Exploring the relationship between job satisfaction and organizational commitment: an instrumental variable approach

Abstract

The possible role of job satisfaction (JS) on organizational commitment (OC) has been a very important and hotly debated topic among experts. However, existing studies have yielded mixed results potentially due to utilization of small datasets, different methodological designs, estimation techniques that do not control for potential endogeneity between the variables, or a combination of these issues. Using a large matched employer-employee data-set from Britain (WERS2011), we find that increases in employees’ JS positively influence OC. We also show that this relationship holds when an instrumental variable framework (IV ordered probit/IV probit) is adopted to take into account the potential endogeneity of JS. However, throughout the analysis, the IV estimates are smaller in magnitude in comparison to where JS is considered as an exogenous variable. Moreover, utilising a two-stage probit least square (2SPLS) estimator, we support our previous findings i.e. increased JS is likely to lead to enhanced OC, but we also show that greater OC leads to higher levels of JS suggesting that JS and OC are likely to be reciprocally related. Overall, the IV estimates confirm the importance of addressing the endogeneity issue in the analysis of the relationship between JS and OC.

Keywords:

• Job satisfaction
• Organizational commitment
• Endogeneity
• Instrumental variable framework
• WERS
• Britain

Acknowledgements

We would like to sincerely thank Paul Tracey, Professor of Innovation and Organization at Judge Business School for useful comments on a prior version of this research.

Notes

1. A large number of measures of JS have been developed but there appears to be no consensus on how to measure JS. There is also debate as to whether single item questions are adequate, or whether it is better to conceptualize JS as multi-dimensional and to employ facet measures (Judge & Kammeyer-Mueller, 2012; van Saane, 2003; Wanous, Reichers, & Hudy, 1997). van Saane (2003) reviewed and evaluated 29 JS measures published between 1988 and 2001 and found only seven met their reliability and validity criteria. The recently renewed debate about how precisely to define JS (Judge & Kammeyer-Mueller, 2012) also implies earlier measures may no longer be considered adequate.

2. Mathieu (1991) and Lance (1991) both found that the influence of satisfaction on commitment was higher than the influence of commitment on satisfaction. Farkas and Tetrick (1989) and Huang and Hsiao (2007), on the other hand, found the relationship to be broadly symmetrical.

3. Rayton (2006) used a bivariate probit estimation technique, which allows for interaction between the error processes of the employee commitment and JS equations. However, the second dependent variable did not appear on the right-hand side of the first equation (recursive bivariate model, see Greene, 2003).

4. The empirical work is extensive and covers an extensive spectrum of methods (see, for example, Bakan et al., 2004; Bateman & Strasser, 1984; Brunetto et al., 2012; Buonocore & Russo, 2013; Curry et al., 1986; Elangovan, 2001; Froese & Xiao, 2012; Huang & Hsiao, 2007; Markovits et al., 2010; Mathieu, 1991; Rayton, 2006; Top & Gider, 2013; Top et al., 2015; de la Torre-Ruiz et al., 2017; Williams & Hazer, 1986; Wong, Chun, & Law, 1995). It can be argued that the empirical strategy adopted here can deal with statistical issues concerning for example, endogeneity and the measurement levels of the examined variables (see for example, Bollen, 2001).

5. Ferrer-i-Carbonell and Frijters (2004) found that assuming ordinality or cardinality of happiness scores makes little difference.

6. 0 < μ₁ < μ₂ < μ₃ < μ₄.

7. In this case u_fi is decomposed into independent components as follows: u_fi = ϑ_f + η_fi where η_fi is a random error term with mean 0 and variance ${\mathit{\sigma }}_{\mathit{\eta }}^2$; ϑ_f is the firm specific unobservable effect capturing differences in satisfaction across firms with mean 0 and variance ${\mathit{\sigma }}_{\mathit{\vartheta }}^2$, and it is assumed to be independent of S_fi and X_fi.

8. This potential for unobserved heterogeneity will result in the error term, u_fi in model (2), being correlated with S_fi. The correlation between u_fi and S_fi may also result in biased estimates of the other coefficients.

9. CMP is a limited-information maximum likelihood (LIML) estimator where the first stage parameters are structural and the second stage parameters are reduced form. The error terms $v_{1,fi}$ and $v_{2,fi}$ are assumed to be jointly normally distributed. For further discussion see Roodman (2011).

10. For the validity of the instruments see the discussion in the next section where the model is re-examined within an IV probit framework.

11. To help identify the simultaneous system of equations in vector X₂ we include the previously discussed instruments along with a set of control variables. In contrast the vector X₁ includes religious denomination (Farrukh, Wei Ying, & Abdallah Ahmed, 2016; Guiso, Sapienza, & Zingales, 2003), type of contract and various control variables. Treating OC as continuous, the Hansen J statistic is found to be 3.559, which is insignificant at the 5% level. Also, standard F-tests indicate the joint significance of these variables in the OC model (F(2, 13,422) = 25.790; p < 0.01).

12. For further discussion of methods to adjust the standard errors see Keshk (2003).

13. If, $\text{cov}(S_{\mathit{fi}},P_{\mathit{fi}})$ is zero the bias term disappears.

14. This rules out omitted variables or measurement error in u_fi that are correlated with S_fi.

15. The asymptotic bias in the estimate of the coefficient (γ) of S_fi will be also positive.