Common wave behavior for mergers and acquisitions in OECD countries?

a unique analysis using new Markov switching panel model approach
Chung-Hua Shen
Department and Graduate Institute of Finance, National Taiwan University

Shyh-Wei Chen
Department of Finance, Da-Yeh University

Mei-Rong Lin
Department of Money and Banking, National Chengchi University

Abstract
This paper investigates whether or not there is co-waved merger and acquisition (Mactivity in 26 OECD countries. We apply the Markov Switching model to panel data (MSP hereafter), an approach which has not previously been attempted. Two distinct regimes are recognized in emerge from MAdata: the wave merger regime and normal merger regime. Our MSP captures the co-wave pattern of the sample countries and has a much better fit than either the univariate Markov Switching model or the conventional linear panel model.

Citation: Shen, Chung-Hua, Shyh-Wei Chen, and Mei-Rong Lin, (2008) "Common wave behavior for mergers and acquisitions in OECD countries? a unique analysis using new Markov switching panel model approach." Economics Bulletin, Vol. 7, No. 8 pp. 1-12 Submitted: March 23, 2008. Accepted: April 21, 2008. URL: http://economicsbulletin.vanderbilt.edu/2008/volume7/EB-08G30007A.pdf

923). thereby causing total M&A activity to behave like a cycle. That is. where heated M&A activity occurs in one regime (wave regime) and inert M&A activity in the other (normal regime). empirical studies have. M&A activity echoes clear-cut life cycles. M&As from 1895-1979. M&A activity oscillates between high and low levels. The model typically separates data into two regimes. He provided evidence that the MS model fits the M&A data much better than linear ARIMA models.A. Recent anecdotal evidence documents that a fifth wave has since occurred. Among the first to explore the wave issue was Town (1992) who employed the univariate Markov Switching (MS hereafter) model of Hamilton (1989). April 6. Golbe and White’s (1993) finding of a sine-curve for merger behavior gave further evidence in favor of wave patterns.– one clustered around the turn of the century.A. p. and this at the end of the1990s. was likely a product of the fact the MS model imposes fewer limitations on data and can endogenously generate wave behavior through Hamilton’s filter. and this. however. Audretesch (1989) fully lent their support for the wave pattern theory by demonstrating that U. 1959. 1 . Scherer and Ross (1990) described aggregate merger activity as episodic and pointed out four prominent waves. business cycle effects (Nelson. whether M&A activity is characteristically in clear-cut waves. 2006. or peaks. he concluded. Post. reached mixed results.2 While the occurrence of such episodes seems to support the existence of M&A waves. 2 Bloomberg News. 846) do clearly advocate that “mergers come in waves” and consider the lack of any explanation for this phenomenon among the ten most important unsolved mysteries in the field of finance (p. stock market booms (Shleifer and Vishny (2003) and more.1 In the same vein. oligopolistic behavior reaction. By contrast.S. Shugart and Tollison (1984) rejected the notion of wave patterns because by using annual data on U. which they interpret as a clear sign of the wave phenomenon. Resende (1999) also used the MS model to capture merger wave behavior in sixteen sectoral M&A 1 Factors that cause M&A waves may include government encouragement and deregulation policy. two others in 1929 and 1968 and a fourth during the early and mid-1980s. 1996. Brealey and Myers (1991.1 Introduction Studies on the wave patterns of merger and acquisition (M&A) activity have recently received wide attention as researchers increasingly delve into whether there are oscillations between high and low levels of M&A activity. they found that M&A levels had been characterized by a unit root process.S. that is. Yagil. 1994).

We assume there are n = 1. 2. and the data span is t = 1. and found that the wave phenomenon exists in each sectoral data set. It is assumed that the unobserved state variable St follows a first order two-state Markov chain 2 . Furthermore. share a common pattern among all of these countries or not. To explore these issues. The present research departs from that trend by employing quarterly M&A data for 26 selected OECD countries. . and whether such M&A waves. The aim of this paper is to examine M&A wave activity among 26 selected OECD countries. or the “wave” period.. The exercise is worthwhile because scholarly interest in merger time series continues unabated. see Meese and Rogoff (1983) for exchange rates and Shugart and Tollison (1984) for mergers. . Our goal is to determine whether or not M&A activity in OECD countries occurs in intermittent waves. and regime two (St = 2) the low mean regime.A. . Supposing that the true data generating process comes from a two-regime Markov Switching model. . .data sets for the U. . or the “normal” period. the extant literature related to mergers makes it clear that there are two common characteristics which are shared by empirical studies concerning time series studies of exchange rates. if they exist. the focused series was once characterized by a random walk. for examples. for examples. Since the estimations from both the MS model and the MSP model have a higher mean in State 1. Interestingly. we refer to it as the wave period. T. . and U. Secondly. Past studies on M&A waves have exclusively focused on M&A activity in the U. while Section 4 presents the empirical results.K. Section 2 discusses the Markov Switching Panel Model we employ. . The remainder of this paper is organized as follows. the sixteen sectoral M&A waves appear to display co-movements. 2. recent studies show that the series is characterized by waves. see Engel and Hamilton (1990) for exchange rates and Town (1992) for mergers. we adopt the conventional univariate MS model as well as the MSP model. First. 26 which represent 26 individual OECD countries.K. Section 3 explains the data we use. the Markov Switching Panel (MSP) Model. Section 5 summarizes the findings and offers some concluding remarks. 2 Methodology Asea and Blomberg (1998) are credited with being the first to develop the MSP model which is briefly described below. and it achieves this by using a newly-devised model.S. we label regime one (St = 1) the high mean regime.

i. µ( j) = µ1 ( j) = µ2 ( j) = · · · = µ N ( j).    If the individual effect for each country is the same.  y Nt µ ( j) 1 0 ··· 0   1        µ ( j)   0 1 ··· 0   .e. σ2 ( j)) IN ×1 . and µn ( j) is the mean of the value of the M&A activity of the nth country at regime j.   ε Nt ( j)    . Thus: yt = Dµ( j) + ε t ( j).  . The “individual effect” in the MSP model is denoted as µn ( j). µ( j) =  2 .   . (1) Pr[St = 2|St−1 = 1] = 1 − p11 .Square Dummy Variable model (hereafter LSDV). Then the specification for the Markov Switching Panel model is written as: ynt = µn ( j) + ε nt ( j). . The term ε nt ( j) is an i.  . i. nt Because equation (2) is a modification of the Least. when St = j. 3 . (2) where ynt is the number of M&As for the nth country at period t.. Because the MSP-Pooled model is a newly-devised derivative of the MSP-LSDV model. σ2 ( j)) IN ×1 . We re-write equation (2) in a more compact form. D.     . ε t ( j) ∼ N (0.   .i. D =  .d random variable with mean E[ε nt ( j)] = 0 and 2 variance E[ε2 ( j)] = σε ( j). ε t ( j) =  2t .. Pr[St = 1|St−1 = 2] = 1 − p22 .     .and has the following transition probabilities: Pr[St = 1|St−1 =] = p11 . . It is also assumed that the probability for the initial state is ρ j = p(S0 = j). µ( j) and ε t ( j) are as follows:         y  1t   y2t yt =  . . ε t ( j) ∼ N (0.     µ N ( j) 0 ··· 0 1 ε ( j)   1t     ε ( j)  . . then the MSP-LSDV model is reduced to: yt = µ( j) D + ε t ( j). we can simply apply the likelihood ratio test to perform model selection. We refer to this as the Markov Switching Panel Pooled model (hereafter MSP-Pooled). .e. where the definitions of yt . and the error term for the ith equation in State j after the fixed effect is taking into account as ε nt ( j). . Pr[St = 2|St−1 = 2] = p22 . we refer to it as the Markov Switching Panel Least Square Dummy Variable model (hereafter MSP-LSDV). the ergodic probability for S0 = j.

and the U. Ireland. Norway. It is not easy to evaluate the variances of the two regimes since some countries. may be superficial because. Luxembourg. Spain. Germany. the Netherlands Antilles.85 to 0. Worth bearing in mind is that according to Shugart and Tollison (1984). They are Austria. 549 and 108 for Japan. are around 0. and these are all taken from Thomson Securities Data Company (SDC) data bank. The estimated transition probabilities. ADF) to check the stationary properties of the M&A data. the basic statistics and the results of the ADF test of the M&A activity in each of the 26 OECD countries are available from the authors upon request. respectively. Hungary. however.. the Slovak Republic. Irregardless. which would imply the rejection of wave patterns. albeit not all. Mexico. South Korea.S. 4 . Turkey. We examine whether the mean of the two regimes are the same or not. A switch between two regimes. p11 and p22 . however. We first apply the unit root test (which is the Augmented-Dickey Fuller test.3 Data and Results Twenty-six OECD countries are taken into account in this study. indicating that the duration period for remaining in a particular regime is quite long. we cannot reject the null hypothesis of the existence of the unit root. Poland. the variances across remaining countries are markedly different. Our null of the equal means (Engle and Hamilton. For example. This result. Denmark. There is no question that the mean of “the wave regime” (µ1 ) is substantially different from that of “the normal regime” (µ2 ). Iceland. Sweden. as stressed by Perron (1989). In this study.K. Finland. the ADF test has a tendency to accept the null when a series has experienced one or more structural breaks. 3 The scatter plots. have similar standard errors.A. 1187 and 844 for the U. They are also significantly different from zero at the conventional level. the Czech Republic. Italy.3 Table 1 summarizes the estimated results from the univariate MS model. and 3757 and 2345. because of a small t-ratio. respectively. µ1 and µ2 are 645 and 398.K. Quarterly data from 1990:Q1 to 2005:Q2 are used in our empirical study.98. is also one type of structural change which. as we claimed here. Greece. This motivates us to conduct our estimations from the regime switching model. there can be no wave if a series has a unit root. Portugal. the U. New Zealand.S. reported by Engel and Hamilton (1990). Japan. may lower the power of the ADF unit root test. 1990) is: ′ H0 : µ1 = µ2 . This property is similar to the “long swing” phenomenon for the nominal exchange rate. Switzerland. for Germany. for the U. in fact.

ˆ ˆ ˆ var ( p11 ) + var ( p22 ) + 2cov (ˆ11 . σ1 = σ2 . This may result in 4 Many tests in switching models suffer from the nuisance problem. New Zealand and Norway.and the Wald test statistic is: ˆ ˆ [ µ 1 − µ 2 ]2 ∼ χ 2 ( 1) . We next employ the MSP model to examine the co-movement feature of M&A activities. The null hypothesis of dependence is: ′′ 2 2 H0 : p11 = 1 − p22 . However. whether the number of M&A activities change for a period that cannot be completely independent of the regime that prevailed in the previous period. p22 ) The Wald test results are reported in the last column of Table 1. Third. i. Except for Austria. the wave regime and the normal regime. the estimated filtered and smoothed probabilities reach unity in 2000 for all OECD countries. the high estimated transition probabilities suggest that a M&A wave has the “long swing” property. Recall that Table 1 shows that variances across countries are substantially different.4 Estimated results of the null of the equal means ( H0 ) are reported in the second column from the right in Table 1. We omit the graphs of the filtered and smoothed probabilities of M&A for each country because of space limitations and they are available from the authors upon request. the merger activities are rather few and far between for the remaining OECD countries in 1997. three interesting findings are worth noting. Second. Engel and Hamilton (1990) suggested using the following the Wald statistic: ˆ ˆ [ p11 − (1 − p22 )]2 ∼ χ 2 ( 1) . Accordingly. Resende (1999) also found that sectoral patterns of M&A waves had co-movement. In addition. ˆ ˆ ˆ ˆ var (µ1 ) + var (µ2 ) − 2cov (µ1 . Because the Wald statistics are overwhelmingly significant. Readers are referred to Hansen (1992. We also examine whether the merger activities follow the Markov mechanism–that is. it is concluded that M&A activity does indeed switch between the two regimes. 1996) and Garcia (1998) for details. though the graphs are based on a univariate study. as discussed in Engel and Hamilton ′ (1990). 5 . and µ1 = µ2 .. the null hypothesis of equal means is overwhelmingly rejected. except for Mexico.e. indicating that it is a wave regime. there is a similar pattern of co-movement among the 26 OECD countries. First. we reject the notion that the transition probabilities are dependent. indicating the M&A data do have a Markov mechanism. µ2 ) This test statistic is free of the nuisance parameter problem.

Table 3 reports the estimated results for the MSP-Pooled and MSP-LSDV models. we first use the LR statistics to examine the existence of the individual effects.94.0 and −2274. the null might still be rejected. and distinct preference is given to the LSDV model over the Pooled model. With the log-likelihood values for the MSP-Pooled and MSP-LSDV models being −2261. The null hypothesis of no individual effects is: H0 : d1 = d2 = · · · = d25 = 0. the MSPLSDV is preferable to the conventional LSDV model. The log-likelihood values for the Pooled and LSDV models are −2797.5 + 2261. the discussion below is based on the MSP-LSDV model. respectively.93. respectively. Thus. which are 4. Figure 1 illustrates the average filtered probabilities from the MSP-LSDV model. 6 . Therefore.65 at the 5% level. these high estimated transition probabilities indicate that M&A waves have the “long swing” property. The LR statistic is 2 × (−1726. Once again. thus confirming the existence of common M&A waves. respectively.strong heteroscedasticity even though we control for the individual effect when applying the MSP model. Thus.4. which far exceeds the critical value of the Chi-square distraction of 37.5.2) = 1095.6 . which also exceeds the critical value of 42. as this is greater than the critical value of =67. thus.56.71 and 2. the MSP-Pooled model is outright rejected. values which are significantly different from zero. However. Again.2.5 and −1726. Again.5 The last two columns of Table 3 report the means of the wave regime and the normal regime. making the log-likelihood ratio (LR hereafter) 2 × (−2274. We also calculate the LR statistics using the conventional LSDV model as the null hypothesis (no switching) and the MSP-LSDV model as the alternative.96 and p22 = 0. We then examine the existence of the individual effect. respectively. the null of the no individual effect is decisively rejected. we first adopt the weighed least-square to remove the heteroscedasticity by dividing each country’s standard deviation.5 + 2274.5) =1070. The estimated transition probabilities for the two regimes are p11 = 0.50. Table 2 reports the estimated results from the conventional Pooled and LSDV models. the LR statistic is 2 × (−1726. This is highly indicative of the fact that the sample countries follow a Markov Switching process.2 + 2797. because the LR statistics are far greater than the critical values.0) = 1045. All of the coefficients of the dummy variables of all countries are significant except for d13 (Netherlands Antilles). According to 5 Note ′′′ that the LR test here suffers from the problem of the non-identified nuisance parameters.

there are three small peaks: at the beginning of 1992. P. Brealey. B. 33. 763–88. R. References Andrestsch. A. and S. C. 4th ed. 1995 and 2000. (1998). McGraw-Hill. B. The highest value falls in 2000. the standard line of 0. Given the weight of the evidence. Asymptotic null distribution of the likelihood ratio test in Markov-switching model. This implies that if industries decide to merge. Bloomberg (1998). in the 1997 sample period. That is.Hamilton (1989). The determinants of conglomerate mergers. the duration period for staying in a particular regime is relatively long as shown in the high estimated transition probabilities. the probability drop is similar to that of the univariate Markov switching model. K. our MSP model has a far better fitting than the univariate MS model and the conventional linear Panel model. 7 . Asea. There can be no doubt that the MS model and the MSP-LSDV model roughly show the same trend in most years. 687–713 Garcia. Journal of Econometrics. which coincides with the results from the univariate Markov Switching model. New York. P. D. American Economist. Hamilton (1990). 4 Concluding Remarks This paper investigates the co-waved merger and acquisition activity in 26 OECD countries. Dempster. Myers (1991). Also found is that M&A activity is characterized by long swing behavior. Principles of Corporate Finance. bf 83. and this is especially so in 2000 and 1997. Lending cycles. 89–128. 80. and B. an approach which has not been attempted in the past.. 39. Maximum likelihood from incomplete data via the EM algorithm. we believe the MSP can really successfully describe the M&A co-movement phenomenon in our sample. then the merger wave could persist for quite some time. Apart from this. 5–18.5 is placed. 52–60. Also. R. Two regimes are characterized in the M&As data: the wave and normal regimes. Our MSP cannot reject the co-wave pattern of the sample countries. Laird and D. The novelty of this paper is that we apply the conventional Markov Switching model to merger panel data (MSP here). (1989). Rubin (1977). and J. Engel. Long swings in the dollar: Are they in the data and do markets know it? American Economic Review. International Economic Review.. 28. D. As shown in the Figure.M. N. Journal of The Royal Statistical Society Series.

Journal of Econometrics.: A sectoral study. M. Merger and macro-economic factors. 45. 85–94. 7.K. Resende.Golbe. Industrial Market Structure and Economic Performance. Boston: Houghton Mifflin. 14. M. Perron. Rogoff (1983). 493–99. E. and White. B. (1992). Vishny (2003). (1996). Post. 3rd edition. D. 195–198. Analysis of time series subject to changes in regime. Scherer F. Empirical exchange rate models of the seventies: Do they fit out of sample? Journal of International Economics. Erratum: The likelihood ratio test under nonstandard conditions: Testing the markov switching model of GNP. Merger waves and the structure of merger and acquisition time-series. A. 8 . (1994). 70(3). L. (1992). Journal of Applied Econometrics. Town. Journal of Applied Econometrics. 3–24. (1999). Anatomy of A Merger. Econometrica. 39–70. Review of Financial Economics. London: Prentice Hall. J. The great crash. A New Approach to the Economic Analysis of nonstationary time series and the business cycle. 1361-1401. Hamilton. J. B. (1989). 181–90. 11. Stock market driven acquisitions. R. Review of Economics and Statistics. D. (1990). D. A. S61–S82. Hansen. R. (1989). 7. 57. Hamilton. 57. Oxford Bulletin of Economics and Statistics. J. The likelihood ratio test under nonstandard conditions: Testing the markov switching model of GNP. and R. (1993). 75. 5(2). Hansen. Journal of Applied Econometrics. A. 295–311. Meese. Yagil. 83–100. and K. 61. E. Catch a wave: The time series behavior of mergers. Wave behavior of mergers and acquisitions in the U. 357–384. P. J. Econometrica.M. Journal of Financial Economics. the oil price shock and the unit root hypothesis.153–159 Shleifer. David (1990). and R. (1996).

88)* 30.68)* 27.74)* 27.94)* 1.25 (1.90 (0.43* 158.02)* 0.08 (3.39)* ′ H0 ′′ H0 2.64* 244.16 (6.84 and 2.06)* 0.09* 381.48 (6.98 (4.85 (0.45)* 21.A.32 (0.23 (0.72 (1. * denotes significance at the 5% level.04* 56.98 (0.98 (0.77 (0.96)* 2354.58* 237.02) * σ1 39.92 (15.30* 98.41)* 72.66 (4.4)* σ2 13.72)* 844.91)* 31. U.98 (0.96 (0.95 (0.02 (1.56 (0.70)* 5.89 (6.05)* 0.00 (4.54)* 78.44* 83.53)* 12.90 (0.37 (4.06)* 0.82)* 24.89)* 27.78 (0.20 (9.05)* 0.03)* 0.95 (0.98)* 17.80 (3.12)* 0.09)* 0.37 (1.17)* 96.96 (0.08)* 0.02)* 165.00)* 67.98 (0.59)* 89.04)* 0.02 (0.18* 72.24* 115.98 (0.96 (0.44 (3.06 (2.06* 58.55* 9.18* 55.57 (1.92 (0.08)* 0.63 (0.68 (1.40* 86.04)* 0.85 (1.83 (2.34 (2.00)* 8.97* 69.02)* 0.47 (11.92 (0.22)* 4.86 (4.70)* 28.31 (13.08)* 0.01* 86.24 (5.31)* 135.49 (9.13)* 0.12 (4.85* 59.03)* 0.K.55)* 225.02)* 0.85 (0.76 (10.27)* 34.27)* 122.63)* p11 0.31 (3.08)* 0.71 (0.04)* 6.04)* 15.50 (3.11 (16.98 (0.94)* 16.08)* 18.05)* 0.83 (3.96 (0.81)* 61.08)* 0.92 (0.03)* 0.22* 704.92 (14.35 (1.68 (6.43)* 34.70)* 25.86* 163.59 (2.93 (0. respectively.90 (0.92 (0.16 (2.03)* 0.02)* 0.03)* 0.98 (2.28)* 398.78)* 9.22 (8.33 (8.85)* 77.78 (20.23)* 152. 3.76)* 44.44 (10.86 (0.46)* 49.58 (62.90)* 69.96 (0.35 (1.17 (10.36 60.24 (3.54)* 22.03)* 0.74 (9.54 (4.84 (8.98 (0.82 (110.55)* 22.68)* 11.12)* 81.93)* 84.67* 73.08)* 0.03)* 0.27 (31.96 (0.92* 136.09* 156.03)* 0.82* 97.71.45 (2.23 (43.98 (0.31 (7.36* 933.15* 141.97 (0.69)* 129.07 (2.63.41)* 1.21)* 108.39 (1.58 (4.14)* 28.36 (0.10)* 0.58* 34.43 (3.92 (0.25* 59.80)* 21.00)* 22.13* 114.77)* 35.96* 118.09)* 0.86)* 199.02)* 0.04* 11.76)* 103.28 (0.06)* 549.06)* 12.90* 20.16)* 44.08)* 250.95 (0.S.12* 60.06)* 0.96 (0.48)* 45.42)* 1.03)* 0.07* 51.97 (0.06 (2.05* 166.54 (0.70)* 17.98 (0.04)* 0.63 (16.43) 46.51 (3.79* 187.75)* 7.76)* 146. 9 .92* 163.02)* 0.92* 92.03)* 0.85* 46.88 (3.63 (3.35)* 15.75 (5.33)* 153.03)* 0.05)* 0.90 (0.93 (23.09)* 61.54 (2.48)* 58.42* 86.61)* 4. 5% and 10% level are 6.48)* 23.89 (0.10)* 0. µ1 73.22)* 12.96 (0.98 (0. ′ ′′ The critical values for H0 and H0 at the 1%.89 (0.12)* 20.09 (1.79 (6.87)* 28.12 (5.54* 161.61 (0.73)* 16.87* 11.90 (0.50 (3.42)* 11.77 (6.90 (0.91)* 15.85 (2.06)* 0.04)* 0.98)* 458.91)* µ2 58.66)* 128.02)* 0.10* 92.60 (4.84* 68.08)* 0.96 (0.84* 46.05)* 0.54 (0.04)* p22 0.28)* 645.39 (2.24)* 15.24 (5.99* 149.33)* 55.68)* 8.09)* 68.94 (0.26 (15.05)* 0.96 (0.01)* 90.07 (6.65)* 5.96 (0.90 (5.60 (2.34* The numbers in parentheses are standard errors.07* 33.02 (1.87* 102.90 (8.66* 78.27)* 9.94 (0.04) * 0.Table 1: Univariate Markov Switching model (1990:Q1–2005:Q2) Countries Austria Czech Republic Denmark Finland Germany Greece Hungary Iceland Ireland Italy Japan Luxembourg Netherlands Antilles Mexico New Zealand Norway Poland Portugal Slovak Republic South Korea Spain Sweden Switzerland Turkey U.74 (1.76 (0.56 (1.82)* 171.22* 58.81 (31.93 (0.09)* 0.12)* 11.00 (0.95 (0.71 (17.05 (23.79)* 217.03)* 0.56 (3.36)* 351.61)* 62.08)* 0.02)* 0.93)* 12.87 (0.92 (0.10)* 11.93 (0.03)* 0.94* 808.44)* 50.67)* 104.80 (0.29 (78.80)* 4.82)* 1187.21* 43.06)* 3757.48)* 97.36 (1.

18) −2.03)* LSDV Model 3.18)* −2.33 (0.17 (0.18)* −2.28 (0.11 (0.04 (0.38 (0.18)* −0.39 (0.2 Numbers in parentheses are standard errors.18)* 0.18)* −2.0 −2274.86 (0.18)* −1.59 (0.29 (0.18)* 1.Table 2: Conventional Panel model (1990:Q1–2005:Q2) parameter µ d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 d11 d12 d13 d14 d15 d16 d17 d18 d19 d20 d21 d22 d23 d24 d25 σ Log-L 1.75 (0.18)* −0.18)* −1. 10 .87 (0.09 (0.18)* −1.59 (0.30 (0.18)* −0.18)* −1.37 Pooled Model 2.78 (0.18)* −2.13)* −1.18)* −0.41 (0.07 (0.18)* −1.29 (0.18)* −1.68 (0. * denotes significance at the 5% level.18)* −2.86 (0.18)* −2.33 (0.57 (0.42 (0.73 (0.18)* 1 −2797.18)* −1.18)* −2.02 (0.18)* −0.18) −0.18)* −1.86 (0.

15 (0.57 (0.99 (0.55 (0.98 (0.13)* −1.26)* 0.24)* −1.05)* 0.96 (0.97 (0.01)* 0.62 (0.57 (0.13)* −0.95 (0.13)* −1.24 (0.56 (0.73 (0.31)* −2.82 (0.48 (0.40 (0.89 (0.27)* 1.68 (0.02)* −1.90 (0.94 (0.93 (0.23)* −1.22)* −1.27) −0.02)* 0.62 (0.16)* −1.87 (0.29)* −1.28)* −2.29) −2.56 (0.62 (0.01)* 0.14)* −1.Table 3: Markov Switching Panel Model (1990:Q1–2005:Q2) parameter MSP-Pooled Model MSP-LSDV model St = 1 µ(1) µ(2) d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 d11 d12 d13 d14 d15 d16 d17 d18 d19 d20 d21 d22 d23 d24 d25 σ (1) σ (2) p11 p22 Log-L 1.13)* −0.26)* −0.01 (0.13)* −1.26)* 0.47 −1726.23)* −0.21)* −0.75 (0.17)* −2.64 (0.17 (0.13)* −2.29) −1.08 (0.71 (0.47 (0.29)* −0.83 (0.82 (0.68 (0.30)* −2.66 (0.93 (0.37 (0.14)* −1.25 (0.18)* −0.17)* 2.13)* −2.13)* −2.01)* 2.16 (0.32) −0.37 (0.29)* −0.13)* 1. * denotes significance at the 5% level.61 (0.24)* −0.09)* St = 2 −0.91 (0.04 (0.24)* −2.13)* 0.14)* −0.68 (0.01)* 0.12)* −1.76 (0.98 (0.42 (0.64 (0.74 (0.01)* −2261.17 (0.13)* −2.62 (0.14)* 0.30)* −2.50 (0.50 Numbers in parentheses are standard errors.50 (0.23)* −2.13)* −2.39 (0.03)* 0.23)* −1. 11 . 55 (0.37) −2.05)* 4.76 (0.14)* −0.13) −2.08 (0.49 (0.58 (0.19)* −0.13)* −0.46 (0.13) −0.94 (0.

0 0.8 0.2 0.0 1990 1992 1994 1996 1998 2000 2002 2004 Figure 1: Average filtered probabilities from the MSP-LSDV model (1990:Q1–2005:Q2).5 rule 0.The average filtered probability from the MSP-LSDV model 1. 12 .4 0.6 0.