Multivariate Regression Interpretation pt 2

Visualizing Multivariate Regression Lines

  • Conceptual Difficulty: Envisioning the regression line in a multivariate context is challenging because the graph involves multiple axes representing multiple variables rather than a simple relationship between two axes.

  • Variable Interaction: Even if a primary variable establishes a specific slope, a second variable can change the direction of that line.

  • Multidimensional Space: Visualizing these lines accurately requires thinking in multidimensional space, which is typically reserved for advanced mathematics courses.

  • Pedagogical Note: The instructor suggests students "suspend belief" for the time being regarding the visual geometry of these lines if they find it confusing to wrap their heads around.

Multivariate Regression: Nate Silver’s Manchester Union Leader Study

  • Research Question: Nate Silver investigated whether an endorsement from the Manchester Union Leader during the New Hampshire presidential primaries has a measurable effect on a candidate’s share of the vote.

  • Operationalization of the Endorsement:     * The endorsement occurs when the newspaper writes an op-ed and selects a preferred nominee during competitive primaries.     * Variable type: Binary (11 or 00), where 11 indicates endorsement and 00 indicates no endorsement.

  • The Need for Control Variables:     * To accurately measure the effect of the endorsement, one must control for prior public opinion/polling.     * The best predictor of a candidate's actual vote share is their polling average (e.g., three weeks before the primary).     * The logic is to see how the endorsement affected the vote share given their pre-existing polling performance.

  • Statistical Output and Variables:     * Independent Variable (IV): Manchester Union Leader endorsement (00 to 11 unit change).     * Control Variable: Polling average prior to the primary.     * Dependent Variable (DV): Candidate’s actual vote share in the New Hampshire primary.

  • Findings and Magnitude:     * The analysis found the endorsement to be highly statistically significant based on the pp-value.     * Magnitude: The endorsement is pegged as having roughly an 1111 percentage point impact (10.7710.77 unit increase).     * Complete Interpretation: The endorsement has a significant positive effect of approximately 1111 percentage points on the result in the New Hampshire primary, "taking into account how candidates were already doing in the New Hampshire polls previously."

Worksheet Example: Murders and Poverty

  • Model Overview: A regression predicting annual murders per million people based on the percentage of the population living in poverty in randomly selected cities.

  • Variable Interpretations:     * Coefficient Estimate (2.5592.559): For every one unit increase (1%1\% increase) in the percentage of people living in poverty, there is a corresponding increase of approximately 2.5592.559 murders per million people.     * Note on Control: In this specific bivariate case, there were no other variables to control for.

  • R-Squared (R2R^2) Interpretation:     * The R2R^2 value is 0.70520.7052.     * This means changes in the poverty percentage explain about 70.52%70.52\% of the variation in annual murders per million people in the sample.     * The remaining 29.48%29.48\% (approximately 29.5%29.5\%) of the variation remains unexplained, likely due to other variables not included in the study.

Comparative Study: Terrorist Attacks and Executive Approval (Mary DeHolman et al.)

  • Hypothesis Context: The literature suggests a "rally around the flag" effect where executive approval rises during crises, but the authors hypothesized this might differ for women leaders.

  • Interaction Variable: The study looked at the interaction between an international terrorist attack and the executive being a woman.

  • Key Findings:     * General Effect: Across all executives, there was a positive, statistically significant "bump" of about 33 percentage points in approval following a terrorist attack.     * Women Executives Effect: The coefficient was −4.333-4.333, which was statistically significant (p<0.05p < 0.05, indicated by a star/asterisk).     * Full Interpretation: When a terrorist attack occurs and the executive is a woman, her approval rating declines by 4.333%4.333\%, holding constant the existence of the terrorist attack itself.

  • Evaluation Framework: When interpreting these tables, one must look at:     1. Direction: Positive or negative.     2. Magnitude: The change in yy for a one-unit increase in xx.     3. Statistical Significance: Indicated by stars or pp-values.     4. Multivariate Context: Interpreting the results while "holding constant" other information in the table.

Political Favorability Case Study: Theresa May

  • Context: Predicting favorability toward Conservative Prime Minister Theresa May following the Manchester attack.

  • Model Comparison:     * Model 1 (Bivariate/Simplified): Includes "surveyed after Manchester attack" and "Labour Party member" status.     * Model 2 (Multivariate/Full): Includes the above plus ethnicity (British), gender, party ID, income, and political ideology.

  • Labour Party Impact (Model 1):     * Coefficient for Labour membership: −2.815-2.815.     * Interpretation: Being a Labour Party member (moving from 00 to 11) leads to a statistically significant 2.8152.815 decrease in favorability toward Theresa May compared to other respondents, holding constant when they were surveyed.     * R2R^2: 0.1430.143 (14.3%14.3\% of the variation explained).

  • Impact of Manchester Attack (Model 2):     * Variable: Surveyed after the Manchester attack.     * Coefficient: −3.51-3.51 (described by the professor as a decline of about "a third of a percentage point" in the context of the evaluation scale).     * Significance: Statistically significant negative effect.     * Interpretation: Even after controlling for ethnicity, gender, party ID, income, and ideology, being surveyed after the attack resulted in a statistically significant negative perception increase/favorability decline of −3.51-3.51.     * R2R^2: 0.4420.442 (44.2%44.2\% of the variation explained).

  • Understanding the Change in R2R^2:     * The jump from 14.3%14.3\% to 44.2%44.2\% occurs because adding more control variables (identity markers, ideology, etc.) provides more information to help predict the outcome.     * Adding variables will almost always increase R2R^2, even if the variables are relatively meaningless, though in this case, the demographic factors explained roughly an additional 30%30\% of the variation.

Statistics Discussion: R-Squared and Significance

  • Questions & Discussion:     * Student Question (Nana): Does the lack of a "star" on the R2R^2 jump mean it isn't statistically significant?     * Professor Response: No. Statistical significance tests are used to reject a null hypothesis for specific variables. There is no null hypothesis for R2R^2. It is simply an overall measure of how much variation is explained by the whole model.     * Reliability: Significance levels apply to individual variables to see if they have a meaningful impact on the DV (e.g., gender might be included but not be significant), but R2R^2 is just a summary statistic for the model's explanatory power.

Course Evaluations and Logistics

  • Evaluation Process:     * Access via Moodle dashboard under "Course Evaluations and Surveys."     * Students are urged to complete these as they are anonymous and provide "voice and vote."

  • Future Curriculum Changes:     * Course 4010140101 is being phased out and replaced by half-semester, half-credit lab courses focused on different issue tools.     * Student feedback is needed for revising this curriculum.

  • Feedback Guidelines:     * Focus on constructive and thoughtful feedback regarding content and the TA (Cameron).     * The professor asks to avoid comments on physical appearance, fashion sense, or height as they are not helpful.

  • Bias Awareness:     * Research indicates that white cisgendered males (like the professor) often receive higher scores due to identity bias.     * Students are asked to be cognizant of this bias when evaluating instructors of different identities and to focus strictly on course content and teaching quality.

  • Upcoming Schedule:     * Friday: IS Symposium.     * Next Monday: Final class/exam discussion.