We revisit identification based on timing and information set assumptions in structural models, which have been used in the context of production functions, demand equations, and hedonic pricing models (e.g. Olley and Pakes, 1996, Blundell and Bond, 2000). First, we demonstrate a general under-identification problem using these assumptions in a simple version of the Blundell–Bond dynamic panel model. In particular, the basic moment conditions can yield multiple discrete solutions: one at the persistence parameter in the main equation and another at the persistence parameter governing the regressor. We then propose a possible solution in the simple setting by enforcing an assumed sign restriction and discuss more general practical advice for empirical researchers using these methods.

We study productivity differences in vertically integrated steel facilities using equipment-level information on inputs and output for each of the main stages in the value chain. We obtain stage-level productivity estimates by estimating a multistage production system and then integrate them into estimates for integrated facilities. At this level, we do not find statistically significant differences in productivity by ownership. This conceals important differences upstream and downstream in the value chain: private firms outperform in pig iron and steelmaking but lag in sintering. Inferior access to higher-quality raw materials and use of less automated technology are likely sources of these differences.

Summary We study Ackerberg, Caves, and Frazer's ( Econometrica , 2015, 83 , 2411–2451; hereafter ACF) production function estimation method using Monte Carlo simulations. First, we replicate their results by following their procedure to confirm the existence of a spurious minimum in the estimation, as noted by ACF. In the population, or when sample sizes are sufficiently large, this “global” identification problem may not be a concern because the spurious minimum occurs only at extreme values of capital and labor coefficients. However, in finite samples, their estimator can produce estimates that may not be clearly distinguishable from the spurious ones. In our second experiment, we modify the ACF procedure and show that robust estimates can be obtained using additional lagged instruments or sequential search. We also provide some arguments for why such modifications help in the ACF setting.