1/120
Comprehensive review flashcards covering simple linear regression, OLS properties, assumptions, inference, heteroskedasticity testing, robust standard errors, multiple regression, omitted variable bias, multicollinearity, nonlinear specifications, logarithms, control variable selection, interaction effects, and joint F-tests from EKHM65 Econometrics I.
In the regression line E(Y∣X)=β0+β1X, what does the intercept β0 represent?
The expected/average value of Y when X=0.
In the regression line E(Y∣X)=β0+β1X, what does the slope β1 represent?
The change in the expected/average value of Y when X increases by one unit.
Why does the linear regression model assume a straight line?
It is a simple way to summarize how the average value of Y changes with X, assuming that the average relationship is approximately linear.
What is the error term u in the population regression model Y=β0+β1X+u?
The difference between the observed value Y and the conditional mean of Y given X, defined as u=Y−E(Y∣X).
What is the sample residual ei (or u^i) in Ordinary Least Squares (OLS)?
The difference between the observed value and the fitted value: ei=Yi−Y^i=Yi−β^0−β^1Xi.
What mathematical criterion does the Ordinary Least Squares (OLS) estimator minimize?
It minimizes the sum of squared residuals: ∑i=1nei2=∑i=1n(Yi−β^0−β^1Xi)2.
What is the formula for calculating the OLS slope estimator β^1?
β^1=∑i=1n(Xi−Xˉ)2∑i=1n(Xi−Xˉ)(Yi−Yˉ).
What is the formula for the OLS intercept estimator β^0?
β^0=Yˉ−β^1Xˉ.
In the college GPA dataset with Yˉ=3.057 and Xˉ=1.076, what is the calculated value of β^0?
β^0=3.057−(−0.0895×1.076)=3.153.
What does the coefficient of determination R2 measure?
The fraction of the sample variation in Y that is accounted for by variation in the fitted values Y^, calculated as R2=TSSESS.

In regression analysis, what does the bracketed quantity in this diagram represent?
Total variation in Y, or Total Sum of Squares (TSS=∑i=1n(Yi−Yˉ)2).
What is the mathematical definition of Explained Sum of Squares (ESS)?
ESS=∑i=1n(Y^i−Yˉ)2.
What is Least Squares Assumption #1 for simple linear regression?
The conditional distribution of u given X has a mean of zero: E(u∣X=x)=0.

What condition does this graph illustrate regarding Least Squares Assumption #1?
Zero conditional mean violated: E(u∣X)=0, where the actual conditional mean departs from the assumed regression line.
What is Least Squares Assumption #2?
(Xi,Yi),i=1,…,n are independently and identically distributed (i.i.d.) draws from the population.
What is Least Squares Assumption #3?
Extremely large or small values of X and Y are sufficiently rare (outliers are unlikely).

What econometrics concept does this figure demonstrate regarding OLS regression?
The sensitivity of OLS regression to an outlier, showing how a single extreme observation rotates the regression line.
What are the three core asymptotic properties of the OLS estimator β^1 under the standard assumptions?
Unbiasedness (E(β^1)=β1), Consistency (β^1pβ1 as n→∞), and Asymptotic normality in large samples.
In the Stata output for reg colGPA skipped, what were the sample size, R2, and slope coefficient for skipped?
Number of observations = 141, R2=0.0685, and coefficient = −0.0895215.

Using the prediction equation Y^=3.15−0.09X, what is the predicted college GPA for a student skipping 2.0 lectures per week?
Y^=3.15−0.09(2.0)=2.97.
What do degrees of freedom represent in regression estimation, and what is the formula?
Degrees of freedom indicate the amount of independent information remaining after estimating parameters: df=n−k, where k is the number of estimated parameters (n−2 for simple OLS).
In the college GPA regression where β^1=−0.0895 and SE(β^1)=0.028, what is the t-statistic for testing H0:β1=0?
t=0.028−0.0895−0=−3.19, which leads to rejecting H0 at the 5% level (∣−3.19∣>1.96).
What is the assumption of homoskedasticity in regression analysis?
The variance of the error distribution is constant across all values of X: Var(u∣X=x)=σ2.
What does the Gauss-Markov theorem state about OLS under homoskedasticity (BLUE)?
OLS is the Best Linear Unbiased Estimator (BLUE), meaning it has the minimum variance (most efficient) among all linear unbiased estimators.
What happens to the OLS coefficient estimates in the presence of heteroskedasticity?
They remain unbiased, consistent, and asymptotically normal.
What is the primary damage caused by heteroskedasticity in OLS regression?
The conventional standard errors are invalid (typically underestimating the true parameter variance), which inflates t-statistics and makes hypothesis tests unreliable.
What is the mathematical formula for variable standardization?
Zx=sxX−Xˉ, rescaling X to have a mean of zero and standard deviation of one.
How is the slope coefficient interpreted when the independent variable X is standardized?
The expected change in Y associated with a one standard deviation increase in X.
How is the constant interpreted when the independent variable X is standardized?
The expected value of Y for an individual with the sample-average value of X (since Zx=0 at the mean).
In the regression of chborn on std_educyears (n=33,587), what were the coefficient on std_educyears and the constant?
Coefficient = −0.7651867 (one SD increase in education reduces children by 0.77) and constant = 3.131569 children.
What are the null and alternative hypotheses for the Breusch-Pagan test for heteroskedasticity?
H0: Homoskedasticity (constant residual variance) vs. H1: Heteroskedasticity (non-constant residual variance).
What are the five steps to conduct the Breusch-Pagan test for heteroskedasticity?
(1) Estimate main regression, (2) compute squared residuals u^2, (3) regress u^2 on the explanatory variables, (4) compute LM=n×R2 from that auxiliary regression, (5) compare LM to χcrit2(q).
Why does the Breusch-Pagan test use squared residuals u^2 instead of raw residuals?
Because error size (dispersion) matters rather than sign; large positive and negative residuals both indicate high variance.
In the Breusch-Pagan auxiliary regression for chborn on std_educyears, what was the sample size and the auxiliary R2?
Sample size n=33,587 and auxiliary R2=0.0245.
What was the calculated Breusch-Pagan test statistic LM in the chborn regression, and what was the conclusion at the 5% level?
LM=33,587×0.0245=822.9. Since 822.9>χcrit2(1)=3.84, the null hypothesis of homoskedasticity is rejected.
What are heteroskedasticity-robust standard errors (Eicker–Huber–White standard errors)?
Standard errors that give more weight to observations with larger squared residuals when estimating parameter variance, correcting inference without altering point estimates.
When , robust is specified in Stata, what happens to the point estimates of the coefficients?
The coefficients remain completely identical; only the standard errors, t-statistics, p-values, confidence intervals, and F-statistic change.
What happens to the Stata output header when the , robust option is used?
The ANOVA table (Source, SS, MS) is replaced by 'Linear regression', and the Adjusted R2 line is omitted from the printed output.
How can Adjusted R2 be displayed in Stata after running a regression with , robust?
By executing the command display e(r2_a).
What causes omitted variable bias (OVB) in multiple regression?
OVB occurs when an omitted variable is correlated both with the included regressor X and with the dependent variable Y, leading to E(u∣X)=0.
What is the population regression function for a model with two explanatory variables?
E(Y∣X1=x1,X2=x2)=β0+β1X1+β2X2
How is the coefficient β1 interpreted in the model Y=β0+β1X1+β2X2+u?
The expected change in Y per one-unit change in X1, holding X2 constant (a partial effect).
Why does plain R2 mechanically increase every time an additional predictor is added to an OLS model?
Because OLS minimizes the sum of squared residuals, it can never perform worse when additional information is available.
What is the formula for Adjusted R2?
Adj. R2=1−[n−k−1n−1]×(TSSSSR), where k is the number of regressors excluding the constant.
What is the difference between perfect multicollinearity and imperfect multicollinearity?
Perfect multicollinearity occurs when one regressor is an exact linear function of others (causing Stata to drop a variable); imperfect multicollinearity occurs when regressors are highly but not perfectly correlated.
What are textbook symptoms of severe imperfect multicollinearity in regression output?
Coefficients with unexpected signs or magnitudes, dramatic shifts in estimates when variables are added/dropped, and large standard errors/insignificant t-tests despite a high R2.
What is the Variance Inflation Factor (VIF), and what are its common rules of thumb?
VIF quantifies how much a coefficient's variance is inflated by correlation with other regressors: VIF=1 means no correlation, VIF≥4 warrants scrutiny, and VIF>10 indicates severe multicollinearity.
In the occupational score example, what accounting identity linked age, labor market experience, and years of education?
agei≈experiencei+edui+7
In the occupational score regression with age, experience, and edu, what were the VIF values for these three variables?
age = 1266.08, experience = 1320.36, and edu = 93.47 (Mean VIF = 383.59).
Why did edu have a VIF of 93.47 despite modest pairwise correlations with age (0.056) and experience (−0.210)?
Because multicollinearity reflects how well a regressor is predicted by a linear combination of all other regressors jointly, not just pairwise correlations.
How did dropping age from the occscore model affect the coefficients on edu and experience?
edu returned to a positive, significant 1.118 (t=18.05), and experience became positive and significant (0.103, t=7.49).
In the census fertility model, how many categories did urban/rural and region of birth have, and how many dummy variables were included?
Urban/rural had 2 categories (1 dummy), and region of birth had 16 categories (15 dummies), following the k−1 rule.
In the census fertility dataset (n=33,587), what percentage of the sample lived in urban versus rural areas?
Urban = 61.49% (20,651 observations), Rural = 38.51% (12,936 observations).
Which categories were chosen as the omitted reference categories in the census fertility specification?
New England (for region of birth) and rural (for urban residence).
Why did foreign-born regions like Africa (n=5) and Oceania (n=6) yield insignificant p-values (0.986 and 0.883) in the census regression?
Their extremely small sample sizes resulted in huge standard errors (0.77 and 0.70) and very low statistical power, preventing precise estimation.
In the full census fertility model, what does the constant β^0=3.4436 represent?
The predicted number of children (3.44) for a rural woman born in New England with the sample-average years of education.
How did the coefficient on std_educyears change between the simple model and the multiple regression model controlling for urban residence and region of birth?
It shrank from −0.765 in the simple model to −0.639 in the multiple model.
Why did the coefficient on std_educyears shrink in magnitude when urban residence and region of birth were controlled for?
Because more educated women were disproportionately urban and concentrated in lower-fertility regions; the simple model suffered from omitted variable bias.
What was the estimated coefficient on urban in the simple model vs. the multiple regression model?
−1.177 in the simple model, shrinking to −1.055 in the multiple regression model.
How did the coefficient for South Atlantic change when comparing the full model to the region-only model?
It more than doubled, increasing from 0.503 in the full model to 1.137 in the region-only model.
What lesson does the shift in region coefficients provide about omitted variable bias?
OVB does not push every coefficient in the same direction or by the same amount; it depends on how the omitted variable correlates with each specific regressor.
What is the formula for predicted children chborn in the full census model?
chborn=3.4436+βregion+βurban×(1 if urban, else 0)+(−0.6387)×std_educyears
What is the predicted number of children for a rural woman born in the Mountain region (β=1.0224) with average education?
chborn=3.4436+1.0224+0+(−0.6387×0)=4.47 children.
What is the predicted number of children for an urban woman born in the Pacific region (β=−0.3508, βurban=−1.0553) with education 1 SD above average?
chborn=3.4436−0.3508−1.0553−0.6387(1)=1.40 children.
What is the predicted number of children for a rural woman born in South/Central America (β=1.2471) with education 1 SD below average?
chborn=3.4436+1.2471+0+(−0.6387×−1)=5.33 children.
In the occupational score model with region dummies (reference: Northeast), what is the predicted difference between Midwest (β^=−1.93) and South (β^=−3.06)?
β^Midwest−β^South=−1.93−(−3.06)=1.13 points higher in the Midwest.
In the linear fertility model with age_marriage, what was the estimated coefficient on age at marriage?
−0.2113 (t=−34.06), meaning each additional year delayed is associated with 0.21 fewer children.
What restrictive assumption does a single linear term for age_marriage impose?
It forces the decline in children to be identical (0.21 children per year) regardless of whether marriage is delayed from 20 to 21 or from 30 to 31.
What is a polynomial regression model?
A regression specification that includes powers of a continuous regressor (such as X2,X3,X4) to model nonlinear curves.
Why can OLS estimate a polynomial model directly without departing from the linear regression framework?
Because the model is nonlinear in the variables, but remains linear in the parameters (β coefficients).
In the quadratic marriage model chborn=10.95−0.43×age_marriage+0.003×age_marriage2, how does the marginal effect differ between age 20–21 and age 30–31?
The predicted decline is 0.28 children between ages 20 and 21 (3.84−3.56), but only 0.21 children between ages 30 and 31 (1.34−1.13).
What is the sequential rule for determining how many polynomial terms to include in a regression?
Add the first polynomial X2 and retain if significant; then add X3 and retain if significant; continue sequentially.
What curve shape does adding a cubic term (X3) allow the regression function to take?
An S-shaped curve, allowing the relationship to change direction twice.
What are two dummy-variable alternatives to polynomial regression for modeling nonlinear relationships?
(1) A categorical (binned) specification grouping continuous values into ranges, and (2) A fully flexible specification with a dummy for every unique value.
In the binned age at marriage model (reference: 21–25), what were the coefficients for ages 15–20 and 26–30?
am_15_20 = 1.1984 (t=27.31) and am_26_30 = −0.9675 (t=−27.22).
What is the primary advantage and primary drawback of the fully flexible dummy specification?
Advantage: it imposes zero shape assumptions on the relationship; Drawback: it consumes many degrees of freedom, balloons standard errors at sparsely populated values, and risks fitting noise.
What mathematical base does the natural logarithm (ln) use, and for what values is it defined?
Base e≈2.71828, and it is defined strictly for positive values (X>0).
What is the mathematical approximation for small changes in natural logarithms?
Δln(X)≈100%ΔX, or ln(x+Δx)−ln(x)≈xΔx.
What is the equation and coefficient interpretation of a linear-log model?
Y=β0+β1ln(X); a 1% increase in X is associated with a 0.01×β1 unit change in Y.
In the linear-log model Income=−912.53+501.79×ln(age), what is the predicted effect of a 1% increase in age?
An increase of approximately 501.79×0.01=$5.02 in income.
What is the equation and coefficient interpretation of a log-linear model?
ln(Y)=β0+β1X; a one-unit increase in X is associated with an approximately (100×β1)% change in Y.
In the log-linear model ln(Income)=5.958+0.021×age, what is the effect of one additional year of age?
An approximately 2.1% increase in income.
Why did the sample size fall from 930,339 to 705,151 when switching from the linear-log to log-linear income model?
Because ln(Y) is undefined at Y=0, so logging income dropped all observations with zero recorded earnings (about 25% of the sample).
What happened to the age coefficient when zero-income observations were artificially recoded to 1 (Model 4)?
The age coefficient turned negative (−0.009), because non-earners anchored at ln(1)=0 skewed older, creating an artificial negative slope.
What is the equation and coefficient interpretation of a log-log model?
ln(Y)=β0+β1ln(X); a 1% change in X is associated with a β1% change in Y (elasticity).
In the log-log model ln(Income)=3.882+0.798×ln(age), what is the estimated elasticity?
0.798 (a 1% increase in age is associated with an approximately 0.8% increase in income).
What are the five criteria that make an econometric model good?
(1) Plausible substantive relationship, (2) functional form fits data reasonably well, (3) interpretable coefficients, (4) avoids unnecessary complexity, (5) performs well according to relevant diagnostics.
Why can Adjusted R2 not be used to compare Model 1 (Y=Income) and Model 2 (Y=ln(income))?
Because Adjusted R2 measures variation explained in Y; differently scaled dependent variables have different total variations and are not directly comparable.
Why is Adjusted R2 non-comparable between Model 1 (n=930,339) and Model 3 (n=705,151)?
Because the sample size changed; the jump in R2 reflects the composition of the restricted sample rather than superior functional form.
What is the strict rule for when Adjusted R2 can be used as a valid model selection tiebreaker?
It can only compare models that share both the exact same dependent variable and the exact same estimation sample.
In causal modeling, what is a confounder and how should it be handled?
A variable Z associated with X that directly affects Y; it should generally be included as a control to prevent omitted variable bias.
In causal modeling, what is a mediator and why can controlling for it be problematic?
A variable lying on the causal pathway between X and Y (X→Z→Y); controlling for it absorbs part of the total causal effect of X on Y.
In causal modeling, what is reverse causality and why should such variables be excluded?
When the outcome directly affects the control variable (Y→Z), making the control endogenous and biasing estimates.
Why was occupational score excluded as a control when estimating the earnings disadvantage of foreign-born men?
Because occupational attainment is a primary mediator through which immigrant status affects earnings; controlling for it eliminates the disadvantage you seek to measure.
Why was number of children excluded as a control variable in the men's earnings model?
Fertility responds to migration and economic status, and affects labor supply, making it a downstream endogenous variable.
In the earnings analysis, what was the foreign-born coefficient in Model 1 (without residential controls) vs. Model 2 (with residential controls)?
Model 1: +79.38USD (foreign-born advantage); Model 2: −89.91USD (foreign-born disadvantage).
Why did the foreign-born coefficient flip from positive (+79.38) to negative (−89.91) when residential controls were added?
Immigrants were disproportionately concentrated in higher-paying urban areas and regions (Northeast); controlling for geography removed this locational advantage.
In Model 4 (including citizenship), what were the predicted earnings differences for foreign-born citizens and non-citizens relative to native-born men?
Foreign-born citizens earned −50.26USD less, while foreign-born non-citizens earned −182.55USD less (−50.26+(−132.29)), a citizenship gap of 132.29USD.
Why is an additive regression model unable to evaluate whether the returns to education are greater in urban areas?
An additive model constrains the education coefficients to be identical across locations, allowing urban residence only to shift earnings by a constant parallel amount.
What is an interaction term in regression, and how is it constructed?
A regressor constructed by multiplying two independent variables together (X1×X2), allowing the effect of X1 on Y to depend on the level of X2.