OLS Problems Cheatsheet
Problem What it means Causes Effects Detection Solution
Heteroscedasticity Variance of the 1. Omission of important **Results are **Graphical method Robust the SE
error term is not variables unreliable as (run reg r.)
constant ie. obsv 2. Misspecification of coefficients do not ** Ho : var (e) = o^2
are not spread functional form meet the minimum (alt – not equal to) **weighted least
evenly along the 3. Outliers in the data variance (best) squares
line of best fit property Compare prob value to
significance level
**Biases the standard
errors & distorts t-stats If prob < sign, reject Ho
leading to wrong and conclude
decision
Autocorrelation When the error 1. Omission of ** Unreliable **Graphical method **First
terms in a important variables hypothesis testing differencing
regression model 2. Misspecification of ** Durbin Watson
are correlated functional form (exact values of DW) ** Cochrane
across 3. Manipulation of data ** Durbin’s alt (prob Orcutt Two Step
observations — 4. Data Transformation approach) Procedure
usually over time through differencing ** Durbins H statistic
(time series (solve for h, use stats
data). approach)
Endogeneity Explanatory 1. Error in ** Causes **Hausman Test Use a different
variable(s) measurement parameter(s) to be (requires you to run model: Indirect
correlated with 2. Omission of key bias and inconsistent reg. but results will be least squares or
the error term. variables provided) IV estimation
3. Dual causality
(education/wage) Ho : The regressor is **see IV section
exogeneous for steps
Compare prob value to
significance level
If prob < sign, reject Ho
and conclude i.e model
is endogeneous
Multicollinearity Independent 1. Constraints on the **Perfect: it becomes **High R^2 but few **Drop
variables in a model impossible to estimate significant coefficients problematic
model are 2. Model specification the coefficient of one variables
correlated: error of the explanatory **Pairwise correlation
perfectly or 3. Indiscriminate use of variables (values above .7 could **Increase
imperfectly variables indicate sample size
4. Overdetermined **Imperfect: multicollinearity)
model Large variance and SE **Transform the
5. Limited sampling Low t-stats (less **Auxiliary regression model
6. Constant significant) (r values > 0.7 could
explanatory variable Increases sensitivity to indicate **substitute
small changes in the multicollinearity) | if rj > coefficients from
data r original, j might be previous studies
the cause
**VIF > 10 indicates
high multicollinearity
**Tolerance level < o.1
indicates
multicollinearity
** Condition number >
15 indicates
multicollinearity
Note:
** Null Hypothesis is the absence of the problem. Rejecting means the problem exists