Unit I: Causal Inference

Randomized Controlled Trials (RCT)

Q: What is an RCT and why is it important for causal inference?
A: An RCT randomly assigns subjects to a treatment or control group,
ensuring that, on average, both groups are statistically equivalent except for
the treatment. This breaks any confounding, yielding an unbiased estimate
of the treatment effect (the Average Treatment Effect, ATE)[1]. For example,
a medical trial randomly giving a new drug to half the patients and a placebo
to the other half can isolate the drug’s effect on health. By design, an RCT
satisfies the exogeneity condition (randomization guarantees no omitted
confounders) and often yields the Intention-to-Treat (ITT) and Local
Average Treatment (LATE) estimates of effect.
Figure: Flowchart of an RCT illustrating random assignment to treatment vs.
control. (From Wikimedia Commons.)
Exam-Style Question (12 marks): “Explain how random assignment in an
RCT ensures valid causal inference. Describe the potential outcomes
framework and define the Average Treatment Effect (ATE).”
Answer: In an RCT, each subject $i$ has two potential outcomes: $Y_i(1)$ if
treated and $Y_i(0)$ if not. We observe only one of these for each
individual. Random assignment ensures that treatment status $D_i$ is
independent of the potential outcomes $(Y_i(1), Y_i(0))$. Thus, the observed
difference in average outcomes equals the causal effect:

ATE=E [ Y ( 1 ) −Y ( 0 ) ] =E [ Y ∨D=1 ] − E [ Y ∨D=0 ] .

Since assignment is random, covariates (observed and unobserved) are

balanced, so this difference is unbiased[1]. The intent-to-treat (ITT) effect
measures $E[Y|D=1] - E[Y|D=0]$, and the Local Average Treatment
Effect (LATE) may be estimated via instrumental variables if only compliers
are affected by encouragement. Graphically, RCT results are often shown as
side-by-side boxplots or bar charts with confidence intervals.
Numerical Example: Suppose 100 students are randomly assigned to a
new teaching method or standard method. The treated group’s mean score
is 75, and the control group’s mean is 70. The estimated effect is $75 - 70 =
5$ points. Because of randomization, we attribute this 5-point difference to
the teaching method, not other factors. (Statistical inference would compute
the standard error for this difference.)
Instrumental Variables (IV)
Q: What is an instrumental variable and when is it needed?
A: IVs are used when the key regressor $X$ is endogenous (correlated with
the error term). An instrument $Z$ is a variable that (1) affects the
treatment $X$ (relevance) and (2) is as good as randomly assigned or
exogenous, having no direct effect on the outcome $Y$ except through
$X$ (exclusion restriction)[2]. Under these conditions, one can use two-stage
least squares (2SLS) to recover a causal effect. The first stage regresses $X$
on $Z$ (and controls); the second stage regresses $Y$ on the predicted $\
hat X$. In the exactly-identified case (one instrument per endogenous
regressor), the IV estimator has the closed form:

^β =( Z ′ X )− 1 Z ′ Y ,
IV

where $X$ is the matrix of regressors and $Z$ includes the instrument(s)
(and exogenous controls)[3]. Intuitively, this is the “ratio” of the jump in $Y$
to the jump in $X$ induced by $Z$.
Exam-Style Question (12 marks): “Explain the requirements for a valid
instrument and derive the IV estimator. Use the example of a randomized
scholarship as an instrument for education.”
Answer: A valid instrument $Z$ must satisfy: (i) Relevance: $\
operatorname{Cov}(Z,X)\neq0$ so $Z$ shifts $X$ (e.g. scholarship offer
increases schooling); (ii) Exogeneity: $Z$ is independent of other factors
affecting $Y$ (scholarship is as-if random); (iii) Exclusion: $Z$ has no direct
effect on $Y$ except through $X$[2]. Under these, we can write $X_i=\pi_0+\
pi_1 Z_i+u_i$ (first stage) and $Y_i=\alpha+\beta X_i + \varepsilon_i$
(structural). The 2SLS solution gives $\hat\beta=(Z'X)^{-1}Z'Y$[3]. For
example, if being offered a scholarship ($Z=1$) raises average years of
education $X$ by 2 years and raises average log-wages $Y$ by 0.1, the IV
estimate of the return to education is $0.1/2 = 0.05$ (5% per year).
Numerical Problem: A policy gives extra tutoring ($Z$) to some students,
raising their math score $X$ from 60 to 70 on average. The treated students’
test score ($Y$) rises from 80 to 85. Using the Wald (ratio) estimator: effect
= (85–80)/(70–60) = 5/10 = 0.5 points per unit of $X$.
Difference-in-Differences (DiD)
Q: What is DiD and what key assumption does it rely on?
A: DiD compares changes in outcomes over time between a treated and a
control group. With one pre- and one post-period, the DiD estimator is:

DID=( Ý treated,post − Ý treated,pre ) − ( Ý control,post − Ý control,pre ) ,

which nets out common trends[4]. Equivalently, estimate regression

$Y_{it}=\beta_1+\beta_2 d_i+\beta_3 t_t+\beta_4 (d_i\times t_t)+u_{it}$;
the DiD effect is $\beta_4$. The parallel trends assumption requires that,
absent treatment, the treated and control groups would have followed the
same time trend[4].
Exam-Style Question (12 marks): “Define the DiD estimator. Explain the
parallel trends assumption and how it can be assessed. Provide an example.”
Answer: Let $t=0$ be before and $t=1$ after a policy that affects group
$d=1$ (treated) but not $d=0$ (control). The DiD estimate is $(\bar
Y_{1,1}-\bar Y_{1,0}) - (\bar Y_{0,1}-\bar Y_{0,0})$[4]. Graphically, parallel
trends means lines of $Y$ over time for $d=1$ and $d=0$ would be parallel
before $t=0$. To check it, one may compare pre-treatment trends (if
multiple periods) or use placebo tests. For instance, if healthcare reform was
implemented in State A in year 2020 (treated) but not in State B (control),
DiD estimates the effect on health outcomes by comparing A’s pre-post
change minus B’s pre-post change.
Numerical Example: Suppose in 2019 (pre) the treated region’s average
income was 100 and control’s was 90; in 2021 (post) they are 110 and 95,
respectively. Then $\text{DID}=(110-100)-(95-90)=10-5=5$. This suggests
a 5-unit increase attributable to the policy (beyond the baseline trend).
Regression Discontinuity Design (RDD)
Q: What is RDD and how does it identify causal effects?
A: RDD exploits a cutoff in a continuous “running variable” $X$ that
determines treatment. If treatment is assigned when $X\ge c_0$, then units
just above and just below $c_0$ are assumed similar except for treatment.
The sharp RDD effect at the cutoff is
δ=lim E [ Y ∨X =x ] − lim E [ Y ∨X =x ] .
x ↓ c0 x ↑ c0

In a fuzzy RDD, treatment probability jumps at $c_0$ but not from 0 to 1, so

the local average treatment effect (LATE) is given by the Wald estimator[5]:

lim E [ Y ∨ X=x ] − lim E [ Y ∨ X=x ]

x↓ c 0 x↑ c 0
δ fuzzy = ,
lim E [ T ∨ X=x ] − lim E [ T ∨X =x ]
x ↓ c0 x↑ c 0

where $T$ indicates actual treatment[5]. The key assumption is continuity

of potential outcomes at $c_0$ (no other factor changes exactly at the
cutoff)[6].
Exam-Style Question (12 marks): “Explain the difference between sharp
and fuzzy RDD. State the assumptions needed and how the treatment effect
is estimated.”
Answer: In a sharp RDD, $T_i=1$ iff $X_i\ge c_0$. The treatment effect is
identified by the discontinuity in the conditional mean of $Y$ at $c_0$. In a
fuzzy RDD, $P(T_i=1|X_i)$ jumps at $c_0$ (not deterministically). Then we
use an IV-like ratio of jumps (Wald estimator)[5]. The key conditions are that
(i) the cutoff is exogenous (no manipulation of $X$ around $c_0$) and (ii)
$E[Y(0)|X]$ would be continuous at $c_0$ in absence of treatment[6]. In
practice, one plots $Y$ vs. $X$ and fits separate local regressions on each
side of $c_0$ to visualize the jump.
Numerical Example: Suppose a scholarship is awarded to students with
test scores $\ge$70. Students scoring 69 have an average GPA of 3.0; those
scoring 70 have 3.5. This jump of $0.5$ suggests the causal effect of
eligibility. If only 80% of students above 70 actually enroll in the scholarship
program (fuzzy case), and only 20% below enroll (due to appeals), then the
jump in enrollment is $0.8-0.2=0.6$, and the fuzzy RDD estimate is
$0.5/0.6 \approx 0.83$ (points of GPA per 100% increase in enrollment)[5].
Matching Methods
Q: What is matching and when is it used?
A: Matching pairs treated units with similar control units based on observed
covariates (or propensity scores), creating a balanced comparison group. For
example, in evaluating a job training program, each participant (treated)
might be matched to a non-participant with similar age, education, and prior
earnings. The idea is to mimic randomization by ensuring covariate balance.
A common approach is propensity score matching, where one first
models $P(T_i=1|\mathbf{X}_i)$ and then matches on this score. If done
well, the average difference in outcomes across matched pairs estimates the
treatment effect.
Exam-Style Question (12 marks): “Describe propensity score matching
and discuss its strengths and weaknesses.”
Answer: Propensity score matching involves: (1) estimating the probability
(propensity) of treatment given covariates, e.g. via $\Pr(T=1|X)=p(X)$; (2)
pairing each treated unit with one or more controls having similar $p(X)$;
and (3) comparing average outcomes between matched groups. This
controls for observed confounders. Strengths: Can reduce bias when
treatment selection is based on observables; only needs a low-dimensional
score instead of many covariates. Weaknesses: It cannot account for
unobserved confounders (selection on unobservables). If the propensity
model is misspecified or common support is limited (no suitable matches for
some treated), bias remains. Matching also may discard data (unmatched
units).
Numerical Problem: Suppose five treated units have covariate $X=(x_1,\
dots,x_5)$ and five potential controls have $X=(y_1,\dots,y_5)$. If $|x_i - y_j|
$ is minimized by matching $x_1$–$y_3$, $x_2$–$y_1$, etc., compute the
average treatment effect on the treated (ATT) as the mean difference in
outcomes across these pairs. Solution: After matching, ATT =
(1/5)∑ₖ(Y_{treated,k} – Y_{control,k}) for the matched pairs. (One would
plug in the numbers given in an exam scenario.)
Case Studies (Applications): Common examples include Angrist and
Krueger’s RDD (school-entry cutoff and later earnings), Card and Krueger’s
DiD (minimum wage change), or experimental studies like Duflo et al.’s
scholarship programs[7][6]. Such cases illustrate the methods above.

Unit II: Panel Data and Binary Models

Pooled OLS vs. Panel Models
Q: What are the differences between pooled OLS, fixed effects, and random
effects models?
A: Pooled OLS treats panel data as simple cross-sectional by ignoring
individual heterogeneity. If unobserved individual effects are correlated with
regressors, OLS is biased. The fixed effects (FE) model allows each
individual $i$ to have its own intercept $\alpha_i$:

Y i t =α i+ β X i t +u it .

Estimating by “within” (demeaning each $i$’s data) removes $\alpha_i$,

controlling for all time-invariant differences[8]. FE effectively uses only
within-unit variation. The random effects (RE) model assumes $\alpha_i$ is
random and uncorrelated with $X_{it}$, combining between- and within-
variation:

Y i t =α + β X i t + ui t +ϵ i t , ui t ∼ ( 0 , σ 2u ) , ϵ i t ∼ ( 0 , σ 2ϵ ) .

If the assumption $\Cov(\alpha_i, X_{it})=0$ holds, RE is consistent and

more efficient; otherwise FE is consistent. A Hausman test compares FE vs.
RE estimates to decide.
Exam-Style Question (12 marks): “Explain how the fixed-effects
estimator controls for unobserved heterogeneity. When is the random-effects
model appropriate? Describe the Hausman test.”
Answer: The FE estimator subtracts each unit’s time-mean: $Y_{it}-\bar Y_i
= \beta (X_{it}-\bar X_i) + (u_{it}-\bar u_i)$. Thus any time-invariant $\
alpha_i$ drops out, eliminating bias from omitted individual traits. The RE
model can use the between variation if $\alpha_i$ is uncorrelated with
$X_{it}$, gaining efficiency. The Hausman test checks consistency: under
the null of no correlation, FE and RE estimates coincide. A significant
difference suggests $\alpha_i$ is not exogenous, so prefer FE.
Numerical Example: A panel of 3 individuals over 2 years:
| i | t | $Y_{it}$ | $X_{it}$ |
|----|---|-------|-------|
| 1 | 1 | 10 | 5 |
| 1 | 2 | 12 | 6 |
| 2 | 1 | 15 | 7 |
| 2 | 2 | 17 | 9 |
| 3 | 1 | 20 | 10 |
| 3 | 2 | 19 | 8 |
FE would regress $(Y-\bar Y)$ on $(X-\bar X)$ (within each i). If individual
means are (11,16,19.5), then the demeaned series are e.g. for i=1: (–1, +1),
i=2: (–1, +1), i=3: (+0.5, –0.5). The FE slope is estimated from these within
changes. Pooled OLS would use all six points (ignoring $\alpha_i$) and may
give a different slope if $\alpha_i$ correlate with $X$.
Binary Outcome Models (LPM, Logit, Probit)
Q: Compare the Linear Probability Model (LPM) with Logit/Probit.
A: The LPM simply applies OLS to a 0/1 outcome. It has the form $P(Y=1|
X)=X'\beta$[9]. Advantages: simplicity and easy interpretation of $\beta$.
Drawbacks: predicted probabilities can lie outside [0,1], and
heteroskedasticity is inherent (Var$(Y|X)=p(1-p)$)[10]. Logit and Probit
models instead assume

X′β
e
P ( Y =1∨ X )= X′ β
( Logit ) , P ( Y =1∨ X )=Φ ( X ′ β )( Probit ) ,
1+e

where $\Phi$ is the standard normal CDF. These map $X'\beta$ into $(0,1)$
with an S-shaped curve. Estimation is by maximum likelihood. They allow for
nonlinear effects and properly restrict probabilities, but coefficients are less
directly interpretable (only odds-ratios or marginal effects).
Exam-Style Question (12 marks): “State the LPM and explain its
limitations. Contrast with a Probit model. When is each appropriate?”
Answer: The LPM: $Y_i = \beta_0+\beta_1X_i + \varepsilon_i$ with $Y_i\
in{0,1}$ implies $\Pr(Y_i=1|X)=X_i'\beta$[9]. Limitations include possible $\
hat p>1$ or $<0$, and non-constant variance. The Probit uses $\Pr(Y=1|X)=\
Phi(\beta_0+\beta_1X)$, ensuring predictions in (0,1). It assumes an
underlying latent normal variable. If the true relationship is nonlinear or
probabilities near 0/1, Logit/Probit is preferable. LPM may suffice for rough
interpretation or large samples (especially when marginal effects vary little).
One often runs a test or checks if many predictions of LPM fall outside [0,1]
[10].
Numeric Illustration: Suppose $\hat P(Y=1|X)=0.2+0.1X$. For $X=10$,
LPM predicts $P=1.2$ (impossible). In Logit, $\beta_0=−1.386,\;\
beta_1=0.1$ would imply $\Pr=1/(1+e^{-(-1.386+1)})=0.5$, always valid.
High-Frequency Data: Modern economics sometimes uses high-frequency
(e.g. intraday financial) data. Such data exhibit autocorrelation, volatility
clustering, and microstructure noise. Specialized models (e.g. GARCH for
volatility) are used, but this is beyond basic panel/binary models.

Unit III: Time Series Analysis

Stationarity and Unit Roots
Q: What is stationarity, and how do unit roots affect time-series?
A: A time series is stationary if its statistical properties (mean, variance,
autocovariance) do not change over time[11]. Formally, the joint distribution
of $(X_{t_1},…,X_{t_k})$ is the same as that of $(X_{t_1+\tau},…,X_{t_k+\
tau})$ for all $\tau$. Nonstationarity often arises from trends or unit roots. A
unit root means the series has a root equal to 1 in its characteristic
equation, implying shocks have permanent effects[12][13]. For example,
$X_t = X_{t-1} + \varepsilon_t$ is a unit root process (a random walk). Such
series can be differenced once ($\Delta X_t$) to become stationary. The
Augmented Dickey-Fuller test is commonly used to test for a unit root.
Exam-Style Question (12 marks): “Define (weak) stationarity. Explain
what a unit root is and how to test for it. Why is stationarity important for
time-series modeling?”
Answer: A series is (weakly) stationary if $E[X_t]=\mu$ constant and $\
Cov(X_t,X_{t+h})$ depends only on $h$ (not $t$)[11]. If instead $X_t$
contains a unit root (e.g. AR(1) with $\phi=1$), it is nonstationary: shocks
change the mean permanently[12][13]. Stationarity ensures that sample
moments are stable over time and that models (like ARMA) have well-defined
behaviors. To test for a unit root, one performs the Dickey-Fuller regression
$\Delta X_t = \alpha + \gamma X_{t-1} + u_t$ and tests $\gamma=0$. If we
fail to reject $\gamma=0$, the series likely has a unit root. If present, we
often difference the data before fitting ARMA/ARIMA models.
Autoregressive (AR) Models
Q: What is an AR(p) model?
A: An AR($p$) model regresses the series on its own lagged values[14]. For
example, an AR(1) is

X t =φ1 X t −1 + ε t , ε t ∼ WN ( 0 , σ 2 ) .

More generally, AR($p$): $X_t = \varphi_1X_{t-1} + \dots + \varphi_pX_{t-p}

+ \varepsilon_t$, where $\varepsilon_t$ is white noise[14]. Stationarity of AR
models requires the roots of the lag polynomial to lie outside the unit circle.
For instance, AR(1) is stationary only if $|\varphi_1|<1$. The AR model
captures persistence: e.g. in AR(1), the mean of $X$ is 0 and $\Var(X_t)=\
sigma^2/(1-\varphi_1^2)$ if $|\varphi_1|<1$.
Moving Average (MA) Models
Q: What is an MA(q) model?
A: An MA($q$) expresses $X_t$ as a linear combination of current and past
shocks[15]. The MA($q$) model is

X t =μ+ ε t +θ 1 ε t −1 +⋯+θ q ε t − q ,

with $\varepsilon_t$ i.i.d. white noise[15]. Unlike AR models, a finite MA is

always stationary[16]. The MA handles short-term correlations: the
autocorrelation function (ACF) of an MA($q$) cuts off after lag $q$. For
example, MA(1): $X_t = \mu + \varepsilon_t + \theta_1\varepsilon_{t-1}$.
ARIMA Models and Forecasting
Q: What is an ARIMA model and how is it used for forecasting?
A: ARIMA stands for Autoregressive Integrated Moving Average. It
generalizes ARMA to nonstationary series by differencing[17]. An
ARIMA($p,d,q$) model difference the series $d$ times to make it stationary,
then fits AR($p$)+MA($q$) to $\Delta^d X_t$[17]. For example, an
ARIMA(1,1,1) on $X_t$ is equivalent to ARMA(1,1) on $\Delta X_t$. Once
fitted, one can forecast future values. Forecasting uses the model’s
equations to generate $\hat X_{t+h}$ and associated confidence intervals.
For example, in an AR(1): $\hat X_{T+1} = \varphi_1 X_T$, and forecast
variance grows over horizon. Model selection often uses AIC/BIC or the
[Link] procedure.

Exam-Style Question (12 marks): “Describe the Box-Jenkins approach to

time series modeling. How does one build an ARIMA model and use it for
forecasting?”
Answer: Box-Jenkins modeling involves: (1) Identification: test stationarity
(ADF test) and difference if needed ($d$). Examine ACF/PACF of (differenced)
series to choose $p,q$. (2) Estimation: Estimate AR($p$) and MA($q$)
parameters (often by MLE). (3) Diagnostics: Check residuals for whiteness
and no autocorrelation. (4) Forecasting: Use the model to predict
$X_{T+h}$. For example, an ARIMA(1,1,1) is $\Delta X_t = \varphi_1 \Delta
X_{t-1} + \varepsilon_t + \theta_1\varepsilon_{t-1}$. After estimation, a 1-
step forecast from time $T$ is $\hat X_{T+1} = X_T + \hat\varphi_1 (X_T -
X_{T-1}) + \hat\theta_1 \hat\varepsilon_T$. One also constructs forecast
intervals assuming normal errors. This method is widely used for economic
forecasting (e.g. GDP, inflation).
Numerical Example: A series follows $X_t = 1.2\,X_{t-1} + \varepsilon_t$
(an AR(1) with $\varphi_1=1.2$, nonstationary). Differencing: $\Delta X_t =
1.2\,\Delta X_{t-1} + \varepsilon_t$. Estimate $\hat\varphi_1=0.8$ after
differencing. If $X_{100}=200$, the 1-step ahead forecast is $\hat
X_{101}=200 + 0.8\,(200-199) = 200.8$.

Unit IV: Data Collection & Unstructured Data

Survey Design and Sampling
Q: What are key issues in designing a survey?
A: Survey design involves (i) Defining the population and frame: Decide
who to survey (e.g. all households in a region) and ensure a complete
sampling frame. (ii) Sampling method: Common methods include simple
random, stratified, cluster sampling. Stratification (e.g. by region) can
improve precision. (iii) Sample size: Larger samples reduce sampling
error (the difference between sample and true population values)[18]. (iv)
Questionnaire design: Clear, unbiased questions avoid measurement error
or response bias. (v) Fieldwork: Minimize nonresponse and interviewer
effects.
A sampling error is the deviation of a sample statistic from the true
population value due to randomness[18]. It can be quantified by the
standard error formula (e.g. $\sigma/\sqrt{n}$). To reduce sampling error,
increase $n$ or improve representativeness.
Exam-Style Question (12 marks): “Explain the difference between
probability and non-probability sampling. Discuss how to minimize bias in
survey data.”
Answer: Probability sampling (e.g. simple random, stratified) ensures
each unit has a known chance of selection, allowing statistical inference.
Non-probability sampling (e.g. convenience samples) does not, risking
bias. To minimize bias, use random or stratified sampling, pre-test
questionnaires, train interviewers, and follow up nonrespondents. The
sampling error is $E[\bar X_{\text{sample}}]-\mu_{\text{pop}}$, ideally
zero in expectation[18]. Other errors include measurement (poor questions)
and coverage (missing subgroups).
Application Example: A nationwide opinion poll might use multistage
cluster sampling: randomly select districts, then villages, then households.
Within each, select one respondent (maybe stratified by age/gender). The
design effect of clustering is accounted for in standard errors.
Text as Data (Unstructured Textual Data)
Q: How is textual data used in economics?
A: Economists increasingly analyze text (news, social media, speeches) to
quantify sentiment, topics, or networks. For example, one can create a
sentiment index from newspaper articles or measure inflation expectations
from central bank minutes via word counts. Techniques include bag-of-
words, topic models, and embeddings. The general idea is to convert
unstructured text into numeric features that feed into statistical models (e.g.
using Python’s NLP libraries). Case examples: Caldara and Iacoviello’s
economic policy uncertainty index, or Sentiment Analysis of Tweets to
predict stock markets.
Exam-Style Question (12 marks): “Discuss how you would incorporate
unstructured textual data into an economic study. Provide an example using
sentiment analysis.”
Answer: One approach is to assign a sentiment score to each document.
For instance, use a lexicon of positive/negative words to score newspaper
articles on economic outlook. These scores become a time series that can be
regressed on macro variables (e.g. consumption). Alternatively, use machine
learning (word embeddings) to classify tone. Example: An economist might
compute a “consumer confidence” index by tallying words like “good”, “bad”
in consumer surveys over time. Challenges include ensuring reliability and
dealing with high dimensionality.
Spatial and Network Data
Q: How are spatial and network data used in quantitative economics?
A: Spatial data involve geographic information (e.g. GIS coordinates).
Economists use it to study regional development, environmental resources,
or spatial spillovers. For example, one might include distance to city center
or neighboring region outcomes in regressions. Geostatistical tools (e.g.
spatial lag models) account for spatial autocorrelation. Network data
capture relationships (e.g. trade links, social networks). Analysis might
involve adjacency matrices and network regressions (like peer effects
models). For instance, an economist could model how technology adoption
spreads through a network of farmers using graph theory. These data
expand traditional “observations” to include interactions and space.
Exam-Style Question (12 marks): “Explain how you would use GIS
(geographic information systems) in an economic analysis. Discuss spatial
autocorrelation and one method to address it.”
Answer: GIS enables mapping each unit (e.g. household or district) to
coordinates or regions. One can overlay economic data on maps to visualize
patterns. In regression analysis, spatial autocorrelation means nearby
units have similar outcomes (violating i.i.d.). To account for this, include
spatial lags (e.g. average outcome of neighbors) or use spatial error models.
For example, when studying house prices, distance to amenities (schools,
parks) can be included. Testing (Moran’s I) can detect spatial clustering.
Final Remarks: Together, the above methods—spanning causal inference,
panel regressions, time series, and data collection—form the toolkit of the
Quantitative Economics course. They ensure that students can critically
analyze data patterns and draw credible conclusions about economic
phenomena.
Sources: Key concepts and formulas are drawn from course materials and
standard references (MIT OCW, econometrics texts) as cited above[1][4][5]
[6][10][2].

[1] Randomized controlled trial - Wikipedia

[Link]
[2] [3] [7] Endogeneity_IV_MIT (1).pdf
[Link]
[4] Difference in Difference (1).pdf
[Link]
[5] Regression Discontinuity Design_Slides 2 (1).pdf
[Link]
[6] 6 Regression Discontinuity – Causal Inference
The Mixtape
[Link]
[8] Fixed effects model - Wikipedia
[Link]
[9] [10] Linear probability model - Wikipedia
[Link]
[11] [12] [13] Stationary process - Wikipedia
[Link]
[14] Autoregressive model - Wikipedia
[Link]
[15] [16] Moving-average model - Wikipedia
[Link]
[17] Autoregressive integrated moving average - Wikipedia
[Link]
[18] Sampling Errors in Statistics: Definition, Types, and Calculation
[Link]