model the distinct types of structural breaks. The optimal value of $1\le \rho \le 4$ minimizes the Akaike information criteria for model selection (Banerjee et al., 2017). We follow an iterative procedure to identify $\rho$. We first set the limits of $\rho$ from 1 to 4 with steps of 0.1. After identifying the first level optimal value of $\rho$ at 1 decimal place, we then use the intermediate $\rho$ parameter with steps of 0.01 and iteratively repeat the procedure until we achieve convergence in the value of $\rho$. We continue this procedure until $\rho$ converges with successive iterations, whilst ensuring that the adjustment coefficient also shows convergences.

The non-linear variables are computed as partial sum decompositions given as follows: (3) $\ln tu{r}_t=\ln tu{r}_0+\ln tu{r}_t^{+}+\ln tu{r}_t^{-}$(3) where $\ln tu{r}_0$ is an arbitrary value, $\ln tu{r}_t^{+}$ and $\ln tu{r}_t^{-}$ denote the partial sum processes which accumulate positive and negative changes in $\ln tu{r}_t$, and are defined as: (4) $\ln tu{r}_t^{+}=\sum _{j=1}^t\mathrm{\Delta }\ln tu{r}_j^{+}=\sum _{j=1}^t\mathrm{m}\mathrm{a}\mathrm{x}\left(\mathrm{\Delta }\ln tu{r}_j,0\right)$(4)

and. (5) $\ln tu{r}_t^{-}=\sum _{j=1}^t\mathrm{\Delta }\ln tu{r}_j^{-}=\sum _{j=1}^t\mathrm{m}\mathrm{i}\mathrm{n}\left(\mathrm{\Delta }\ln tu{r}_j,0\right)$(5)

Asymmetric effects of tourism on real GDP per worker are confirmed by evaluating the null hypothesis that ${\text{β}}_4^{+}$ is statistically equivalent to ${\text{β}}_4^{-}$ via a WALD test. If asymmetric effects are rejected, then the appropriate model is the linear Fourier ARDL model. Cointegration testing is done via the Bounds test by testing whether the lagged level variables in Eq. (2) are significantly different from zero. The resulting F-statistic is checked against the relevant critical bounds. Cointegration is confirmed if the F-statistic exceeds the upper critical bound. Cointegration is rejected if the F-statistic is below the lower critical bound. The outcome of cointegration is inconclusive if the F-statistic is between the upper and lower bounds.

In the final step, the asymmetric cumulative dynamic multipliers are derived as follows: (6) $m_h^{+}=\sum _{j=0}^h{\displaystyle \frac{\mathrm{\partial }\ln y_{t+j}}{\mathrm{\partial }lntu{r}_t^{+}};m_h^{-}=\sum _{j=0}^h{\displaystyle \frac{\mathrm{\partial }\ln y_{t+j}}{\mathrm{\partial }lntu{r}_t^{-}},h=0,1,2,}}$(6) As$h\to \mathrm{\infty }$, then $m_h^{+}$ and $m_h^{-}$ approach the long-run coefficients. The adjustment path and duration of disequilibrium provide useful information on the pattern of asymmetric adjustment.

4. Results

We confirm that tourism has asymmetric effects on economic growth only in Vanuatu (Table 1). The models are cointegrated at the 5 percent level (Table 2). Specifically, the outcome of cointegration is accepted after structural breaks are accounted for in the model using the Fourier function. Notably, cointegration is absent from the ARDL models which does not consider structural breaks in the specification (Shahzad et al., 2017; Kumar et al., 2018). The results indicate that a 1 percent increase in visitor arrivals per worker increases real GDP per worker by 0.17 percent in Fiji (Table 2). Tourism is insignificant in Tonga (Table 2). The results for Vanuatu indicate that a 1 percent decline in visitor arrivals per worker reduces real GDP per worker by 0.10 percent. Positive shocks in tourism have insignificant effects on growth. Our results agree with Kumar and Patel (2022) that tourism has symmetric growth effects in Fiji and that negative tourism shocks reduce output growth in Vanuatu.² We differ in the size of the effect of tourism on economic growth for Fiji. Our estimate for the effect of tourism on growth is less than half of the estimate of Kumar and Patel (2022). Our results concerning Tonga disagree with Kumar et al. (2021). We find that tourism does not have any significant effect on economic growth. The estimates satisfy the various diagnostic tests implying robust inferences (Table 2, Panel c). The progression of the dynamic multiplier for Vanuatu agrees with the long-run estimates in Table 2 (Figure 1).

Figure 1. Dynamic multiplier. Note: The solid black (dashed) line indicates the response of real GDP per worker for a unit increase (decrease) in visitor arrivals per worker.

Table 1. Asymmetric effects test.

Null hypothesis	Fiji	Tonga	Vanuatu
Tourism has symmetric effects on growth	${\chi }^2\left(1\right)$ = 0.2190 [0.6398]	${\chi }^2\left(1\right)$ = 0.3175 [0.5731]	${\chi }^2\left(1\right)$ = 3.5712** [0.0488]

Note: ** indicates asymmetric effects at 5 percent. P-value in [.].

Table 2. Main results.

Variable	Fiji		Tonga		Vanuatu
	FARDL	ARDL	FARDL	ARDL	FNARDL	NARDL
Panel a: long-run coefficients
$\ln \mathrm{t}{\mathrm{r}}_{\mathrm{t}}$	0.1718*** [0.0426]	0.0854 [0.1783]	0.0654 [0.0648]	−0.0502 [0.3332]
$\ln {\mathrm{t}\mathrm{r}}_{\mathrm{t}}^{+}$					−0.0309 [0.0843]	0.2951 [0.4146]
$\ln {\mathrm{t}\mathrm{r}}_{\mathrm{t}}^{-}$					0.1041*** [0.0297]	0.2883 [0.2268]
$\ln {\mathrm{k}}_{\mathrm{t}}$	0.4346*** [0.0649]	0.6757** [0.3084]	0.4173*** [0.0457]	0.4974*** [0.2381]	0.3086** [0.1452]	0.1542 [0.6417]
$\cos 2\pi \rho \mathrm{t}/\mathrm{T}$	0.0122 [0.0108]		−0.0342*** [0.0062]		−0.0464*** [0.0063]
$\sin 2\pi \rho \mathrm{t}/\mathrm{T}$	−0.0518*** [0.0096]		0.0336*** [0.0052]		0.0106 [0.0101]
Intercept	4.6249*** [0.6431]	2.2665 [3.0375]	4.6327*** [0.4131]	3.9621* [0.1428]	5.4062*** [1.5114]	9.6298 [5.8467]
Panel b: short-run coefficients
$\mathrm{\Delta }\ln \mathrm{t}{\mathrm{r}}_{\mathrm{t}}$	0.1005*** [0.0148]	0.1090*** [0.0183]	0.0641 [0.0637]	−0.0131 [0.0847]
$\mathrm{\Delta }{\mathrm{l}\mathrm{n}\mathrm{t}\mathrm{r}}_{\mathrm{t}}^{+}$					−0.0197 [0.0084]	0.0907 [0.0924]
$\mathrm{\Delta }{\mathrm{l}\mathrm{n}\mathrm{t}\mathrm{r}}_{\mathrm{t}}^{-}$					0.0665*** [0.0112]	0.0886** [0.0373]
$\mathrm{\Delta }\ln {\mathrm{k}}_{\mathrm{t}}$	0.9364*** [0.2501]	0.6039** [0.2952]	0.4082*** [0.0868]	0.1299* [0.0689]	0.1975 [0.1285]	0.0474 [0.1790]
$\cos 2\pi \rho \mathrm{t}/\mathrm{T}$	0.0071 [0.0067]		−0.0335*** [0.0096]		−0.0276*** [0.0079]
$\sin 2\pi \rho \mathrm{t}/\mathrm{T}$	−0.0303*** [0.0061]		0.0329*** [0.0061]		0.0068 [0.0070]
$\mathrm{E}\mathrm{C}{\mathrm{T}}_{\mathrm{t}-1}$	−0.5852*** [0.0927]	−0.2956*** [0.1050]	−0.9782*** [0.1766]	−0.2611* [0.1424]	−0.6397*** [0.1137]	−0.3075 [0.1791]
Panel c: model statistics and diagnostics
Specification	FARDL (1,0,0)	ARDL (1,0,0)	FARDL (1,0,0)	ARDL (1,0,0)	FNARDL (1,0,0,0)	NARDL (1,0,0,0)
adj. R^2	0.9763	0.9695	0.9641	0.9357	0.7423	0.5767
Rho $(\rho )$	1.7500		1.3100		2.7809
Standard error of regression	0.0261	0.0317	0.0183	0.0262	0.0235	0.0301
Bounds test	26.2105 (5.855)	3.3445 (5.855)	8.3885 (6.265)	1.2422 (4.140)	5.7308 (4.850)	1.4965 (5.610)
Serial correlation	2.3121 (0.3147)	2.3428 (0.3099)	0.2067 (0.6494)	0.7202 (0.6976)	2.2196 (0.3296)	3.5335 (0.1709)
Heteroscedasticity	8.3433 (0.2140)	4.6267 (0.4631)	2.9022 (0.7151)	5.1770 (0.1593)	1.6981 (0.9453)	3.0348 (0.5520)
Functional form	1.5185 (0.2266)	0.5108 (0.4796)	0.8037 (0.3825)	1.6761 (0.2110)	1.4200 (0.2498)	0.4518 (0.5096)
Normality	2.1011 (0.3497)	1.5118 (0.4696)	2.0380 (0.3609)	0.8141 (0.6656)	0.9633 (0.6168)	4.2628 (0.1186)
CUSUM and CUSUM of Squares	Stable	Stable	Stable	Stable	Stable	Stable

Notes: ***, **, * indicate significance at 1, 5, and 10 percent. Standard error in [.], p-value for the diagnostic and 1% upper critical value for bounds tests in (.). Diagnostic test degrees of freedom automatically determined.

Structural breaks influence the results, hence the validity of the TLGH (Shahzad et al., 2017). Excluding the effects of structural breaks appears to influence the outcome of cointegration and influences the growth effects of tourism demand. Our results support Higgins-Desbiolles (2020) who argues that the tourism-led growth hypothesis may be a sub-optimal strategy for economic growth. The Fourier approach helps answer issues that relate to emerging themes in the tourism-growth literature. The first is publication bias in the TLGH. Nunkoo et al. (2020) and Fonseca and Sánchez-Rivero (2020) note that studies are more inclined to report positive and significant effects. The erroneous conclusion derived from these studies is that tourism significantly increases economic growth. We find contrary results after including the effects of structural breaks using the Fourier function. The second issue that our study relates to is the identification of nonlinearity in the TLGH. Our results indicate that the appropriate modelling of the effect of structural breaks without assigning dummy variables could nullify the outcome of nonlinearity. This questions the view that tourism has nonlinear effects on growth (Balsalobre-Lorente et al., 2021; Kumar et al., 2020).

5. Conclusions, policy implications, limitations, and research outlook

This study demonstrates the efficacy of Fourier ARDL models in modelling structural breaks in the tourism-led growth hypothesis. Appropriate modelling of structural breaks may influence the effect of tourism on economic growth (Shahzad et al., 2017). The benefit of the Fourier approximation is that it can resolve the issue of multiple structural breaks while avoiding potential overfitting problems. The outcome is that estimates are robust to the effects of structural breaks. The Fourier approximation is flexible enough to accommodate other variants of ARDL models. One such version is the nonlinear ARDL model developed by Shin et al. (2014) which considers directional asymmetries in the effect of tourism on growth. Overall, this paper discusses a simple approach to model the effects of structural breaks in models that assess the TLGH without the resulting overfitting problem. This is important because the standard ARDL model, which is common to the TLGH research (Song & Wu, 2022) may not automatically accommodate structural breaks (Pesaran et al., 2001; Cho et al., 2023).

The findings demonstrate that tourism does not always positively or significantly affect economic growth contrary to earlier studies (Kumar & Patel, 2022). Our results indicate that research on the TLGH is sensitive to the specification of structural breaks in regression models. The results also indicate that research on the TLGH could be subject to publication bias in favour of positive and significant effects (Nunkoo et al., 2020; Fonseca & Sánchez-Rivero, 2020). If tourism has insignificant effects on growth, this implies that economic policies that encourage further investments beyond maintenance levels may not be value-adding. The appropriate policy for Vanuatu would be to avoid declines in visitor arrivals per worker. Policymakers in Tonga would have to identify other drivers of growth. Fiji needs to ensure that the rate of growth of visitor arrivals exceeds the rate of population growth.

One limitation of our study is that we have only focused on three Pacific Island countries. The availability of suitable data for analysis of all variables limits our focus on these countries. We pay appropriate attention to issues relating to the appropriate theory to guide the analysis. We use the Solow growth model to determine the growth effects of tourism. Using the Solow growth model in tourism raises the issue of whether tourism is compatible as a factor input and whether tourism leads to permanent economic growth or only fuels economic fluctuations (Song & Wu, 2022). We bypass these issues by arguing that tourism, as a source of demand-led growth, influences aggregate productivity through learning by doing in an accommodating supply framework (c.f. Fazzari et al., 2020). Overall, future research can apply the methodology described in this study to re-visit the TLGH for other countries and regions.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

We use a total of 42 years annual data from 1979–2020 for Fiji. The data for real GDP is available from 1960-2021. The data for investment is proxied by gross fixed capital formation (GFCF) and is available from 1980-2020. For Tonga, we use 25 years annual data from 1995-2019. Real GDP is available from 1981–2021 and GFCF is available from 1975-2021. For Vanuatu, real GDP is available from 1979–2021 and GFCF is available from 1983-2021. We use the perpetual inventory method to calculate the capital stock series. The initial capital stock is set to 1.5 times the initial real GDP value. The initial real GDP value is set to the year before the start of the investment series. The depreciation rate is set to 5 percent. Labour force participation rate data is available from 1990–2021 for all countries. The average of this multiplied with population derives the labour supply series. The data for visitor arrivals is available from 1975 to 2021 for Fiji, 1995–2019 for Tonga, and 1995–2020 for Vanuatu. Real GDP and GFCF data are in constant 2015 US dollars. All data is from the World Development Indicators database (2023). Visitor arrivals for Fiji is from the Fiji Bureau of Statistics (2023).

Notes

1 We thank an anonymous reviewer for recommending using the per-worker value of tourism in Eq. (1).

2 For an interpretation of the NARDL coefficients, see Balsalobre-Lorente et al. (2021) and Kumar et al. (2020).