Gruppo CNR di economia internazionale Torino, 22-23 Febbraio 2007 “The decline in Italian productivity: new econometric evidence” by S. Fachin & A. Gavosto Discussion by Carlo Altomonte (Università Bocconi) The main contributions of the paper • The paper address the source of the productivity slowdown in Italy through panel cointegration methods • Three relevant contributions: 1. The use of panel cointegration methods allows to relax the assumption of CRS and perfect competition in factor markets, typical of the growth accounting approach 2. The use of a flexible form for the production function (Kmenta linearisation of the CES) allows to discuss the relevance of the labour/capital elasticity of substitution 3. The results are comparable with other, more traditional studies, as in their Figure 6, and thus more prone to a generalisation 1. Panel cointegration methods - I • The authors start from the labour productivity equation • They assume that technical progress pit can be decomposed as: pit = q t + i i and thus they can write: where the industry-specific technological progress is incorporated in a’ • But if the shock is not entirely separable (i.e. not every industry follows the same growth rate of technological progress over time), then the term becomes: pit = q t + ii + hit and thus eit in (3) should incorporate the term hit, leading to an error term potentially correlated with the input choices (simultaneity bias). 1. Panel cointegration methods - II • They implicitly discuss this issue by estimating • Here they assume that the term φ in (4) captures not only the trend in the technical progress but also the effect of any other random shock. • As a result, they have to impose an assumption on the distribution of these shocks: they are (log) additive and generated by a symmetric distribution, so that qs = E(φs| t=s) and φ can be estimated non-parametrically • However, again they implicitly assume that the shocks are entirely separable across industries, which in the panel structure is not likely  openly discuss the issue of simultaneity bias  elaborate more on the actual restrictiveness of the assumptions on q, and the potential correlation of shocks across time/industries (e.g. by discussing the characteristics of the VCov Matrix of their panel estimation) 2. Capital-labour elasticity of substitution • Their equation for log labour productivity (2) does not impose a unitary K/L elasticity of substitution r = 1 as in Cobb-Douglas • And yet, they cannot reject but for one industry (Rubber) the null of r = 1, although the confidence intervals of the point estimates are rather large • This result mimics Balistreri, McDaniel and Wang (2003) for the US economy (1947-1998, 28 industries), but the authors are still skeptical on r = 1 • Large confidence intervals are really caused by different underlying elasticity of substitution, or by the model design ? • One implication of aggregate equations like (2) is that, with r = 1, any redistribution of inputs across plants (i.e. any linear combination of b’x) results in the same aggregate output, which is not true: if firms are heterogeneous in productivity levels and new inputs flow to the most productive firms, different linear combinations would yield different results => “noise” in the estimates • Here the assumption on the elasticity of substitution is relaxed, but a potential bias from the aggregation problem persists => large confidence intervals could be due to the aggregation bias, while r = 1 holds. 3. Relevance of the results • If r = 1 (i.e. a Cobb-Douglas specification) cannot be excluded, what do we learn from the Kmenta specification ? Would the results dramatically change using a Cobb-Douglas ? And a translog ? • The broad dynamics of TFP recovered through the panel cointegration are similar to standard growth accounting analyses, but the order of magnitude is rather different: more than double of the growth rate in the period 1986-90 and almost a third in 1991-95: • Is this due to the fact that TFP estimates in this paper are valid only for the long-run ? • If this is the case, how big is the error incurred in trying to use these estimates in order to explain the business cycle ? • If TFP estimates are ok also for the short run, how to explain the much higher volatility of the business cycle implied by these TFP figures ?