7. Results of the empirical analysis
7.1. Results of the Logit estimation on the whole panel
Table 10 present the results for the whole panel of 27754 manufacturing firms.
Three alternative models (with a different vector of explanatory variables) have been estimated. The first one use CR4 as a indicator of market concentration11 in a simultaneous specification; the second and third models use (respectively) a one-year and two-years lag of CR4 instead, which allows to reduce potential simultaneity biases. The effects of the “simultaneous CR4” and of the “1-year lagged CR4” on the probability to innovate are almost identical: both are significantly positive at the 1%
level, with a low magnitude (the parameter value is 0.008, which means that the odd ratio increases proportionally with the CR4).
The similar effects of the ‘simultaneous’ and ‘1-year lagged’ CR4 probably come from the fact that the concentration ratio stays very stable over time (around 27% on average each year) in our panel, which is not surprising: on a short observation period such as 1992-1995, the degree of concentration in an industry is unlikely to experience important variations. Moreover, the effect of the 2-years lagged CR4 is also very similar in terms of magnitude and standard deviation; it is, however, less significant (10% level).
A second important result regards the effect of firm size on the probability to innovate: in all three models, we find that very small firms (with less than 50 employees) have a significantly lower probability to innovate. Then, as firms size grows from 50-99 employees to 100-499 employees and 500-999 employees, the
11 We also used more sophisticated indicators such as the Herfindhal index (in both contemporaneous and lagged specifications), but this doesn’t significantly change the final results.
probability to innovate increase. Larger firms (1000 employees or more), however, have a lower probability to engage in innovation. To summarize these results, it can be said that firm size has a significant non-linear effect on the probability to innovate:
this probability first increases as firm size increases, but begins to decrease after a certain threshold has been reached.
Table 10: parameter estimates of the panel Logit model (27754 manufacturing firms)
Variables Model 1 Model 2 Model 3
Constant -3.587 (0.44)*** -4.396 (0.503)*** 0.685 (1.707) Grexp86-91 0.119 (0.03)*** 0.087 (0.053)* 0.195 (0.555)***
Grexp91-96 0.005 (0.005) 0.0006 (0.006) 0.012 (0.009)
Et-1986 -0.146 (0.035)*** -0.201 (0.041)*** 0.451 (0.183)**
CR4_lag1 0.006 (0.002)***
CR4_lag2 0.006 (0.003)*
D1 -1.187 (0.271)*** -0.872 (0.295)*** -0.950 (0.543)*
Observations 103744 77808 51872
Log Likelihood -21576.66 -15741.58 -10160.32***
Chi-Square 4283.55*** 3844.71*** 2552.55
Rho 0.773 (0.001)*** 0.829 (0.001)*** 0.916(0.001)***
Standard errors are reported in brackets. Significance levels are: *10%, **5%, and ***1%.
NB: gray cells indicate that variables are not included in the regression model.
The one before last line reports the Chi-Square associated with the LR test of the null hypothesis H0: “β = 0” ; the stars indicate the level of significance at which H0 is rejected.
These findings are consistent with those observed in the empirical literature; in particular, we find, as Cohen et al. (1987), that the size of a business unit affects its decision to innovate. That this effect may be non linear should not be surprising, if one considers that a non-linear relationship between R&D intensity and firm size has
been identified in a number of previous studies (Scherer, 1965a, 1965b; Malecki, 1980; Link, 1981; Bound et al., 1984).
Our third important result regards firms’ age: on the average, older firms have a significantly lower (1% level) probability to innovate. Again, this result is consistent across all three models. Let us now focus on the results related to the economic context of Taiwan in the 1990s and the late 1980s.
First of all, our indicator of the fluctuations of Taiwan’s currency (the variation in the exchange rate of the US $ / NT $ between 1896 and year t) has a significantly negative effect on the probability to innovate in models 1 and 2. This effects become positive, but less significant, in the third model. A plausible explanation of that result is that, by construction, the indicator of the fluctuations of the NT $ captures some of the time specific effect. For this reason, the results may change when the panel of firms is observed on a shorter period of time (the third model, with a 2-years lag in CR4, can only be run for years 1994 and 1995).
Second, the growth of the export market has over the 1991-1996 period has no significant effect on the probability to innovate: thus, firms’ decision to innovate may not depend on the current growth of the external market. However, the past growth of the external market, captured here by the growth of Taiwan’s exportations over the 1986-1991 period, seems to matter: it has a significantly positive effect on the probability to innovate in the first and the third models. In the second model, however, the effect is not significant, although the parameter remains positive
This puzzling result may be partially explained by the relative importance of each industry: the foreign demand is important only for some industries, whereas some industries, like the chemical of metal industries, have always sold their product on the domestic market only. This explanation is reinforced by further empirical investigation: when running the Model 2 on all industries but the Metal and Machinery category, the parameter associated to the “Growth of Exportations86-91” variable becomes significant and positive again. In order to better understand this result, we will examine the results of these regressions by category of industry in Sub-Section 7.2.
Finally, the last line of the Table 10 features a parameter Rho which captures the contribution of the variance the individual (random) effect to the total variance of the dependent variable. Here, Rho is significantly different from zero (at the 1% level)
in all three models, which means that the individual effect does capture some unobserved heterogeneity.