multi-product firms, we consider single-product firms only.15 The underlying latent
model for the sector switching decision is presented in equation (4), where sijt
represents the sector switching decision for firm i, in sector j, at time t. Vectors xijt-1 and
zjt-1 are one-period lagged firm specific and sector specific explanatory variables,
respectively, and the unobserved heterogeneity is treated as a random effect, υi .
sijt*=α0+xijt-1'α1+zjt-1'α2+υi+eijt (4)
where sijt = 0 if sijt*≤ 0 and sijt = 1 if sijt*> 0 and α0 , α1 and α2 are parameters to be
estimated.
4. Data
Data come from the Vietnamese Census of Production for 2001-2004 provided by the
General Statistics Office of Vietnam GSO (2005).16 The dataset includes all registered
enterprises at the end of each year, and we consider 19 two-digit level sub-groups of
the manufacturing sector (detailed in Table 1). The total sample consists of 61,510
observations on 23,916 manufacturing firms. We exclude, as alluded to above, firms
alternative, the unobserved effects can be treated as a random effect. It should be noted, however, that
consistent estimation of the random effects model by maximum likelihood requires the assumption that
the errors are independent of other regressors in the model. Since much of the focus here is on sector
specific explanatory factors, correlations between the regressors and the individual effects are of less
concern.
15 Bernard et al. (2006b) examine the implications of unobserved product-mix variation and product
switching for the measurement of firm and sector level productivity. They demonstrate that production-
technology differences across products and product-choice variation across firms interact to bias standard
production function based estimates of firm productivity. When firms endogenously choose between
products with heterogeneous techniques, standard measures of TFP will be systematically biased (Bernard
et al., 2006b). Olley and Pakes (1996) and Levinsohn and Petrin (2003) control for several sources of
measurement error, but they do not eliminate the bias in productivity measures due to endogenous product
choice. We correct for this by sorting firms into groups that make a single product, and measure
productivity across firms making the same product. We also note that multi-product firms may have
different switching behaviour than single product firms.
16 A more thorough description of the data is given in the Appendix and in GSO (2005).
12