The role of investments in export growth

Abstract

In an increasingly globalised world, exporting plays a central role for economic growth and poverty reduction, particularly in small open economies. In this study, we test the hypothesis that a rise in investment favours entry into export markets and increases exports among firms that are already exporting. We address causal links through impact evaluation techniques for observational data. We examine the binary case and also continuous analysis of investment as treatment. We analyse a panel of Uruguayan manufacturing firms in the period 1997–2008, and we find evidence that investments “cause” exports and export orientation, and this provides a rationale for carefully designing investment promotion policies rather than focusing on other export support policies.

Appendix: Total factor productivity estimation

The estimation of firms’ TFP is carried out using structural techniques: the Olley and Pakes (1996), Levinsohn and Petrin (2003) and Ackerberg et al. (2006) methodologies. All these techniques use observed input decisions to control for unobserved productivity shocks, thus addressing one of the main endogeneity problems that usually arises in empirical estimations of production functions at the micro-level, the so-called simultaneity bias (i.e. the fact that firms’ input choices may respond to productivity shocks). Estimating TFP using Olley and Pakes (OP) and Levinsohn and Petrin (LP) involves different proxy variables: while OP use investments, LP use either materials or energy. Nevertheless, LP have been questioned due to serious collinearity problems, which are addressed by means of the ACF technique.

The OP and ACF methodologies differ in the choice of the state variable: it can be materials used or energy—electrical energy or fuels—as a proxy for unobserved productivity shocks. In OP, the state variable is capital, chosen in period t − 1, and labour adjusts freely in period t. While in ACF, labour is not considered to freely adjust in t, but somewhere else before, since there may be rigidities in the labour market like labour regulations that prevent free adjustment.

We estimate the following Cobb–Douglas production function:

$$y_{it} = \beta_{sl} sl_{it} + \beta_{ul} ul_{it} + \beta_{k} k_{it} + \beta_{m} m_{it} + \varpi_{it} + \eta_{it}$$

(1)

where y _it is gross output (we also tested value added), sl _it skilled labour, ul _it is unskilled labour, m _it is materials and inputs, and k _it capital stock of firm i at time t (all variables in logarithms); and ω _it and ω _it are firm- and time-specific unobserved shocks (ω _it is a productivity shock that affects the firm’s input choices, while η _it is an i.i.d. shock that has no impact on the firm’s decisions).

The residual of Eq. (1) is the firm’s TFP, retrieved from the estimated coefficients as:

$$TPF_{it} = y_{it} - \hat{\beta }_{sl} sl_{it} - \hat{\beta }_{ul} ul_{it} - \hat{\beta }_{m} m_{it} - \hat{\beta }_{k} k_{it}$$

(2)

We estimate the production function for the full set of observations since the number of firm-year observations is small and the estimation of TFP may be sensitive to the number of firm-year observations. In the upper panel of Table 20, we present the coefficients of the production function for the three techniques when the dependent variable is value added; in the lower pane, we present the results for gross output as dependent variable. As proxy variables, we use investments for OP and electrical energy for ACF and LP.

Table 20 Production function estimation

The estimated coefficients of the production function are not directly comparable to previous works such as LP (2003) or ACF (2006) since in those works the estimations of the production functions are performed at the 3-digit ISIC industry level. Nevertheless, we find reasonable estimates for capital and labour, though smaller than those found by ACF for three industries and alike the Wood industry in ACF (2006). Furthermore, the standard errors (SEs) and returns to scale are also smaller. Moreover, we find some differences in the estimated coefficients depending on the methodology used. Nevertheless, the estimated TFP are highly correlated, as can be seen in Table 21.

Table 21 Correlation matrix of TFP estimated using different methodologies