Comment on Rust

Was reading Rust’s paper on engine replacement. i was thinking this is a great paper to test whether dynamic programming works for actual human decision makers.

How good is stochastic dynamic programming in explaining the decision to replace a bus engine? the best metric it has is an estimated hazard function, that gives probability of replacement as a function of mileage. he also has confidence intervals around the hazard function.

how does this stand up to the data? can dynamic programming predict zucker’s behavior?

rust doesnt directly show this. he does estimate an empirical (non-parametric) estimate of the hazard function (fig 3). In another figure, fig 5, we have the 1 stnd dev bands, as well as the 95% CI bands. unfortunately, its not superimposed, but i think that the CI bands capture some of the empirical hazard, but not all…

most of the observations are around the 200k mileage, the mean. the total number of buses available is 162, but they at the end look at only a subset of that.  i wonder how good the prediction is at the mean?  not as good, but rust doesn’t provide the guidance at all as to how well…

the empirical hazard becomes erratic high levels of mileage, coz of a lack of observations. at these levels of mileage, there are only a few obs, which are either replacement or not. meanwhile the estimated hazard levels out at high levels (non-myopic) levels off at 7%.

here’s the interesting part, the lack of observations makes it hard to determine which cost function mr zurcher was using. after speaking with zurcher, rust chooses a linear specification.

this is unfortunate coz, if zurcher is not around, how do we know which specification is correct? how do we choose?

A specific question, rust says the method is ‘efficient’. does he mean minimum variance for a given sample size? i’m not sure this is proven, altho i’m willing to believe it.

Provincial and Elections Data

i want to know the empirical relationship between govt policy (spending and taxation) and elections.

One question is: do politicians myopically increase spending and lower taxes in the lead up to an election year? in a regression setup, we can estimate the following model:

Y_it=fixed effect(i)+timedumy(t)+eps

in the philippines, election years are fixed, and therefore perfectly collinear with the year in which they are held. So can cannot seperately estimate this model:

Y_it=fixed effect(i)+timedumy(t)+ElectionYeardummies(t)+eps

as the time and electionyeardummy would be exactly collinear in the year of the elections.

in what sense is this a problem? i cant separate year shocks and elections. although, time varying factors like GDP may help us here.

there is a study in india that looks at this question. in india,  elections and the year dummies are not perfectly correlated, coz states can have elections when they want. These are called midterm elections. the fact that it is whenever they want, implies endogeneity between the timing of actual elections and the shocks that cause states to hold midterm elections.

the author (stuti) solves by using scheduled elections (vs. actually elections). scheduled elections are an IV coz its correlated with actual elections, and its uncorrelated with shocks that coincide with the election year (and induce a midterm election).

work in the philippine case doesn’t have  endogeneity, in stuti’s sense, but there is an identification problem.

i’d be more confident about this if we had enough election years. in theory, we have: ‘92, 95, 98 2001 04 07 (as yet uncollected, but theoretically available), so we have 6 elections. i think this is enough to counter the criticism that we can’t separate year shocks from election effects, coz what are the chances that all the election year dummies are significantly different from other years just coz of some cyclical shock that just happens to correspond the the election cycle? very low indeed.

we need fiscal data from 84 to 2006 to use all 6 elections. We also need province-time level variation. in the philippines, its very hard to get time series of provincial level variables for 20 years. the main source of provincial data is fies, and apis. i don’t think there is data for either prior to the late 90s, or even apis and fies are available. but lets see…