I model the strategic interactions between the manager of a firm and an outside investor in a dynamic cheap talk game with two-sided asymmetric information. Each period, the investor selectively discloses his information to influence the manager's capital investment decision. While the manager knows that she can learn from the investor's disclosures, she also knows that the investor is trying to manipulate her; in equilibrium, the investor's incentives to mislead the manager constrain the credibility of his disclosures, leading to a mutually-deleterious loss of information. I compare the set of cheap talk equilibria with the Bayesian persuasion equilibrium. My model has implications for short-termism, management guidance, and investor credibility over the business cycle.
Available upon request.
I present a cheap talk model in which an informed sender strategically communicates with an uninformed receiver. This model nests Crawford and Sobel’s (1982) lauded uniform-quadratic environment. Counterintuitively, I show that if the sender’s optimal action is less sensitive to the state of nature, then communication will be least informative in states where the sender’s and receiver’s preferences are most aligned.
Instrumental variable (IV) estimates are often much larger than their ordinary least squares (OLS) counterparts, suggesting implausible values of the omitted variable bias. We show that a meta-regression that uses the IV and OLS estimates can separate omitted variable bias from measurement error, potentially resolving the puzzle. Meta-regression can also be used to assess instrument validity and the relevance of heterogeneous treatment effects. We apply our method to three published papers: in the first two, the omitted variable bias is much less important than measurement error, while in the third it appears the instrument is invalid.
We highlight the importance of measurement error in applied empirical work using 2,185 instrumental variable regressions from 326 papers published in top economics and finance journals. We derive meta-regression estimators that distinguish measurement error from omitted variable bias and simultaneity bias, and find that 50–80% of the variance of the average published regressor is noise. While we estimate that statistically significant coefficients are 12 times likelier to be published than statistically insignificant coefficients, publication bias does not quantitatively explain our results. Our results suggest measurement error is a severe, pervasive, and understated source of bias in economics and finance.
Instrumental variables estimators are commonly used in economics and finance to establish causal relationships. Although instruments that fail the exclusion restriction do not reliably estimate parameters of interest, testing the exclusion restriction is uncommon, due to the difficulty of finding multiple valid instruments. We derive closed-form instrumental variable estimators that allow for tests of over-identifying restrictions even for the case of a single valid instrument. We also derive estimators that are consistent when instruments and regressors are mis-measured with correlated errors. Monte Carlo simulations suggest that our estimators have power to reject even in relatively small samples. We also apply our estimators to the IV regressions of Mian & Sufi (2014) and cannot reject the null hypothesis that the exclusion restriction holds.