Does NEEQ’s tiered system enhance corporate labour share？
新三板分级制度是否提升了企业劳动份额？

4.2.1. Heterogeneous treatment effect

The staggered DID methodology has become an essential toolkit for evaluating the causal effect of policy intervention. Scholars regard it as an equivalent estimator of DID, whose estimation of average treatment effect (ATE) is unbiased under the assumption of a parallel trend. However, recent studies point out that the variations over groups (control and treatment groups) and across time (ex-ante and ex-post period) lead the staggered DID to be unable to identify the unbiased ATE (Callaway & Sant’anna, Citation2021; Goodman-Bacon, Citation2021). Therefore, we further apply two methods to mitigate this concern.

(1) Propensity score matching (PSM) + DID test. Following Jiang and Zhu (Citation2022), we choose the implementation of NEEQ’s tiered system (in 2016) as the one-off exogenous shock event, applying the sample from 2013 to 2019 (ex-ante and ex-post 3 years, respectively), to conduct the PSM + DID analysis as follows.

We first match treatment firms (those that entered and remained in the innovative layer from 2016 onward, TREAT = 1) with control firms (those that never accessed the innovative layer, TREAT = 0) in 2016. Specifically, we perform PSM to identify control firm with near identical propensity scores for each treatment firm. In addition to all control variables in EquationEquation (8)(8) $\mathit{LSi},\mathit{t}={\beta }_0+{\beta }_1\mathit{ILA}{Y}_i{{,}_t}_{\mathit{\hspace{0.17em}}}+{\sum }^C\mathit{ontroli}{,}_t+\mathit{Y}\mathrm{ear}+\mathit{Firm}+\mathit{\epsilon i}{,}_t$(8) , we add variables that determine whether the firms can enter the innovative layerFootnote¹ (ISSUE, level of cumulative stock financing; EQUITY, value of shareholder equity; MV, market value; LNNONZERO, number of trading days) as the covariates. Panel A in Table 3 presents the results of the covariate balance test. PSM reweights each observation of the control group so that the mean and variance of all covariates are balanced (i.e. not statistically different) across the treatment and control firms. In particular, the covariate deviations are less than the standardised difference (20%) (Rosenbaum & Rubin, Citation1985), demonstrating an effective matching result.

We additionally set dummy variable POST that equals 1 for year 2016 onwards and 0 otherwise. Panel B in Table 3 presents the result of estimating the corresponding DID model. The coefficient of interest in this test is TREAT*POST. This coefficient is positive and significant at the 1% level for LS and LNLS, which is consistent with our hypothesis.

(2) Two-stage DID. Following Gardner (Citation2021), we adopt the two-stage DID model as follows. In the first stage, we identify and estimate the heterogeneous group and period effects with EquationEquation (9)(9) $Y_{0\mathrm{gpit}}={\lambda }_{\mathrm{g}}+{\gamma }_{\mathrm{p}}+{\epsilon }_{\mathrm{gpit}}$(9) by using the subsample of untreated observations. In the second stage, by subtracting the estimated group and period effects from the observed outcome, we regress the adjusted outcomes on treatment status (D_gp) with EquationEquation (10)(10) ${\hat{Y}}_{\mathrm{gpit}}=Y_{\mathrm{gpit}}-{\hat{\lambda }}_{\mathrm{g}}-{\hat{\gamma }}_{\mathrm{p}}={\beta }_{\mathit{gp}}{D}_{\mathit{gp}}+{\epsilon }_{\mathrm{gpit}}$(10) .

(9) $Y_{0\mathrm{gpit}}={\lambda }_{\mathrm{g}}+{\gamma }_{\mathrm{p}}+{\epsilon }_{\mathrm{gpit}}$(9)

(10) ${\hat{Y}}_{\mathrm{gpit}}=Y_{\mathrm{gpit}}-{\hat{\lambda }}_{\mathrm{g}}-{\hat{\gamma }}_{\mathrm{p}}={\beta }_{\mathit{gp}}{D}_{\mathit{gp}}+{\epsilon }_{\mathrm{gpit}}$(10)

where subscripts i, t, g, p represent firm, year, group and period, respectively. λ_g and γ_p represent the group and period fixed effects, respectively. Using the treatment among successive groups, we divide firms and time into treatment groups g and periods p, so that firms within group 0 are untreated in all periods, only firms in group 1 are treated in period 1, firms of groups 1 and 2 are treated in period 2, et cetera. Therefore, Y_gpit and Y_0gpit separately represent the observed and untreated potential outcomes for firm i of group g during time t of period p.

Table 4 presents the result of estimating the two-stage DID. When the dependent variable is LS (LNLS), the coefficient of ILAY is 0.0025 (0.0146). The magnitude of the coefficient is lower than the one 0.0108 (0.0638) in Table 2, yet significant and positive at the 1% level. These results indicate that our conclusion remains robust after considering the group and period effects.

4.2.2. Parallel trend

We follow Moshirian et al. (Citation2021) and set up a dynamic DID model to test whether the corporate labour share satisfies the parallel trend assumption before firms enter the innovative layer. Specifically, we set up dummy variables: ILAY (≤-4) equals 1 if the observation is 4 years or more before the firm becomes the innovative layer and 0 otherwise; ILAY (-i) (0<i ≤3) equals 1 if the observation is i year before the firm enters the innovative layer and 0 otherwise; ILAY (0) equals 1 when the firm accesses the innovative layer and 0 otherwise; similarly, ILAY (i) (0<i ≤3) equals 1 if the observation is i year after the firm enters the innovative layer and 0 otherwise; ILAY (≥3) equals 1 if the observation is 3 years or more after the firm enters the innovative layer and 0 otherwise. To avoid the trap of setting dummy variables, we excluded ILAY (≤-4). Table 5 Panel A presents the results of estimating the parallel trend. The coefficients of ILAY(-i) are not significant before the firm entered the innovative layer, verifying the parallel trend assumption. In contrast, the coefficients of ILAY(i)(i > 0) are significantly positive, further proving the causal relationship between NEEQ’s tiered system and corporate labour share.

It is worth noting that changes within the sample structure may unfavourably affect the dynamic DID test. However, the PSM sample with a one-time exogenous shock event (in Section 4.2.1) can alleviate this concern. With the above PSM sample, we set up the following dummy variables: ILAY (-i) equals 1 if the observation is i year before the firm enters the innovative layer and 0 otherwise; ILAY (0) equals 1 when the firm accesses the innovative layer and 0 otherwise; similarly, ILAY (i) equals 1 if the observation is i year after the firm enters the innovative layer and 0 otherwise. To avoid the trap of setting dummy variables, we excluded ILAY (−3). Table 5 Panel B presents the results of estimating the parallel trend with the PSM sample. The difference in labour share between treatment and control firms significantly increases only after the implementation of NEEQ’s tiered system, consistent with the result of Panel A.

4.2.3. Placebo test

A potential concern in the DID model is that some unobservable variables could lead to spurious causal inferences. Therefore, we further apply two placebo test methods to mitigate this concern.

(1) Random permutation. When the exogenous policy is fully affected by observable factors, OLS regression can estimate an unbiased coefficient of β₁ (EquationEquation 8(8) $\mathit{LSi},\mathit{t}={\beta }_0+{\beta }_1\mathit{ILA}{Y}_i{{,}_t}_{\mathit{\hspace{0.17em}}}+{\sum }^C\mathit{ontroli}{,}_t+\mathit{Y}\mathrm{ear}+\mathit{Firm}+\mathit{\epsilon i}{,}_t$(8) ) (Zhou et al., Citation2018). However, when we allow for the existence of unobservable influential factors, the estimate coefficient of β₁ in EquationEquation (8)(8) $\mathit{LSi},\mathit{t}={\beta }_0+{\beta }_1\mathit{ILA}{Y}_i{{,}_t}_{\mathit{\hspace{0.17em}}}+{\sum }^C\mathit{ontroli}{,}_t+\mathit{Y}\mathrm{ear}+\mathit{Firm}+\mathit{\epsilon i}{,}_t$(8) is as follows:

(11) ${\hat{\beta }}_1={\beta }_1+\mathit{\theta }\frac{cov(ILAY,\epsilon |\mathrm{z})}{var(ILAY,\mathrm{z})}$(11)

where z represents the vector of control variables. θ = 0 indicates an unbiased estimate coefficient of β₁.

Because θ = 0 cannot directly be predicted in the actual estimation, we replace ILAY with a variable irrelevant to labour share (i.e. β₁ = 0), and retrodict θ = 0 since ${\hat{\beta }}_1$ = 0. Specifically, by randomly assigning ILAY to each observation, we conduct 500 permutations of the data. Figure 1 depicts the distribution of the regression coefficients (${\hat{\beta }}_1$) for LS and LNLS. Both regression coefficients cluster approximately 0; thus, we can retrodict θ = 0, excluding the unobservable factors’ impact on our conclusion.

Figure 1. Distribution of regression coefficients for random permutation.

Display full size

(2) Random experiment. Following Li and Zhang (Citation2021), we randomise 8.6% of the sample as the treatment group (ILAY = 1) with the remaining as the control group (ILAY = 0) to estimate EquationEquation (8)(8) $\mathit{LSi},\mathit{t}={\beta }_0+{\beta }_1\mathit{ILA}{Y}_i{{,}_t}_{\mathit{\hspace{0.17em}}}+{\sum }^C\mathit{ontroli}{,}_t+\mathit{Y}\mathrm{ear}+\mathit{Firm}+\mathit{\epsilon i}{,}_t$(8) and replicate the above procedure 500 times. Figure 2 presents the distribution of regression coefficients and p-values for LS and LNLS. When the dependent variable is LS, the regression coefficient obtained by random simulation clusters approximately 0, with the mean value not significantly different from 0. It is worth noting that the coefficient (p-value) of our baseline regression is 0.0108 (0.002), which is entirely independent of the random distribution. When the dependent variable is LNLS, the results are consistent. Taken together, the placebo test reinforces our assertion that the tiered system has a significant impact on corporate labour share, and the treatment effect is not a statistical artefact.

Figure 2. Distribution of regression coefficients and p-values for random experiment.

Display full size

4.2.4. Firms withdrew from the innovative layer

According to our hypothesis, firms’ labour share will decrease after they withdraw from the innovative layer. Therefore, we apply the sample of firms exiting from the innovative layer to reversely verify the effect of the tiered system on labour share. We define the variable EXITFootnote² (firm exits from innovative layer equals 1 and 0 otherwise) and modify EquationEquation (8)(8) $\mathit{LSi},\mathit{t}={\beta }_0+{\beta }_1\mathit{ILA}{Y}_i{{,}_t}_{\mathit{\hspace{0.17em}}}+{\sum }^C\mathit{ontroli}{,}_t+\mathit{Y}\mathrm{ear}+\mathit{Firm}+\mathit{\epsilon i}{,}_t$(8) by replacing the independent variable ILAY with EXIT. Given that labour remuneration has strong cost stickiness (Chen et al., Citation2010), we use LS_t + 1 (LNLS_t + 1) and LS_t + 2 (LNLS_t + 2) as the dependent variables. Table 6 presents the results of relevant regressions. When the dependent variable is LS_t + 2(LNLS_t + 2), the coefficients of EXIT are negative and significant at the 1% level. The reverse analysis provides additional support for verifying the impact of the tiered system on corporate labour share.

4.2.5. Other endogeneity problemsFootnote³

(1) High-dimensional fixed effect. Although we control for year and firm fixed effects in EquationEquation (8)(8) $\mathit{LSi},\mathit{t}={\beta }_0+{\beta }_1\mathit{ILA}{Y}_i{{,}_t}_{\mathit{\hspace{0.17em}}}+{\sum }^C\mathit{ontroli}{,}_t+\mathit{Y}\mathrm{ear}+\mathit{Firm}+\mathit{\epsilon i}{,}_t$(8) , it is worth noting that some macro factors varying across different provinces or industries within different years may affect our conclusions. To mitigate this concern, following Dong et al. (Citation2020), we include Province*Industry*Year fixed effect to control for these unobservable macro factors. Table 7 presents the corresponding results, which are consistent with our main conclusion.

(2) Oster test. Oster (Citation2019) notes that the estimator β*=β* (Rmax, δ) is a consistent estimator of the actual coefficient when the model contains unobservable variables. δ is the coefficient of proportionality, which measures the relative explanatory power (R2) of observable and unobservable factors for the coefficient of interest. For instance, a δ of 3 indicates that for unobservable factors to overturn the result, they need to be three times as important as observables. Rmax is the maximum R2 when unobservable factors are all included. The technique has been widely used in the accounting literature (Call et al., Citation2018; Sun & Liu, Citation2022; Zhang et al., Citation2021).

Following Oster (Citation2019), we set R_max equal to 1.3 times R² in the baseline regressions. When β = 0, values of δ greater than 1 suggest a robust result. Table 8 presents the corresponding results. When the dependent variable is LS (LNLS), the value of δ is significantly greater than 1, mitigating concerns that unobserved heterogeneity drives our results.

(3) Excluding reverse causality. We address the possibility of reverse causality in which firms with higher labour share are more willing to enter the innovative layer. Specifically, following prior literature (Cao & Zhang, Citation2020; Glaum et al., Citation2018; Sun & Liu, Citation2022), we examine the impact of labour share (LS/LNLS) on firm’s decision to enter the innovative layer (ILAY). Given that the tiered system was implemented in 2016, we apply the sample from 2016 to 2020 in this test. Table 9 presents the corresponding results. The coefficient of LS (LNLS) is statistically insignificant, indicating that the labour share is not an important determinant for firms to enter the innovative layer.

4.2.6. Exclude potential alternative interpretation

(1) Exclude the influence of “firms’ different development stages”. Zhang et al. (Citation2012) argue that firms of different development stages may feature disparate levels of labour share. Meanwhile, the tiered system is designed to meet the differentiated financing, trading, service and regulatory needs of SMEs at different development stages. Therefore, a potential alternative interpretation is that the development stages, rather than the tiered system, lead to an increase in the labour share. To mitigate this concern, in addition to the existing control variable LNAGE in EquationEquation (8)(8) $\mathit{LSi},\mathit{t}={\beta }_0+{\beta }_1\mathit{ILA}{Y}_i{{,}_t}_{\mathit{\hspace{0.17em}}}+{\sum }^C\mathit{ontroli}{,}_t+\mathit{Y}\mathrm{ear}+\mathit{Firm}+\mathit{\epsilon i}{,}_t$(8) , we control for the following variables: LIFECYCLE, firm-level life cycle (Dickinson, Citation2011; Liang et al., Citation2019); LNESTAGE, natural logarithm of one plus the number of years a firm has been established. Table 10 presents the corresponding results. After controlling for the differences within firms’ development stages, our conclusion is still valid and robust.

(2) Exclude the influence of “the protection of workers’ rights and interests”. When firms enter the innovative layer, they face stricter regulations. Therefore, they may enhance the protection of workers’ rights and interests, such as improving employee welfare, providing five social insurance and one housing fund and standardising employees’ dismissal procedures. These protections would also lead to an increase in the labour share.

To mitigate this concern, we split the full sample into two subsamples based on the labour protection intensity of firm’s registered locale. Following Huang et al. (Citation2020), we apply the index of market-intermediate institutions and legal systems, LAW, to measure the intensity. Firm i belongs to the high (low) protection group if its LAW is greater than (less than or equal to) the annual median level of its industry. If the above logic holds, the impact of the tiered system on corporate labour share is expected to be more (less) significant for the low (high) protection group.

Table 11 presents the corresponding results. The coefficients of ILAY are all significantly positive with no significant difference between the subsamples, excluding the alternative interpretation.

5.3.1. The effect of financing constraints

In our hypothesis development, we point out that reducing the financing constraints is a channel through which the tiered system increases the labour share. Therefore, the effect of the tiered system and labour share may disparate at firms with different levels of financing constraints. According to EquationEquation (6)(6) $\mathit{r}=\mathit{p}\frac{\mathrm{\partial Q}}{\mathrm{\partial K}}-\frac{\mathrm{\partial R}}{\mathrm{\partial K}}\mathit{wL}$(6) , firms with low financing constraints are less likely to generate significant financing gains; even if they do not enter the innovative layer, they can determine labour shares approximate to the optimal marginal output returns. Nevertheless, the financing gains from capital investment may distort the income distribution of firms with high financing constraints. The effect of tiered system on the cost of capital reduction, however, can raise sufficient and timely capital to alleviate firm’s financing constraints. Therefore, firms allocate more investment resources towards labour factors, enhancing the corporate labour share. In this case, the impact of the tiered system on labour share is expected to be more (less) significant for firms subjected to higher (lower) financing constraints.

To verify this premise, we split the full sample into two subsamples based on firms’ levels of financing constraints. We calculate the financing constraints level, KZ, for each firm (Kaplan & Zingales, Citation1997). Firm i belongs to the high (low) financing constraints group if its KZ is greater than (less than or equal to) the annual median level of its industry.

Table 14 presents the corresponding results. The coefficients of ILAY are statistically significant only in the subsample of high financing constraint firms. The evidence suggests that the tiered system visibly affects labour share for firms facing higher financing constraints.

5.3.2. The effect of corporate information transparency

Financing constraints increase with the enhancement of information asymmetry (Hale & Santos, Citation2009). Therefore, the effect of entering the innovative layer on alleviating financing constraints may be more pronounced for firms with low information transparency. In this case, the impact of the tiered system on labour share is expected to be more (less) significant for firms with lower (higher) information transparency.

To verify this premise, we split the full sample into two subsamples based on firms’ levels of information transparency. We calculate the information transparency level for each firm (ABSDA, the absolute value of abnormal accruals with the modified Jones model) (Zhou et al., Citation2017). Firm i belongs to the high (low) information transparency group if its ABSDA is greater than (less than or equal to) the annual median level of its industry.

Table 15 presents the corresponding results. The coefficients of ILAY are statistically significant only in the subsample of low information transparency firms. The evidence suggests that the tiered system visibly affects labour share for firms featuring lower information transparency.

5.3.3. The effect of complementarity between labour and capital factors

Jiang and Zhu (Citation2022) point out that for firm with a higher complementary effect between labour and capital factors, reducing financing constraints is more likely to increase their labour share. This is because firm with a greater substitution effect between capital and labour investments, compared with labour investment, prefers to invest more capital goods represented by fixed assets. The unbalanced allocation of capital and labour leads to a decrease in production efficiency. Although reducing financing constraints cuts down the financing gains from the mortgage value of the fixed assets, firms can solve the deficiency by purchasing more capital goods rather than alleviating the labour shortage. Thus, firms may further invest more in fixed assets, hindering the tiered system in increasing their labour share. In this case, the impact of the tiered system on labour share is expected to be more (less) significant for firms with higher (lower) complementarity between labour and capital factors.

To verify this premise, we split the full sample into two subsamples based on firms’ levels of factor complementarity. We calculate the complementarity level (COMPLE) for each firm as follows:

We apply the time-series data of each sample firm from 2013 to 2020 and estimate using EquationEquation (15)(15) $\mathit{GF}{A}_{\mathit{i},\mathit{t}}={\beta }_0+{\beta }_1\mathit{GPA}{Y}_i{,}_t+\mathit{\epsilon i}{,}_t$(15) :

(15) $\mathit{GF}{A}_{\mathit{i},\mathit{t}}={\beta }_0+{\beta }_1\mathit{GPA}{Y}_i{,}_t+\mathit{\epsilon i}{,}_t$(15)

where GFA_i,t is the growth rate of net fixed assets of firm i in year t; GPAY_i,t is the growth rate of cash payments to and on behalf of employees of firm i in year t. COMPLE is the regression coefficient β₁. Firm i belongs to the high (low) complementarity group if its COMPLE is greater than (less than or equal to) the annual median level of its industry.

Table 16 presents the corresponding results. The coefficients of ILAY are statistically significant only in the subsample of high-complementarity firms. The evidence suggests that the tiered system visibly affects labour share for firms with a higher complementary effect between labour and capital factors.