interval for E(Y X ) : 1 1. Prediction Equation VIF(X j ) 1 R 2j 1 ( X i X )2 yˆ i ˆ0 ˆ1 xi Yi t / 2 S e R 2j is the coefficient of n SS X 2. Sample Slope determination for the Here Y is the E(Y X). ˆ1 SS xy x x y y i i regressionof X j as x x 2 10. Prediction interval SS xx i dependent variable 2 SSx= ∑ x - (∑ x) /n 2 for a randomly chosen value of Y If VIF >10, it is co SSxy= ∑ xy- ∑ x*∑ y/n given x linearity
3. Sample Y Intercept A (1 - )100% prediction 14. Tolerance Factor: 1-Rj2
interval for Y is : 14. Beta Weights ˆ0 y ˆ1 x Sx 4. Coeff. Of Determination 1 (Xi X ) 2 Beta i Yi t / s S e 1 Sy n SS x R2 SSR 1 SSE Sx Std dev of X SST SST where Xs are observed values Sy Std dev of Y 5. Std. Error of Estimate of independent variable . 15. Partial F Test 2 Y is the estimate of Y, n is ( SSER SSEF ) / r Y Y i the sample size and Se is the Fr ,n ( k 1) MSEF Se n k 1 standard error of Y SSER - sum of squares of Se Yi 2 0 Yi 1 X i Yi error of reduced model 11. Coeff. of Correlation n k 1 SSEF - sum of squares of 6. Standard Error of 0 and SS XY error of full model r ( signb1) R 2 1 SS XX SSYY r – no. of variables dropped from full model. S (0 ) Se x 2
Forward Regression Or,
nSSxx Fr,n-k-1 = (R2R+1-R2R)/[(1- Fdefaultin > 3.84 Se R2R+1)/n-k-1] Pin < 0.05 S ( 1 ) Where, R2R+1 = R2 + r2 SSxx Backward Regression Fout < 2.71 Pout > 0.10 16. Outliers 7. Test Statistic Estimate Parameter 12. Adjusted R2 Measure Potential Outliers t( n 2) Est. std. error of estimate SSE /(n k 1) 1 1 RA2 1 Standardized > 3 (3 sigma level) SST /(n 1) S8.e ( 1Confidence ) Interval residual, n 1 Studentized of 0SS and 1 RA2 1 (1 R 2 ) residual t( n 2) 1 x n (k 1) Se R 2A The adjusted coefficient 1 t( / 2,n2) Se (1 ) of determination 0 t ( / 2 , n 2 ) S e ( 0 ) R 2 Unadjuste d coefficient 9. Confidence interval of determination for mean value of Y given x n number of observations k no. of explanatory variable s Mahalanobis > Critical chi-square value (i ) Lt Yt PARTIAL CORRELATION distance with df = number of Correlation between y1 and x2, explanatory variables (1 ) ( Lt 1 Tt 1 ) when the influence of x3 is (Outliers in independent (ii) Tt ( Lt Lt 1 ) removed from both y1 and x2. variable) (1 ) Tt 1 Cook’s distance > 1 implies potential outlier r12 r13r23 r12,3 (iii) Ft 1 Lt Tt (1 r132 )(1 r232 ) Leverage > 2(k+1)/n, then the point values is influential (k is the (iv) F t m Lt mTt r212 = (R2 – R21/2)/(1-R21/2) number of independent variables and n is the 3. Theil’s Coeff Semi-partial (or part) correlation sample size) sr12,3 is the correlation between y1 n
Y F and x2 when influence of x3 is
SDFBeta > 2/n 2 t t t 1 partialled out of x2 (not on y1). SDFFit U1 , 2 (k 1) n n n r12 r13r23 Y t 2 F t 2 sr12,3 17. Mahalanobis t 1 t 1 1 r232 2 Distance Ft 1 Yt 1 n -1
Mi = [(Xi – X)/ Sx]2
t 1 Or, S2rk = R2-R2Y/K U2 Yt Outlier if value >10 2 Square of part correlation = n -1 Yt 1 Yt 18. Cook’s Distance t 1 Yt unique contribution of the explanatory variable to R2 Di = ∑j (Yj – Yj(i))2/k x MSE (coefficient of multiple 23.P.I. for multiple 19. Durbin Watson Test *determination). determination). regression (Note: the second Sr12 = R2Y12 - r2Y2 Durbin Watson value close to one is actually C.I.): r2Y2 – zero order correlation 2 implies no auto-correlation Sr12 = semi partial correlation U1 is bounded between 0 and 1, with values closer to Upward and downward BIAS: Durbin Watson value close to 0 implies positive auto- zero indicating greater correlation accuracy.
Durbin Watson value close to If U2 = 1, there is no
4 implies negative auto- difference between naïve correlation forecast and the forecasting technique 20. Relationship between F and R2 If U2 < 1, the technique is UP – Bias: Coefficient of X better than naïve forecast increases in the absence of OV F = (R2/1- R2) x DOWN – Bias: Coefficient of X ((n-(k+1))/k) If U2 > 1, the technique is no decreases in the absence of OV better than the naïve 22. Standarised x =( xi- forecast. xbar)/σx A (1 - ) 100% prediction interval for a value of Y given values of X : i FORECASTING yˆ t s ( yˆ ) MSE 2 1. Exponential ( ,(n(k 1))) 2 Smoothing Lt * Yt 1 (1 ) * Lt 1 A (1 - α) 100% prediction interval for the conditional mean of Y given Ft Lt values of X : i 2. Double Exponential yˆ t s[Eˆ (Y )] Smoothing ( ,(n(k 1))) 2