Вы находитесь на странице: 1из 3

Homework 7

Nathaniel Higgins nhiggins@jhu.edu, nhiggins@ers.usda.gov

Discrete dependent variable problems


The next few problems use data from the National JTPA Study (NJS). JTPA stands for Job Training Partnership Act. Until recently, when the Workforce Investment Act replaced it, the programs funded under JTPA constituted the largest federal training program in the United States. Variable name bd accept black hisp othrace age edlt10 ed10 11 ed13 15 edgt15 Description observation ID number 1 if person was accepted into JTPA; 0 otherwise 1 if person is black; 0 otherwise 1 if person is hispanic; 0 otherwise 1 if permson is not white, black, or hispanic; 0 otherwise age in years 1 if person has fewer than 10 years of completed schooling; 0 otherwise 1 if person has 10 or 11 years of completed schooling; 0 otherwise 1 if person has 13 or 15 years of completed schooling; 0 otherwise 1 if person has more than 15 years of completed schooling; 0 otherwise

The omitted racial/ethnic group is white, and the omitted years of schooling category is exactly 12 years of completed schooling.

5.
Using the Corpus Christi data, estimate a linear probability model of the probability of acceptance into JTPA at Corpus Christi conditional on race/ethnicity, age and years of completed schooling. Report and discuss the coecient estimates. Use the following commands to generate results: use corpus.dta reg accept black hisp othrace age ed* The dummy for less than 10 years of eduction has a strong, statistically signicant negative eect on the probability of acceptance. Relative to having 12 years of schooling (relative to the omitted category), having less than 10 years of schooling reduces the probability of acceptance by about 0.085. 1

6.
Using the estimated coecients from the Question 5, generate predicted participation probabilities. Do any of these probabilities lie outside [0,1]? If so, do the observations corresponding to these values show any particular patterns in the values of the variables? Use the following commands to generate results: predict phatOLS, xb sum phatOLS Nope. Although the LPM is most certainly capable of being naughty and predicting outside of the [0,1] interval, in this case it plays nice and gives us predictions we can interpret.

7.
Estimate probit and logit models of acceptance into JTPA in Corpus Christi conditional on race/ethnicity, age and years of completed schooling using these data. Are the probit and logit coecient estimates similar to one another and to the LPM estimates? Should they be? Explain why or why not. Use the following commands to generate results: probit accept black hisp othrace age ed* logit accept black hisp othrace age ed* We should not expect the coecients in the LPM to be the same as the coecients in a logit or a probit model. We should expect signs and signicance to be similar, and we should expect that the marginal eects, i.e. the partial derivatives of the logit/probit models w.r.t. the appropriate coecients to be roughly similar to the LPM coecients.

8.
Calculate the marginal eects of each included variable in the probit, logit, and linear probability models for the same model described in Question 7. Are the probit, logit, and linear probability model marginal eects similar to one another? Should they be? Use the following commands to generate results: * Probit model probit accept black hisp othrace age ed* margins, dydx(*) * Logit model logit accept black hisp othrace age ed* margins, dydx(*) * Finally, the LPM reg accept black hisp othrace age ed* The marginal eects of each model are similar. In each case the signs of all coecients are identical. The scale of the eects is similar, though maybe more similar between the 2

logit and LPM than to the probit model. We expect this in many cases. We expect this because we are calculating the average marginal eects. If we inspected the marginal eects at certain values of the independent variables say where the curvature of the probit function was particularly pronounced then we might get very dierent answers.

Вам также может понравиться