Donner index
5. The long-term impact of mergers and the role of macroeconomic shocks
5.4 What drives merger during the first phase of globalization in Germany?
5.4.1 How should I model the driving forces for mergers?
Before conducting a VAR analysis to measure the impact of mergers on share prices, dividends, and the nominal capital, it is essential to model the decision to merge. Note that mergers cannot be regarded as exogenous microshocks – but are the result of a decision process within a company and their shareholders. Company specific characteristics could play a crucial role for encouraging or preventing future expansion plans. One may argue that inserting the merger decision as additional endogenous variable into a broader VAR framework would work. However, as far as I know, VAR models that allow a binary choice variable as endogenous variable are not yet developed.155 Thus, I propose the following two-step approach: First, applying a panel probit model uncovers how the merger decision depends on the variables used in the VAR model. Second, I make a prediction with regard to mergers in the following year based on today’s knowledge. This enables to identify two types of errors, namely unanticipated mergers and falsely forecasted ones. Note that I mean with the term `falsely forecasted´ mergers that a merger is expected by market participants – but in fact is not executed. Both errors can be handled as exogenous shocks in the VAR framework;
therefore, I eliminate the endogenous nature of mergers.
5.4.2 Merger activity during the investigation period 1870 to 1913
Only a few studies on mergers during this time period are available for Germany. Moreover, as mentioned in section 5.2.1 the interest rests on the role of external growth and not on the success of mergers. Among these studies, Tilly’s (1982) contribution is the most noteworthy because he covered the period 1880 to 1913 and focused on large-scale enterprises. Thus, my sample should be comparable to Tilly’s (1982) data set. Note that in both cases annual data are used; however, Tilly (1982) worked with a different methodology to choose the cross-sectional units. Consequently, he claimed that only surviving companies were considered - but the number of companies in his sample increased over time. It stays unclear for the reader whether this varying number of observations stems from the inclusion of newly founded firms or is due to lacking information on the expansion of the respective enterprises.
Figure 5.5 provides an overview with regard to the number of executed mergers in my sample. This graph also embeds the number of takeovers provided by Tilly (see p.634, table 1, 1982). Compared to Tilly (1982), my sample shows that the mergers are centered around a peak in the year 1906. Note that Tilly (1982) did not include the banking industry. Moreover,
155 Note that it is possible to use the standard linear regression model to explain the merger decision – but it is well known that this is misleading for binary variables (see Greene (2000)).
he covered only the period from 1880 to 1913, and his sample size varied from 38 to 49. To take the changing sample size into account, figure 5.5 reports the number of mergers divided by the sample size. This can be regarded as a measure for the merger activity in the sample.
The difference could stem from the fact that Tilly (1982) excluded the banking industry. The exclusion matters because in my sample, banks account for 70% of all mergers. In addition, my short-term study emphasized the importance and the success of mergers among banks.
Hence, neglecting the banking industry, Tilly (1982) can only tell a part of the story.
Figure 5.5: Merger activity in the period 1880 to 1913
To compare my sample with the results of Tilly (1982), I calculate the number of mergers divided by the sample size. Thus, this is a measure for the merger intensity.
0%
5%
10%
15%
20%
25%
1880 1883 1886 1889 1892 1895 1898 1901 1904 1907 1910 1913
mergers (Kling) mergers (Tilly, 1982)
Noteworthy, Huerkamp’s (1979) figure about the merger activity is very close to my finding even though she excluded the banking industry. She uncovered a strong increase in the number of mergers from 1887 to 1907 – but her analysis is limited to a pure descriptive study because additional firm specific information was not collected. In contrast to Tilly (1982) and my study, she also included firms that are not listed on stock exchanges.
5.4.3 Model selection of a dynamic panel probit model with random effects
The decision to undertake a merger within one year can be regarded as a count data model if more than one merger is conducted. However, acquirers undertook rarely more than one merger within one year. Hence, it seems to be appropriate to model the binary decision of a company: to merge or not to merge. Because I deal with panel data, a panel discrete choice model should be applied. Fortunately, in recent years considerable progress in estimating panel probit respectively logit models was made.156 To account for company specific effects, I propose a random effects model of the following shape. Note that this specification can be justified later by running log-likelihood ratio tests.157
it
Obviously, mit takes the value one if company i executes a merger in year t and zero otherwise. The column vector ∆zi(t-j) consists of the first differences in real share prices, real dividends, and real nominal capital. Note that I take the natural logarithm before calculating the respective first difference.158 To determine the lag length of this dynamic model, I carry out the maximum likelihood estimation of (5.1) with different lag specifications. Thereafter, the Akaike and the Schwarz criterions are calculated together with the log likelihood of the respective model. To make a comparison reliable, the relevant sample is fixed.
Based on McFadden (1974), one can derive a pseudo R2; thereby, a probit model that includes only a constant term have to be estimated to obtain a reference basis for the log likelihood. Hence, the likelihood of every model specification j is compared to the log-likelihood of the reference model with constant term denoted ll0.
0
2 1
ll
RPSEUDO = −llj (5.2)
156 Hsiao (1992) provided an excellent overview of panel probit models.
157 Note that my distributed lag model could be affected by multicollinearity among lagged explanatory variables – but I concentrate on predicting mergers. Correspondingly, a high explanatory power is more important than precisely determined partial impacts.
158 Note that I take always natural logarithms when I refer to share prices, dividends, or nominal capital.
Table 5.4 reports the outcome of model (5.1) with different lag length p. The Akaike criterion indicates that the lag length should be set equal to two, whereas the Schwarz criterion favors the reference model with constant term. Regardless which lag structure is used, random effects matter indicated by log-likelihood ratio tests. The null hypothesis of no random effects can be rejected in all cases. This models point out that current and former changes in share prices and dividends as well as the growth rate of the net national product denoted gdpit
possess no impact on mergers. In contrast, higher inflation rates and mergers executed one or two years ago lead to a higher probability that during the year t company i announce a merger.
Using specification tests like log-likelihood ratio tests (LR),159 the number of explanatory variables can be further reduced. Finally, the probit model has the following structure.
All coefficients are significant on the 1% level of significance. How can one interpret this result? To facilitate the interpretation, one can calculate marginal effects. A merger in the prosecuting period increases the probability for an additional expansion by 8.08 percentage points. Mergers occurring two periods ago still influence the current merger activity with a marginal effects of about 9.20. Hence, there is evidence for the emergence of merger waves during the investigation period. Furthermore, an increase in inflation rates by one standard deviation adds 1.51 percentage point to the probability of mergers. To obtain an impression whether these marginal effects are essential, one should have in mind that the average forecasted probability for merger reaches 2.93%. Most notably, the current and past changes in company characteristics like share prices and dividends do not matter.
159 The test statistic reaches 8.58 and the corresponding p-value 0.804.
Table 5.4: Selecting the appropriate dynamic panel probit model
Applying a maximum likelihood estimation procedure, I run random effects probit models with different lags. To calculate the pseudo R2 statistic, a reference model with a constant term is estimated. The number of stars indicate significance. One star represents 10%, two 5%
and three 1% level of significance. The LR statistic tests whether random effects should be assumed; thereby, the null hypothesis states that individual effects can be neglected.
Basic model Without lags One lag Two lags Three lags Constant -2.0584*** -2.2969*** -2.5611*** -2.5159*** -2.5537***
mit-1 - - 0.8363*** 0.7361*** 0.6544***
mit-2 - - - 0.8133*** 0.7060***
mit-3 - - - - 0.2491
∆pit - 0.2763 0.8192 0.7474 0.7536
∆pit-1 - - 0.4246 0.7821 0.8031
∆pit-2 - - - -0.2076 0.0205
∆pit-3 - - - - -0.0096
∆dit - -0.0548 -0.1156 -0.1218 -0.1544
∆dit-1 - - -0.3365 -0.3243 -0.3669
∆dit-2 - - - -0.2160 -0.2714
∆dit-3 - - - - -0.2365
∆nit - 1.0266*** 1.2653*** 1.1530*** 1.2119***
∆nit-1 - - -0.2729 -0.0362 -0.0297
∆nit-2 - - - 0.4997 0.6487
∆nit-3 - - - - -0.4090
inflationit - 0.1895*** 0.2064*** 0.1871*** 0.1874***
inflationit-1 - - 0.0897 0.1160* 0.1364**
inflationit-2 - - - 0.0336 0.0522
inflationit-3 - - - - 0.0008
gdpit - 0.0027 -0.0117 -0.0147 -0.0217
gdpit-1 - - -0.0007 -0.0061 -0.0039
gdpit-2 - - - 0.0011 0.0006
gdpit-3 - - - - -0.0123
AIC 416.95 397.02 385.44 380.20 388.77
SBIC 422.14 428.19 452.98 478.90 518.65
Log-likelihood -207.47 -192.51 -179.72 -171.10 -169.39
Observations 1333 1333 1333 1333 1333
Pseudo R2 - 0.07 0.13 0.18 0.18
LR-test 43.42*** 48.89*** 22.45*** 7.81*** 6.40***
5.4.4 Predicting mergers during the period 1870 to 1913
In line with my descriptive finding that the merger activity increases after 1895, the probit model predicts higher probabilities for this period. However, the peak in the year 1872 seems to be misleading. Distinguishing between banks, which exhibited the highest activity in these years, and all other industries highlights the remarkable discrepancy between banks and other industries after 1900. Figure 5.6 plots the average expected probability for mergers based on the knowledge available in the previous year. Hence, the figure shows the one year forecast.
Figure 5.6: Predicted probability to merge for banks and other industries
I calculated the predicted probability that a firm will initiate a merger in year t based on the available information at t-1. This forecasts are obtained for every observation in the panel data set. Thereafter, I derive the average predicted probability and plot the 95% confidence intervals for this estimate.
year
1912 1907 1902 1897 1892 1887 1882 1877 1872
average merger probability
,3
,2
,1
0,0
-,1
BANK
0
1
To determine the expected merger, one has to specify a cutoff rate that has to be exceeded to expect a merger in the following year. Using a cutoff value around 0.125 enables to reach the highest possible accuracy for correctly anticipated mergers. However, only 32.08% of all mergers are correctly anticipated by the model; hence, many merger are not predictable.
Restricting the relevant time span to the period after 1896, which was characterized by the
new exchange law established in the year 1896, yields better forecasts. Note that the same cutoff value is optimal for the reduced time span; however, the probability of correctly anticipated mergers reaches 40.48%.
Consequently, I can now determine which merger can be anticipated and which one act as a microeconomic shock. Merger shocks can be treated as exogenous events that trigger responses in share prices, dividends, and nominal capital. Thus, by inserting the unpredicted mergers into a VAR framework, impulse responses can be obtained.
Obviously, I have to deal with the problem how to handle incorrectly predicted mergers. When I believe in the discrete choice model, then false predictions should lead to disappointments. Therefore, a negative microeconomic shock can affect prices, dividends, and the nominal capital. By defining two types of shocks, one obtains two types of potential errors that can be inserted into the VAR. The dummy variable denoted mp takes the value one if a merger occurs surprisingly, whereas the dummy mn stands for false forecasts. Embedding both surprising events into a VAR framework enables to test whether these mistakes trigger any consequences.