Econometric Modeling

Abstract

This chapter treats an econometric model as an experiment, or set of experiments, that are embedded within an economic model. It shows how to unpack the “experimental content” of an econometric analysis, and demonstrates how this concept and the modeling principles from the preceding chapters contribute to the development of econometric models. These ideas come to life in applications to orchestra auditions, aluminum recycling, the returns to schooling, economic development in The Gambia, fracking, and more.

Food for Thought

• 1. The opening to this chapter addressed the use and misuse of weighting across cross-sectional units of different “size,” such as states.

  • (a) Set out an individual-level regression with a state-level policy variable and a state-level random effect , as in Eq. (8.1). Work with it to show that, in a state-level analysis, West Bengal should be weighted more than Kashmir, but not seven times more.

  • (b) Chapter 3 addressed the same issue differently. Return to question #3 in that chapter and give it another try. Does our discussion of experimental content make it easier to answer?

• 2. Figure 8.3 contains a somewhat-recent snapshot of the states that had adopted an important policy in the U.S . Its scale of implementation also seems “somewhat geologic.” (It is, in fact, extremely geologic. States that were covered by the Western Inland Sea–look it up–are far less likely to have adopted this policy.) What policy is this?

• 3. The following questions all pertain to the chapter’s discussion of Gruber et al. (1999).

  • (a) Compare my Table 8.1 with the original in Gruber et al. (1999). Note the differences. How do these differences reflect principles of effective description ?

  • (b) The choice to examine the sum of squares of the changes in the fee differential is not inconsequential. Could comparing the magnitudes of these values, rather than their squared magnitudes, be justified? Does examining the weighted sum of squares focus on how California affects the coefficient estimates or the standard errors ?

  • (c) The chapter claims that every “problem” with the experimental content of your analysis conforms to a violation of the classical OLS assumptions . If so, which key assumption is violated in Gruber et al.? Relate your answer to the techniques advocated in Bertrand et al. (2004).

• 4. The following questions all pertain to the chapter’s discussion of Goldin and Rouse (2000).

  • (a) Because orchestra auditions consists of several, sequential rounds, the most natural econometric model for the situation analyzed by Goldin and Rouse (2000) would be an ordered probit , were the data suitable to support it. Assume you possessed an index of auditioner “quality.” Then lay out a suitable ordered probit model to examine the effect of blind auditions on females’ probability of advancement, in which the effect of discrimination against females shifts the thresholds required for advancement to the next round.

  • (b) One way to think about the system’s dynamics is to create a latent variable for “progressiveness ,” which increases at different rates within different orchestras , and which influences some of the coefficients and variables in Eq. (8.4). Incorporate this variable into this equation, and show how bias and serially correlated errors are likely to result.

  • (c) I prefer the following specification to Eq. (8.4), though it is econometrically equivalent, because it better describes the actual process at work. How so?

    P = α+βB+γF · (1 − B)+δX+ϵ

• 5. These questions pertain to Fig. 8.2.

  • (a) The residuals for each orchestra but one signify an econometric issue. Identify the issue associated with each orchestra.

  • (b) In this figure, the errors for each orchestra signify a different econometric problem. Is that likely to happen in practice, or is it more likely that many orchestras would display similar problems?

  • (c) In general terms, explain how the “second-generation model” discussed in the text could be used to conduct the final analysis at the level of the experimental unit, as advocated in this chapter. Then explain why would be necessary to do so, in order to get correct standard errors .

  • (d) (Difficult.) Work out how to implement this approach, in order to get unbiased coefficient estimates and standard errors , under the assumption that the deterministic component of the model is correctly specified, and that there are no problems with serial correlation.

• 6. The main specification in Card and Krueger’s (1994) two-period, two-state difference-in-difference analysis is as follows:

  ΔE = α+βX+γNJ+ε

  where E is restaurant-level employment, X are restaurant-level controls, and NJ is a dummy for New Jersey, the minimum-wage-increasing state. Show that if a random effect is included in this specification at the level of the experimental unit, the true standard error of $\widehat{\gamma}$ is unidentified and its OLS standard error is biased downward.

• 7. Traffic accidents can be considered independent, (statistically) rare events, so the number of fatal accidents in any given place and time has a Poisson distribution. This naturally lends itself to a count data model. However, a pure Poisson regression has the classic problem that the conditional variance of fatalities equals its mean, which does not hold up in practice.

  One method of addressing this problem is to specify a generalized linear model:

  F _s,t~Poisson(f _s,t)

  log(f _s,t) = α+βX _s,t+ε _s,t

  where F _s,t is the observed number of fatal accidents in location s in period t, X contains the explanatory variables, and ε_s,t is a random effect . The second line of this equation predicts the latent variable f _s,t, which serves as the mean (and variance) of the Poisson distribution from which realized F _s,t is “drawn.”

  An alternative method is to specify a Negative Binomial Model:

  F _s,t~NegBin(f _s,t, σ)

  log(f _s,t) = α+βX _s,t

  where σ is the “overdispersion” parameter associated with the Negative Binomial distribution.

  • (a) Which is the more natural model? Why?

  • (b) Which model is more commonly used in economics? Why?

• 8. Cunningham et al. (2018) examine how abortion clinic closures affect abortion rates in Texas , relating the change in counties’ abortion rates to the increase in the distance women must travel to have an abortion when a closer clinic closes (measured using five “increase in distance” categories). There are 254 counties in Texas and 42 abortion clinics in their sample, 18 of which closed during the interval a clinic-closing law took effect, leaving 11 of 16 metropolitan statistical areas without a clinic. There were nearly four million Texas pregnancies during the study period, approximately 10% of which ended in abortion.

  Sketch out the experimental content of this study. Identify the experimental unit and the spatial and temporal scale of variation in the key independent variable. To the nearest power of ten, how many independent experiments does this study contain?

Fig. 8.3

The shaded states in this map had adopted a particular policy, as of a date in the not-too-recent-past. Can you tell what policy this is?