Empirical questions in economics are studied via econometric models. These models contain unknown parameters, which are typically the object of interest in empirical applications. For instance, one might be interested in learning about the average propensity to consume, or the marginal effect of a tax increase on labor supply, and researchers use data and econometrics analysis to estimate that parameter.
Clearly, one would like to use methods of data analysis that allow one to learn as much as possible about these unknown parameters. And for a large set of models, it is well understood how to do so, at least approximately. With many observations from a well-behaved model, the problem of not knowing the parameter is effectively equivalent to the "standard problem" of observing a single Gaussian (bell-shaped) observation with known spread and a mean that corresponds to the unknown parameter. For a non-negligible subset of models, however, there is only an approximate equivalence to a different, "non-standard" problem. These unknown correlation patterns make it impossible to perfectly learn about the spread of the parameter estimator, which leads to a non-standard approximately equivalent problem where the spread of the single Gaussian vector observation is unknown.
The proposal studies such non-standard problems, and seeks to develop ways of learning about the unknown parameters. The derivation of nearly efficient methods involves the development of appropriate analytical results, such accurate upper bounds on the potential quality of inference in non-standard problems, as well as numerical methods that determine a nearly efficient procedure, that is one with an inference quality close to the upper bound.
