MODEL SPECIFICATION SPECIFICATION ERRORS D N Gujurati Basic
MODEL SPECIFICATION & SPECIFICATION ERRORS D. N. Gujurati , Basic Econometrics, Mc. Graw-Hill, 1988, 2 e, Chapter 13. What Constitutes a “Correct” Model? & What May Go Wrong? Dr. C. Ertuna
MODEL SPECIFICATION Quality of a model depends on following 5 criteria: a) Parsimony: (Fewer is Better) b) Identifiability : (One Solution) c) Goodness-of-Fit: (Best Fit) d) Theoretical Consistency: (Plus or minus) e) Post Sample Predictive Power: (Good even after) Dr. C. Ertuna
PARSIMONY (Fewer is Better) A good model should incorporate few key variables that capture the essence of the phenomenon under investigation and left all minor influences to the error term. Dr. C. Ertuna
IDENTIFIABILITY (One Solution) There should be only one estimate for a given parameter. Dr. C. Ertuna
GOODNESS-OF-FIT (Best Fit) • Dr. C. Ertuna
THEORETICAL CONSISTENCY (Plus or Minus) Estimated coefficients should have right signs. Dr. C. Ertuna
POST SAMPLE PREDICTIVE POWER (Good Even After) Predictive power outside the sample period should also be good. Dr. C. Ertuna
SPECIFICATION ERRORS There are 4 major specification errors: a) Omitted Variable Error b) Irrelevant Variable Error c) Functional Form Error d) Measurement Error Dr. C. Ertuna
OMITTED VARIABLE ERROR Consequences of Omitted Variable Error (or Omission Error) are: a) The disturbance variance will be incorrect. b) Standard Error of the parameters will be biased. So, confidence intervals and hypothesis testing will be misleading. Hence, significance level of estimated parameters will be misleading. c) If omitted variable correlates with included variable(s), then intercept and partial regression slopes will be biased as well as inconsistent. d) If omitted variable does not correlate with included variable(s), then intercept will be biased but partial regression slopes will be unbiased. Dr. C. Ertuna
IRRELEVANT VARIABLE ERROR Consequences of inclusion of Irrelevant Variable are: a) Estimated parameters will be inefficient. That means; Standard Error of parameters are larger than the true model, hence, parameters are less precise. b) However: i. All estimators will be unbiased and consistent. ii. Confidence interval and hypothesis testing will be valid. iii. Disturbance variance will be correctly estimated Dr. C. Ertuna
- Slides: 10