Bribes, Rents and Industrial Firm Performance in Albania and Kosovo

Abstract

Using data from a novel representative survey, we examine how corruption affects the performance of industrial-sector firms in Albania and Kosovo, two low–middle-income post-socialist economies in the Western Balkan region. Bribes are costs that firms incur to “grease the wheels” of the bureaucracy and/or to seek rents. Rents, however, may improve firm performance. Thus, to estimate the total effect of corruption, we develop and collect a set of perception-based indicators of both corruption and rents. In addition, we allow both bribery and rents to affect output growth through multiple channels—by influencing the firm’s investment and hiring decisions, by affecting total-factor productivity and by modifying the marginal product of factor additions. We find that, in Albania and Kosovo, bribes and rents have both positive and negative effects on firm performance. The net effect of corruption, however, is negative and large and is not fully offset by the beneficial effects of rents.

Appendices

Appendix 1: Survey Design Characteristics

The survey interviews were conducted face to face by nine enumerators, who were trained by the authors. All enumerators were native speakers of Albanian, expect for one Kosovar enumerator, a native Serbian speaker, who was responsible for conducting interviews in the Serb-majority municipalities of northern Kosovo. The respondent was always either the company’s owner or a top manager. To ensure consistency and data quality, the authors supervised the firm selection process and audited some 100 interviews across the two countries.

The survey employed a two-stage sampling design. Using information from official business registries, the total population of industrial firms in Albania and Kosovo was first partitioned into 22 country sectors, or “strata”. Next, individual firms were selected independently within each stratum. Stratification ensures that firms from more successful sectors, which are naturally more numerous, are not overrepresented in the sample (Levy and Lemeshow 2008, p. 123).

Furthermore, the sampling procedure was such that larger firms (measured by the number of full-time employees) were more likely to be selected into the sample than smaller firms from the same stratum; firms of equal size were equally likely to be sampled. To implement this scheme, the stratum population was ranked by firm size. The interviewers were instructed to first telephone and arrange interviews with larger companies, working their way down in descending order. The randomness of selection arises from the possibility that a firm may not answer the phone call or may decline to be interviewed.

Table 8 provides descriptive information on the 22 country sectors, including the sample ($n_{h}$) and population ($N_{h}$) size for stratum h, the sampling rate ($f_{h} = n_{h} /N_{h}$) and the share of the sample ($n_{h} /\sum n_{h}$) and population ($N_{h} /\sum N_{h}$) pertaining to each stratum.

Table 8 Sample and population characteristics.

Appendix 2: Factor Analysis of the Rent Dummies

The factor analysis estimates the following system of equations, written in matrix form:

$$\overrightarrow {Rent} = \underline{A} \vec{f} + \vec{u}$$

where $\overrightarrow {Rent}$ is the vector of observed rent dummies, $\vec{f}$ is a vector of factor scores, $\underline{A}$ is the pattern matrix and $\vec{u}$ is the uniqueness vector. The elements $a_{mn}$ of the pattern matrix ($\underline{A}$) are known as “factor loadings”, since they measure the partial correlation between each factor and the observed variables. Estimating $\underline{A}$ requires solving the following eigenvalue problem: $\underline{\varSigma } \vec{a}_{mi} = \gamma_{i} \vec{a}_{mi}$, where $\underline{\varSigma }$ is the correlation matrix of the observed rent variables, $\vec{a}_{mi}$ is the vector of partial correlations between $fi$ and the four rent variables, and $\gamma_{i}$ is the eigenvalue.

A critical assumption of factor analysis is that the observed variables are continuous. Our $Rent$ variables, however, are binary. A factor analysis of the Pearson correlation matrix for a set of dummies can be severely misleading. The standard solution is to factor-analyse the matrix of tetrachoric correlations between the observed dummies (StataCorp 2014, p. 2679). A tetrachoric correlation “estimates the Pearson correlation of the latent continuous variables” (ibid.). The assumption is that the respondents’ agreement/disagreement with questions 1.1–1.4 (Table 2) is just a coarse measurement of an underlying variable—the degree to which firms in a given sector receive rents.

Having estimated the pattern matrix $\underline{A}$ using a principal component method, we performed a rotation of the matrix elements, a standard procedure intended to simplify factor structure and aid interpretation of the factor scores (Hamilton 2013, p. 318). We employed an oblique promax rotation, which “simplifies factor patterns while permitting some degree of correlation between the factor [scores]. Correlated factors will be statistically less parsimonious […], but [potentially] more realistic” (Hamilton 2013, p. 340). The results, however, are very similar if we choose an orthogonal varimax rotation instead. We only retain factors whose eigenvalue is greater than 1, implying that they explain more than the equivalent of one observed variable’s variance (Hamilton 2013, p. 316).

Appendix 3: Sampling Design and OLS Estimators

Under our sampling scheme, a stratum is defined as the firm’s country-sector. Each stratum $h$ consists of a finite population of firms ($N_{h} < \infty$). To make a finite-population correction, we define the “sampling rate” $f_{h}$ for stratum $h$ as the ratio of sampled individuals to the size of the population in $h$: $f_{h} = n_{h} /N_{h}$.

Next, we weighted our observations by their probability of being selected into the sample. The “sampling weight” $w_{ih}$ for observation $i$ in stratum $h$ is proportional to the inverse of the probability of i’s being selected into the sample. Under our sampling scheme, the probability of selection depends on the stratum’s population size, but also on the size of the company, measured in terms of the number $L_{i}$ of full-time employees. Thus, following Levy and Lemeshow (2008, p. 350), we define the sampling weight as follows:

$$w_{ih} = \frac{C}{{P_{ih} }} = C\left( {\frac{{N_{h} }}{{n_{h} }}} \right)\left( {\frac{1}{{L_{i} }}} \right)$$

where $P_{ih}$ is the probability of observation i being selected and $C$ is a constant that we set equal to 1. To perform the estimation, we use Stata’s svy suite of commands for survey data (StataCorp 2013).

Appendix 4: Output Growth Models: Additional Specifications

Here, we further investigate the robustness of the results presented in column 4, Table 5. In that model, the fact that the coefficients on $f2$, $\left( {f2 \cdot \Delta \ln K} \right)$ and $\left( {f2 \cdot \Delta \ln L} \right)$ are all insignificant may result from multicollinearity. Thus, model 1 in Table 9 omits the two interaction terms⁵⁰ and recovers the positive and significant coefficient on $f2$ estimated in models 2 and 3 (Table 5). Although the rents captured by $f2$ (import protection and protection from unfair competition) do not exert a moderating influence on the marginal product of labour and capital, they do appear to stimulate TFP growth.

Table 9 Output growth models: additional specifications

Lastly, model 2 in Table 9 tests the robustness of our results to an alternative fixed-effects specification. The ability of firms to generate sales revenue may depend on sector—(e.g. prices) and/or location-specific (e.g. institutional) effects. Omitting these determinants of firm performance may bias the parameter estimates. For instance, an exogenous change in market prices may allow firms to increase revenues without improving product quality or increasing the scale of production; and this effect may be spuriously picked up by our rents or corruption variables. Alternatively, governance quality may vary across regions, and these differences may influence both the performance of local firms and the incidence of corruption (e.g. in the municipal bureaucracy). To address these possibilities, model 2 in Table 9 investigates the sensitivity of our results to replacing the country fixed-effects dummy (Albania) with a full set of country-sector and location (city) fixed effects.⁵¹ The location fixed effects (but not the country-sector fixed effects) are jointly significant. Even so, our results are qualitatively unaltered (although, of course, the OLS standard errors are generally larger in this much more extensively specified model).