ABSTRACT
This study examines the effect of outliers on causal relationship between financial development and economic growth using 48 countries from 1988 to 2014. The dynamic panel model of Levine, Loayza, and Beck (2000) is used to examine this issue. We propose a novel approach by combining the least square dummy variable correction method (LSDVC) to remove the estimates bias in the dynamic panel model and the least trimmed squares (LTS) to control outlier influence. The combination of these two methods is referred to as LSDVC + LTS. Our results show a counter-intuitive evidence that bank development negatively affects economic growth when the outlier influence is ignored. This counter-intuitive evidence holds even when the conventional winsorization method is used to control the outliers. However, bank development exhibits a positive influence on economic growth once the proposed approach LSDVC + LTS is adopted. Also, stock market development exhibits a positive effect on economic growth regardless of the outliers.
Supplemental Material
Supplemental data for this article can be accessed on the publisher’s website.
Notes
1. See Schumpeter (Citation1991); McKinnon (Citation1973) and Pradhan, Arvin, and Bahmani (Citation2015).
2 See Zhu, Ash, and Pollin (Citation2004); Beck, Levine, and Loayza (Citation2000); Arestis, Demetrius’s, and Luintel (Citation2001); Naceur and Ghazouani (Citation2007) and Ghartey (Citation2015).
3. See also Acosta et al. (Citation2008); Beck, Levine, and Loayza (Citation2000); Shen et al. (Citation2015) and Dobbelaere, Lauterbach, and Mairesse (Citation2016).
4. Bruno (Citation2005) extends the LSDV bias approximations, including three corrected levels,
, and
in Bun and Kiviet (Citation2003), which are obtained by modifying the inside operator to accommodate the dynamic selection rule and unbalance the strictly balanced panels acquired from Kiviet (Citation1995). Kurennoy (Citation2015) studies the behavior of the bias-LSDV correction (LSDVC) estimator from Kiviet (Citation1995) in dynamic panel data models with an endogenous regressor, and finds that the LSDVC estimator works poorly in the endogenous case.
5. Based on the Monte Carlo results, Judson and Owen (Citation1996) suggested that the GMM estimator did not outperform the rivals approaches in terms of either the average size of bias or efficiency. They suggested using the LSDVC for panels with small time dimensions (T ≤ 10) while recommending the IV estimator of Anderson and Cheng (Citation1981) for longer panels. The efficiency of the IV estimator improves with T, and the IV estimator is computationally simpler than the LSDVC. Kiviet (Citation1995) conducted the simulation method and found that LSDVC outperformed the GMM method. However, in theory, a key limitation for the LSDVC method is that all its explanatory variables are strictly exogenous.
6. If , we stop the iteration is stopped; otherwise, another iteration yielding
and so on is applied. The sequence
is nonnegative and must converge. In practice, only finite amounts of h-subsets are present; thus, index m must exist, such that
. Hence, convergence is always reached after some finite steps.
7. The estimates of the regression parameters are given together with a corresponding scale estimate. For this purpose, the preliminary scale estimate s°, which is based on the minimal median and multiplied by a finite-sample correction factor (that depends on NT and k) for the case of normal errors, is . Achieving
subset HFINAL obtains the final scale estimate
. Then,
(in absolute value) is defined as the LTS standardized residuals.
8. The following is a simplified example. When we randomly choose the h sample with the smallest h value (NT/2 < h≤NT), we consider the LSDVC estimator based on the h sample to obtain Q1. Simultaneously, we choose the front h observations whose LSDVC residuals have been scored from least to greatest. Similarly, we can achieve Q2 under the same iteration, wherein we can find the final estimator based on the final h sample with the least Qfinal value. We can achieve the expected results.
9. The explanatory variables, Banki,t-1, Stocki,t-1 are assumed to be uncorrelated with future realizations of the error term (see Levine, Loayza, and Beck Citation2000).
10. Maksimovic and Demirguc-Kunt (Citation2001); King and Levine (Citation1993a); Levine, Loayza, and Beck (Citation2000); Benhabib and Spiegel (Citation2000); Shen and Lee (Citation2006).
11. Because our sample abundantly comprises 48 countries over 27 years, missing values must be present in the sample. To accomplish our work, we use the average value to replace the missing observation.