Econometrics Cheat Sheet Assumptions and properties Ordinary Least Squares

By Marcelo Moreno - King Juan Carlos University

Econometric model assumptions Objective - minimizePn the2 Sum of Squared Residuals (SSR):
The Econometrics Cheat Sheet Project
Under this assumptions, the OLS estimator will present min i=1 ûi , where ûi = yi − ŷi
good properties. Gauss-Markov assumptions:
Basic concepts 1. Parameters linearity (and weak dependence in time y
Simple regression model
Equation:
Definitions series). y must be a linear function of the β’s.
yi = β0 + β1 xi + ui
Econometrics - is a social science discipline with the 2. Random sampling. The sample from the population
Estimation:
objective of quantify the relationships between economic has been randomly taken. (Only when cross section)
ŷi = β̂0 + β̂1 xi
agents, test economic theories and evaluate and implement 3. No perfect collinearity.
β1 where:
government and business policies. ˆ There are no independent variables that are constant:
β̂0 = y − β̂1 x
Econometric model - is a simplified representation of the Var(xj ) ̸= 0, ∀j = 1, . . . , k. Cov(y,x)
ˆ There isn’t an exact linear relation between indepen- β0 β̂ 1 = Var(x)
reality to explain economic phenomena.
Ceteris paribus - if all the other relevant factors remain dent variables.
x
constant. 4. Conditional mean zero and correlation zero.
a. There aren’t systematic errors: E(u | x1 , . . . , xk ) = Multiple regression model
Data types E(u) = 0 → strong exogeneity (a implies b).
Cross section - data taken at a given moment in time, an y Equation:
b. There are no relevant variables left out of the model:
static photo. Order doesn’t matter. yi = β0 + β1 x1i + · · · + βk xki + ui
Cov(xj , u) = 0, ∀j = 1, . . . , k → weak exogeneity.
Time series - observation of variables across time. Order Estimation:
5. Homoscedasticity. The variability of the residuals is
does matter. ŷi = β̂0 + β̂1 x1i + · · · + β̂k xki
the same for all levels of x:
Panel data - consist of a time series for each observation where:
Var(u | x1 , . . . , xk ) = σu2
of a cross section. β̂0 = y − β̂1 x1 − · · · − β̂k xk
6. No auto-correlation. Residuals don’t contain infor- Cov(y,resid x )
Pooled cross sections - combines cross section from dif- β0 x 2 β̂j = Var(resid xj )j
mation about any other residuals:
ferent time periods. Corr(ut , us | x1 , . . . , xk ) = 0, ∀t ̸= s. Matrix: β̂ = (X T X)−1 (X T y)
7. Normality. Residuals are independent and identically x 1
Phases of an econometric model
1. Specification. 3. Validation. distributed: u ∼ N (0, σu2 )
8. Data size. The number of observations available must
Interpretation of coefficients
2. Estimation. 4. Utilization. Model Dependent Independent β1 interpretation
be greater than (k + 1) parameters to estimate. (It is
Regression analysis Level-level y x ∆y = β1 ∆x
already satisfied under asymptotic situations) Level-log y log(x) ∆y ≈ (β1 /100)(%∆x)
Study and predict the mean value of a variable (dependent Log-level log(y) x %∆y ≈ (100β1 )∆x
variable, y) regarding the base of fixed values of other vari- Asymptotic properties of OLS Log-log log(y) log(x) %∆y ≈ β1 (%∆x)
ables (independent variables, x’s). In econometrics it is Under the econometric model assumptions and the Central Quadratic y x + x2 ∆y = (β1 + 2β2 x)∆x
common to use Ordinary Least Squares (OLS) for regres- Limit Theorem (CLT):
sion analysis. ˆ Hold 1 to 4a: OLS is unbiased. E(β̂j ) = βj Error measurements Pn Pn
ˆ Hold 1 to 4: OLS is consistent. plim(β̂j ) = βj (to 4b Sum of Sq. Residuals: SSR = i=1 û2i = Pi=1 (yi − ŷi )2
Correlation analysis left out 4a, weak exogeneity, biased but consistent) Explained Sum of Squares:
n
SSE = Pi=1 (ŷi − y)2
Correlation analysis don’t distinguish between dependent n
ˆ Hold 1 to 5: asymptotic normality of OLS (then, 7 is Total Sum of Sq.: SST = SSE + SSR = i=1q (yi − y)2
and independent variables.
necessarily satisfied): u ∼ N (0, σu2 ) Standard Error of the Regression: σ̂u = n−k−1 SSR
ˆ Simple correlation measures the grade of linear associa- a
ˆ Hold 1 to 6: unbiased estimate of σu2 . E(σ̂u2 ) = σu2
p
tion between two variables.Pn Standard Error of the β̂’s: se(β̂) = σ̂u2P· (X T X)−1
r = Cov(x,y) √Pn i=1 ((xi −x)·(y i −y)) ˆ Hold 1 to 6: OLS is BLUE (Best Linear Unbiased Esti- n
i=1 (yi −ŷi )
2
σx ·σy = 2
Pn 2 Mean Squared Error: MSE =
i=1 (xi −x) · i=1 (yi −y) mator) or efficient. Pn n
ˆ Partial correlation measures the grade of linear associa- Absolute Mean Error: AME
|y −ŷ |
= i=1 n i i
ˆ Hold 1 to 7: hypothesis testing and confidence intervals Pn
tion between two variables controlling a third. Mean Percentage Error:
|û /y |
MPE = i=1n i i · 100
can be done reliably.

3.2-en - github.com/marcelomijas/econometrics-cheatsheet - CC-BY-4.0 license

R-squared Individual tests Dummy variables
Tests if a parameter is significantly different from a given
Is a measure of the goodness of the fit, how the regression value, ϑ. Dummy (or binary) variables are used for qualitative infor-
fits to the data: ˆ H0 : β j = ϑ mation like sex, civil state, country, etc.
SSE
R2 = SST = 1 − SSR
SST ˆ H1 : βj ̸= ϑ ˆ Takes the value 1 in a given category and 0 in the rest.
ˆ Measures the percentage of variation of y that is lin- β̂ −ϑ ˆ Are used to analyze and modeling structural changes
early explained by the variations of x’s. Under H0 : t = se(j β̂ ) ∼ tn−k−1,α/2
j in the model parameters.
ˆ Takes values between 0 (no linear explanation of the If |t| > |tn−k−1,α/2 |, there is evidence to reject H0 . If a qualitative variable have m categories, we only have to
variations of y) and 1 (total explanation of the varia- Individual significance test - tests if a parameter is sig- include (m − 1) dummy variables.
tions of y). nificantly different from zero.
When the number of regressors increment, the value of the ˆ H0 : βj = 0 Structural change
ˆ H1 : βj ̸= 0 Structural change refers to changes in the values of the pa-
R-squared increments as well, whatever the new variables
β̂j rameters of the econometric model produced by the effect
are relevant or not. To solve this problem, there is an ad- Under H0 : t = ∼ tn−k−1,α/2
se(β̂j ) of different sub-populations. Structural change can be in-
justed R-squared by degrees of freedom (or corrected R-
If |t| > |tn−k−1,α/2 |, there is evidence to reject H0 . cluded in the model through dummy variables.
squared):
2 The location of the dummy variables (D) matters:
n−1
R = 1 − n−k−1 · SSR n−1
SST = 1 − n−k−1 · (1 − R )
2 The F test
ˆ On the intercept (additive effect) - represents the mean
2 Simultaneously tests multiple (linear) hypothesis about the
For big sample sizes: R ≈ R2 difference between the values produced by the structural
parameters. It makes use of a non restricted model and a
change.
restricted model:
Hypothesis testing ˆ Non restricted model - is the model on which we want
y = β0 + δ 1 D + β 1 x 1 + u
ˆ On the slope (multiplicative effect) - represents the ef-
to test the hypothesis.
Definitions fect (slope) difference between the values produced by
ˆ Restricted model - is the model on which the hypoth-
An hypothesis test is a rule designed to explain from a sam- the structural change.
esis that we want to test have been imposed.
ple, if exist evidence or not to reject an hypothesis y = β0 + β1 x 1 + δ 1 D · x 1 + u
Then, looking at the errors, there are:
that is made about one or more population parameters. Chow’s structural test - is used when we want to analyze
ˆ SSRUR - is the SSR of the non restricted model.
Elements of an hypothesis test: the existence of structural changes in all the model param-
ˆ SSRR - is the SSR of the restricted model.
ˆ Null hypothesis (H0 ) - is the hypothesis to be tested. SSRR −SSRUR n−k−1 eters, it’s a particular expression of the F test, where the
Under H0 : F = · q ∼ Fq,n−k−1
ˆ Alternative hypothesis (H1 ) - is the hypothesis that SSRUR
null hypothesis is: H0 : No structural change (all δ = 0).
where k is the number of parameters of the non restricted
cannot be rejected when the null hypothesis is rejected.
model and q is the number of linear hypothesis tested.
ˆ Test statistic - is a random variable whose probability Changes of scale
If Fq,n−k−1 < F , there is evidence to reject H0 .
distribution is known under the null hypothesis.
Global significance test - tests if all the parameters as- Changes in the measurement units of the variables:
ˆ Critic value - is the value against which the test statistic
sociated to x’s are simultaneously equal to zero. ˆ In the endogenous variable, y ∗ = y ·λ - affects all model
is compared to determine if the null hypothesis is rejected
ˆ H0 : β1 = β2 = · · · = βk = 0 parameters, βj∗ = βj · λ, ∀j = 1, . . . , k
or not. Is the value that makes the frontier between the
ˆ H1 : β1 ̸= 0 and/or β2 ̸= 0 . . . and/or βk ̸= 0 ˆ In an exogenous variable, x∗j = xj · λ - only affect the
regions of acceptance and rejection of the null hypothesis.
In this case, we can simplify the formula for the F statistic. parameter linked to said exogenous variable, βj∗ = βj · λ
ˆ Significance level (α) - is the probability of rejecting R2 n−k−1
Under H0 : F = 1−R 2 · ∼ F k,n−k−1 ˆ Same scale change on endogenous and exogenous - only
the null hypothesis being true (Type I Error). Is chosen k
by who conduct the test. Commonly is 0.10, 0.05 or 0.01. If Fk,n−k−1 < F , there is evidence to reject H0 . affects the intercept, β0∗ = β0 · λ
ˆ p-value - is the highest level of significance by which the
null hypothesis cannot be rejected (H0 ). Confidence intervals Changes of origin
The rule is: if the p-value is less than α, there is evidence The confidence intervals at (1 − α) confidence level can be
Changes in the measurement origin of the variables (en-
to reject the null hypothesis at that given α (there is calculated:
dogenous or exogenous), y ∗ = y + λ - only affects the
evidence to accept the alternative hypothesis). β̂j ∓ tn−k−1,α/2 · se(β̂j ) model’s intercept, β0∗ = β0 + λ

3.2-en - github.com/marcelomijas/econometrics-cheatsheet - CC-BY-4.0 license

Multicollinearity Heteroscedasticity Auto-correlation
ˆ Perfect multicollinearity - there are independent vari- The residuals ui of the population regression function do The residual of any observation, ut , is correlated with the
ables that are constant and/or there is an exact linear not have the same variance σu2 : residual of any other observation. The observations are not
relation between independent variables. Is the breaking Var(u | x1 , . . . , xk ) = Var(u) ̸= σu2 independent.
of the third (3) econometric model assumption. Is the breaking of the fifth (5) econometric model as- Corr(ut , us | x1 , . . . , xk ) = Corr(ut , us ) ̸= 0 ∀t ̸= s
ˆ Approximate multicollinearity - there are indepen- sumption. The “natural” context of this phenomena is time series. Is
dent variables that are approximately constant and/or the breaking of the sixth (6) econometric model as-
there is an approximately linear relation between inde-
Consequences sumption.
ˆ OLS estimators still are unbiased.
pendent variables. It does not break any economet-
ˆ OLS estimators still are consistent. Consequences
ric model assumption, but has an effect on OLS.
ˆ OLS is not efficient anymore, but still a LUE (Linear ˆ OLS estimators still are unbiased.
Consequences Unbiased Estimator). ˆ OLS estimators still are consistent.
ˆ Perfect multicollinearity - the equation system of ˆ Variance estimations of the estimators are biased: ˆ OLS is not efficient anymore, but still a LUE (Linear
OLS cannot be solved due to infinite solutions. the construction of confidence intervals and the hypoth- Unbiased Estimator).
ˆ Approximate multicollinearity esis testing is not reliable. ˆ Variance estimations of the estimators are biased:
– Small sample variations can induce to big variations in the construction of confidence intervals and the hypoth-
the OLS estimations.
Detection esis testing is not reliable.
ˆ Graphs - look u y
– The variance of the OLS estimators of the x’s that are
collinear, increments, thus the inference of the param-
for scatter pat- Detection
terns on x vs. u ˆ Graphs - look for scatter patterns on ut−1 vs. ut or
eter is affected. The estimation of the parameter is x
or x vs. y plots. make use of a correlogram.
very imprecise (big confidence interval).
Ac. Ac.(+) Ac.(-)
x ut ut ut
Detection
ˆ Correlation analysis - look for high correlations be- ˆ Formal tests - White, Bartlett, Breusch-Pagan, etc.
tween independent variables, > |0.7|. Commonly, the null hypothesis: H0 : Homoscedasticity. ut−1 ut−1
ˆ Variance Inflation Factor (VIF) - indicates the in- ut−1
Correction
crement of Var(β̂j ) because of the multicollinearity. ˆ Use OLS with a variance-covariance matrix estimator ro-
1
VIF(β̂j ) = 1−R 2 bust to heteroscedasticity (HC), for example, the one pro-
2
j
ˆ Formal tests - Durbin-Watson, Breusch-Godfrey, etc.
where Rj denotes the R-squared from a regression be- posed by White.
Commonly, the null hypothesis: H0 : No auto-correlation.
tween xj and all the other x’s. ˆ If the variance structure is known, make use of Weighted
– Values between 4 to 10 suggest that it is advisable to Least Squares (WLS) or Generalized Least Squares Correction
analyze in more depth if there might be multicollinear- (GLS): ˆ Use OLS with a variance-covariance matrix estimator ro-
ity problems. – Supposing that Var(u) = σu2 ·xi , divide the model vari- bust to heterocedasticity and auto-correlation (HAC), for
– Values bigger than 10 indicates that there are multi- ables by the square root of xi and apply OLS. example, the one proposed by Newey-West.
collinearity problems. – Supposing that Var(u) = σu2 · x2i , divide the model ˆ Use Generalized Least Squares. Supposing yt = β0 +
One typical characteristic of multicollinearity is that the variables by xi (the square root of x2i ) and apply OLS. β1 xt + ut , with ut = ρut−1 + εt , where |ρ| < 1 and εt is
regression coefficients of the model aren’t individually dif- ˆ If the variance structure is not known, make use of Fea- white noise.
ferent from zero (due to high variances), but jointly they sible Weighted Least Squared (FWLS), that estimates a – If ρ is known, create a quasi-differentiated model where
are different from zero. possible variance, divides the model variables by it and ut is white noise and estimate it by OLS.
then apply OLS. – If ρ is not known, estimate it by -for example- the
Correction ˆ Make a new model specification, for example, logarithmic Cochrane-Orcutt method, create a quasi-differentiated
ˆ Delete one of the collinear variables. transformation (lower variance). model where ut is white noise and estimate it by OLS.
ˆ Perform factorial analysis (or any other dimension reduc-
tion technique) on the collinear variables.
ˆ Interpret coefficients with multicollinearity jointly.
3.2-en - github.com/marcelomijas/econometrics-cheatsheet - CC-BY-4.0 license