Large-N observatio ional l data 1 Large-N observational data The - - PowerPoint PPT Presentation

Large-N observatio ional l data

1

Large-N observational data

The basic idea

• Whenever cases are non-experimental and one wants to analyze several of them,

researcher has to revert to statistical methods to control for confounding variables.

• Association between variables can be established visually (i.e., through scatterplots)

and captured as minimizing sum of the squared distances (OLS regression)

• You need to do the best you can to control for major alternative hypotheses.

Common pitfalls

• Endogeneity
• LOVB
• Measurement error (and crappy data)
• Non-comparable data (e.g., urbanization)
• Causal heterogeneity

Observational data very useful in disconfirming contentions, as correlation is commonly a requisite for causal relationship

2

Omitted variable bias and endogeneity

X Y Q + + X Y + + X Y +

• Effect of X upon Y

appears stronger than it is

X Y +

Effect of X upon Y appears stronger than it is; no actual effect of X on Y True effect of X on Y washes out in the analysis; there appears to be no effect when there actually it one

X Y

More complex forms

Q Z Z² Z³

?

3

Omitted variable bias can inflate coefficients An example

Education Parents’ SES Annual income

Cell entries represent fortnightly income in $K 3 3 4 5 2 2 3 4 1 1 2 3 Education tier 1 2 3 Parents’ SES Apparent effect of education

• n income

True effect of education Education $ 1 2 3 1 2 3 If education and parents’ SES are correlated, then…

4

People use the term “measurement error” to refer to at least two different things

Valid but not reliable (inefficient/imprecise) Example? Reliable but not valid Example?

5

Inefficient measures have different effects

Random error on DV Random error on IV

True Apparent

Note that, in the MV case, measurement error can bias coefficients in unpredictable ways e.g.: British colonial legacy e.g., British colonial legacy e.g., corruption

6

Galton’s coefficient of regression (and the concept of “regression toward the mean”)

Father’s height Son’s height

7

MIT OpenCourseWare https://ocw.mit.edu

17.801 Political Science Scope and Methods

Fall 2017 For information about citing these materials

• r
• ur

Terms

• f Use, visit: https://ocw.mit.edu/terms.