Week 1 8-2

What determines someone's identity when it comes to style

Use variables to predict outcome of the dependent variable
What model do you estimate
What sample do you gather

Path diagram → Model latent variables (predicted by observable variables)

Indicators → Say something about an underlying factor
E.g. Specific bag (indicators). Is that specific Anne fleur or can it also be something else

Y = B0 + B1 *F + E
E = Error term

Why does F explain the indicator

One factor model

Description of study on half A4 form

Week 2 12-2
Logistic regression model
How likely are you to be in a certain group (how likely are you to be a person that identifies
herself as Marie Claire)

Transform coefficient first before interpreting it

Negative → Decreases inclination of being of a certain type
Below 50% → You’re not of that type
Logit prediction always lies between 0 and 1
Probability model is exactly the same as log on model

Y = 1 (does identify as …)
Y = 0 (does not identify as …)
Y* → Measure for latent variable for Gothic Fashion Identity that is modeled for
every individual “i”

Latent variable model

Yi* = B0 + ΣjBjXij + ei
Y* cannot be observed, so you have to make assumptions

Y=1 when Y*>0

Model the errors:

Logistic distribution function → Logit model
Normal distribution → Probit regression model

Maximum likelihood estimation (MLE) → The distribution of your dependent variable,

y, is known (Logit or probit model), but the parameters characterizing the
distribution are unknown.

Likelihood function (based on the known distribution) and choose (estimate)

Predictions should be as closed to the reality as possible

Don't learn likelihood density function for normal distribution by heart.

Week 2.2
Use noisy indicators to come up with a prediction for a latent variable

According to the modern day factor analysis: Spearman's model is a one factor model
Graph of 1-factor model
Indicators are caused by the factor
The factor is driving the indicators

Spearman → There's a latent factor of intelligence but it cannot be observed, but I

can assess you on many different scales which gives you a score.

Factor analysis → Exploit correlation to estimate betas to measure intelligence.

SPSS does this.
Confirmatory factor analysis

Errors are independent

LOOK SLIDES FOR THE 1 FACTOR MODEL

Correlate a lot → Covariance is high → Have a lot of common variance

Split up variance into two parts → Common variance and unique variance
Reliability is R2 of the factor model

Wisdom of the crowd → The crowd average will do better than the average person
does

Is the data set suitable?

● The Kaiser-Meyer-Olkin test (KMO) → Sampling adequacy
○ Happy with scores around 0.7
○ <0.5 so a lot of unique variance → Bad, start again
● Bartlett’s test of sphericity → Tests whether correlation matrix looks like an
identity matrix.
○ Expect that the dataset is highly suitable for factor analysis → Reject
the null, because null says they are zero
○ Small value of the significance level → Highly suitable for factor
analysis

Cronbach's alpha → How consistent the questions are with each other

Analyze → Dimension reduction → Factor → Put indicators in variables box

Extraction → Principle nog wat → Maximum likelihood → How many factors want to
extract (fixed number of factors and choose 1) → Save as variables

Week 3.1
The amount of variance you can model → Want as high as possible
Want to minimize a number of factors…

2 methods we need to learn

1. Yin = BiFin + ein
2. Yin = BiF1in+YF2in + ein

B = A vector of coefficients
Eigenvalue → Of a factor explains how much variance is explained by a factor
● The larger the eigenvalue → The more relevant the underlying factor is in
explaining common variance shared by the indicator
● The cattell scree test and the kaiser criterion → Two common tests to
examine the relative size of the eigenvalues of the factors.

Cattell’s Scree test → Plot with ranked indicators and shows each of the factors the
size of the eigenvalue. Count number of factors before the elbow (turn)

Choose the solution for which the model is best interpretable

Rotation matrix
● Implies that we can do whatever we want in order to facilitate interpretation
● Orthogonal rotation → E.g. VARIMAX, the underlying factors are assumed to
be uncorrelated
● Oblique rotation → E.g. OBLIMIN, allows for correlation between factors

BIG 5 → Multiple factor model

Week 3.2
Empirical strategy → Research plan with identification and ….

Regression discontinuity design

Potential outcomes

The LATE theorem

Zi is the instrument, indicating treatment assigned. D1i indicates subject i’s treatemtn status
when Zi=1 and D01 indicates subject i’s treatment status when Zi=0
Y1i and Y01 → potential outcomes for subject i in the treated and untreated states
Y1i - Y01 is the causal effect of treatment on individual i, but this we can never see

IV chain reaction
Winning lottery → charter attendance → math scores
First stage = Effect of the Winning lottery on charter attendance
Second stage = Link between attending a charter and an outcome variable (math scores)
Instrumental variable → Variable that initiates the chain reaction. (winning lottery)
Reduced form = Effect of the instrument on the outcome
First stage * second stage = Reduced form
Second stage is not directly observed and is therefore the causal effect we’re trying to
determine

Substantial first stage → Instrumental variable must really change the variable
we’re interesting in
Independence assumption → Winning lottery should be as good as random
Exclusion restriction → Require the instrument change outcomes solely through the
variable of interest.

Attendance effect = reduced form / first stage

LATE → Local average treatment effect

Four types of children
● Always taker
● Complier
● Never taker
● Defier

Noble lottery offers → noble enrollment → ACT scores

1.
Angrist’s work improves our understanding of causality in the social sciences because it uses
LATE to look at causations without the traditional way without having a control and treatment
group that is randomly selected, but by the use of RD and IV which still makes it possible to look
at causation without randomization.
It is impossible to have a full controlled trial without lurking variables, so Angrist mentions to
always be aware of the lurking variables. Use random aspects in society in order to address
causal questions: Regression discontinuity, local average treatment estimates,
instrumental variables and natural experiments.
The correlation between education level and income is well-known, but establishing causation
requires considering factors like directionality, third variables, selection bias, and mediating
factors. Rigorous research methods are needed to determine if education directly influences
income or vice versa. Understanding these causal mechanisms helps policymakers and
educators develop effective strategies to reduce income inequality

2.
It is important to have reliable methods like those developed by Angrist to measure the effects
of policies or interventions

5.
Instrumental variable → Initiates chain reaction for the random parts of the
individuals you’re studying. Not related to any other variable. What makes the
instrument interesting is that this part is random.
Exclusion restriction

Week 4.2
Frequentist
Beyesian approach → Deals with the probability of a hypothesis based on new
evidence and a belief of the truth of the hypothesis (e.g. past evidence)

Basetheorum

Parameter → Characteristic of the population

Probability rules:
● In between 0-1
● P(S) = 1
● Complement rule: A: P(Ac) = 1-P(A)
● Addition rule
● Multiplication rule

Week 5.1
Guess who game
P(H) = 1/22
P(E∨H )∗P(H )
P(H ∨E)=
P (E∨H )∗P (H)+ P( EH )∗P (H)
P(E)=P (E∨H )∗P ( H)+ P( EH )∗P ( H)

Q1 White sweater → No
P(no white sweater given that' s Nynke)∗P(H )
P(Nynke∨no white sweater)=
P(E∨H )∗P( H )+ P(EH )∗P( H )
P (E|H) = 1
P(H) → 1/22
P(E) → 17/22
1/22
So = 0.0588
17/22
Q2 Piercing → Yes
P(has a piercing giventhat' s Nynke)∗P (H)
P(Nynke∨has a piercing)=
P(E∨H )∗P(H )+ P(EH )∗P(H )
P (E|H) = 1
P (H) = 1/17
P (E) = 12/17
1/17
So =0.0833
12/17

Q3 Brown hair → Yes

P(has brown hair giventhat' s Nynke)∗P (H)
P(Nynke∨has brown hair)=
P(E∨H)∗P(H )+ P (EH )∗P(H )
P (E|H) = 1
P (H) = 1/12
P (E) = 8/12
1/12
So =0.125
8/12

Q4 Necklace → Yes
P (has a necklace given that' s Nynke)∗P(H )
P(Nynke∨has a necklace)=
P( E∨H)∗P(H )+ P( EH )∗P(H )
P (E|H) = 1
P (H) = 1/8
P (E) = 6/8
1/8
So =0.166
6/8

Q5 Sneakers → Yes
P(has sneakers giventhat' s Nynke)∗P (H)
P(Nynke∨has a sneakers)=
P(E∨H )∗P(H )+ P(EH )∗P(H )
P (E|H) = 1
P (H) = 1/6
P (E) = 5/6
1/6
So =0.2
5/6

Q6 Shoulder length hair → Yes

P(has shoulder length hair giventhat' s Nynke)∗P (H )
P(Nynke∨has shoulder length hair )=
P (E∨H )∗P (H)+ P( EH )∗P (H)
P (E|H) = 1
P (H) = 1/5
P (E) = 2/5
 1/5
So =0.5
2/5

Q7 Is it Nynke → Yes
P(is Nynke giventhat' s Nynke)∗P(H )
P(Nynke∨is Nynke )=
P(E∨H )∗P(H )+ P(EH )∗P(H )
P (E|H) = 1
P (H) = 1/2
P (E) = 1/2
1/2
So =1
1/2

Only consider the first 4 people

Alice has 60% of liking paper books
So then P(E|H) isn’t 1 but it’s 0.6 if it’s her in the game

Inference → Observe sample data and say meaningful things about the population

Posterior probability → P(H|E)

P(E) → Normalizing constant

Outcome of P(E|H) * P(H) is proportional to P(H|E)

Likelihoods

Week 5.2
Credible intervals

Beta > alpha → right skew

The higher # of alpha and beta → the higher the peek of the distribution