Export behaviour of SMEs in transition countries

The SME definition follows the European Community definition, based on the number of employees: small firms (including micro-firms) have up to 50 employees; and medium firms have up to 250 employees.

See, for example, Bartlett and Prasinkar (1995), Futo et al. (1997), Scase (1998), McMillan and Woodruff (2002), Hoshi et al. (2003), Iakovleva (2005) and Estrin et al. (2006) among many other contributions.

We conceptualise firms’ export behaviour by taking into account not only the level of export activity but also the likelihood that firms will export at all.

Studies using BEEPS data include Carlin et al. (2001b), Vagliasindi (2001, 2006), Svejnar and Commander (2007), Gorodnichenko et al. (2010); and EBRD Transition Report (2005).

A detailed analysis of two recent large-scale surveys (Keupp and Gassmann 2009, surveying 179 papers; and Terjesen 2010, surveying 200 papers) shows that very few of them were related to transition countries, and none of them employed the large BEEPS databases, Melitz’s theoretical framework, or the methodology employed in this paper.

For any missing value in the dataset, we lose all other information related to a surveyed entity (as we have to drop the entire observation). This fact is usually ignored in empirical investigations.

For an extensive survey of this literature, see Greenaway and Kneller (2007).

There are other measures of assessing firm’s involvement in export markets. For instance, White et al. (1998) use three measures of export performance other than export intensity: number of foreign countries served by a firm, management’s perceptions of export profitability, and management’s satisfaction with export performance. Their discussion is inconclusive with regards to the best export performance measure. In their empirical investigation, they are rather pragmatic; they apply all four indicators to measure export performance in a sample of US service firms. Unfortunately, the dataset BEEPS is not so generous with information on export performance: the export intensity variable is the only information provided in all three rounds of BEEPS. Of course, export profitability also has its own additional drawbacks as a measure of export performance.

Changes in the organisational structure indicate organisational innovations. As these changes have at their core the human factor and its better utilisation, we have decided to place them within this category of factors.

For reasons that will be explained below, investment in R&D and gross investment can be used only in estimations from the 2002 dataset.

The three surveys are not consistent regarding the years or periods in which technology related variables are measured, thus causing much confusion. We summarise the situation as follows.

The use of the variable indicating the introduction of new or upgraded products and processes (in all three surveys) and the variables ‘gross investment’ and ‘investment in R&D’ for 2002 do not cause any endogeneity problem; these will have some effect on export intensity in a later period, but the current value of export intensity cannot affect the previous values of these variables. In cases where these variables and export intensity are measured contemporaneously, the problem of endogeneity precludes using those variables in the estimation process.

Of course, as Syverson (2011) explains, spatial proximity is not a prerequisite for generating productivity spillover effects. According to him (p. 349), “producers are likely to attempt to emulate productivity leaders…regardless of whether they share a common input market”.

We acknowledge that a dummy variable for location in a capital city cannot capture the full range and richness of agglomeration hypotheses. However, this variable does relate to the marked development of capital cities under transition. Unfortunately, the BEEPS dataset does not support more comprehensive proxies for agglomeration.

There are two questions on capacity utilisation in BEEPS: (1) In your judgement, what is your firm’s current output in comparison with the maximum output possible using its facilities/man power at the time? (2) What was the capacity utilisation 36 months ago? We use the second, backward-looking measure.

Our firm-level investigation and modeling strategy is not the appropriate platform for estimating the effects of national-level influences on firms’ export behavior such as free-trade agreements, macroeconomic developments (including policy) and institutional influences. Even a minimal specification to this end would require country (country-group) dummies, period dummies, and country (country-group)-period dummies to model political developments such as regional free-trade associations (especially where such developments come into force during the period of the sample). However, observations on these variables are available only in small numbers (there are 25 countries in our panel samples) and would be collinear with one another by construction, thereby precluding estimation with any useful degree of precision. Instead, we attempt to control for such influences in order to address potential sources of omitted variables bias. Here, our strategy rests on the ability of the firm-level fixed (i.e. time invariant or constant) effects in our model (see Section 3.1) to capture the influence not only of time-invariant variables (such as geographical characteristics) but also of “slowly moving variables”. Here, we follow Plümper and Troeger (2007, pp. 126), who cite Beck and Katz (2001): “… although we can estimate (…) with slowly changing independent variables, the fixed effect will soak up most of the explanatory power of these slowly changing variables”. This applies, in particular, to “politically relevant variables” such as trade agreements, macroeconomic policies and institutions. Even if such variables were not formally in force for the whole of the sample period, anticipated (leading), current and lagged effects—recognized, for example, in the literatures on trade agreements and macroeconomic policy—suggest that it is reasonable to think of their effects as sufficiently “slow-moving” over the sample period to be aggregated by time invariant effects at firm and/or country level. Accordingly, our panel estimates control for otherwise unmodelled systematic influences on the dependent variable at the firm level, which is the appropriate level for our investigation; in addition, country dummies control for any remaining systematic influences that vary between countries; and period dummies control for any remaining systematic influences that are common across all firms in the sample in a particular period. In the cross-section estimates, the country dummies control for otherwise unmodelled systematic influences on the dependent variable that occur in the period covered by the sample.

Maddala (1977, pp. 162–163) and Wooldridge (2002, pp. 518–519) discuss the use of tobit models to estimate models where the dependent variable is generated by, in effect, a dual decision making process: in our case, firms’ decisions as to whether or not to export and, if so, how much to export. The advantage of tobit estimation is that zero observations, which potentially yield useful information, are incorporated into the model as the outcome of a decision-making process. Moreover, truncation at one is unlikely to affect our estimates in a substantial manner: in our pooled sample, for example, only 1.35 % of firms generate 100 % of their sales from exports (4 % when the upper limit is set at 95 %). Nonetheless, we implemented two robustness checks to address residual concerns on this issue. We replicated our preferred model using our pooled sample: first, we implemented tobit estimation with censoring at both zero and one, and; second, we implemented the generalised linear model recommended by Baum (2008, p. 301) for modelling “proportions data in which zeros and ones may appear as well as intermediate values”. In neither case were the estimates substantially different from those reported below. Finally, we note that in Tobin’s (1956) original presentation of what came to be known as the tobit model, his dependent variable is a proportion. For these reasons, we disagree with Hobdari et al. (2009, p. 12) who criticise the tobit estimation of export intensity because this variable is “bounded by definition”. In our view, this neglects the dual decision-making process that informs the construction of the tobit estimator.

Random effects (RE) estimation is defined by the assumption that the independent variables are exogenous with respect to the group-specific (time invariant or fixed) effects. To minimise potential endogeneity of this kind, we specify a model in line with a wide range of theoretical influences in order to include in the estimated part of the model as many time-invariant determinants of firms’ export intensity as possible (Wooldridge 2006, pp. 481 and 493). However, we have stressed the limitations of theory, which suggests that we might not have captured all possible influences. Yet, many of our variables of interest are dummy variables, and these, according to investigation by Monte Carlo methods, may be estimated with correct coefficients and standard errors. Greene (2003a, p. 26) finds that: “In spite of the high intercorrelation of the (group-specific) effects and the regressors, the dummy variable coefficient and its standard error are estimated essentially correctly… Surprisingly, the marginal effect of the dummy variable is also well estimated…”. Table 2 establishes that the panel model includes 17 dummy variables and 10 continuous variables. Moreover, the groups of variables of particular interest—human capital and innovation/technology—both contain dummy variables, so that analysis does not depend only on continuous variables. There are, of course, remaining doubts concerning the validity of RE estimation. For this reason, we do not rely solely on panel analysis but also report cross-section estimates for three individual waves as well as for a pooled dataset.

Only a short description of the content of the BEEPS dataset is provided here. Various sample specific information—general and country specific—are provided in the reports accompanying the survey and datasets (see shttp://​www.​ebrd.​com/​pages/​research/​analysis/​surveys/​beeps.​shtml; accessed February, 2011). See also the EBRD Transition Report (2005).

BEEPS was conducted also in 1999, but this survey is omitted from our analysis as nonconformities with later rounds are too great; many variables covered in the later rounds were not included in the 1999 round.

From the dataset, we have dropped firms with over 250 employees (i.e. large firms). In addition, to preserve the randomness of the sample, we have dropped also the panel component of firms for 2005 and 2008/2009 and the so-called ‘manufacturing overlay’ (a group of additional companies surveyed outside the normal sample stratification in several countries in order to increase the weight of their manufacturing sectors). The SME component for different countries ranges from 80 to 85 %.

With regards to the panel sample, we employ only the “balanced panel component”, as imputing the unbalanced panel would mean violating the Missing Completely at Random assumption, crucial to the Multiple Imputation technique.

In addition to the usual descriptive statistics, we also examined the correlation matrix between our variables, paying particular attention to those related variables grouped together as “human capital”, “technology” or productivity-enhancing “spillover” variables. On conventional criteria (Taylor 1990, p. 37), only one correlation coefficient across all of our samples, and across all categories of interest, can be characterised as a “modest or moderate” correlation (i.e. between 0.36 and 0.67), otherwise, the largest correlations in each category are all “low or weak” (i.e. ≤0.35).

The missing values in our case are treated as non-response items, resulting from two sources: first, the interviewee did not know the answer or refused to reply, and second, the interviewer neglected to ask the question or did not record the answer.

Rubin (1987, p. 2) suggests m in a range of 2–10. However, Kenward and Carpenter (2007, p. 208) show that in some cases a larger m is required for reliable estimation and inference, especially in cases when the proportion of missing data is high. Because the percentage of missing data for some of our variables is relatively large, we apply m = 20. For practical implementation of MI, we use the routines written for STATA (see Royston 2005a, b, 2007; Carlin et al. 2008). The syntax written to implement MI for this paper is available on request.

Although this assumption cannot be tested, Schafer and Graham (2002) show that small violations of MAR usually have only a minor impact on estimates and standard errors.

The corresponding estimated conditional marginal effects are available on request.

Because of limited space in the table, the bootstrapped standard errors (using 50 replications) are not reported; they are available on request.

Greene and Wooldridge suggest that tobit estimates should be divided by the estimated standard error of the regression and then compared with the respective parameters of the probit model. If the tobit model is valid, then the ratios should be close—they cannot be equal due to sampling error—to the corresponding coefficient estimates in the probit model; otherwise the tobit estimates might be unreliable.

The detailed comparisons of tobit and probit estimates are reported for the panel and pooled samples in the Appendix, Tables 7 and 8. For reasons of space, these comparisons are not reported for the other three samples, but are available on request. Henceforth, the same applies to all empirical results referred to but not reported in detail.

In an attempt to find out whether majority foreign ownership has a different effect from any foreign ownership, the model was respecified using a dummy variable for majority foreign ownership, taking a value of one for companies with 50 + 1 % foreign capital and zero otherwise. The models in Tables 2 and 3 were then reestimated. The results were similar to those reported in Tables 2 and 3, where foreign ownership is measured by a continuous variable. (These additional results are available on request.).

See, for instance, Yudaeva et al. (2003) in the case of Russia; Konings (2001) for Bulgaria, Poland and Romania; and Djankov and Hoekman (2000) for the Czech Republic.

For a detailed discussion of the effects of financial crisis in TCs, see EBRD Transition Report (2009). A collection of papers in Bartlett and Monastiriotis (2010) concentrate on the effects of the crisis on SEE countries.