Help with Homework Stata

Soc 362, Spring, 2016

Take-home Exercise 1: Chapter 17 Models

April 14, 2017, JA Logan/Chris Holmes

Scenario

You are a junior researcher on a project evaluating a church-affiliated inpatient drug rehab center. Your

research team is preparing to present preliminary results to the board of supervisors tomorrow. Before the

meeting, your boss wants to discuss your preliminary analyses. Prepare a brief report to share with your

boss.

Assignment overview

This assignment is intended to be done in three hours, as though it were an overnight analysis for the research

team. So be brief and to-the-point in your responses. Cut and paste (with the computer, of couse) the Stata

output you need to report, instead of making formal tables like you would for a term paper or article. Don’t

write an overall narrative; this is not a term paper. Just get the job done in a form that you could confidently

take to a research group meeting the next day.

We wrote the assignment instructions (beginning on the next page) as a memorandum to you. Submit

your meeting notes in pdf format only, with supporting Stata output, to the Dropbox now ready on the

Learn@UW site.

There is a strict page limit: Four pages of double spaced text, in 11-point typeface. Plus Stata output as

needed. The Stata output can be placed inside the text, in which case it will not count against the page limit.

If we determine that the text itself has exceeded the limit, you won’t get any credit for the assignment.

1

Assignment

Memorandum

To: [you]

From: Research director

Date: April 14, 2017

I’d like you to do a preliminary analysis of two of our most important outcome variables for tomorrow’s

meeting:

• epcount: Number of aggressive incidents performed by patients while in the rehab center.

• dropped: Coded 1 if the patient dropped out of rehab before end of his/her scheduled treatment

period, 0 otherwise.

The working set of explanatory variables I’d like you to consider is:

• scheduled: number of days of scheduled treatment

• completed: number of days completed (less than scheduled if dropped out)

• pastoral: 1 = pastoral referral (i.e., from a religious cleric); 0 = other

• intake: intake inventory score (at admission)

• age: age (in years) at admission

• services: 1 = (voluntary) attendance at evening religious services; 0 = none

Analysis of program drops. Analyze the dropped outcome – whether the client dropped out of the

program before its completion – as a function of other appropriate variables. Determine if age (either linear

or quadratic) is significant and what its final specification should be. I know we will have an economist and

a graduate of that sociology statistics course in the room, so you should estimate a linear probability model

(LPM) and a probit, in addition to a logit model. State the differences in interpretation among those three

models.

Using the logit model, you should prepare predictions for the dropped outcome as a function of intake

score, holding pastoral at 1 and then at 0, and doing something appropriate with the other variables.

Make sure your predictions are expressed in the probability scale, and that you include confidence intervals.

Either a nice graph comparing the models or else side-by-side comparisons using appropriate values of

intake should be fine.

Board members have noticed that clients from pastoral referrals have been dropping out less frequently

than the others, and wonder whether a greater participation in the evening services by clients with pastoral

referrals could help explain this. Estimate two logit regressions, with and without services in the model,

to see if you can find evidence of this possibility. Then explain what feature of logit (and probit) regression

tends to make it hard to find weak effects of this kind using such comparisons.

2

Analysis of episode counts. First, analyze the factors that explain epcount, the number of aggressive

episodes that caused staff to intervene in a case. Present results from OLS regression (some of the supervi-

sors seem to believe linear regression is always best) and compare them to results from Poisson regression.

Make sure to account for the different numbers of days clients were resident in the program. Which analysis

is better for these data? Justify your answer.

Using Poisson regression, consider these questions:

• Is age a significant predictor, in either a linear or a quadratic specification? What is the best specifica-

tion to use for age?

• Does pastoral referral make a difference in this outcome? What would be the best way to explain the

effect to the board, assuming you find one?

The board is especially interested in the relationship with the intake inventory score since they’re hoping

they might be able to predict which clients would be most helped. Interpret the coefficient for intake

in this model, and make a graph of predicted counts (don’t forget to include confidence intervals), or you

could make some specific predictions for particular values of intake. Since the board is considering cutting

off admissions for clients under 25 years old, restrict your predictions to clients who are at least 25. State

whether you included current clients under 25 in estimating your prediction model, and what test (if any)

you used to justify the choice you made.

Finally, do a check of whether a negative binomial regression might fit better, and interpret your result.

3