CV

I study how new technologies diffuse into a market when they can replace old technologies on some margins but not others. I analyze this in the Australian electricity market where renewables cannot provide many of the ancillary services that fossil-fueled generators can. The grid operator needs to balance falling costs from renewables while meeting requirements for ancillary services. I evaluate whether a A$166 million investment in a complementary technology to provide these ancillary services was cost-effective. I estimate the costs of running fossil-fueled generators with a structural model of the market to be A$47 million per year, but find these costs did not decrease following the investment, implying that the investment was not cost-effective. This shows that fossil-fueled generators can have a persistent role in electricity markets, even when there are new technologies to replace them.

Maximum likelihood estimation of dynamic discrete choice models in the tradition of Rust (1987) requires solving a nonconvex program, and consistency and asymptotic normality hold only at the global maximizer, which no existing method guarantees. For models with linear-in-parameters payoffs and fixed shock-distribution parameters, we modify the MPEC of Su et al (2012) by relaxing the equality that defines the ex-ante value function and penalizing the slack. The modified program is globally convex for logit and binary probit, the most common distributions in the literature. With probability approaching one the relaxation is exact, so every maximizer is a global maximizer of the original MPEC and converges to the unique population maximum. A simple sufficient condition, positive Lagrange multipliers on the relaxed constraint, certifies exactness ex post at any candidate solution for any estimation method. We show that the relaxed program can be expressed as a convex conic program when the shock distribution is T1EV. We extend the relaxed model to incorporate k-step updating for efficiency, and to discrete unobserved heterogeneity. In Monte Carlo experiments we certify global optimality in over 99.5% of samples at the largest sample sizes.