Tutorials

Question 1

Laddering vs Probing

Answer 1

Laddering : asking ‘why?’ questions. Up > more abstract, down > more concrete

Probing: asking about concrete / specific / sensory-based examples. ‘what else?’

Question 2

formulas to remember

Answer 2

• t-value:
mean 1 - mean 2 / standard error

• Marginal effects for OLS moderated regression analysis: 𝝏𝒀𝝏𝑿⁄ = a1 + a3Z

• Direct/Indirect/Total effect in mediation models
total effect = c=c’ + ab
direct effect = c’
indirect effect = ab

Question 3

Reliability

Answer 3

Degree to which measure is consistent + free from random errors

Question 4

Calculate reliability

Answer 4

Single-item > impossible calculate variance true score, thus reliability
Multi-item > cronbachs alpha

Question 5

Validity

Answer 5

Measure what it intends to measure?

Question 6

Key elements validity

Answer 6

1. Content validity > reflect conceptual domain, judged by experts
2. Face validity > reflect conceptual domain, judged by non-experts
3. Convergent validity > different measures of same construct should diverge
4. Discriminant validity > measures of different construct should NOT diverge

Question 7

Regression choice

+ solution correlations

Answer 7

Depends on data structure

1. Cross-sectional data: sample of units, 1 observation each
2. Time series data: 1 unit, multiple observations over time period 
3. Panel data: sample of units, multiple time periods
4. Clustered / nested data: hierarchical levels (employees, subs, MNE) Problem: observations are correlated Solution 1: cluster standard error  Solution 2: generalized estimating equations  Solution 3: include dummies as control variables  Solution 4: multilevel modeling

Question 8

OLS requirements

Answer 8

1. Linear
2. Random
3. Mean = 0
4. No multi-collinearity
5. Homotheticity
6. Error terms are normal

Question 9

Limited Dependent Variable Models

Answer 9

range of possible values is restricted in some important way

Question 10

marginal effect

Answer 10

How DV changes when a IV changes

When X is continuous >
When X is NOT continuous > Y,X=1 - Y,X=2

Question 11

Difference marginal effects vs corresponding coefficients

Answer 11

In OLS: yes
In the simple OLS model with linear effects, estimated coefficients = marginal effects

In MLS: no

Question 12

marginal effect of an interaction term

Answer 12

There is no marginal effect of an interaction term or any higher order terms

Question 13

STATA lessons

Answer 13

• When higher order terms / non-OLS model used > regression coefficients harder to interpretate
• Create higher order terms within regression command, otherwise imported as separate variables when calculating marginal effects
• Inform Stata about indicator variable = dummy by adding = i.

Question 14

Panel data

Answer 14

TSCS = sample of units over multiple time periods

Time series = periods can be daily, weekly, monthly etc.
Cross-sectional = units can be individuals, families, cities etc.

Question 15

Balanced vs unbalanced panel data

Answer 15

Balanced = every unit has all X + Y readings for T periods
Unbalanced = some units have missing values

Question 16

FE vs RE models

Answer 16

Fixed-effects model: separate fixed intercept controlling unobserved heterogeneity
> allows arbitrary correlation: YES
> not feasible when T is small and change is slow

Random-effects model: separate fixed intercept randomly draws from N
> allows arbitrary correlation: NO
> when interested in effects time-invariant covariates

Question 17

Outliers:

Problem + solution

Answer 17

Few observations which disproportionately influence results, especially when n is small
Refer to observations, not variables

Solution:

• Remove outliers > no
• Robust regression > only possible for OLS
• Winsoziring variables = limiting values 1/2%

Question 18

Control variables + note

Answer 18

explain the dependent variable, not those associated with independent variables

Note:

• Include “enough” controls to isolate the effect of our key variables, but should be randomized controlled experiment
• Data availability (missing values)
• Sample size (n is small, less control v)

Question 19

How avoid overcontrolling

Answer 19

1. Do not include 2 variables measuring same construct, different dimension is possible
2. Do not control mechanisms how X influences Y > not possible research total effect