Final Paper Assignment

Sociology 229A:  Event History Analysis

 

Due:  Start of class (2:00) June 3

 

Length:  4-5 Pages

 

Overview

 

The goal of this project is to conduct some simple event history analyses of a dataset and (briefly) write up the results.  I will provide a dataset and some general hypotheses to test.  Your task will be to analyze the data, conduct diagnostics, and accurately describe your findings.

 

The dataset includes information about the founding and demise of “licensed lenders” in New York.  Licensed lenders were a form of bank that became popular in the Great Depression, which have since been largely out-competed by more familiar (to us) types of banks and credit unions. 

 

The unit of analysis is the organization.  The event of interest is “organizational mortality,” the failure or dissolution of a bank.  The dataset contains multiple-record-per-case data with information on organizational founding and death, as well as constant and time-varying covariates.

 

CHOOSE FIVE OUT OF THE FOLLOWING SEVEN RESEARCH QUESTIONS TO EXAMINE IN YOUR PAPER.  Some questions cannot be answered with a Cox model, depending on the time-clock that you choose.  If you use a Cox model, be sure to choose questions that you can actually answer!

 

Research Questions:

 

  1. It has been argued that resources are critical to organizational survival.  What is the effect of organizational assets on mortality?  (There are two measures – raw assets, and assets rescaled in 100K dollars.  Both are fine options, but I’d use the latter to ease interpretation.)

 

  1. It is also commonly argued that large organizations have better prospects than small ones.  Do organizations with many offices survive longer than organizations with fewer offices?

 

  1. The World Wars were a difficult time for many kinds of financial institutions.  What effect did those time periods have on the failure of licensed lenders?  NOTE:  WILL NOT WORK WITH COX MODELS IF THE TIME-CLOCK IS HISTORICAL TIME, DUE TO ABSENCE OF VARIATION AT POINTS IN ANALYSIS TIME.

 

  1. In many cases, young organizations prove more vulnerable than older, well-established ones.  This is often operationalized by an organization’s age.  Do you find evidence to support this?  NOTE:  WILL NOT WORK WITH A COX MODEL IF THE TIME CLOCK IS CHRONOLOGICAL AGE OF ORGANIZATION, DUE TO ABSENCE OF VARIATION AT POINTS IN ANALYSIS TIME.

 

  1. Organizational research often focuses on the concept of density.  Density refers to the total number of organizations that exist.  To the extent that high density represents intense competition, density may encourage the demise of licensed lenders.  Do you find this to be the case (either for density overall, and/or “regional” density, which refers to density within a local area)?

 

  1. Another measure of organizational size is whether a licensed lender is purely local versus having branches in other parts of the country.  Use the variable “out_of_state_branch” to see if large multi-state organizations fare better than local organizations.

 

  1. Licensed lenders became popular during the great depression because of the weakness of traditional banks during that difficult period.  How did the licensed lenders fare as the economy grew and the financial system stabilized over subsequent decades?  You can choose any of the GDP measures to test this hypothesis, but I suggest trying the simple variable “gdp” (either raw or logged).  NOTE:  WILL NOT WORK WITH COX MODELS IF THE TIME-CLOCK IS HISTORICAL TIME, DUE TO ABSENCE OF VARIATION AT POINTS IN ANALYSIS TIME.

 

Issues & Diagnostics:

 

  1. You must choose the type of hazard model and the time-clock to use.  I suggest chronological age, but historical time is also an option.  You must also choose between a Cox Model and various parametric options.  Plots of the hazard curve over time may prove useful in making the decision.  (Discrete time methods would work with this dataset… but most diagnostics would be unavailable.  So, I suggest you use a continuous time model.)
  2. Some of the variables exhibit non-proportionality (depending a bit on the model chosen).  Address non-proportionality in the assets variable.  Piecewise models and/or time-interactions may be helpful to identify the non-proportionality.  If you use a Cox model you can use diagnostics like “estat phtest, detail.”  Construct a time-interaction to reduce the non-proportionality.  (You can decide what kind of time interaction to use.)
  3. Construct at least one residual plot to identify outliers in the data.

 

 

The Paper

 

Start with an abstract that briefly summarizes the topic and results.  That will serve as an introduction to your paper.  Write up your paper in the form of a “Methods” and “Results” section of a journal article.  (You need not include a background, intro, lit review or theory section.  Nor do you need a discussion section or conclusion.)

 

In the methods section:  Start with a description of the basic dependent variable – the outcome of interest, and the way that the analysis is set up.  Indicate the time-clock, whether you have time-varying covariates, and whether or not you have repeating events.  Next, describe the kind of event history model you chose (Cox or various parametric options) and the justification for your decision.  If plots or diagnostics were used to make your choice, include them in an appendix.  Provide the relevant formula for the hazard rate.  You may omit a description of the independent variables.  Assume the reader is familiar with them.  Also:  Describe your efforts to address non-proportionality in the assets variable, and your findings regarding outliers.  Indicate if any cases were removed from the analysis.  Include the outlier plot as an appendix.  (You may optionally include other diagnostics in appendices, if you wish.)

 

In the results section:  Provide a table of results.  You may use your discretion about how many models to present.  A single model with all the variables (described above) is sufficient.  I suggest presenting raw coefficients, but you can add hazard ratios if you wish.  Describe your findings and indicate your answers to the arguments/hypotheses raised above.  Be sure to provide clear interpretations of coefficients.  For at least one independent variable I want you to go beyond merely interpreting the effect of a 1-unit change… For instance, you might discuss the impact on the hazard rate of moving from the bottom quartile in assets to the top quartile (or some other meaningful increment).

 

Finally, conclude by listing three ways to further improve your study.  These could include additional variables to include in the analysis, other problems identified in the diagnostics, problems with model fit, measurement issues, etc.  Just take a few minutes to show that you can think critically about the data and models that you are using. 

 

I’m guessing you’ll have (roughly) a .5 page abstract, 1.5 pages of “Methods” and 2-3 pages of “Results.”  Plus a table.  Plus appendices with a hazard plot, a residual plot, and optionally some additional diagnostics.  You may write more if needed, by try to keep things tight/focused.