Regression Analysis
Methodology of Conflict and Democracy Studies
December 17

Regression Analysis
•A variety of techniques with the same aim
•
•Identification of effects of one or more IVs on DV
•
•What it allows:
•Identify effect of each independent variable
•Control of effects of other independent/control variables
•Predict values of DV based on specific values of IVs
•

Which Regression?
•Everything depends on your dependent variable
•
•Linear (OLS) regression:
•Scale variable (or long ordinal)
•
•Logistic regression:
•Binary variable (0/1) – binary logistic regression
•Nominal (0/1/2/3) – multinomial logistic regression
•
•No limits on independent variables (all types allowed)

Examples
•OLS regression:
•How do age, gender and education affect income of people?
•Does attendance on lectures increase % amount of obtained points in your courses?
•
•Logistic regression:
•Do men have higher chances to end up in jail than women?
•Does attendance on lectures increase your chances to pass the course?

OLS Regression - Requirements
•Dependent variable:
•Exactly one variable
•
•Independent variable:
•One or more variables, all types without limits
•
•Some further requirements:
•Independence of observations
•No collinearity between independent variables
•
•

What is OLS Regression about?
•Basically, it is about searching for ideal lines that best describe the relationship between
independent and dependent variable
•
•The best line is the one that is the least inaccurate of all possible lines
•
•Accuracy measured using sum of squares of vertical differences between predicted and observed data




R square
•Provides information about the overall fit of the model
•How well our model (= our IVs) explains the dependent variable
•Comparison of improvement of regression line compared to mean
•
•Ranges from 0 to 1 (zero to hundred per cent)
•
•Show how much of the variance of dependent variable we are able to explain using our set of
independent variables
•Use Adjusted R square to control for inflation of number of IVs

The Outcomes of OLS Regression
•OLS regression estimates:
•Intercept
•Effects of each independent variable
•
•y = b0 + b1*x + b2*y + b3*z + …
•
•y stands for predicted value of dependent variable
•b0 stands for intercept
•b1, b2, b3 etc. stand for slopes of independent variables x, y, z etc.

Example
•Is turnout in local elections affected by town population?
•
•Hypothesis: Turnout decreases as population increases
•Null hypotheses: There is no relation between population size and turnout
•
•Dependent variable:
•Turnout – turnout in % (scale)
•
•Independent variable:
•Population_th - town population in thousands of people (scale)

How to Perform the OLS Regression
•Analyze > Regression > Linear
•
•Select the variables:
•Turnout into ‘Dependent’
•Population_th in the section for independent variables
•
•
•

•Model Summary:
•Our model explains 7 per cent (0,07 * 100) of variance of dependent variable
•
•ANOVA:
•Our model is a significant improvement in predicting the dependent variable and our results can be
applied to the population







Predictions Based on Results
•y = b0 + b1*x
•Turnout = 60.8 + 0.591*Population_th
•
•
•
•
•
Population
Population in thousands
Formula
Predicted turnout
Town 1
500
0.5
60.8 – 0.591*0.5 = 60.8 – 0.296
60.5
Town 2
1,000
1
60.8 – 0.591*1 = 60.8 – 0.591
60.2
Town 3
5,000
5
60.8 – 0.591*5 = 60.8 – 2.955
57.8
Town 4
10,000
10
60.8 – 0.591*10 = 60.8 – 5.91
54.9
Town 5
25,000
25
60.8 - 0.591*25 = 60.8 - 14.775
46,0

Example 2
•Is turnout in local elections affected by town population, the local financial situation and
whether there is a true competition?
•
•Dependent variable:
•Turnout – turnout in % (scale)
•
•Independent variables:
•Population_th - town population in thousands of people (scale)
•Fin_Index – indicator of financial situation in town (0-6; 0 = worst, 6 = best) (scale)
•Competition – 1 for at least two competitors or 0 for only one competitor (binary)
•

How to Perform the OLS Regression
•Analyze > Regression > Linear
•
•Select the variables:
•Turnout into ‘Dependent’
•Population_th in the section for independent variables
•
•Because we have more than one IV:
•Statistics > Collinearity Diagnostics
•
•
•

•Our model explains 45.5 per cent of variance of dependent variable
•Substantial improvement compared to model that included only one independent variable
•
•Our model is a significant improvement in predicting the dependent variable and our results can be
applied to the population
•













Unstandardized B Coefficient
•Scale v. Binary Variables
•
•Same definition for scale and binary variables:
•Shows how the value of DV changes if the value of an IV increases by one unit
•
•BUT
•
•Binary (dummy) variables have only two values – 0 and 1
•Unlike scale variables, there is only one possible increase by one unit
•The estimated effect is thus completely exhausted by this one increase

•Competition:
•0 – no competition (only one candidate)
•1 – competition (at least two candidates)
•Shift from 0 to 1 means that towns with competition are predicted to have a nearly 18 percentage
points higher turnout than towns without competition
•
•Population_th:
•Shift of population from 1 thousand to 2 thousand leads to drop of turnout by 0.77 percentage
points
•Shift of population from 1 thousand to 5 thousand leads to drop of turnout by 3.08 percentage
points (4 times decrease of 0.77)
•Shift of population from 5 thousand to 12 thousand leads to drop of turnout by 5.39 percentage
points (7 times decrease of 0.77)
•
•

Standardized Beta Coefficient
•Provide information about importance of independent variables
•
•Measured in standard deviation units à allow to easily compare the IVs
•
•Higher distance from zero (both positive and negative) indicates higher importance of the
independent variables

•
•Results show that Competition is the most important predictor of all three independent variables
•
•Population_th is less important and Fin_Index is the least important
•
•
•

Predictions Based on Results
•
•
Population
Fin_Index
Competition
Formula
Predicted turnout
Town 1
1,000
3
0
55.569 – 0.77*1 – 1.382*3 + 17.995*0
50.7
Town 2
1,000
3
1
55.569 – 0.77*1 – 1.382*3 + 17.995*1
68.6
Town 3
5,000
3
0
55.569 – 0.77*5 – 1.382*3 + 17.995*0
47.6
Town 4
10,000
6
1
55.569 – 0.77*10 – 1.382*6 + 17.995*1
57.6
Town 5
25,000
6
0
55.569 – 0.77*25 – 1.382*6 + 17.995*0
28.0

Control of Assumptions
•Outliers – cases with extreme values
•Collinearity – association between independent variables
•
•How to do that:
•Analyze > Regression > Linear
•Statistics > Collinearity diagnostics + casewise diagnostics

•VIF above 5 (10) or Tolerance below 0.2 (0.1) constitutes a problem
•Solution – more models or dropping one of the variables
Collinearity

Outliers
•The data should contain up to:
•5 % of cases with residual above 2 (below -2)
•1 % of cases with residual above 2.5 (below -2.5)
•
•If we find outliers we can rerun the model without these cases and compare whether the results
change