AAEC 4302 ADVANCED STATISTICAL METHODS IN AGRICULTURAL RESEARCH Chapters 6.4-6.5, 7.4 Variables & Model Specifications 1 The Reciprocal Specification (6.4) • The reciprocal model specification is: 1 ˆ ˆ ˆ Yi B0 B1 Xi Reciprocal Model Specification 7.00 6.00 5.00 Y 4.00 3.00 2.00 1.00 0.00 1 2 3 4 5 6 7 8 9 10 X 2 The Reciprocal Specification 1 ˆ ˆ ˆ ˆ Yi B0 B1 B2 X 2i ... Bˆ k X ki i 1,...,n X 1i • Relationship between Y and the transformed independent variable XT 1 is linear ji X ji ˆ B ˆ XT B ˆ X ... B ˆ X e Yi B 0 1 1i 2 2i k ki i 3 The Reciprocal Specification • Therefore, the standard OLS method (i.e. formulas) can be used to fit this line 1 • Instead of X ji , XTji i 1,...,n is used as the X ji jth independent variable in the OLS formulas or in the data set given to the Excel program for calculating the OLS parameter estimates Geometry of the reciprocal relation is shown on page 116 4 The Reciprocal Specification • Model specified relation between inflation and unemployment as reciprocal, observations for 15 observations (1956-1970): 1 ei Yi B0 B1 X 1i • UINVi = 1/UMPLi • INFLi = B0 + B1*UINVi + ui • The estimated regression is: INFLi = -1.984 + 22.234*UINVi +ei R2= 0.549 SER=0.956 5 The Reciprocal Specification • B0 =-1.984 • As UNEML increases, INFL decreases and approaches the lower limit of -1.984 percent • Quantitative implications are understood when we compare diff. predicted values of INFL for diff. rates of unemployment • If UNEMPL = 3%, INFL = -1.984 +22.234*(1/3) = 5.43 % • If UNEMPL = 4%, INFL = -1.984 +22.234*(1/4) = 3.57 % 6 The Log-Linear Specification (6.5) • A special type of non-linear relations become linear when they are transformed with logarithms • Specifically, consider Yi e X1 X 2 ...X k B0 B1 B2 Bk • We take natural logs of both sides of this equation: ln Yi B0 B1 ln X1i B2 ln(X 2i ) ... Bk ln(X ki ) ui • This is also known as the Log-Log or Double-Log specification, because it becomes a linear relation when taking the natural logarithm of both sides 7 The Log-Linear Specification • The former implies that the usual OLS formulas (or the standard Excel program) can be used to estimate the coefficients of a Log-Linear model specification, but they are applied to lnYi and ( X 1i , ln X 2i ,...,ln X ki , instead of Yi and X 1i , ln (X 2i ,..., X ki (i.e. let Yi lnYi and (X 1i ln X 1i , X 2i ln X 2i ,..., X ki ln X ki for all i’s and then use the OLS formulas or the Excel program) 8 The Log-Linear Specification • A disadvantage of the log-linear specification is that one has to assume that all of the Y-Xj relations in the model conform to this type of non-linear specification (i.e. one needs to take the ln of Y and of all of the independent variables in the model) • In short, in the Log-Linear model specification: Figure 6.3 page 120 9 The Log-Linear Specification • An important feature is that Bˆ j directly measures the elasticity of Y with respect to Xj; i.e. the percentage change in Y when Xj changes by one percent • Notice in this model specification the slope (i.e. the unit change in Y when Xj changes by one unit) is not constant (it varies for different values of Xj), but the elasticity is constant throughout! 10 The Log-Linear Specification • B1 =1.2 • 5% change in X will cause Y change by: 1.2 * 0.05 = 0.06, or 6% 11 The Log-Linear Specification • Model of aggregate demand for money in the US • We construct variable - real quantity of money: Mi= NMi / (PrIndGNPi /100) • Ln Mi= Bo + B1 ln GNPi + Ui • Estimated regression: LnMi= 3.948 + 0.215 LnGNPi R2 = 0.78 SER=0.0305 12 The Log-Linear Specification • B1= 0.215, or 0< B1<1 the elasticity of M with respect to GNP is 0.215 • 5% increase in GNP leads to 0.215*5=1.075% increase in predicted M • Predict demand for money when GNP = 1000: ln1000=6.908 lnM = 3.948 + 0.215*6.908 = 5.433 Antilog of 5.433 = 222.8 bill $ 13 The Polynomial Specification (7.4) • A polynomial model specification (with respect to X 1 only) is: 2 ˆ ˆ ˆ ˆ ˆ X ... Bˆ X Yi B0 B11 X1i B12 X1i B 2 2i k ki Polynomial Specification 60 50 Y 40 30 20 10 0 1 2 3 4 5 6 X1 7 8 9 10 11 14 The Polynomial Specification • In a polynomial specification, as Xj increases, Y can increase or decrease at an increasing or at a decreasing rate: it is a very flexible nonlinear model specification • An advantage of the polynomial model specification is that it can combine situations in which some of the independent variables are non-linearly related to Y while others are linearly related to Y 15 The Polynomial Specification • A polynomial model can be estimated by OLS, 2 viewing X j as any other independent variable in the multiple regression 16 The Polynomial Specification Multiple regression : EARNSi B0 B11Ed1i B12 EXP1i B13 EXPSQ Cross-sectional DB with 100 observations Estimated EANRS function: EANRSi = -9.791 +0.995 EDi + 0.471EXPi – 0.00751EXPSQi R2=0.329 SER4.267 B 1= 0.995 – holding the level of experience constant one additional year of education increases earnings by $995 EANRSi = constant + 0.471EXPi – 0.00751EXPSQi where the “constant” depends of the particular value chosen for ED 17 The Polynomial Specification • Slope = 0.471 + (2)(-0.00751)EXP • If EXP = 5 years, then slope = 0.471 + (2)(-0.00751)(5) = 0.396thou $ • To find the level of experience corresponding with the pick of earnings 0 = 0.471 + (2)(-0.00751)EXP , EXP = 31.4 • To assess the effect on earnings by change in experience ∆EARNS = {0.471+2(-0.00751)EXP}∆EXP Worker with EXP=10, add 2 add. years of experience ∆EARNS = {0.471+2(-0.00751)10}2 = 0.6416 thousand $ • To assess the effect on earnings by change in experience ∆EARNS = 0.471∆EXP -0.00751∆EXPSQ ∆EARNS = (0.471)* (2 ) - 0.00751* 44 = 0.6116 thousand $ 18 2 2 12 – 10 = 44