1LIR 832 – Lecture 8The Good, the Bad and the Ugly:Tufte: Good and Bad Analysis of Quantative DataLoury: Gender Earnings GapQualitative DataTufte Group 1: What did John Snow do in his analysis of the cholera epidemic that allowed him to support his theory (causal relation)him to support his theory (causal relation) about water and cholera? Group 2: What ended up obfuscating the relationship between temperature and O-ring failure in the analysis pre and post Challenger Disaster?2Classic Approach to Research How to write a research paper, an article and what to look for in statistical researchz Find an interesting or pressing questionzReview literature and develop a theoretic modelzReview literature and develop a theoretic model What is a theoretic model?z b. Specify the empirical model: selection of independent variables and functional form. What do we mean by an empirical model? How is this different from a theoretic model?z c. Hypothesize the expected signs of the coefficientsz d. Collect the dataz e. Estimate and evaluate the equationz f. Document the resultsOpening of the Gender Earnings Gap What do we learn in the opening two paragraphs?ihi dd dbhi. The issue addressed by the paper ii. What has gone before iii. How this paper contributes to the literature (knowledge) iv. Why is this important?Review Literature and Develop a Theoretic Model Relevance of the Research Question College educated women made up a large and growing segment of the population While most studies of the decline in the gender earnings gap present aggregate results across different years of schooling, factors that are responsible for narrowing the gender gap may differ by skill level. An analysis of the effects of college selectivity and academic performance makes it possible to determine whether gains have been uniform for all college educated women or whether some individuals have benefitted relative to others.3Explanations of Changes in Gender Differences in Earnings relative gain in the number of years worked by women compared to men.changed in perceptions about appropriate role ofchanged in perceptions about appropriate role of women in the labor market changes in industrial structure & the desired skill mix of workers within industry Note that each reason is developed and often supported by other informationData and Empirical results Description of data used in the study NLS data setsz longitudinal (follow individuals over time)g( )z when and how collectedz types of data collectedz easy to find the data and obtain additional information (standard data sets) Data is limited to white employees, why?  How does this limit what we can learn from the study? What would be the natural next step?4 Table information:z name in form readily understoodz mean and standard deviation by genderdiff i ( h i i i f 10)zdifference in means (what is missing: see fn10) focused discussion of factors most relevant to this study:z male female contrasts between 1979 and 1986:z raw weekly earningsz changing gap in number of business and education majors by genderz other germain factors:5Display of Regression Results Lay-Out of Table 2 name, coefficient and SE of coefficient by gender and year. size of sample and r2Implicit Discussion of Specification ground on wage equations is very well plowed. education variables: college graduate years of college 2 – 3 College GPAg Business Major Engineering Major Other Major College Characteristics:z Median College Sat Score Other Characteristics:z Marriedz Union Memberz Tenure on Current Job How complete is this specification?The Results: 1979 start by showing that the results conform with prior work (conventional wisdom) on factors which we are not immediately concerned with Cut to the chase: returns to getting a degree are higher for women than for men: base is one year of college: estimate return for 2 - 3 years with no degree estimate return to degree compare return to degree to: those with one year of college6Return to College Education Coefficient on College grad men: .1214 women .3173 compare return to degree to those with 2 - 3 years of college (calculate significance of the difference:z men: .1214 - .1004 = 0.02z women .3173 - .1115 = 0.21 show that this difference is statistically significantReturn to College Quality Women’s earnings are more sensitive to the quality of the school than men’s earnings:z Median SAT: men 0.0187 (.0214) women 0.0565 (.0282)Repeat for 1986 estimates and compare to 1979 Return to graduating fell for women both absolutely and relative to men Return to women’s GPA rose substantially Larger relative returns to Business major, larger returns to engineering major. Change in the structure of returns to college to women. In 1979 the return was to graduating from college, now it has more to do with performance in college and the major.z Suggests that the changes in returns reflect within occupation changes in worker demand which favored women.7Conclusion Brief summary of results and major points, doesn’t beat a dead horse.What You Need To Ask In Analyzing Statistical Research What is the question being asked by the researchers? How does this fit into prior research? Does the current research come out of prior research, or is it de novowith little base in previous work? De novo work requires closer examination than work which is firmly based in prior, established research. What is the ‘theoretic’ model is used by the researchers and what are the implications of this model? What hypotheses are derived about the relationship of interest.Questions to Ask: The Empirical Model What is the dependent variable? How similar is it to the dependent variable in the theoretic model and how well measured in the variable (is the definition similar to that in the theoretic model)? What are the variables of interest? Are they measured in a fashion which is compatible with the theoretic model? What controls (other variables) are included in the model? Are they derived from prior literature? Are they included because of the implications of the theoretic model? Is the model an extension of prior research? If so, does it follow previously models reasonably closely? Is the specification reasonable given the problem (are we using a