ECONOMETRICS: RELAXATION OF CLRM ASSUMPTIONS
ASSUMPTIONS
HETEROSCEDASTICITY
MULTICOLLINEARITY
AUTOCORRELATION
NATURE
NON-CONSTANT VARIANCE
LINEAR RELATIONSHIP AMONG EXPLANTORY VARIABLES
CORRELATION AMONG ERROR TERMS
CAUSES
Following error learning models
Discreationary income
Improvement in data collection techniques
Outliers
Model specification
Skewness in distribution of regressors
Data collection method employed
Constrains on the model
Model specification
Overdetermined model (K>n)
Inertia
Model specification
Cobweb phenomena
Lags
Transformation of Data
Manipulation of Data
Estimators are still Unbiased
Large variance hence inefficient
Relatively large confidence interval
Large variances and covariances, making precise estimation difficult
Large confidence interval leading to Type II error
Statistically insignificant T-ratios of one or more coefficients
High R-squared althogh t-ratios may be statistically insignificant
OLS estimators may be sensitive to small data changes
The residual variance is likely to be underestimated
An overestimated R-squared
Var(B) could be underestimated
Usual test of significance may not be valid
Graphical Method
Park Test
Glejser Test
Spearman’s Rank Correlation Test
Goldfeld–Quandt Test
BPG Test
White's General Test
Graphical Method
High Pairwise correlation
High R-squared but insignificant T-ratios
Examination of partial correlation
Run Auxiliary regression and obtain R-squared J to test significance
Tolerance and VIF
Scatter plots
Graphical Method
Runs Test
Durbin–Watson d Test
The Breusch–Godfrey Test
Apply GLS if σ2i is know
Apply Robust Standard errors if σ2i is unknown
Do nothing
Using known information
Combining cross-sectional and time series data (pooling data)
Dropping a variable(s) and specification bias (if MC is severe)
Multiply ρ by the lagged model and subtract the product from actual regression and apply OLS
First difference approach *
ρ Based on Durbin–Watson d Statistic (Preferably for large samples
ρ Estimated from the Residual
CONSEQUNCES
DETECTION
RESOLVING