ECON3208/5048, Applied Econometric Methods, 2020 T3
ECON3208/5408 Course Project Description, 2020T3
The Topic and Data
The topic is based on Kenkel and Terza (2001): The effect of physician advice on alcohol consumption
(http://dx.doi.org/10.1002/jae.596, also included in the kit), where a major task is to estimate the effect of
advice on drinks. The data (KTDATA.DTA) and a do-file (kt-temp.do) for reading the data are provided.
This topic involves various issues that may be encountered in empirical research. The issues
include endogeneity and some special data features. Mostly, these issues have been discussed in
ECON3208 and two assignments. You should carry out this project using the tools and techniques
covered in our course (up to the end of Ch17.2, and up to p18 of Slides-W5-1a) although they may not be
perfect for the data.
You are not required to replicate the above Kenkel and Terza (KT) article (as some techniques
and methods there are not covered in ECON3208). You should use this article to gain a good
understanding of the topic, motivation, questions of interest, issues involved, and data to be analysed.
The Report
You should read Chapter 19 of Wooldridge to get insights about how to proceed with an empirical
project. You should report your analysis in the following 7 sections. You should limit your report to 8
pages (excluding the cover sheet).
- Introduction (1 page). You may discuss why the topic is of interest and how it is related to previous
literature (referring to two or three related articles discussed in KT). You should outline the
econometric issues, your modelling strategies, and provide a summary of your findings. - Data (0.5 page). You may briefly describe the data, including the data source, variable definitions,
important descriptive statistics, and the main features of key variables. You should let readers see
what you see as important. - Conceptual Model (1 page). You may very briefly describe the empirical economic model, on which
your econometric models are based. This can motivate your choice of regressors in the econometric
models. You should read Section 2 of KT for this part. - Econometric Models (2 pages). You may describe your econometric models in detail, and discuss
how you address various issues in econometric analysis (such as suspected endogeneity and data
features β drinks being nonnegative with many zeros and advice being binary). The main assumptions
and estimation method for each econometric model should be briefly discussed. You may need to
complete this section in conjunction with your computation in Stata, which could involve many trialand-error iterations. See also the βEconometric Analysisβ section below. - Empirical Results (2 pages). Your results and findings of econometric analysis should be presented
in detail in this section. You may use tables for your presentation (e.g., similar to Table 17.3 of
Textbook). You should interpret your results properly, using the tools covered in ECON3208.
Comparing results from different models is a good way to check if your findings are robust or
insensitive to the variations in models and assumptions. You may also want to present the results of
relevant tests, which may justify or reject the models and assumptions you use. It is important to
comment on the merits and drawbacks of your econometric models, and discuss possible violation of
your main assumptions and biases in your findings.
Page 3 of 4 - Conclusions (0.5 page). You may reiterate your main findings here, and comment on possible policy
implications. You may discuss briefly the remaining issues that you are unable to resolve, and you
may comment on how you would like to tackle them. - References (0.5 page). You should list your textbook (if it is used) and articles you have read and
used as references.
Econometric Analysis
(a) A goal of this project is for you to explore and apply the knowledge and tools you have learned so far
(up to the end of Ch17.2) in a research project. You should be able to comment on the strength and
weakness of your models and methods.
(b) You should briefly explain why some variables are included in, and others are excluded from, an
equation. Always pay attention to endogeneity: Is there endogeneity? Do I have valid instruments?
Can I test the validity of instruments? Does endogeneity make a difference?
(c) You should start with linear models. While not perfect, linear models can be regarded as a linear
approximation to the true model. It can also serve as a benchmark for comparisons. In particular, we
understand well how endogeneity is handled in linear models.
(d) The method we test for endogeneity (see Ch15.5a) can also be used to estimate the regression
coefficients in the presence of endogeneity. This approach, known as βcontrol functionβ method (see
p10-13 of Slides-W2-1b and p13 of Slides-W4-2b), can be extended to nonlinear models. Suppose we
want to use (π₯π₯, ππ1) to explain π¦π¦, where ππ1 is exogenous, and π₯π₯ is possibly endogenous. Note that ππ1
may involve two or more variables (i.e., it can be a vector). You can think of π¦π¦ = ππππππππππππ and
π₯π₯ = ππππππππππππ in this context.
Assume that the reduced-form equation for π₯π₯ can be either linear with π₯π₯ = ππ1π π 1 + ππ2π π 2 + π£π£, or
probit with π₯π₯ = Ξ¦(ππ1π π 1 + ππ2π π 2) + π£π£. Here, (ππ1, ππ2) are exogenous, (π π 1,π π 2) are parameters, Ξ¦(β ) is
the standard normal CDF, and π£π£ is an error term with πΈπΈ(π£π£|ππ1, ππ2) = 0. Note that ππ2 may involve two
or more variables (i.e., it can be a vector).
Further, assume that the structural equation for π¦π¦ can be either linear with π¦π¦ = π₯π₯π₯π₯ + ππ1π·π· + π’π’, or
Tobit with π¦π¦ = max{0, π₯π₯π₯π₯ + ππ1π·π· + π’π’). For Tobit, π’π’ is an error term that is conditionally normal with
π’π’ = ππππ + ππ, πΈπΈ(ππ|π£π£, π§π§1, π§π§2) = 0, ππ βΌ ππ(0, ππ2), and (πΎπΎ,π·π·, ππ) are parameters. The structure of π’π’ here
takes into account the possible correlation between π’π’ and π£π£. The parameter ππ can be used to test
whether π₯π₯ is exogenous (when ππ = 0, π’π’ and π£π£ are uncorrelated) or endogenous (when ππ β 0, π’π’ and π£π£
are correlated).
It follows that the structural equation for π¦π¦ can be expressed as π¦π¦ = π₯π₯π₯π₯ + ππ1π·π· + π£π£π£π£ + ππ for the
linear model, and π¦π¦ = max{0, π₯π₯π₯π₯ + ππ1π·π· + π£π£π£π£ + ππ) for the Tobit model, where ππ is normally
distributed and uncorrelated with (π₯π₯, ππ1, ππ2). Hence, if we were able to observe π£π£, the OLS estimation
would be applicable to the linear model and the maximum likelihood estimation would be applicable
to the Tobit model.
As we do not observe π£π£, we use a two-step approach (control function approach). If models are
correct, (πΎπΎ,π·π·, ππ) can be consistently estimated in two steps:
Page 4 of 4
(i) estimate the reduced-form equation for π₯π₯, either the linear model π₯π₯ = ππ1π π 1 + ππ2π π 2 + π£π£ or
the probit model π₯π₯ = Ξ¦(ππ1π π 1 + ππ2π π 2) + π£π£, and save the residual π£π£οΏ½;
(ii) estimate the structural equation (either linear or Tobit) replacing π£π£ by π£π£οΏ½.
However, the standard errors from Step (ii) can be incorrect because they are based on the first step
estimation. As we did not cover how to correct such standard errors, you may assume the standard
errors from Step (ii) are good approximates to the true standard errors, and acknowledge this
weakness.
For brevity, the above presentation does not include an intercept in the models. In your report,
however, all models should include an intercept.
Stata Commands
For Stata commands, you may consult the Stata do-files (from Weeks 2 to Week 5) deposited in the
βTutorialsβ folder on Moodle. You may also consult the do-files for Assignments 1 and 2. The following
points should also be useful.
β’ OLS estimation of linear model
regress x z1 z2
predict xhat //Save fitted values
predict vhat, residuals //Save residuals in vhat
test z2 //Test null hypothesis that coef on z2 is zero
β’ 2SLS estimation of linear model
ivregress 2sls y z1 (x=z2 z3) //2SLS using z2 and z3 as instruments for x
predict yhat //Save 2SLS fitted values
predict uhat, residuals //Save residuals in uhat
β’ Probit estimation
probit x z1 z2
predict xhat //Save fitted values in xhat
generate vhat=x-xhat //Save residuals in vhat
β’ Tobit estimation
tobit y x z1
predict yhat, ystar(0,.) //Save fitted values in yhat
margins, dydx(x) predict(ys(0,.)) //Find partial effect of x
display r(rho)^2 //Display R-squared
β’ Tobit estimation: Prefix a binary regressor x with βi.β
tobit y i.x z1
margins, dydx(i.x) predict(ys(0,.)) //Find partial effect of binary x