Using Excel For Eco No Metric. Book

Using Excel
For Principles of Econometrics, Third Edition
Version 1.0
i
Using Excel
For Principles of Econometrics, Third Edition
ASLI K. OGUNC
Texas A&M University-Commerce
R. CARTER HILL
Louisiana State University
JOHN WILEY & SONS, INC

New York / Chichester / Weinheim / Brisbane / Singapore / Toronto
ii
Asli Ogunc dedicates this work to Patara
Carter Hill dedicates this work to Todd and Peter
ACQUISITIONS EDITOR Xxxxxx Xxxxxxxx

MARKETING MANAGER Xxxxxx Xxxxxxxx
PRODUCTION EDITOR Xxxxxx Xxxxxxxx
PHOTO EDITOR Xxxxxx Xxxxxxxx
ILLUSTRATION COORDINATOR Xxxxxx Xxxxxxxx
This book was set in Times New Roman and printed and bound by.XXXXXX XXXXXXXXXX. The cover
was printed by.XXXXXXXX XXXXXXXXXXXX
This book is printed on acid-free paper. ∞

The paper in this book was manufactured by a mill whose forest management programs include sustained
yield harvesting of its timberlands. Sustained yield harvesting principles ensure that the numbers of trees cut
each year does not exceed the amount of new growth.
Copyright © John Wiley & Sons, Inc. All rights reserved.

No part of this publication may be reproduced, stored in a retrieval system or transmitted in any form or by
any means, electronic, mechanical, photocopying, recording, scanning or otherwise, except as permitted
under Sections 107 or 108 of the 1976 United States Copyright Act, without either the prior written
permission of the Publisher, or authorization through payment of the appropriate per-copy fee to the
Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923, (508) 750-8400, fax (508) 750-
4470. Requests to the Publisher for permission should be addressed to the Permissions Department, John
Wiley & Sons, Inc., 605 Third Avenue, New York, NY 10158-0012, (212) 850-6011, fax (212) 850-6008, E-
Mail: [email protected].
ISBN 0-471-xxxxx-x
Printed in the United States of America
10 9 8 7 6 5 4 3 2 1
iii
PREFACE
This book is a supplement to Principles of Econometrics, 3rd Edition by R. Carter Hill, William
E. Griffiths and Guay C. Lim (Wiley, 2008), hereinafter POE. This book is not a substitute for
the textbook, nor is it a stand alone computer manual. It is a companion to the textbook, showing
how to perform the examples in the textbook using Excel 2003. This book will be useful to
students taking econometrics, as well as their instructors, and others who wish to use Excel for
econometric analysis.
In addition to this computer manual for Excel, there are similar manuals and support for the
software packages EViews, Excel, Gretl, Shazam and Stata. In addition, all the data for POE in
various formats, including Excel, are available at http://www.wiley.com/college/hill.
Individual Excel data files, errata for this manual and the textbook can be found at
http://www.bus.lsu.edu/hill/poe. Templates for routine tasks can also be found at this web site.
The chapters in this book parallel the chapters in POE. Thus, if you seek help for the examples in
Chapter 11 of the textbook, check Chapter 11 in this book. However within a Chapter the sections
numbers in POE do not necessarily correspond to the Excel manual sections.
We welcome comments on this book, and suggestions for improvement. *
Asli K. Ogunc
Department of Accounting, Economics and Finance
Texas A&M University-Commerce
Commerce, TX 75429
[email protected]
R. Carter Hill
Economics Department
Louisiana State University
Baton Rouge, LA 70803
[email protected]
.
*
Microsoft product screen shot(s) reprinted with permission from Microsoft Corporation. Our use does not directly or indirectly imply
Microsoft sponsorship, affiliation, or endorsement.
iv
BRIEF CONTENTS
1. Introducing Excel 1
2. Simple Linear Regression 21
3. Interval Estimation and Hypothesis Testing 39
4. Prediction, Goodness of Fit and Modeling Issues 49
5. Multiple Linear Regression 68
6. Further Inference in the Multiple Regression Model 80
7. Nonlinear Relationships 104
8. Heteroskedasticity 125
9. Dynamic Models, Autocorrelation, and Forecasting 142
10. Random Regressors and Moment Based Estimation 155
11. Simultaneous Equations Models 169
12. Nonstationary Time Series Data and Cointegration 178
13. An Introduction to Macroeconometrics: VEC and VAR Models 188
14. An Introduction to Financial Econometrics: Time-Varying Volatility
and ARCH Models 200
15. Panel Data Models 204
16. Qualitative and Limited Dependent Variable Models 215
17. Importing Internet Data 219
Appendix B. Review of Probability Concepts 223
v
4.2 Measuring Goodness-of-Fit 53
4.2.1 Calculating R2 54
CONTENTS
4.2.2 Covariance and correlation
analysis 54
CHAPTER 1 Introduction to Excel 1 4.3 Residual Diagnostics 57
1.1 Starting Excel 1 4.3.1 The Jarque-Bera test for
1.2 Entering Data 4 normality 59
1.3 Using Excel for Calculations 6 4.4 Modeling Issues 60
1.3.1 Arithmetic operations 8 4.4.1 Scaling the data 60
1.3.2 Mathematical functions 9 4.4.2 The log-linear model 61
1.4 Excel Files for Principles of 4.4.3 The linear-log model 62
Econometrics 15 4.4.4 The log-log model 62
1.4.1 John Wiley & Sons 4.5 More Examples 63
website 15 4.5.1 Residual analysis with wheat
1.4.2 Principles of Econometrics data 63
website 16 4.5.2 Log-linear model with wage
1.4.3 Definition files 16 data 65
1.4.4 The food expenditure data 16 4.5.3 Generalized R2 67
CHAPTER 2 The Simple Linear Regression CHAPTER 5 Multiple Linear Regression 68

Model 21 5.1 Big Andy’s Burger Barn 68
2.1 Plotting the Food Expenditure Data 21 5.2 Prediction 70
2.2 Estimating a Simple Regression 28 5.3 Sampling Precision 71
2.3 Plotting a Simple Regression 31 5.4 Confidence Intervals 76
2.4 Plotting the Least Squares 5.5 Hypothesis Testing 77
Residuals 35 5.6 Goodness-of-Fit 78
2.5 Prediction Using Excel 36
CHAPTER 6 Further Inference in the

CHAPTER 3 Interval Estimation and Multiple Regression Model 80
Hypothesis Testing 39 6.1 The F-test 80
3.1 Interval Estimation 39 6.2 Testing the Overall Significance of the
3.1.1 Automatic interval Model 84
estimates 39 6.3 An Extended Model 85
3.1.2 Constructing interval 6.4 Testing Some Economic
estimates 41 Hypotheses 86
3.2 Hypothesis Testing 43 6.4.1 The Significance of
3.2.1 Right-Tail tests 43 advertising 86
3.2.2 Left-Tail tests 45 6.4.2 Optimal level of
3.2.3 Two-Tail tests 46 advertising 87
6.5 Nonsample Information 90
CHAPTER 4 Prediction, Goodness-of-Fit 6.6 Model Specification 93
and Modeling Issues 49 6.6.1 Omitted variables 93
4.1 Prediction for the Food Expenditure 6.6.2 Irrelevant variables 94
Model 49 6.6.3 Choosing the model 94
4.1.1 Calculating the standard error 6.7 Poor Data, Collinearity and
of the forecast 49 Insignificance 99
4.1.2 Prediction interval 52
vi
CHAPTER 7 Nonlinear Relationships 104 9.3 Estimating an AR(1) Model 145
7.1 Nonlinear Relationships 104 9.3.1 Least squares 145
7.1.1 Summarize data and estimate 9.4 Detecting Autocorrelation 148
regression 104 9.4.1 The Durbin-Watson
7.1.2 Calculating a marginal test 148
effect 106 9.4.2 An LM test 149
7.2 Dummy Variables 107 9.5 Autoregressive Models 150
7.2.1 Creating dummy 9.6 Finite Distributed Lags 151
variables 107 9.7 Autoregressive Distributed Lag (ARDL)
7.2.2 Estimating a dummy variable Model 153
regression 107
7.2.3 Testing the significance of the
CHAPTER 10 Random Regressors and
dummy variables 109
Moment Based Estimation 155
7.2.4 Further calculations 109
10.1 Least Squares with Simulated
7.3 Applying Dummy Variables 110
Data 155
7.3.1 Interactions between
10.2 Instrumental Variables Estimation with
qualitative factors 110
Simulated Data 157
7.3.2 Adding regional dummy
10.2.1 Correction of IV standard
variables 114
errors 159
7.3.3 Testing the equivalence of two
10.2.2 Corrected standard errors for
regressions 116
simulated data 160
7.4 Interactions Between Continuous
10.3 The Hausman Test: Simulated
Variables 119
Data 162
7.5 Dummy Variables in Log-linear
10.4 Testing for Weak Instruments: Simulated
Models 121
Data 163
10.5 Testing for Validity of Surplus
CHAPTER 8 Heteroskedasticity 125 Instruments: Simulated Data 164
8.1 The Nature of Heteroskedasticity 125 10.6 Estimation using Mroz Data 164
8.2 Using the Least Squares Estimator 126 10.6.1 Least squares regression 164
8.3 The Generalized Least Squares 10.6.2 Two-stage least squares 165
Estimator 128 10.7 Testing the Endogeneity of
8.3.1 Transforming the model 128 Education 167
8.3.2 Estimating the variance 10.8 Testing for Weak Instruments 167
function 130 10.9 Testing the Validity of Surplus
8.3.3 A heteroskedastic Instruments 168
partition 131
8.4 Detecting Heteroskedasticity 134
CHAPTER 11 Simultaneous Equations
8.4.1 Residual plots 134 Models 169
8.4.2 The Goldfeld-Quandt 11.1 Truffle Supply and Demand 169
test 135 11.2 Estimating the Reduced Form
8.4.3 Testing the variance Equations 170
function 139 11.3 2SLS Estimates of Truffle Demand and
Supply 171
CHAPTER 9 Dynamic Models, 11.3.1 Correction of 2SLS standard
Autocorrelation, and Forecasting 142 errors 173
9.1 Lags in the Error Term 142 11.3.2 Corrected standard errors in
9.2 Area Response for Sugar 143 truffle demand
vii
and supply 174 CHAPTER 16 Qualitative and Limited
11.4 Supply and Demand of Fish 175 Dependent Variable Models 215
11.5 Reduced Forms for Fish Price 16.1 Models with Binary Dependent
and Quantity 176 Variables 215
16.1.1 The linear probability
model 216
CHAPTER 12 Nonstationary Time-Series
16.1.2 Least squares estimation of the
Data and Cointegration 178
linear probability model 216
12.1 Stationary and Nonstationary
Data 178
12.2 Spurious Regression 179 CHAPTER 17 Importing Internet Data 219
12.3 Unit Root Test for
Stationarity 181 Appendix B Review of Probability
12.4 Integration and Concepts 223
Cointegration 185 B.1 Binomial Probabilities 223
12.5 Engle-Granger Test 185 B.1.1 Computing binomial
probabilities directly 223
CHAPTER 13 An Introduction to B.1.2 Computing binomial
Macroeconometrics: VEC and VAR probabilities using
Models 188 BINOMDIST 225
13.1 VEC and VAR Models 188 B.2 The Normal Distribution 227
13.2 Estimating a VEC Model 188
13.3 Estimating VAR 197
CHAPTER 14 An Introduction to Financial

Econometrics: Time-Varying Volatility and
ARCH Models 200
14.1 ARCH Model and Time Varying
Volatililty 200
14.2 Testing, Estimating and
Forecasting 202
CHAPTER 15 Panel Data Models 205

15.1 Sets of Regression Equations 205
15.2 Seemingly Unrelated Regressions 209
15.2.1 Breusch-Pagan test of
independence 209
15.3 The Fixed Effects Model 210
15.3.1 A dummy variable model 211
15.4 Random Effects Estimation 214
viii
CHAPTER 1
Introduction to Excel
CHAPTER OUTLINE
1.1 Starting Excel 1.4 Excel Files for Principles of Econometrics
1.2 Entering Data 1.4.1 John Wiley & Sons website
1.3 Using Excel for Calculations 1.4.2 Principles of Econometrics website
1.3.1 Arithmetic operations 1.4.3 Definition files
1.3.2 Mathematical functions 1.4.4 The food expenditure data
1.1 STARTING EXCEL
Start Excel by clicking the Start menu and locating the program, or by clicking a shortcut, such
as,
Excel opens. Click on the New Workbook icon.
The worksheet looks like this
1
2 Chapter 1
There are lots of little bits that you will become more familiar with as we go along. The active
cell is surrounded by a border and is in Column A and Row 1. We will refer to cells as A1, B1
and so on.
Across the top of the window is a Menu bar. Sliding the mouse over the items opens up a
pull down menu, showing further options.
Perhaps the most important of all these is Help.

Introduction to Excel 3
Microsoft Office has a cute (or annoying) feature. You can have a little assistant showing on the
screen. If you click your assistant you can type in a question or search.
If you right-click on the critter you can choose to hide him or change is personality. We will hide
him.
The Standard Toolbar has the usual Microsoft functions New, Open, Save, Print, Print
preview, Copy and Paste. The AutoSum key is a feature of Excel¸ and the Sort buttons allow
you to order data according to the magnitude of one of your columns.
The Formatting Toolbar has the usual functions. The use of Grid lines can clarify a worksheet,
as can the use of colored fonts and filling in cells for emphasis.
4 Chapter 1
1.2 ENTERING DATA
We will use Excel to analyze data. To enter data into an Excel worksheet move the cursor to a
cell and type. First enter X in cell A1 and Y in cell B1. Navigate by moving the cursor with the
mouse, or use the Tab key to move right or left, or Arrow keys to move right, left, up or down.
In A2, enter the number 1. Press Enter and it will take you to the next cell, fill in the rest as
shown.
To nicely center the data in the cells, highlight cells A1:A6. There are several ways to highlight
the cells. For small areas the easiest way is to place cursor in A1, hold down the left mouse button
and drag it across the area you wish to highlight. For larger areas, using a key-stroke combination
is very convenient.
• To highlight a column—place cursor in A1. Hold down Ctrl-key and Shift-key

simultaneously, which we will denote as Ctrl-Shift. Press the down arrow ↓ on the
keyboard.
• To highlight a row—place cursor in A1. Press Ctrl-Shift and right arrow → on the
keyboard.
• To highlight a region—place cursor in A1. Press Ctrl-Shift then the down arrow and
then the right arrow.
After selecting A1:A6, click Center
The result is
Repeat this process for B1:B6. This centering is just for appearance, and has no affect on any
functionality.
To highlight the entire worksheet, click the square in the upper left corner.
6 Chapter 1
1.3 USING EXCEL FOR CALCULATIONS
What is Excel good for? Its primary usefulness is to carry out repeated calculations. We can add,
subtract, multiply and divide; and we can apply mathematical and statistical functions to the data
in our worksheet. To illustrate, highlight Columns A & B, down to row 7, as shown below. Click
AutoSum. This will sum the rows and place the sum in the final row.
The result its
Carry out a similar sequence of steps to sum the columns. Highlight rows 2:6 and columns A:C.
Click AutoSum.
The column sum (X + Y) is now in column C.

To Insert Columns hold the cursor over the A in column A. The cursor turns into a down arrow
and the entire first column is highlighted.
On the Excel Menu select Insert/Columns. This will insert a new column to the left of the
highlighted column.
Enter a column head, which will serve to identify what is in the first row, and enter “Sum” to
identify the contents of row 7.
8 Chapter 1
Add a header in column D.
1.3.1 Arithmetic operations

Standard arithmetical functions are defined as follows (from Excel Help, type arithmetic
operators)
Arithmetic operators: To perform basic mathematical operations such as

addition, subtraction, or multiplication; combine numbers; and produce numeric
results, use the following arithmetic operators.
Arithmetic operator Meaning (Example)

+ (plus sign) Addition (3+3)
– (minus sign) Subtraction (3–1)Negation (–1)
* (asterisk) Multiplication (3*3)
/ (forward slash) Division (3/3)
% (percent sign) Percent (20%)
^ (caret) Exponentiation (3^2)
To create a new variable Y2, select cell E2 and enter the formula =C2^2. This command instructs
Excel to square the value in cell C2.
The new value is Y2 = 4. Select this cell, and click Copy (or the shortcut Ctrl+C)
The border of the cell begins to rotate. Move the cursor to the lower right corner of the cell until
the “plus” sign appears.
Hold the left mouse button down and drag the plus down to cell E6.
Release the left button and you will find the new values of Y2. What you have done is Copy the
formula in E2 to the new cells, and Excel calculates the square of each Y value.
1.3.2 Mathematical functions

There are a large number of mathematical functions, most of which you will never use. Find the
right function is fairly simple using Help. Search for math functions.
Among the results returned is “Find functions and enter arguments” which is an online training
session video, which you might find useful.
10 Chapter 1
Another entry is “List of worksheet functions”. A partial list of the functions available is listed on
the next page. These are taken from the Excel help result.
Math and trigonometry functions

Function Description
ABS Returns the absolute value of a number
CEILING Rounds a number to the nearest integer or to the nearest multiple of significance
COMBIN Returns the number of combinations for a given number of objects
EVEN Rounds a number up to the nearest even integer
EXP Returns e raised to the power of a given number
FACT Returns the factorial of a number
FACTDOUBLE Returns the double factorial of a number
FLOOR Rounds a number down, toward zero
GCD Returns the greatest common divisor
INT Rounds a number down to the nearest integer
LCM Returns the least common multiple
LN Returns the natural logarithm of a number
LOG10 Returns the base-10 logarithm of a number
ODD Rounds a number up to the nearest odd integer
PI Returns the value of pi
POWER Returns the result of a number raised to a power
PRODUCT Multiplies its arguments
QUOTIENT Returns the integer portion of a division
RAND Returns a random number between 0 and 1
RANDBETWEEN Returns a random number between the numbers you specify
ROUND Rounds a number to a specified number of digits
ROUNDDOWN Rounds a number down, toward zero
ROUNDUP Rounds a number up, away from zero
SERIESSUM Returns the sum of a power series based on the formula
SIGN Returns the sign of a number
SQRT Returns a positive square root
SQRTPI Returns the square root of (number * pi)
SUBTOTAL Returns a subtotal in a list or database
SUM Adds its arguments
SUMIF Adds the cells specified by a given criteria
SUMPRODUCT Returns the sum of the products of corresponding array components
SUMSQ Returns the sum of the squares of the arguments
Returns the sum of the difference of squares of corresponding values in two

SUMX2MY2
arrays
SUMX2PY2 Returns the sum of the sum of squares of corresponding values in two arrays
SUMXMY2 Returns the sum of squares of differences of corresponding values in two arrays
TRUNC Truncates a number to an integer

12 Chapter 1
Of course there is no way to remember all these. Again however Excel makes the functions easy
to use. There is an Insert Function (or Paste Function) button.
Create a new heading in F1, LOGY, which will contain the natural logarithm of Y. All logs used
in POE are natural logs. Highlight F2 and click Insert Function.
The Insert Function dialog box opens. Type in a description of the function you want, and press
Go.
Excel will return some suggestions. Scroll down the list and note that the definitions of the
functions appear at the bottom.
Click OK. A Function Arguments dialog box opens. Enter the number you wish to take the
logarithm of, or to locate a cell click on the Data range box.
With this box open select the cell C2. This enters the command =LN(C2) into F2. Press Enter.
Back in the Function Arguments dialog box we find that the natural log of 2 is 0.693. Click OK.
14 Chapter 1
The value is returned to the worksheet in F2. Once again we can Copy the formula to compute
the natural log of the remaining Y values.
Once you know the function you wish, you can of course just enter the formula into a cell and
press Enter.
Now that you have put lots of effort into this example, it is a good idea to save your work. On the
Excel Menu bar, select File/Save as
In the resulting dialog box, find the folder in which you plan to save your work for POE. We will
use the path c:\data\excel. The standard extension for Excel files is *.xls. Name the file and click
Save.
1.4 EXCEL FILES FOR PRINCIPLES OF ECONOMETRICS
The book Principles of Econometrics, 3e, uses many examples with data. These data files have
been saved as workbooks and are available for you to download to your computer. There are
about 150 such files. The data files and other supplementary materials can be downloaded from
two web locations. You should download not only the *.xls files, but also the definition files,
which have the extension *.def. Download these files from either the publisher John Wiley and
Sons, or from the book website maintained by the authors.
1.4.1 John Wiley and Sons website

Using your web browser enter the address www.wiley.com/college/hill. Find, among the authors
named “Hill” the book Principles of Econometrics.
16 Chapter 1
Click on the book title and follow the link to Supplements. Click on Supplements. There you
will find links to many supplement materials, including a link that will allow you to download all
the data files at once.
1.4.2 Principles of Econometrics website

Alternatively, you may wish to download individual files.
• Go to the site www.bus.lsu.edu/hill/poe for the data, errata and other supplements.
• For the Excel data files go to www.bus.lsu.edu/hill/poe/excel.htm.
1.4.3 Definition files

There is a data definition file for each data file used in the book. These are simple “text” or
“ASCII” files that can be opened with utilities like Notepad or Wordpad, or a word processor.
Locate food.def. Its contents are:
food.def
food_exp income
Obs: 40
1. food_exp (y) weekly food expenditure in $

2. income (x) weekly income in $100
Variable | Obs Mean Std. Dev. Min Max

-------------+--------------------------------------------------------
food_exp | 40 283.5735 112.6752 109.71 587.66
income | 40 19.60475 6.847773 3.69 33.4
The definition files contain variable names, variable definitions, and summary statistics.
1.4.4 The food expenditure data

In the first few chapters you will use data on household food expenditure. Locate the file food.xls
and open it. To illustrate, click on the Open icon on the Menu.
Navigate to the file you wish to open.

The worksheet should appear as
So as to not alter the original file, you may want to save the file with a new name, such as
food_chap01.xls. Select File>Save As from the Excel Menu.
In the dialog box enter

18 Chapter 1
Compute the summary statistics, to make sure they match the ones in food.def. Select
Tools>Data Analysis.
In the resulting dialog box choose Descriptive Statistics, then OK.
In the Descriptive Statistics dialog box we must enter the Input Range of the data. Click on the
Data Range box.
The following box will open.
While it is open, highlight columns A & B, rows 1 to 41. You can do this by
• Click A1, hold down left mouse button, and drag over desired area; or
• Click A1, hold down Ctrl+Shift. Press right arrow → and then down arrow ↓.
In the resulting window click the Data Range box again.

Note that now the input range is filled in.
• The Excel data range is $A$1:$B$41. This range locates the upper left corner of the
highlighted area (A1) and lower right corner (B41). The $ makes this an Absolute Cell
Reference, that will not be changed if the data are moved.
• Tick the box Labels in First Row so that these cells will not be treated as data.
• Select the radio button New Worksheet Ply and enter a name for the new worksheet
page.
• Tick the box Summary Statistics so that Excel will print the results.
Click OK
The resulting worksheet is not formatted well. Select Format>Column>AutoFit Selection

20 Chapter 1
One of the nice things about Microsoft Office is that information from one application can be
transferred to another. With the Descriptive Statistics highlighted, enter Ctrl+C to copy. In an
open document enter Ctrl+V to paste.
Now you have a nice table of statistics in your document that can be edited in the usual way.
food_exp income
Mean 283.5734993 Mean 19.60475005
Standard Error 17.81551026 Standard Error 1.08272795
Median 264.479996 Median 20.0299995
Mode #N/A Mode #N/A
Standard Deviation 112.6751802 Standard Deviation 6.847772819
Sample Variance 12695.69623 Sample Variance 46.89199259
Kurtosis -0.002430221 Kurtosis 0.48455582
Skewness 0.511465877 Skewness -0.651185498
Range 477.949974 Range 29.710002
Minimum 109.709999 Minimum 3.69
Maximum 587.659973 Maximum 33.400002
Sum 11342.93997 Sum 784.190002
Count 40 Count 40
You may wish to save this file by clicking
The food expenditure data will be used extensively in the next chapter.
CHAPTER 2
The Simple Linear Regression
Model
CHAPTER OUTLINE
2.1 Plotting the Food Expenditure Data 2.4 Plotting the Least Squares Residuals
2.2 Estimating a Simple Regression 2.5 Prediction Using Excel
2.3 Plotting a Simple Regression
In this chapter we introduce the simple linear regression model and estimate a model of weekly
food expenditure. We also demonstrate the plotting capabilities of Excel and show how to use the
software to calculate the income elasticity of food expenditure, and to predict food expenditure
from our regression results.
2.1 PLOTTING THE FOOD EXPENDITURE DATA
We will use Chart Wizard to scatter plot the data. Open food.xls file in Excel.
21
22 Chapter 2
Place the cursor on the Chart Wizard icon and click. In the dialog box, choose the chart type as
XY (Scatter) and click next.
To define the Data Range, highlight the data after clicking on the space provided.
Select the data columns and click Data range again
Aside: Referencing Cells

Select Help. Search the phrase “Cell Reference.” One of the resulting hits is
The Simple Linear Regression Model 23
Click this link to find the following description of cell references:
A reference identifies a cell or a range of cells on a worksheet and tells Microsoft Excel where to
look for the values or data you want to use in a formula. With references, you can use data
contained in different parts of a worksheet in one formula or use the value from one cell in
several formulas. You can also refer to cells on other sheets in the same workbook, and to other
workbooks. References to cells in other workbooks are called links.
The A1 reference style
By default, Excel uses the A1 reference style, which refers to columns with letters (A through IV,
for a total of 256 columns) and refers to rows with numbers (1 through 65536). These letters and
numbers are called row and column headings. To refer to a cell, enter the column letter followed
by the row number. For example, B2 refers to the cell at the intersection of column B and row 2.
To refer to Use
The cell in column A and row 10 A10
The range of cells in column A and rows 10 through 20 A10:A20
The range of cells in row 15 and columns B through E B15:E15
All cells in row 5 5:5
All cells in rows 5 through 10 5:10
All cells in column H H:H
All cells in columns H through J H:J
The range of cells in columns A through E and rows 10 through

A10:E20
20
Reference to another worksheet In the following example, the AVERAGE worksheet

function calculates the average value for the range B1:B10 on the worksheet named Marketing in
the same workbook.
Note that the name of the worksheet and an exclamation point (!) precede the range reference.
Aside: Relative vs. Absolute References

On the same Help page, you will find the following useful information:
The difference between relative and absolute references
Relative references A relative cell reference in a formula, such as A1, is based on the relative
position of the cell that contains the formula and the cell the reference refers to. If the position of
24 Chapter 2
the cell that contains the formula changes, the reference is changed. If you copy the formula
across rows or down columns, the reference automatically adjusts. By default, new formulas use
relative references. For example, if you copy a relative reference in cell B2 to cell B3, it
automatically adjusts from =A1 to =A2.
Copied formula with relative reference

Absolute references An absolute cell reference in a formula, such as $A$1, always refer to a
cell in a specific location. If the position of the cell that contains the formula changes, the
absolute reference remains the same. If you copy the formula across rows or down columns, the
absolute reference does not adjust. By default, new formulas use relative references, and you need
to switch them to absolute references. For example, if you copy a absolute reference in cell B2 to
cell B3, it stays the same in both cells =$A$1.
Copied formula with absolute reference
2.1 (Continued)
Excel assumes that the first column is the X-variable. Select the Series tab.
On the Series tab define the columns that are X and Y variables again using the Data range box.
• Click Next.
• Add or modify labels and title and click Next.
• The default legend is not informative. To delete it go the Legend tab and uncheck box.
26 Chapter 2
• The last step is to place your chart. You can have the chart on the current worksheet or in
a new one. Make your choice and click Finish.
If you chose to print it in a new worksheet, your plot will appear in the new worksheet.
Food_E xp
700
600
500
400
300
200
10 0
0 5 10 15 20 25 30 35 40 45
in c o m e
To print a worksheet, first click on File/Page Setup.
There you can choose the page layout, either Portrait or Landscape, and adjust the size. The
Print Preview allows you to see what will be printed before actually doing so, which saves time
and paper.
Click on File/Print to open the print dialog window. In the Print dialog box, make sure the
printer is correctly specified. Here you can specify Print range, pages to be printed etc.
28 Chapter 2
Alternatively, if the Print Preview is satisfactory, click on the printer icon on the main toolbar.
If you are creating a document, click inside the graph and enter Ctrl+C which copies the figure
into the “clipboard.” Transfer to an open document, place the cursor in the document where you
want the figure and enter Ctrl+V to paste the figure into the document, where it can be resized by
dragging a corner inward. Give it a try.
2.2 ESTIMATING A SIMPLE REGRESSION
To estimate the parameters b1 and b2 of the food expenditure equation, place cursor in an empty
cell and click on Tools/Data Analysis.
The Data Analysis tool may not automatically load with a default installation of the program, if
Data Analysis tool doesn't appear on the menu, click on Add-Ins under Tools:
Check the box next to the Analysis ToolPak add-in and click OK.
Data Analysis should now appear in the Tools menu, and you should not have to run the add-ins
again for this option.
When the Data Analysis dialog box appears, click on Regression, then OK.
The Regression dialog box will appear with lots of user-defined inputs and options.
30 Chapter 2
To define the input ranges, first click on the Input Y Range box. The box will minimize.
Highlight the data in the y column, including the label. Do the same procedure to input the X
Range.
Other options:
• Don’t forget to check the Label box if you are including labels.
• Do not check the Constant is Zero box. This would suppress the intercept.
• The output can be sent to the current page or another page in the workbook.
Name the new worksheet Food Expenditure and hit Enter.
Since you chose to place the output in a separate worksheet, a new worksheet will appear as a tab
in the lower left corner of the work area. If you click on the Food Expenditure tab, you will
notice that the columns are not wide enough to show the cells completely. Highlight the data, or
the entire sheet, and click on Format/Column/Autofit Selection.
The output contains many items that you will learn about later. For now it is important to note
that the Coefficients corresponding to Intercept and income are the least squares estimates b1 and
b2.
2.3 PLOTTING A SIMPLE REGRESSION
In order to plot the regression function we must re-estimate the food expenditure equation and
choose the Line Fit Plots option in the regression dialog box.
32 Chapter 2
Click OK. The graph will be produced, and placed on the same worksheet as the regression
output. If you can't find it on the worksheet, click on the File/Print Preview or click on the Print
Preview icon and look for it.
You will need to enhance your graph. Here are a few suggestions:
• Right click on the legend and delete, if desired.
• Left click once on the title to highlight it. Left click once more to edit. Name
appropriately. If you double click the box surrounding the title, a dialog box opens that
allows you more formatting options.
Repeat for the Y and X axes if you want to change the names of the variables.
Notice that both the actual values of Y and the predicted values are plotted. To include the
estimated regression function, place the cursor over one of the predicted Y points (the pink ones)
until the caption "Series Predicted Y" appears.
Right Click and choose Add Trendline.
Under the Type tab, choose Linear.

34 Chapter 2
Under the Options tab, check the box next to Display equation on chart. Click OK.
If your figure is small, and begins to get cluttered, increase its size by clicking inside the border.
Place the mouse arrow over the black square in the corner, until a double arrow appears. Then
drag the mouse, with the left button held down, to increase (or decrease) the figure size.
Your figure should look something like the one below.
2.4 PLOTTING THE LEAST SQUARES RESIDUALS
The least squares residuals are defined as
eî = y i − yˆ i = y i − b1 − b2 xi
In order to plot the least squares residuals, we must re-estimate the food expenditure equation and
choose the Residual Plots option in the regression dialog box.
If you wish to enhance your graph, you can do so by right clicking on chart area or plot area.
36 Chapter 2
2.5 PREDICTION USING EXCEL
Obtaining predicted values from an estimated regression function can be done either by doing the
computations or using the TREND function.
Insert a new worksheet to the workbook select Insert/Worksheet on the main menu.
Rename the worksheet Predictions by right clicking on the tab.

In the Predictions worksheet, create a template for prediction by copying the estimated
coefficients from the regression and labeling them appropriately. Enter the income value for
which you want to predict food expenditures. Create the formula for the predicted value of y,
yˆ = b1 + b2 Income , using the cell references.
The results in cell B6 will be 287.6088614.
Excel has a built in function that computes predicted values from simple regressions. The form of
the Trend function is
TREND(range of Y variable, range of X variable, value of x0)

38 Chapter 2
The value x0 is the value at which the prediction is desired. To use this function, return to the
worksheet page containing the data. Type in the following command,
The result will be in cell D2.
As a final step, save your file. We recommend saving it under a new name, like food_chap02.xls,
so that the original data file will not be altered.
CHAPTER 3
Interval Estimation and Hypothesis
Testing
CHAPTER OUTLINE
3.1 Interval Estimation 3.2 Hypothesis Testing
3.1.1 Automatic interval estimates 3.2.1 Right-Tail tests
3.1.2 Constructing interval estimates 3.2.2 Left-Tail tests
3.2.3 Two-Tail tests
In this chapter we continue to work with the simple linear regression model and our model of
weekly food expenditure.
3.1 INTERVAL ESTIMATION
For the regression model y = β1 + β2 x + e , and under assumptions SR1-SR6, the important result
that we use in this chapter is given by
bk − βk
t= ~ t( N − 2) for k = 1,2
se ( bk )
Using this result we can show that the interval bk ± tc se ( bk ) has probability 1− α of containing
the true but unknown parameter βk , where the “critical value” tc from a t-distribution such that
P ( t ≥ t c ) = P ( t ≤ −t c ) = α 2
3.1.1 Automatic interval estimates

To construct the confidence interval estimates, we will use the Regression function in Excel. To
do that, open the workbook containing the food expenditure regression that we considered in
39
40 Chapter 3
Chapter 2. Excel provides 95% confidence interval for the Least Squares estimates by checking
the Confidence Level box in the Regression Dialog box.
Excel will report the confidence interval in the Summary Output next to the coefficient
estimates. The results show the lower and upper values for the 95% confidence interval for the
coefficient estimates.
To have Excel calculate a different confidence interval, estimate a regression, and after checking
the Confidence Level box, type in the desired level, for example, 98 for 98% confidence interval.
Interval Estimation and Hypothesis Testing 41
3.1.2 Constructing interval estimates

While using the automatic interval estimation feature in Excel is the quickest and easiest way to
obtain interval estimates, a general template can be created for calculating interval estimates.
To construct the interval estimates we require the least square estimates bk, their standard
errors se(bk) and the critical value for t-distribution, tc. We already know we can find the least
squares estimates bk , their standard errors se(bk) in summary output. We also need to find values
tc, such that α/2 of the probability is in either tail. As an example, the critical values that mark off
α/2 = .025 of the probability in each tail of a t-distribution with 38 degrees of freedom. Checking
Table 2 at the front of your book, we find that the value is tc = 2.024 on the positive side and,
using the symmetry of the t-distribution, − 2.024 on the negative side. Excel makes it easy to
compute critical values from the t-distribution using the TINV function.
To generate 95% confidence interval, first open the workbook containing the food
expenditure regression and insert a worksheet. Move the cursor over the tab with the default
name, right click, and rename the sheet t-values.
Go to the newly created worksheet and select a cell. Click on the drop down menu next to Sum
(∑).
Find TINV from the Statistical Function category, and click OK.
In case you don’t see Statistical Function on your menu, choose More Functions. Then, find
TINV from the Statistical Function category, and click OK.
42 Chapter 3
Alternatively, click on the Paste or Insert Function (f*) icon to open up the same search.
Either way, in the TINV dialog box, fill in the box as shown below.
The Probability value you need to fill in is the α in two-tails of the t-distribution. Enter the
degrees of freedom, 38, and click OK. The resulting value, 2.024394147 will appear in your
worksheet when you click OK.
We can now create a Confidence Interval Template. Cells with bold border require user
input, some obtained from the regression results.
Once you have the worksheet, the Least Squares Estimates and the Standard Error values can
be typed in, or can be copied from the regression results on the food expenditure worksheet.
3.2 HYPOTHESIS TESTING
Inference in the linear regression model includes tests of hypotheses about parameters which also
depend upon Student’s t-distribution. One- or two-tail general tests can be calculated by methods
similar to the confidence interval construction. The required ingredients are results from the least
squares estimation and the ability to use the Excel functions TINV and TDIST.
3.2.1 Right-Tail tests

To test the hypothesis H 0 : β2 = 0 against the alternative that it is positive (>0), as described in
Chapter 3.4.1a of POE, we use a one-tail significance test. For this purpose, we must compute the
value of the critical values that define the rejection region, the test statistic and the p-value of the
test.
44 Chapter 3
• If we choose the α = .05 level of significance, the critical value is the 95th percentile of
the t(38) distribution which can be computed by the TINV function discussed earlier.
TINV(0.10, 38) = 1.685954461. The value returned is the right tail critical value.
Beware that the Excel function TINV(probability, degrees of freedom) returns the value
such that the area in the two-tails of the t(df) distribution equals the given probability
value. Thus if we want a critical value such that α = .05 is in the right tail, we must
provide the TINV function with probability = .10.
• The test statistic is the ratio of the estimate b2 to its standard error, se(b2).
• For the p-value, we use the TDIST function where the amount of probability in the right-
tail of a t-distribution with 38 degrees of freedom can be computed. To obtain this value,
we find TDIST from the Statistical Function category under either Sum drop down
menu, or Paste Function.
This work can be simplified by using a template that can be used when needed.
The t-statistic value 4.8773 falls in the rejection region, and p-value is less than the level of
significance α, thus we reject this null hypothesis.
To test an economic hypothesis such as H 0 : β2 ≤ 5 against H1 : β2 > 5 , the same steps are
followed except for the construction of the t-statistic. This can be accomplished by replacing the
0 with 5 in our template.
The t-statistic value 2.4887 falls in the rejection region, and p-value is smaller than the level
of significance α, thus we reject this null hypothesis.
3.2.2 Left-Tail tests

To test the significance, we test H 0 : β2 ≥ 0 against H1 : β2 < 0 . The value of the t-statistic for this
null and alternative hypothesis is the same as for a right-tailed test.
• If we choose the α = .05 level of significance, the critical value is the 5th percentile of the
t(38) distribution which can be computed by the TINV function discussed earlier.
−TINV(0.10, 38) = −1.685954461. The value returned is the right tail critical value.
• The test statistic is the ratio of the estimate b2 to its standard error, se(b2).
• The p-value, is the area to the left of the calculated t-statistic.
Let’s plug in the left tail values into our template:

46 Chapter 3
To test the null hypothesis that β2 ≥ 12 against the alternative β2< 12, we use the template and
plug in the input numbers. The t-statistic value −1.6859 does not fall in the rejection region, and
p-value is greater than the level of significance α, thus we fail to reject this null hypothesis.
3.2.3 Two-Tail tests

For the two tail test of the null hypothesis that β2 = 0 against the alternative that β2 ≠ 0 the same
steps are taken. We can plug the necessary information into the template.
Since the t-statistic value 4.8773 falls in the rejection region, and p-value is smaller than the level
of significance α, thus we reject this null hypothesis. This test is also carried out by Excel within
the Regression Summary Output labeled t-Statistic and p-value.
To test the null hypothesis that β2 = 12.5 against the alternative β2 ≠ 12.5, we use the two-tailed
test template and plug in the input numbers. The t-statistic value −1.6859 does not fall in the
rejection region, and p-value is greater than the level of significance α, thus we fail to reject this
null hypothesis.
48 Chapter 3
CHAPTER 4
Prediction, Goodness-of-Fit, and
Modeling Issues
CHAPTER OUTLINE
4.1 Prediction for the Food Expenditure Model 4.4 Modeling Issues
4.1.1 Calculating the standard error of the 4.4.1 Scaling the data
forecast 4.4.2 The log-linear model
4.1.2 Prediction interval 4.4.3 The linear-log model
4.2 Measuring Goodness-of-Fit 4.4.4 The log-log model
4.2.1 R2 4.5 More Examples
4.2.2 Covariance and correlation analysis 4.5.1 Residual analysis with wheat data
4.3 Residual Diagnostics 4.5.2 Log-linear model with wage data
4.3.1 The Jarque-Bera test 4.5.3 Generalized R2
4.1 PREDICTION IN THE FOOD EXPENDITURE MODEL
We have already illustrated how to obtain the predicted values for the food expenditure for a
household in Chapter 2. In this chapter, we will calculate the standard error of the forecasted
value and construct a prediction interval.
4.1.1 Calculating the standard error of the forecast

Recall from Section 2.6 of this manual, the forecasted value of household food expenditure for a
household with income of $2000 per week is calculated as $287.6088614. Now, we will compute
the standard error of the forecasted value where forecast error is calculated as
f = y0 − y0 = (β1 + β2 x0 + e0 ) − (b1 + b2 x0 )
The estimated variance for the forecast error is
49
50 Chapter 4
⎡ ( x0 − x ) ⎤
2
1 σˆ 2
var ( f ) = σˆ ⎢1 + + + ( x0 − x ) var ( b2 )
2
2
⎥ = σ
ˆ 2
+
⎢⎣ N Σ ( xi − x ) ⎥⎦
2
N
And the square root of the estimated variance is the standard error of the forecast.
se ( f ) = var ( f )
To calculate the forecast error for the food expenditure data, open the food.xls file. In addition to
the summary regression output (see Section 2.3 of this manual), we will need the sample mean of
income. To obtain the sample mean of income select Tools/Data Analysis and from the menu
choose Descriptive Statistics.
In the dialog window specify the data range and ask for summary statistics.
Excel will provide the univariate summary statistics for Food Expenditure and Income variables.
We will use the mean of income and sample size in the calculation of the standard error of the
forecast.
Prediction, Goodness-of-Fit, and Modeling Issues 51
Now, we can go back to Regression Output, and plug in the necessary numbers into our formula
and calculate the standard error of the forecast. Recall from above that the formula is
⎡ ( x0 − x ) ⎤
2
1 σˆ 2
var ( f ) = σˆ ⎢1 + + + ( x0 − x ) var ( b2 )
2
2
⎥ = σ
ˆ 2
+
⎢⎣ N Σ ( xi − x ) ⎥⎦
2
N
and
se ( f ) = var ( f )
The formula indicates that, the farther x0 is from the sample mean x , the larger the variance of
the prediction error, the smaller the sample size, the larger the forecast error and the less reliable
the prediction is likely to be.
52 Chapter 4
To calculate the standard error of the forecast for x0 , we will need the σ2 which is the Mean
Squared Residual (MSR) from cell D13 in the ANOVA table, the var ( b2 ) which is the square
of the standard error of income from cell C18 and x0 which is $2000. Once we type in the
formula, it is possible to make the same calculation for different values of income.
4.1.2 Prediction interval

We construct a 100(1 − α ) % prediction interval as
y0 ± tc se ( f )
Since forecasted value and the standard error of forecast have been already calculated,
constructing the confidence interval is very straightforward. Recall that we can obtain the tc
values as shown in Section 3.1.2 and the forecasted values in Section2.6 of this manual.
We can also create a worksheet, name it predictions and calculate standard error of forecast and
the prediction intervals for specified values of income.
4.2 MEASURING GODDNESS-OF-FIT
The ANOVA table in the regression output provides the goodness-of-fit measures.
54 Chapter 4
4.2.1 Calculating R2
In the simple regression model, the R 2 is the square of the sample correlation between the x and y
variables. It is calculated as R 2 = SSR SST = 1 − SSE SST is reported in standard regression
output.
4.2.2 Covariance and correlation analysis

The covariance and correlation can tell us about the linear relationship between two variables, a
primary concern of linear regression. Specifically, the covariance tells us the direction of the
linear relationship, while the correlation is a measure of the strength (and direction) of the linear
relationship. Multiple R, in the simple regression output, gives us the square root of R 2 which is
the correlation between x and y.
A more general way to calculate covariance and correlation can be achieved by utilizing the Data
Analysis under the Tools menu.
The sample correlation coefficient, r, measures the direction and strength of the linear
relationship between two variables and is between −1 and 1. To obtain the sample correlation
coefficient, choose correlation from the Tool/Data Analysis menu and click OK. The
Correlation dialog box will appear. Fill in the appropriate input range and be sure to check the
Labels in First Row box since labels are included. Place the output on a new worksheet named
Correlation and click OK.
The correlations are:
The results will appear in the new worksheet, Correlation. You may need to format the
worksheet by choosing Format/Column/AutoFit Selection. The estimated correlation between
food expenditures and weekly income is 0.620485 which is the value given as Multiple R in the
regression output summary. Values on the diagonal of the correlation matrix will always equal 1.
56 Chapter 4
To find the sample covariance matrix, click on Tool/Data Analysis menu again and highlight
Covariance and click OK.
The Covariance dialog box will appear. Fill in the appropriate input range and be sure to check
the Labels in First Row box since labels are included. Place the output on a new worksheet
named Covariance and click OK.
The covariance matrix will appear on the new worksheet, but needs to be formatted. Choose
Format/Column/Auto Fit Selection.
The diagonal elements of the covariance matrix are the estimated sample variances. The
covariance between food expenditures and weekly income is positive, suggesting a positive linear
relationship. The value of the covariance, 466.7817, does not, however, tell you the strength of
that linear relationship.
4.3 Residual Diagnostics

Every time a regression model is estimated, certain regression diagnostics need to be carried out.
By analyzing the residuals of the fitted model, we may be able to detect model specification
problems. A histogram of the residuals can suggest the distribution of the errors, and the Jarque-
Bera test statistic can be used to formally test for normality. Both of these functions are important
since hypothesis testing and interval estimations are based on distributional assumptions.
In order to create a histogram of the residuals, we need to rerun the food expenditures model
and choose the Residuals output option.
Excel will provide the residuals for each observation, in addition to the standard Regression
output. Examine the values of the residuals, noting the lowest and highest values. Create a BIN
column next to the residuals column and determine the category values for the histogram. In this
column, enter the values -250, -200, -150, -100, -50, 0, 50, 100, 150, 200 and 250.
58 Chapter 4
Once you have entered the BIN column, choose Tools/Data Analysis/Histogram from the menu
bar, and click OK.
The Histogram dialog box will appear.
Fill in the data ranges, by highlighting the residuals for the Input Range, and highlight the values
created in the BIN column for the Bin Range. Place the output on a new worksheet called
Histogram. After checking the Chart Output box, click OK.
The Bin values and Frequencies appear, along with the histogram. Format the histogram graph
as needed. (Remove legend, resize, rename title, etc). The residuals seem to be centered around
zero, but the symmetry of the distribution seems questionable.
4.3.1 Jarque-Bera test for normality

A formal test of normality, the Jarque-Bera test, uses skewness and kurtosis, which can be easily
estimated with Excel.
As discussed earlier in Section 4.1.1, Excel has a tool for calculating the Descriptive
Statistics, which can be found on the Tools/Data Analysis menu. Skewness and kurtosis can also
be found among the statistics calculated.
Skewness is a measure of asymmetry of a distribution about its mean. For a sample
x1 , x2 ,…, xT an empirical measure of skewness is
Σ ( xi − x )
3
S= T
σ3x
where
1 T
∑ ( xi − x )
2
σx =
T i =1
Kurtosis measures the peakedness, or flatness, of a distribution. An empirical measure of

Kurtosis is
Σ ( xi − x )
4
K= T
σ 4x
T
In the case of the least squares residuals, ê , the formulas simplify because ∑ eˆ
i =1
i = 0 , making
eˆ = 0 . Thus the formulas for skewness and kurtosis of the least squares residuals are
S=
∑ eˆ i
3
T
, K=
∑ eˆ i
4
T
σ 3
σ 4
respectively where
1 T 2
σ= ∑ eî
T i =1
Skewness measures the symmetry of the data, a value of zero indicating perfect symmetry.
Kurtosis refers to the "peakedness" of the distribution, with a value of 3 for a normal distribution.
Using these measures, the test statistic for the Jarque-Bera test for normality is
60 Chapter 4
T⎛ ( K − 3) ⎞
2
JB = ⎜ S 2 + ⎟
⎜
6⎝ 4 ⎟
⎠
where S is skewness and K is kurtosis. This test statistic follows a chi-square distribution with 2
degrees of freedom. Below is preparation of the worksheet for this calculation for the food
expenditures model.
Return to the worksheet containing the regression results and the residuals. Copy the column
containing the least squares residuals to a new worksheet and create three additional columns and
label them ‘ehatsquared’, ‘ehatcubed’ and ‘ehat to the fourth’. The formulas for these functions
are ^2, ^3 and ^4 respectively and is illustrated in the below figure. Once you copy the formulas
down the columns, you will be in a position to compute the S, K and JB statistics. With formulas
showing, this portion of the worksheet should look like
4.4 MODELING ISSUES
4.4.1 Scaling the data

Change the scale of the variables can easily be done on the worksheet containing the data. Label
the empty column to the right of the independent variable INCOME*. In the first empty cell of
this column, type =B2*100. Copy the formula down the column.
Estimate a regression, using the new independent variable, INCOME* instead of the original
independent variable, INCOME.
The coefficient on INCOME has changed and so has its standard error. Everything else in the
regression output remained the same. When reporting results, be sure to note the appropriate units
of measure for both food expenditures and weekly income.
4.4.2 The log-linear model

The use of logarithmic transformations are very common with economic data. Transforming the
dependent variable using the log function will result in changing the interpretation of the
regression equation. To estimate the log-linear version of the food expenditure model, we need to
transform the dependent variable.
ln ( y ) = β1 + β2 x + e
First, create a new column and label it ln(Food_Exp). In this column, calculate the natural log of
food expenditures by typing =ln(A2) in the first empty cell. Copy the formula down the column.
62 Chapter 4
Estimate the regression model using the ln(Food_Exp) as the dependent variable instead of
Food_Exp.
The interpretation for will be, an increase in income of $100 leads to a 4% increase in the food
expenditure.
4.4.3 The linear-log model

In linear-log model, the independent variable is transformed but not the dependent variable.
y = β1 + β2 ln ( x ) + e
The interpretation for the linear-log model is, 1% increase in income, leads to about $1.32
increase in the food expenditure.
4.4.4 The log-log model

In log-log model, both the dependent and the independent variables are transformed. Estimated
β2 represents the elasticity indicating % change in the y variable, when x variable increase by
1%.
ln ( y ) = β1 + β2 ln ( x ) + e
In the food expenditure model, a 1% increase in income, is expected to increase food expenditure
by about 0.56%.
4.5 MORE EXAMPLES
4.5.1 Residual analysis with wheat data

For this example, open the excel file wa-wheat.xls. The dataset contains 48 observations on
average annual wheat yield in four different shires of Western Australia; Northampton, Chapman,
Mullewa, Greenough and TIME. First, estimate the simple regression of Greenough on time, and
ask for the residual plot.
Excel will produce the scatter plot of the residuals through time in addition to the standard
regression output.
64 Chapter 4
You can format the plot or change the plot type by right-clicking on the picture and selecting
Chart Type. You can then pick the desired chart type. Let’s pick columns for this example.
Excel will change the chart type of the residual plot, and provide a bar chart.
Based on the residual analysis, if you think that is the incorrect functional form, you can generate
a new column in Excel and transform time variable and rerun the regression using the
transformed variable as your independent variable. To generate the cubic equation results
described in the text, generate a transformed column under column F and replace TIME with
TIME CUBE as your independent variable.
4.5.2 Log-linear model with wage data

We will use the cps_small data set to illustrate a log-linear model. Open the cps_small.xls data
set. In this chapter, we will only use EDUC and WAGES, so delete the rest of the columns. You
can delete the columns by highlighting the entire column and selecting Delete under Edit.
After deleting the unwanted columns, save the Excel sheet as Ch4_Wage. Then, transform the
wage variable into log by typing the function in the first row below the label and copying down
the entire column.
66 Chapter 4
You can now estimate the log-linear function by using the new variable, ln(WAGE) as the
dependent variable.
The forecasted value from the log-linear wage equation is
lnyhat = b1 + b2 x
In order to obtain a prediction for the dependent variables y, we need to use the exponential
function to get the “natural predictor” back.
yˆ n = exp ( lnyhat ) = exp(b1 + b2 x)
In large samples a more precise predictor is obtained by correcting that “natural predictor” by
multiplying it by exp(σˆ 2 / 2) . Using the coefficient estimates, the raw forecasted values can be
calculated as shown in the cell B23 of the below Excel output. Then, the natural predicted and the
corrected predictors are calculated.
The standard error of the forecast can be found from the regression output using the formula
provided in 4.1.1. After the predicted values and the standard error of forecast is found, it is very
straight forward to construct the prediction interval. Remember that the corrected predictor will
always be greater than the natural predictor since the correction factor is always greater than one.
4.5.2 Generalized R2
The generalized R2 is the appropriate measure of fit for this model is the square of the correlation
between the “best” predictor and the wage variable. Remember that the corrected predictor and
the natural predictor only differ by the constant so they have the same correlation to the wage
variable. You can calculate the generalized R2 by using the Tools>Data Analysis>Correlation
and choose the column of Wage and the column of predicted (either corrected or natural) wage or
you can use the CORREL statistical function:
CORREL(wage,wage_c)
CHAPTER 5
Multiple Linear Regression
CHAPTER OUTLINE
5.1 Big Andy’s Burger Barn 5.4 Confidence Intervals
5.2 Prediction 5.5 Hypothesis Testing
5.3 Sampling Precision 5.6 Goodness-of-Fit
5.1 BIG ANDY’S BURGER BARN
The multiple linear regression model is the extension of the simple model where there is more
than one explanatory variable. We will use Big Andy’s Burger Barn to transition to the multiple
regression model. This is a multiple regression model where the dependent variable, SALES is a
linear function of PRICE charged and the level of advertising, ADVERT.
SALES = β1 + β2 PRICE + β3 ADVERT + e
Using Excel to perform multiple regression is just like estimating a simple regression model,
except we will include all explanatory variables in the X range. Open andy.xls.
68
Multiple Linear Regression 69
The first column is the monthly sales in $1,000 in a given city, the second column is the price of
hamburgers (actually a price index for goods sold) measured in dollars and the third column is the
advertising spending, also measured in $1,000. To estimate this model, go to Tools>Data
Analysis>Regression and fill in the regression dialog box so that the Y Range is the dependent
variable, SALES and the X Range includes both the PRICE and the ADVERT columns. Make sure
to check the Labels box and hit OK.
The results look very similar to what we've seen before, except now we have parameter estimates
and other information on PRICE and the ADVERT. This portion of the output appears as
70 Chapter 5
5.2 PREDICTION
Using the estimated regression equation, we can now forecast values of sales for different values
of price and advertising like we did in the simple regression case using the below formula.
n = b + b PRICE + b ADVERT
SALES 1 2 3
= 118.9136131 − 7.907854804 * PRICE + 1.862583787 * ADVERT
You can transfer this formula to Excel in the following fashion:
These formulas will yield the sales forecast for specified values of price and the advertising
expenditure as shown below.
5.3 SAMPLING PRECISION
To estimate the error variance, we will use the ANOVA (Analysis of Variance) table. Recall hat
the error variance of the regression equation is:
N
∑ eˆ 2
t
SSE
σˆ 2 = i =1
=
N −K N −K
The formula is the sum of squared errors divided by the degrees of freedom. This quantity is
important and is reported automatically in the ANOVA table when a regression is estimated.
The square root of the estimated regression variance is the Standard Error of the regression and
is reported in the Regression statistics.
Standard Error = MSE = 23.87420813 = 4.886124039

72 Chapter 5
The estimated least squares variance/covariance matrix can be represented as
⎡ var ( b1 ) cov ( b1,b2 ) cov ( b1,b3 ) ⎤

⎢ ⎥
Cov(b1 , b2 , b3 ) = ⎢ cov ( b1,b2 ) var ( b2 ) cov ( b2,b3 ) ⎥
⎢ ⎥
⎣⎢ cov ( b1,b3 ) cov ( b2,b3 ) var ( b3 ) ⎦⎥
The estimated variance/covariance matrix of the least squares estimators are not directly reported
by Excel. However, in the simple model they are easily obtained. The estimated variance of b2 is
σˆ 2
ˆ ( b2 ) =
var T
∑( x − x)
2
t
t =1
The standard error of the estimated coefficient is
σˆ 2
se ( b2 ) = var
ˆ ( b2 ) = T
∑( x − x)
2
t
t =1
In the Excel output we are given the values of the standard errors for the least squares estimates.
The Standard Errors are reported in the column next to the Coefficient estimates. The
estimated variances can be obtained by squaring the standard errors.
So, we can fill in the variances in the covariance matrix as follows:
⎡ 40.3433 cov ( b1,b2 ) cov ( b1,b3 ) ⎤

⎢ ⎥
Cov(b1 , b2 , b3 ) = ⎢ cov ( b1,b2 ) 1.2012 cov ( b2,b3 ) ⎥
⎢ ⎥
⎢⎣ cov ( b1,b3 ) cov ( b2,b3 ) .4668 ⎥⎦
The formula for cov ( b2,b3 ) is:

− r23σˆ 2 T
cov ( b2,b3 ) = ∑ xtk*2 = ∑( x − xk ) .
2
where
(1 − r ) ∑ x ∑ x
tk
2 *2 *2
t =1
23 t2 t3
Since the estimated covariances are not reported by Excel directly, we need to translate this
formula into Excel. Let’s start by creating a worksheet by copying our data worksheet. Choosing
Edit>Move or Copy Sheet.
Move/Copy dialog box will open, choose the data sheet and make sure to click the Create a
copy box.
74 Chapter 5
You will see another tab added to your three existing worksheets. Rename the new sheet
covariance by highlighting the name of the tab. First calculation will be the sum of square
calculations in the denominator,
T 2
∑ xtk*2 = ∑ ( xtk − xk )
t =1
To accomplish this, go to the bottom of PRICE column and write the formula to get the average
x , and repeat it for ADVERT column.
Next, create 2 new columns and name them PRICE - MEANP and ADVERT- MEANA. In these
columns, we will store the difference from the mean.
Recall that we store the mean of price in cell B77. By putting the $ signs, we are making sure,
EXCEL will use the same cell for the mean even when we copy our formula to another cell. Now,
copy and paste the formula to the rest of the column and repeat the process for the ADVERT-
MEANA column, too.
Once you have calculated the differences from means, we will now calculate the sum of squares.
Recall that we can do Sum of Squares calculations using the Insert function
or simply by typing the formula = SUMSQ(…).
Once you have the squares calculated, we can now create columns to calculate the deviation, all
we have to do is to put the formula together. Recall that the formula is
−r23σˆ 2
cov ( b2,b3 ) =
(1 − r ) ∑ x ∑ x
2
23
*2
t2
*2
t3
76 Chapter 5
We already know that σˆ 2 = MSE = 23.8742 from our regression output. We have calculated the
correlation coefficient between PRICE and ADVERT to be 0.0264. Plug all the numbers into the
formula will give us the covariance of
cov ( b2,b3 ) = -0.01974 .
5.4 CONFIDENCE INTERVALS
The 95% confidence interval for each parameter is provided by default in the Excel regression
output. If a different confidence interval is needed, Excel will also provide that. We can return to
the worksheet containing the original andy.xls data. Run the regression using Tools/Data
Analysis/Regression. We can then check the Confidence Level box and set the level to 90. Set
all other desired options and click OK.
Both the 95% and 90% confidence intervals are reported for the β1, β2, and β3.
The 90% confidence interval for β3 suggests that an additional $1000 in advertising expenditures
leads to an increase in total revenue that is between $724 and $3001. Note that a 95% interval is
always wider than the 90% confidence interval.
5.5 HYPOTHESIS TESTING
The t-test to test the significance of β2, that the coefficient is zero, against the two sided
alternative that it is not, is
b2 − 0 −7.908
t= = = −7.215
se ( b2 ) 1.096
Excel provides the t-stats and p-values in the regression output is the simplest way to test the
significance of a single coefficient. The results are
Recall that the t-stat is the coefficient divided by its standard error. Based on the reported p-
values, both b2 and b3 are significant at the 5% as well as 1% level.
Sometimes we will need general tests about our parameters, such as the elasticity of demand
where the null and alternative hypotheses are H0: β2 ≥ 0: a decrease in price leads to a decrease in
total revenue (demand is price inelastic) and H1: β2 < 0: a decrease in price leads to an increase in
total revenue (demand is price elastic).
Or we may want to test an economic hypothesis such as
H 0 : β3 ≤ 1
H 1 : β3 > 1
To test whether an increase in advertising expenditures is "worth it", that is, total revenues
increase enough to cover the increased cost of the advertising, we use the t-statistic
b3 − 1
t=
se ( b3 )
If the null hypothesis is true, it suggests that a dollar increase in advertising expenditures leads to
less than a dollar increase in total revenue. In this case, it doesn't make sense to spend that extra
dollar. On the other hand, if we reject the null hypothesis we conclude that there is sufficient
statistical evidence to suggest that an increase in advertising expenditures is "worth it" in terms of
the increase in total revenue. The value of the test statistic is
b3 − 1 1.8626 − 1
t= = = 1.263
se ( b3 ) 0.6832
Since 1.263<1.666, we do not reject the null hypothesis. The critical value for this one-tail test is
obtained as
78 Chapter 5
5.6 GOODNESS-OF-FIT
The goodness-of-fit of the regression model is based on the ANOVA table. The coefficient of
determination, R2 and the ANOVA table are reported for the multiple regression model as they
are for the simple model. The R2 measures the percent variation in dependent variable explained
by the regression model. We already know how to decompose the sums of squares as
SST=SSR+SSE
And the coefficient of determination, R2 is
SSR SSE
R2 = =1−
SST SST
The results from the regression are

The value R2 = .448 means that 44.8% of the total variation in total revenue is explained by price
and advertising expenditures. However, there can be a problem with this measure in the multiple
regression model since addition of each additional explanatory variable will inflate the R2 even if
there is no economic basis for the variables to appear in the model.
An alternative measure is the “adjusted R2”, denoted by Excel as Adjusted R Square is
reported just below R Square. Adjusted R Square imposes a penalty for adding explanatory
variables so it can never be larger than R2. The Adjusted-R2 can also be calculated from the
ANOVA table.
SSE /( N − K )
R2 = 1−
SST /( N − 1)
While this solves the problem associated with R2 (which has a particular interpretation!), the
adjusted-R2 has no interpretation. It is no longer the percent of the variation in total revenue that
is explained by the model. It should NOT be used as a device for selecting appropriate
explanatory variables; good economic theory should determine the model.
CHAPTER 6
Further Inference in the Multiple
Regression Model
CHAPTER OUTLINE
6.1 The F-test 6.5 Nonsample Information
6.2 Testing the Overall Significance of the Model 6.6 Model Specification
6.3 An Extended Model 6.6.1 Omitted variables
6.4 Testing Economic Hypothesis 6.6.2 Irrelevant variables
6.4.1 The significance of advertising 6.6.3 Choosing the model
6.4.2 Optimal level of advertising 6.7 Poor Data, Collinearity and Insignificance
6.1 F-TEST
The t-test is used to test a specific null hypothesis, such as a single test of significance. With the
multiple regression model, we might be interested in testing whether two or more explanatory
variables are jointly important to the model. The F-test allows for testing joint hypotheses, and is
based on a comparison of the sum of the squared errors from an unrestricted (full, or "original")
model to the sum of squared errors from a model where the null hypothesis has been imposed. In
the Big Andy’s Burger Barn example, we estimated the model
Si = β1 + β2 Pi + β3 Ai + ei
We can use the t-test to test the hypothesis H 0 : β2 = 0 against H1 : β2 ≠ 0 . Another way to test
this hypothesis is in terms of the models each hypothesis implies using an F-test. If the null
hypothesis is true, then the restricted model is formed as
Si = β1 + β3 Ai + ei
The F-test compares the sums of squared errors from the restricted model and the unrestricted
model. A large difference will signal that the restrictions are false. The F-statistic we will use is
80
Further Inference in the Multiple Regression Model 81
( SSER − SSEU ) J
F=
SSEU ( N − K )
where SSER and SSEU are the sum squared errors from the ANOVA tables of the restricted and
unrestricted models, respectively. J is the number of hypotheses or restrictions in the null
hypothesis. N is the sample size of the unrestricted model, and K is the number of parameters in
the unrestricted model. If the null hypothesis is true, this test statistic follows the F-distribution
with J numerator and N−K denominator degrees of freedom.
Unfortunately, Excel will not report these values automatically, so we will prepare a template
for the F-test. Let’s start with Insert>Worksheet in the Andy.xls file. Name the new worksheet
F Test. Type a heading such as "Hypothesis Testing - F-Test" in cell A1. Type Data Input
labels and Computed Values labels in column A. For appearances, left justify and set to bold
font the labels Data Input and Computed Values and right justify the sub-labels as shown
below.
In column B, we will type the formulas necessary to calculate the F-statistic, the appropriate
decision, and the p-value associated with the calculated F-statistic. The commands are similar to
those used to create the t-test template in Chapter 5. To calculate the F-statistic for a particular
test, see the formula in cell B14. The functions FINV and FDIST are used to find the F-critical
value and the p-value associated with the calculated F-statistic, respectively. The syntax of these
functions are FINV(α,df_n,df_d) and FDIST(F-stat, df_n,df_d).
To obtain the information needed in the Data Input section of the template, we need two
regressions; the unrestricted model and the restricted model. We will use the SSE's from the
ANOVA tables of each model. Now we can conduct the test for the restricted and unrestricted
models mentioned above.
Since we already have the unrestricted regression model for Andy.xls, we only need to run
the restricted model by going to Tools>Data Analysis>Regression. This time include only
advertisement (ADVERT) as the explanatory variables and save the worksheet as “Restricted
Model” and click OK.
82 Chapter 6
The regression output will be stored in the “Restricted Model” worksheet and will contain the
ANOVA table.
We can now transport the SSEs from the two models into our F-test template and test the
hypothesis. First, highlight the cell that contains the SSE from the restricted model. With cell C13
highlighted, choose Edit/Copy from the main Menu. A scrolling border appears around the cell.
The value is now stored on the Excel's clipboard and be pasted as desired. Go to the “Restricted
Model” worksheet and Edit>Paste the SSE number from the restricted model under the Data
inputs at cell B7.
Now repeat this procedure for the unrestricted SSE from the original regression and paste it in cell
B8. Fill in all the Data Input with appropriate information. Type "0.05" in cell B5 for testing at
the 5% level. The computed values should now appear, and the appropriate decision reported.
Since the F-statistic > F-critical value, and the p-value < α, we reject the null hypothesis at the
5% level and conclude the price does have an effect on total revenue.
84 Chapter 6
6.2 TESTING THE OVERALL SIGNIFICANCE OF THE MODEL
In the application of the F-test, the model significance at the desired α-level is determined.
Consider a general linear model with K regressors with K−1 explanatory variables and K
unknown parameters.
yi = β1 + xi 2β2 + xi 3β3 + .... + xik βk + ei
We test whether all of the coefficients on the K−1 explanatory variables are jointly equal to zero,
versus the alternative that at least one of coefficients is not zero. If the explanatory variables have
no effect on the average value of y, then each of the slopes will be zero, leading to the following
null and alternative hypothesis:
H 0 : β1 = β2 = β3 = .... = βk = 0
H1 : At least one of the βk is nonzero for k=2,3,…K
We can use the template created for any F-test of interest including this one. For jointly testing
the significance of all of the explanatory variables in Andy’s Burger Barn, we test that all the β's
are zero except β1, the intercept. Note that when there are NO explanatory variables in the model
other than the intercept. The SSER is equal to the SST from the unrestricted model. The results and
appropriate decision are
We reject the null hypothesis and conclude that our model is significant at the 5% level; price or
advertising expenditures, or both have a significance effect on total revenue.
Alternatively, we can obtain the result to this F-test from Excel's ANOVA table. Notice that
the F and p-values associated with the F-Test are identical.
If we compare the F-statistic to a critical value, or more easily, compare the p-value to α, we
reject the null hypothesis and conclude the model is statistically "important".
6.3 AN EXTENDED MODEL
The concept of diminishing marginal returns is an important one in economics, and you should
carefully consider this when modeling economic relationships. In our Sales for Andy’s Burger
Barn model, it seems reasonable that each and every dollar increase in advertising expenditures
would not lead to the same increase in sales; that is, the possibility of diminishing marginal
returns to advertising should be considered. To allow for this possibility, we include the
explanatory variable squared in the model. For simplicity we will rename SALES = S, PRICE = P
and ADVERT = A.
Si = β1 + β2 Pi + β3 Ai + β4 A2i + ei
To extend our model to include this additional regressor, open the worksheet Andy.xls. Label
column D "A2" and in cell D2, type the formula =C2^2. Copy this formula down the column.
Estimate the regression and save the output in the “Extended Model” worksheet.
We will get the following results:

86 Chapter 6
Price P is now insignificant at the 5% level when squared advertising expenditures (A2) are
included in the model. The estimated coefficient on A2 is negative as expected, and comparing the
p-value to (any level of) α we conclude that it is significantly different from zero. This suggests
that there are diminishing returns to advertising.
6.4 TESTING SOME ECONOMIC HYPOTHESES
Using this expanded model for total revenue, we will now examine several hypotheses of interest,
using both t-tests and F-tests.
6.4.1 The significance of advertising

To test for the significance of all levels of advertising requires a joint hypothesis test; we must
now consider the significance of both β3 and β4 using H0: β3 = 0, β4 = 0 against the alternative
that at least one of the coefficients is not zero. We already have the unrestricted model in the
previous section, we will need to estimate the restricted model, where PRICE will be the only
explanatory variable. The ANOVA results are
We can now transfer the SSEs from this restricted model and the model we labeled “Extended
Model” and use our template to test the significance of level of advertising at 5% significance
level.
We reject the null hypothesis and conclude that advertising expenditures do significantly affect
total revenue.
6.4.2 Optimal level of advertising

We have already illustrated that the returns to advertising diminish. Then the optimal level of
advertising will occur at the point where the marginal cost is equal to the marginal benefit of
advertising. In other words, the optimum level is when the next dollar spent on advertising equals
only one more dollar of sales. Taking the derivative of expected SALES with respect to
Advertising expenditures A will give us the marginal benefit, which is
ΔE ( S )
= β3 + 2β4 A = 1
ΔA
Solving for A gives A* = (1 − b3 ) 2b4 , where b3 and b4 are the least square estimates. We can then
substitute in our estimates for β3 and β4 and solve for the optimal level of advertising
1 − 12.15123398
A* = = 2.014
2(−2.767962762)
which is $2014.
Suppose Andy wants to test if the optimal level of advertising is $1,900. If we substitute 1.9
(since advertising data is in $1000), leads to the following hypothesis:
H0: β3 + 2β4(1.9) = 1 against the alternative H0: β3 + 2β4(1.9) ≠ 1

88 Chapter 6
or equivalently
H0: β3 + 3.8β4 = 1 against the alternative H0: β3 + 3.8β4 ≠ 1
A t-test could be used to test this single hypothesis, using the test statistic,
(b3 + 3.8b4 ) − 1
t=
se(b3 + 3.8b4 )
However, this test would require a calculation using the covariance between b3 and b4. Since
Excel does not report the estimated covariance matrix for the LS estimators, we will instead use
the F-test. We can construct the restricted model by plug in the restriction
( β3 + 3.8β4 = 1 ⇒ β3 = 1 − 3.8β 4 ) into our equation ( S = β1 + β2 P + β3 A + β 4 A2 ) such that the
restricted model will become:
S − A = β1 + β2 P + β4 ( A2 − 3.8 A) + e
To run the restricted model, open Andy.xls.
• Highlight column B. Insert a column by choosing Insert>Columns from the menu bar.
Label this new column S-A indicating Sales - Advertising.
• In the first empty cell of this column, type =A2−D2 and copy this formula down the
column. This column will represent our new dependent variable for the restricted model.
• Next, highlight column D and insert a new column. Label this A2-3.8A. In the first empty
cell, type =F2−(3.8*E2) and copy the formula down the column. This column will be our
new explanatory variable.
• Observe that we inserted the A2−3.8A column next to the PRICE column since in Excel,
the columns used for the X-Range (the explanatory variables) must be in adjacent
columns. So, if you ever find that you want to run a regression and the explanatory
variables are not all in adjacent columns, simply highlight and move things around as
needed. Now we are ready to run the restricted model. Choose Tools>Data
Analysis>Regression. Use S-A as the Y-Range. Use PRICE and A2−3.8A as the X-
Range. Perform this regression as usual.
The regression summary output is

This sum squared residual can now be used in the F-test template. Recall that we are testing H0:
β3 + 2β4(1.9) =1 or that $1,900 is the optimal level of advertising or not. After copying SSEs for
the restricted and unrestricted model, set J=1 since we have only a single null hypothesis. The
results from this test are
We cannot reject the hypothesis that $1,900 is the optimal level of weekly advertising
expenditures at the 5% level and conclude that Andy’s advertising strategy is compatible with the
data.
We can also conduct a joint test of two of Big Andy’s suppositions. Let's say in addition to
proposing that the optimal level of monthly advertising expenditure is $1,900, Andy is assuming
that P = 6 will yield sales revenue of $80. The joint hypothesis will be:
H 0 : β3 + 3.8β4 = 1 and β1 + 6β2 + 1.9β3 + 3.61β4 = 80
H1 : at least one hypothesis is not true

90 Chapter 6
Since J = 2, we must perform an F-test. The restricted model is found by substituting both the
hypotheses in the null on the model and rearranging terms to form a model suitable for
estimation. It can be shown that the equation used to estimate the restricted model is
S − A − 80 − 1.9 = β2 ( P − 6) + β 4 ( A2 − 3.8 A + 3.61) + e
To test this joint hypothesis in Excel, return to the worksheet containing the original data in
andy.xls.
• Create three new columns labeled YSTAR, P−6, and ASTAR.

• In the first empty cell of YSTAR, type the formula =B2−78.1, where cell B2 contains S-A.
• In the first empty cell of P−6, type =C2-6, where cell C2 contains PRICE.
• In the first empty cell of ASTAR, type =F2−(3.8*E2)+3.61, where cell F2 contains A2 and
E2 contains A.
• Highlight the three cells containing these new formulas. Place the cursor on the lower
right-hand corner of this selection until it turns into a cross-hatch. Left click, hold, and
drag down to row 79. Release and the values appear in the cells.
• Estimate a regression using YSTAR as the Y-Range. Use P−6, and ASTAR as the X-
Range.
• Use the SSE from this restricted model and SSE from the unrestricted model to conduct
the F-test, where K = 4, J = 2.
6.5 NONSAMPLE INFORMATION
Often times we have information about a particular model that does not come directly from the
data. The information may come from past experience or from economic tenets. If correct the
nonsample information improves the precision with which you can estimate the remaining
parameters. To illustrate how we might go about incorporating the nonsample information,
consider a model designed to explain the demand for beer, we will use a model of demand for
beer (Q) based on its price (PB), the price of other liquor (PL), the price of all other remaining
goods and services (PR), and income (I). The nonsample information is that consumers do not
suffer from "money illusion"; that is, when all prices and income go up by the same proportion,
there is no change in quantity demanded.
We will use a log-log functional form for the model, and then impose restrictions that
incorporate our nonsample information. The unrestricted model is
ln(Qt ) = β1 + β2 ln( PB ) + β3 ln ( PL ) + β4 ln ( PR ) + β5 ln ( I ) + e
and will impose the restriction β2 + β3 + β4 + β5 = 0.

Rearranging this restriction, we have β4 = −β2 − β3 − β5, which can be substituted into the
unrestricted model. After some manipulation, the equation we will estimate is
⎛ PB ⎞ ⎛ PLt ⎞ ⎛ It ⎞
ln(Qt ) = β1 + β2 ln ⎜ t ⎟ + β3 ln ⎜ ⎟ + β5 ln ⎜ ⎟+e
⎝ PRt ⎠ ⎝ PRt ⎠ ⎝ PRt ⎠
This equation incorporates the properties of logarithms, as well as the nonsample information.
We will get the Restricted Least Square Estimates by incorporating this restriction.
Open the data file beer.xls. Note that the original data have associated labels already.
• Label the next 3 columns PB/PR, PL/PR, and I/PR.
• For PB/PR, in cell F2, type =B2/D2. Highlight cell F2, left-click on the lower right hand
corner, hold, and drag down the next two cells to right.
• Repeat the process for PL/PR, and I/PR.
• Label the next four columns for the logs of the data, labeling them lnQ, lnPB/PR,
lnPL/PR, and lnI/PR. Calculate the natural log of Q in cell I2 by typing =LN(A2), where
cell A2 contains the first observation on Q.
• Calculate the natural log of PB/PR in cell J2 by typing =LN(F2). Highlight cell J2, left-
click on the lower right hand corner, hold, and drag across the next two cells to right.
This copies the formula for the other variables.
• Highlight the section of cells containing the log formulas (I2 to L2). Left-click on the
lower right hand corner of the selection, hold, and drag down the column to copy the
formulas down to row 31.
We can now run the regression by choosing Tools>Data Analysis>Regression to estimate the
regression model. Use lnQ as the Y-Range and lnPB/PR, lnPL/PR, and lnI/PR as the X-Range.
Include labels by checking the Labels box. Store the regression output on a new worksheet
named “Nonsample Information”.
and click OK for the output.

92 Chapter 6
Recall that our restriction was β4 = −β2 − β3 − β5. To get back the estimate of β4, we need one
more step. Unfortunately, while some statistical packages have options to automatically estimate
restricted least squares estimates, Excel does not.
In cell A21, type b4*= as a label. Next, in cell B21 type = -B18-B19-B20 to calculate b4* =
−b2*−b3*−b5*, where cells B18, B19, and B20 contain the respective estimates. The restricted
least squares estimates are
Recall that the log-log model specification provides estimates of elasticities, not marginal effects.
Substituting these results back into our specification, we have
ln(Qt ) = −4.7978 − 1.2994ln( PB ) + 0.1868ln ( PL ) + 0.1667 ln ( PR ) + 0.9458ln ( I )
From the results, we find that demand for beer is price elastic (b2 < −1), does not seem to be
affected by the price of other liquor (β3 is not statistically significant), and might be an inferior
good (β5 < 1), although this would have to be formally tested.
6.6 MODEL SPECIFICATION
Three essential features of model choice are (1) choice of functional form, (2) choice of
explanatory variables (regressors) to be included in the model, and (3) whether the multiple
regression model assumptions MR1–MR6, listed in Chapter 5, hold. In this section, we will
explore the first two.
6.6.1 Omitted variables

If you omit relevant variables from your model, then least squares estimator will be biased. To
introduce the omitted variable problem, we will consider a sample of married couples where both
husbands and wives work. Open edu_inc.xls and first regress family income (FAMINC) on both
husband’s (HE) and wife’s education (WE). The results are
Omitting wife’s education and regressing family income (FAMINC) on only husband’s (HE)
yields:
And including WE and number of preschool age children (KL6) yields:

94 Chapter 6
6.6.2 Irrelevant variables

Including irrelevant variables in your model impairs the precision of least squares estimator.
Least squares will be unbiased, but standard errors of the coefficients will be larger than
necessary. In this example, two extraneous variables (XTRA_X5 and XTRA_X6) are added to the
model. The results are:
Notice how much larger the estimated standard errors became compared to the last regression in
the previous section.
6.6.3 Choosing the model

Choosing an appropriate functional form for the model is very important. Although theoretical
considerations should be the primary guide to functional form selection, you can also use the
RESET (Regression Specification Error Test) test as a check to determine if you are making an
obvious error in specifying your function or not. RESET is basically an F-test where the
restricted model is the "original" model and the unrestricted model is a polynomial approximation
including the predicted yt's squared, cubed, etc., as explanatory variables. The general foundation
of the test is that if the model is improved by artificially including powers of the predicted values,
then the original model must not have been adequate.
RESET is a simple test with the null hypothesis is that your functional form is correct, the
alternative is that it is not. We will talk about two variants of the RESET test; RESET(1) and
RESET(2). The first adds only ŷ 2 to the model and tests its significance using the t-test. The
second adds both ŷ 2 and ŷ 3 and tests their joint significance.
Estimate the regression assuming the functional form is correct and obtain the coefficient
estimates, calculate the predicted values. We will illustrate these tests using the family income
regression where the family income is the dependent variables and the education of the husband
and wife are the explanatory variables.
FAMINC = β1 + β2 HE + β3WE + e
The unrestricted model includes the squares of the predicted y for RESET(1) and both squares
and cubes of the predicted y for RESET(2).
RESET (1) yt = β1 + β2 HEt + β3WEt + γ1 yˆt2 + e
RESET(2) yt = β1 + β2 HEt + β3WEt + γ1 yˆt2 + γ 2 yˆt3 + e
Open edu_inc.xls. First estimate the “Restricted Model” where FAMINC is the Y-Range, and
HE and WE are the X-Range. Place the output on a new worksheet called “Restricted”. Check
the Residuals box. This is needed to obtain the predicted y-values, then click OK.
From the regression output, we will use the SSE to be used for the F-test. Below the regular
regression output, you will observe the residual output. The residual option in the regression
96 Chapter 6
dialog box provides the predicted values and the errors associated with each observation. Since
we will use the squares and cubes of the predicted column, we need to copy and paste this column
next to our data. Highlight cells B25 to B453, which contain the predicted FAMINC. Choose
Edit>Copy from the menu bar. Move to the worksheet containing the original data. Go to cell G1
and Edit>Paste. Your data now looks like:
Now create two new columns, YHAT2 and YHAT3 in columns H and I. However, before we
create the square and the cube of the predicted values, we will need to make an adjustment for the
RESET test to work. We will first create a column for the adjustment; in cell H1 type adjusted
yhat. In cell H2, type =G2/10,000 and copy it down the entire column.
Now, using the adjusted yhat, we will create two new cells. Insert>Column two new columns
next to variable WE and label tem yhat2 and yhat3, respectively. Notice that the adjusted yhat cell
has moved to column J. In cell D2, type =J2^2. In cell E2, type =J2^3.
Highlight both cells D2 and E2. Copy the formulas down to row 429.
Now we have all the explanatory variables ready. Recall that when you perform a regression
in Excel, all of the explanatory variables must be in adjacent cells, that is why we inserted the
new columns next to the other explanatory variables.
For RESET(1), we will only use YHAT2.
We can now conduct the RESET(1). Since we are testing the significance of yhat2, we can either
do a t-test or and F-test using the SSE from the unrestricted model (above) and the restricted
model.
The t-statistic from the regression output of RESET(1) Unrestricted is 2.02 and the p-value is
0.044. Since t-stat > t-critical and p-value< α, we reject the null hypothesis; we have evidence to
support model misspecification. Also, since in a single hypotheses test, F = t2, F= 4.08.
For RESET(2), go to Tools>Data Analysis>Regression.
98 Chapter 6
This time use FAMINC in the Y-Range, and HE, WE, YHAT2 and YHAT3 as the X-Range. Place
the output on a new worksheet called “Reset 2 Unrestricted”.
This time we are testing the joint significance of YHAT2 and YHAT3. In other words, we test the
hypotheses H0: γ1 = γ2 = 0 versus the alternative that γ1 ≠ 0, γ2 ≠ 0, or both do not equal zero. For
the joint test, open the template labeled “F-Test”.
• Fill in the Data Input values for the F-test. J = 2, T = 428, and K = 4.
• Next, we will copy the SSEs from the restricted and unrestricted models and paste it into
the template. The SSE from the unrestricted model is cell C13 of the output above. So
right-click on cell C13 from the regression output.
• Choose Copy, place the cursor in cell B8 of the F-test template, right click and choose
Paste.
• Next, return to the “Restricted Model” regression output and copy the SSE from the
output and paste it in cell C7.
Below are the results of the F-test; we reject the null hypotheses and conclude that the
specification is inadequate.
Beware that the RESET results are different than those in your book. Although these results are
valid, EXCEL will not be able to provide you with the same RESET values despite the
adjustment we made above.
6.7 POOR DATA, COLLINEARITY AND INSIGNIFICANCE
When two or more explanatory variables are correlated, or collinear, the multiple regression
model is unable to isolate individual affects on the dependent variable. Collinearity can cause
high standard errors for the least squares estimators, resulting in t-tests that suggest the
parameters are not significantly different from zero. Some strange results can occur, and we
should be careful in interpreting our results when collinearity is present.
When there are one or more exact linear relationships between any of the explanatory
variables, the least squares estimation process does not work. Many statistical packages will not
even provide results, and will issue some type of error message. Excel does produce results,
without any warnings or error messages but issues zeros for the standard errors.
More commonly, we face situations where the collinearity is not perfect, but can be
"harmful". When linear relationships between our explanatory variables are strong enough, high
standard errors, low t-statistics, and unstable estimates result. We should, therefore, look to see if
our results are being affected by collinearity. There are several things to look at when trying to
determine the existence of this type of problem, correlation and something we call an auxiliary
regression.
To explore the ways of identifying collinearity, we will use cars.xls. Open the data set and
first estimate the model of miles per gallon (MPG) as a function of the number of cylinders (CYL)
in the engine.
100 Chapter 6
The output from this simple regression indicates a strong negative linear relationship between the
MGP and CYL.
Now add the car’s engine displacement in cubic inches (ENG) and weight (WGT) to the model.
The output now shows a very different relationship between the MPG and CYL.
The value of the coefficient of CYL is much smaller, the standard error much larger and the
variable is not significant anymore. We can also test other hypothesis such as the significance of
ENG and joint significance of CYL and ENG. We will observe that CYL and ENG are individually
not significant at 5% level but jointly they are. Joint test can be conducted by estimating the
restricted model where both of there parameters are zero and using the F-test template as
described in the previous section. This can happen when we can not isolate their individual
impact. The independent variables CYL, ENG and WGT are highly correlated with each other and
therefore highly collinear. A simple way the check the strength of possible linearity is to look at
the correlation matrix. Correlation measures the direction and strength of a linear relationship
between two variables. Choose Tools>Data Analysis>Correlation and indicate the data range in
the dialog box.
The correlation matrix is

102 Chapter 6
These results show the high degree of linearity between all of the variables. However, correlation
only measures the pair-wise linearity between variables. A more complex linear relationship
between several variables at a time is not detected by correlation.
To detect more complex linear relationships, we will use the coefficient of determination, R2,
introduced in chapter six. Recall that R2 is interpreted as the percent of the total variation in the
dependent variable that is explained by the model, or the explanatory variables. This
interpretation is very helpful now.
An auxiliary regression is a multiple regression, but one of the original explanatory variables
is used as the dependent variable. We are not concerned with any of the regression output except
the R2, because it measures how much of the variation in that one explanatory variable is
explained, or being determined by, the other explanatory variables. This is, then, a measure of
collinearity.
We can estimate the auxiliary regressions where each explanatory variable is regressed on all
the others.
Regressing ENG on CYL and WGT
Regressing CYL on WGT and ENG

Regressing WGT on ENG and CYL
Check the R2 from each of these regressions. Any above 80% indicates strong collinearity which
may adversely affect the precision with which you can estimate the parameters of the model
containing all of these correlated variables. In our example, the R2s are approximately 94%, 90%
and 87% respectively which are all well above the 80% threshold. Therefore, it is not surprising
that it is difficult to isolate the individual contributions of displacement and number of cylinders
to a car’s gas mileage.
CHAPTER 7
Nonlinear Relationships
CHAPTER OUTLINE
7.1 Nonlinear Relationships 7.3 Applying Dummy Variables
7.1.1 Summarize data and estimate regression 7.3.1 Interactions between qualitative factors
7.1.2 Calculating a marginal effect 7.3.2 Adding regional dummy variables
7.2 Dummy Variables 7.3.3 Testing the equivalence of two
7.2.1 Creating dummy variables regressions
7.2.2 Estimating a dummy variable regression 7.4 Interactions Between Continuous Variables
7.2.3 Testing the significance of the dummy 7.5 Dummy Variables in Log-linear Models
variables
7.2.4 Further calculations
7.1 NONLINEAR RELATIONSHIPS
The least squares estimation procedure we have been using is based on the assumption that the
model is linear in the parameters, though not necessarily linear in the variables. We saw an
example of nonlinearity in variable in Chapter 6 in the sales model with diminishing marginal
returns to advertising expenditures. To allow for this effect, we included the square of advertising
expenditures as another explanatory variable. By transforming the advertisement variable, we
captured the diminishing marginal returns without violating the assumptions of the linear
regression model.
Models that have parameters that are nonlinear require nonlinear least squares estimation.
Although Excel is a powerful spreadsheet, it is not designed to be a complete econometric
software package, and consequently it does not have the capabilities to estimate models that are
nonlinear in the parameters. If you encounter such a problem, use econometric software such as
Stata, EViews, Shazam, or SAS.
7.1.1 Summarize data and estimate regression

The following example will illustrate the flexibility that the polynomial terms can add to the
linear regression model. We will use the wage equation:
104
Nonlinear Relationships 105
WAGE = β1 + β2 EDUC + β3 EXPER + β4 EXPER 2 + e
Get the summary statistics, Tools>Data Analysis>Descriptive Statistics.
This will provide the summary statistics for WAGE, EDUC, EXPER and EXPER2 variables.
To estimate the wage equation, open cps_small.xls. Highlight column D and insert a column
using Insert>Column. Name the new column “EXPER2” and enter the formula =C2^2 and copy
it down.
106 Chapter 7
Estimate the model using Tools>Data Analysis>Regression.
Estimate a regression and use WAGE as the Y-Range and EDUC, EXPER, EXPER2 as the X-
Range. Name the output worksheet “Polynomial”.
7.1.2 Calculating a marginal effect

Since the wage equation is nonlinear in variable “EXPER”, the marginal effect (slope) must be
calculated as follows:
∂E (WAGE )
= β3 + β4 2 EXPER
∂EXPER
The marginal effect needs to be evaluated at a specific point, such as the median. You can get the
median from the summary statistics. The median for the EXPER variable is 18 from the output
above. The marginal effect at the median is:
∂E (WAGE )
= 0.340949174 − 0.005093062*2 *18 = 0.157598937
∂EXPER
7.2 DUMMY VARIABLES
Dummy variables are binary (or indicator) variables that indicate the presence or absence of a
characteristic. In this section, we will use dummy variables in a real estate example. Open
utown.xls.
7.2.1 Creating dummy variables

In many examples in POE dummy variables have already been created and are ready to use. An
important issue in the real estate industry is how to accurately predict the price of a house, based
on several of its characteristics, including the ever-important "location, location, location".
Economists commonly use a "hedonic" model of pricing based on several characteristics such as
size, location, number of bedrooms, age, etc. Using a dummy variable, Dt, which is equal to 1 if
the house is in a desirable neighborhood and is equal to 0 if the house is not in a desirable
neighborhood captures the qualitative factor of location. Including this variable in the regression
model will allow the intercept to be different for houses in desirable areas compared to the
intercept for houses not in desirable areas. This variable has been stored as variable UTOWN.
We can also allow for different slopes for houses in different areas by including an
interaction variable, the product of the dummy variable and one of the continuous explanatory
variables.
7.2.2 Estimating a dummy variable regression

Estimating a dummy variables model is no different than estimating any other regression model.
We will use the interaction slope dummy variable between the size of the house sqft and the
dummy variable for university town, utown. This will allow for the extra square footage of living
space in a good neighborhood affecting the price differently than a house not in a good
neighborhood. The full model we will estimate is
PRICE = β1 + δ1UTOWN + β2 SQFT + γ ( SQFT × UTOWN )
+ β3 AGE + δ 2 POOL + δ3 FPLACE + e
where PRICE = the price of the house, in dollars

UTOWN = 1 for houses near the university (desirable), 0 otherwise
SQFT = square feet of living area
AGE = age of house in years
POOL = 1 if house has a pool, 0 otherwise
FPLACE = 1 if house has a fireplace, 0 otherwise.
Note that this model contains two continuous explanatory variables (SQFT and AGE), and three
dummy variables, capturing the qualitative characteristics of location; presence of a pool, and a
108 Chapter 7
fireplace in addition to the utown variable. Let's now estimate this model. Open utown.xls. Label
column G sqftXutown. In cell G2 create the interaction variable by typing =B2*D2.
Copy this formula down the column to row 1001.
Zeros appear in column G down to row 482. Then sqft values appear after that. This is because
the variable UTOWN is equal to zero through row 482, then one after that.
Estimate a regression using Tools>Data Analysis>Regression, using column A as the Y-
Range and columns B through G as the X-Range. Don’t forget the include labels by checking the
Labels box. Save to a new worksheet called Dummy and Interaction.
The parameter estimates and corresponding t-stats and p-values are

7.2.3 Testing the significance of the dummy variables

To test the significance of the University Town location, we test the individual null hypothesis
using the t-test. From the output above, we conclude that all of the parameters are significant,
using a one-tailed 5% level test. (Remember that Excel reports p-values for a two-tailed test). To
conclude that the δ1 is statistically different from zero means that there is a shift in the intercept
for houses near the university. Similarly, concluding that γ is different from zero means that the
marginal effect of the size of the house is different for houses near the university. But Excel
doesn't know that we have allowed the intercept and the coefficient on SQFT to differ across
observations. It is our responsibility to correctly determine the intercept and slope estimates.
Looking at the original model that we estimated
Alternatively, we can test the significance of the location by testing the joint null hypothesis
H 0 : δ1 = 0, γ1 = 0 against the alternative that at least one coefficient is not zero. To construct the
F-test, we will run the following restricted model:
PRICE = β1 + β2 SQFT + β3 AGE + δ 2 POOL + δ3 FPLACE + e
And compare the SSEs from the restricted model and the unrestricted model (labeled Dummy
and Interaction) using the F-test template. Since we are testing two hypothesis, J = 2 with the
other input data of N = 1,000 and K = 7.
7.2.4 Further calculations

When the UTOWN is equal to one, the intercept is β1 + δ1 and when UTOWN is zero the constant
is simply β1. Similarly, the coefficient on SQFT when UTOWN is one is equal to β2 + γ, and is
equal to β2 when UTOWN is zero. We can calculate the estimates for these parameters on the
Excel regression output worksheet.
In cell A26, type the label NewIntercept and in cell A27, type the label New Beta2.In cell
B26, type =B17+B20 to calculate the intercept when UTOWN is equal to one and type =B18+B23
to calculate the coefficient estimate for SQFT when UTOWN is equal to one in cell B27.
110 Chapter 7
The estimated regression functions for the houses near the university is
ˆ = 24.5 + 27.453 (1) + 7.6122SQFT (1) + 1.2994SQFT +

PRICE 1
−0.1901AGE + 4.3772POOL + 1.6492FPLACE
ˆ = 51.953 + 8.9116SQFT + −0.1901AGE + 4.3772POOL + 1.6492FPLACE

PRICE
7.3 APPLYING DUMMY VARIABLES
In this section we will illustrate a variety of applications of dummy variables using cps_small.xls.
7.3.1 Interactions between qualitative factors

First we will consider the interaction between two dummy variables, black and female in the
following model:
WAGE = β1 + β2 EDUC + δ1 BLACK + δ 2 FEMALE + γ ( BLACK × FEMALE ) + e
Open cps_small.xls. Label column K blackXfemale. In cell K2 create the interaction variable by
typing =F2*E2.
Copy this formula down the column to row 1001.

Recall that you need the X-Range variables next to each other. Insert columns and organize the
columns by copying and pasting.
Estimate a regression using Tools>Data Analysis>Regression, using column A as the Y-Range

and columns B through E as the X-Range. Don’t forget the include labels by checking the Labels
box. Save to a new worksheet called “Interactions btw dummies”.
112 Chapter 7
The output will be:
To test the joint hypothesis H 0 : δ1 = 0, δ2 = 0, γ1 = 0 , we will need to estimate the restricted model
and carry out an F-Test. Estimate the Restricted model, assuming the null hypothesis is correct.
Estimate a regression using Tools>Data Analysis>Regression, using column A as the Y-Range

and columns B only as the X-Range. Save to a new worksheet called “Restricted.”
We can now open the F-test template and fill in the Input data.
114 Chapter 7
7.3.2 Adding regional dummy variables

Next, we can add dummy variables with several categories such as the regional dummies. Our
model becomes:
WAGE = β1 + β2 EDUC + δ1 SOUTH + δ 2 MIDWEST + δ3WEST
+δ4 BLACK + δ5 FEMALE + γ1 ( BLACK × FEMALE ) + e
Since the dummies are already present in the file, we simply estimate the model using
Tools>Data Analysis>Regression, after arranging the columns for the X-Range variables.
Use column A as the Y-Range and columns B through H as the X-Range. Save to a new
worksheet called “Regions.”
The output will provide us with the unrestricted model to test the joint hypothesis
of H 0 : δ1 = 0, δ2 = 0, δ3 = 0 .
116 Chapter 7
Under the null hypothesis, the restricted model is the full (unrestricted) model from the previous
section. If we fill in the Data Inputs in the F-Test template for this test, we fail to reject the null
hypothesis at 5% level.
7.3.3 Testing the equivalence of two regressions

To test the equivalence of the wage equations for the south region versus the reminder of the
country, we create an interaction variable for each variable in the regression model with the
dummy variable south. In other words, our equation
WAGE = β1 + β2 EDUC + δ1 BLACK + δ 2 FEMALE + γ1 ( BLACK × FEMALE ) + e
becomes
WAGE = β1 + β2 EDUC + δ1 BLACK + δ 2 FEMALE + γ1 ( BLACK × FEMALE ) + θ1SOUTH
+ θ2 ( EDUC × SOUTH ) + θ3 ( BLACK × SOUTH ) + θ4 ( FEMALE × SOUTH )
+ θ5 ( BLACK × FEMALE × SOUTH ) + e
First we have to create the interaction variables as explained in the previous section.
After labeling and writing the formulas, copy the formulas for all 1,000 observations.
Once the interaction variables are created, we estimate the model by Tools>Data
Analysis>Regression, using column A as the Y-Range and columns B through J as the X-
Range. Save to a new worksheet called “FULL.”
The results will give us the unrestricted model results.

118 Chapter 7
To test the hypothesis that there is no difference between the model for the south and the rest of
the nation, we have to test the joint hypothesis H 0 : θ1 = θ2 = θ3 = θ4 = θ5 = 0 against the
alternative that at least one of the five hypotheses is not true. Under the null hypothesis, the
restricted model will be
If we input the Data into the F-Test template, we will get:

7.4 INTERACTIONS BETWEEN CONTINUOUS VARIABLES
When we include the product of two continuous explanatory variables in a model, we alter the
relationship between each of them and the dependent variable. Reporting and interpreting the
results require care.
The model we use here is based on a "life-cycle" model of the effects of age and income on a
person's expenditures on pizza. We believe that as a person ages, the marginal effect of income
will probably change (the marginal propensity to spend on pizza probably falls). Since we assume
that the effect of income depends on age, we include an interaction variable that is the product of
these two variables. The model we will estimate is
PIZZA = β1 + β2 AGE + β3 INCOME + β 4 ( INCOME × AGE ) + e
where PIZZA = individual’s annual expenditure on pizza, in dollars

AGE = the age of the individual in years
Y = the annual income of the individual, in dollars
Open pizza.xls to estimate the above model and create the interaction variable after arranging the
order of the explanatory variables.
120 Chapter 7
Copy the formula to all 40 cells.
Estimate the model.

The effect of age and income are no longer given by their estimated coefficients. Instead we must
calculate the marginal effects.
To find the marginal effects of age and income, we take the first derivative of the pizza
function, with respect to the variable of interest. We find that the marginal effect of age on pizza
expenditures is b2 + b4Y and the effect of income is b3 + b4AGE. These will be estimated and
calculated using Excel. In cell A21, type the label AGE and in cell A22, type the label INCOME.
In cell A24, type the label age effect and in cell A25, type the label income effect. In cell B24,
type =$B$16+(B18*B22) and in cell B25, type =$B$17+($B$18*B21). This template can now be
used to calculate the marginal effects, given different levels of age and/or income. Below is an
illustration of two examples.
The formulas provide us with the marginal effect of AGE for incomes 90,00 and 25,000 and
marginal effect of INCOME at age 30 and 50.
7.5 DUMMY VARIABLES IN LOG-LINEAR MODELS
Consider the model:
ln(WAGE ) = β1 + β2 EDUC + γ1 FEMALE + e
The calculation of the exact effect of a dummy variable in log-linear model is slightly more
complicated. We will use cps_small.xls data to illustrate the calculations. Label cell D1, ln(wage)
and type the formula in D2 =ln(A2). Copy the formula down all the cells.
122 Chapter 7
Estimate the log-linear wage equation by using ln(wage) as the Y-Range and EDUC and
FEMALE as the X-Range variables.
The results are:

We can calculate the exact effect of the female dummy given the output as:
The marginal effect will be a nonlinear function of the parameters:
Similarly, we can calculate other nonlinear marginal effects when we include interaction terms.
Estimate the following model:
ln(WAGE ) = β1 + β2 EDUC + β3 EXPER + γ1 ( EDUC × EXPER) + e
First, we need to create the interaction variable and estimate this log-linear model using the
ln(wage) as Y-Range and EDUC, EXPER and educXexper as the X-Range variables.
Given the coefficient estimates from the regression, we can calculate the approximate marginal
effect of experience using
124 Chapter 7
The marginal effect will be again a nonlinear function of the parameters for education of 16
years.
CHAPTER 8
Heteroskedasticity
Chapter Outline
8.1 The Nature of Heteroskedasticity 8.4 Detecting Heteroskedasticity
8.2 Using the Least Squares Estimator 8.4.1 Residual plots
8.3 The Generalized Least Squares Estimator 8.4.2 The Goldfeld-Quandt test
8.3.1 Transforming the model 8.4.3 Testing the variance function
8.3.2 Estimating the variance function
8.3.3 A heteroskedastic partition
8.1 THE NATURE OF HETEROSKEDASTICITY
In simple and multiple linear regression models of earlier chapters, we had assumed all the
assumptions of the Classical Linear Regression (CLRM) model have been met.
yi = β1 + β2 xi + ei
where yi is the dependent variable, xi is the ith observation on the independent variable, β1 and β2
are the unknown parameters and ei is the random error. The error assumptions of CLRM are:
E ( ei ) = 0 Var ( ei ) = σ2 Cov ( ei ) = 0
One of the above mentioned assumptions of the linear regression model is that the variance of the
error term and of the dependent variable is constant across all of the observations. If this
assumption is not met, OLS estimator is still unbiased and consistent but the least square standard
errors and the tests based on these standard errors are neither consistent nor asymptotically valid.
125
126 Chapter 8
8.2 USING THE LEAST SQUARES ESTIMATOR
In Chapter 3, we introduced a model of food expenditures as a function of income. At each

income level, a household chooses its level of food expenditures. It seems reasonable that
households with higher incomes have more choices, anywhere from cheap and simple to
extravagant food. This suggests that the variance for high-income households is greater than that
of low-income households. The consequences of estimating the parameters using the least square
estimator, if heteroskedasticity is present, are that the standard errors that Excel reports are
wrong. Therefore, the t-statistics, p-values, and confidence intervals are also wrong. Let's return
to that model and our Excel results to reconsider our assumptions about the variance of the
model.
Open food.xls. Estimate the regression, using food expenditures as the Y-Range and income
as the X-Range. Choose Line Fit Plots under the Residuals option to produce a chart of the
estimated regression line. After some formatting, the graph should look like
Food Expenditures Regression

y = 83.4160x + 10.2096
300
Food Expenditures
250
200
150
100
50
0
0 500 1000 1500
As income increases, the data points are scattered further and further from the estimated
regression line. The residuals, the vertical distances between the regression line and the individual
observations, are getting larger as income increases. The graph above, therefore, suggests that σ2
is not constant, but is greater for larger income households indicating heteroskedasticity.
Since, with heteroskedasticity, the reported standard errors are incorrect, we need a method for
determining the correct standard errors for our least squares estimators. We can then recalculate
our t-stats and confidence intervals with these heteroskedasticity adjusted standard errors.
One such adjustment is White's Heteroskedasticity Consistent Standard Errors, which for
the simple regression model is.
Heteroskedasticity 127
∑ ⎡⎣( xt − x )eˆt2 ⎤
2
ˆ b2 ) =
var( ⎦
2 2
⎡ ∑ ( xt − x ) ⎤
⎣ ⎦
The square root of this is the estimated standard error for b2. Some statistical packages will
calculate White's standard errors automatically. Excel does not, but we can use the standard Excel
functions to calculate it "by hand".
Return to the data contained in food.xls. Estimate another regression, choosing the Residuals
option.
• From the regression output, Copy the residuals from the residual output to the worksheet
containing the original data and Paste these residuals in column C. Include the label.
• Label column D “ehat^2” and square the values in column C here by typing =C^2 in cell
D2 and copy the formula down the column.
• Label column E “xbar”. In cell E2 type =AVERAGE(B2:B41) where column B contains
the income.
• Label column F “(x-xbar)^2”. In cell F2 type =(B2-$E$2)^2. Recall that the dollar sign
anchors the cell containing “xbar”.
• Label column G “numerator”. In cell G2 type =F2*D2. Highlight cells F2 and G2, and
copy these formulas down the columns.
• In cell F42, type =SUM(F2:F41) to sum the column;

• in cell G42, type =SUM(G2:G41).
• Label B44, White's var(b2), and B45 White's se(b2).
• In cell C44, type =G42/(F42^2)
• and in cell C45, type =SQRT(C44).
The results are

128 Chapter 8
Excel's regression output reports the standard error for b2 as 2.093263531 which is incorrect due
to the existence of heteroskedasticity. White's standard error is actually different so are the
corrected t-stat and confidence intervals. While it usually the case that the corrected standard
errors, it is not always true as you can see from this example. You should recalculate and report
the corrected t-stat and confidence interval, do not report those produced in the regression output.
8.3 THE GENERALIZED LEAST SQUARES ESTIMATOR
Since the least squares is inefficient in heteroskedastic models, we may wish to use Generalized
Least Square (GLS) which is the Best Linear Unbiased Estimator. GLS estimator works by
transforming the model into a homoskedastic one and applying OLS to the transformed model.
8.3.1 Transforming the model

Since we have var ( ei ) = σi2 , we can get constant error variance by dividing ei by σi . To
transform the model, we will weigh the observations using σi . For the food expenditure model,
yi 1 x e
= β1 + β2 i + i
σi σi σi σi
We saw in the graph of the regression that the estimated error terms seem to get larger as income
increases. So we can assume the variance of the model is proportional to income and can be
modeled as var(ei ) = σi2 xi . It can be shown that, if we transform our original data by dividing all
observations by the square root of xi, the new, transformed model is homoskedastic and we can
estimate the new model using least squares where the t-stats and confidence intervals are correct.
To estimate the food expenditure model, where we want to weight the data by 1/ xi . Go
back to the worksheet containing the original data on food expenditures and income.
• Label column C SQRT(x) and columns D, E and F, INT*, X* and Y* respectively.

• In cell C2, type =SQRT(B2), where cell B2 contains the first observation on income.
• In cell D2, type =1/C2. This creates a new intercept term, not equal to one anymore.
• In cell E2, type =B2/C2.
• In cell F2, type =A2/C2.
• Highlight cells C2 through F2. Copy the formulas down the column.
Estimate a regression, using Y* as the Y-Range and INT* and X* as the X-Range.
• Include labels and check the Labels box.

• Check the Constant is Zero box since we now have our new, transformed intercept term.
• Place the output on the worksheet named GLS and click OK.
The regression results are:
The estimates, b1 and b2, differ from the original regression results. However, the interpretations
are the same. Transforming the data in the manner we did changed a heteroskedastic model to a
homoskedastic model; not the meanings of our estimates. The GLS standard errors are lower than
those calculated using White's approximation. This is to be expected because the GLS procedure
is more efficient and provides smaller standard errors, higher t-stats and narrower confidence
intervals.
130 Chapter 8
8.3.2 Estimating the variance function

In the above example, the observation’s standard error (or what it is proportional to) is known. In
most cases this information will not be known and we will have to estimate it. This turns the GLS
estimator into feasible GLS (FGLS).
The first step is to choose a model for variance that is some function of the independent
variables. A common model of the variance uses the exponential function:
σ 2i = exp ( α1 + α 2 zi 2 + ... + α s zis )
where the zis are independent variables and α ’s are the unknown parameters. Taking the natural
logarithm, substituting the least squares residuals for the unobservable σi2 , and adding an error
term gives you a regression model that can be estimated for α i .
ln ( eî2 ) = ln ( σi2 ) + vi = α i + α 2 zi + vi
where the eî2 are from least squares estimation of the original heteroskedastic regression model.
Let zi = log(income) .
To estimate this model using the food expenditure data, go back to the worksheet containing
the original data on food expenditures and income.
• Label column C log(income) and type =ln(B2) in cell C2.

• Copy and paste the residuals from the initial regression to column D.
• Label columns E and F, resid2 and log(resid2), respectively.
• In cell E2, type =D2^2, where cell D2 contains the first observation on residuals.
• In cell F2, type =LN(E2).
• Highlight cells E2 and F2 and copy the formulas down the column.
Estimate a regression, using log(resid2) as the Y-Range and log(income) as the X-Range.
Include labels and check the Labels box. Place the output on the worksheet named FGLS and
click OK. The regression results are:
8.3.3 A heteroskedastic partition

To illustrate a model with a heteroskedastic partition, we use a model of wages as a function of
education and experience. In addition a dummy variable is included that is equal to one if the
individual lives in a metropolitan area. This is an intercept dummy variable indicating people
living in metropolitan areas make higher wages relative to those living in rural with similar level
of experience and education.
First, open cps2.xls and highlight columns D through I and Edit>Delete the columns.
wagei = β1 + β2 educ + β3exper + β 4 metroi + ei
Estimate the model using Tools>Data Analysis>Regression with WAGE as the Y-Range and
EDUC, EXPER and METRO as the X-Range.
Next, estimate each subset (metro and rural) separately using least squares and save the standard
errors of the regression.
• Highlight all the columns of the data, including labels, choose Data>Sort from the menu
bar.
• Click on the down-arrow in the Sort By box. Choose METRO.
• Check either Ascending or Descending option.
• Check the Header Row option, since we included the labels and click OK.
132 Chapter 8
Looking at the METRO column, you will observe that the column will contain zeros till row 194
and after that row, METRO will be one.
Now we must estimate two regressions, using the first half of the data in the first one (METRO =
0), and the second half of the data in the second regression (METRO = 1). The only output we're
interested in from these regressions is the estimated variance of the model.
Estimate a regression on the data, using cells A2 through A193 for the Y-Range and cells B2
through C193 for the X-Range. DO NOT include labels and don’t include the METRO variable
in column D. Place the output on a worksheet named “regression 1.”
Repeat the procedures above, but now include cells A194 through A1001 as the Y-Range and
cells B194 through C1001 as the X-Range. Save output to a worksheet named “regression 2.”
In order to obtain a homoskedastic model, we will transform the data by using the estimated
variances from each partition. The estimated variances of the models are what we are interested
in; no other output from the regressions is important at this point. We divide the data on the first
193 observations by the square root of the estimated variance from a regression using just those
observations, and divide the last 808 observations by the square root of the estimated variance
obtained from the regression using just these data. Then we "pool" the transformed data and
perform generalized least squares estimation to obtain the correct estimates.
The simplest thing to do at this point is to simply write down the values of the MS Residuals
from the ANOVA tables of the regressions; metro and rural. From metro, the value is
31.8237318; from rural, it is 15.24298659. Transform the first 193 observations using the square
root of the estimated variance from rural and the rest of the observations using the square root of
134 Chapter 8
the estimated variance from metro as shown in Section 8.3.1 of this manual. Estimate a
regression, using wage* as the Y-Range and int*, edu*, exper* and metro* as the X-Range
where the stars represent the transformed values for the intercept (int), years of education (edu),
years of experience (exper) and dummy variable for metropolitan area (metro), respectively.
Include all 1000 observations. Include labels and suppress the intercept by checking the Constant
is Zero box.
8.4 DETECTING HETEROSKEDASTICITY
8.4.1 Residual plots

If the regression errors are homoskedastic, when we plot the residuals, there should be no
systematic pattern evident. If the errors are heteroskedastic, we may be able to detect a particular
pattern in the residuals and perhaps even discover the form of the heteroskedasticity.
Let’s use the wage model to plot the least square residuals against the metro dummy. Open
cps.xls. Delete all the columns except wage, educ, exper, metro and save the file as
cps_modified.xls. Estimate a regression, wage as the Y-Range and educ, exper and metro as the
X-Range. Under the Residuals options, choose Residual Plots. This will produce a graph of the
residuals, plotted against each of the explanatory variables.
Let’s look at the plot against the metro dummy variable. After formatting, it should look similar
to this.
In this plot, the least squares residuals are grouped according to rural (metro=0) or metropolitan
(metro=1). Wider variation of residuals at metro=1 indicates higher variance for these
observations and evidence of groupwise heteroskedasticity.
8.4.2 The Goldfeld-Quandt test

Although plots are helpful in diagnosing heteroskedasticity, the Goldfeld-Quandt formally tests
for equal variances. It is a type of F-test and the steps are to order the data based on variance, split
the data into parts, compute the estimated variances, calculate the test statistic GQ = σˆ 12 σˆ 22 , and
compare to an F-critical value, based on T1-K numerator and T2-K denominator degrees of
freedom. If heteroskedasticity is present, the GQ test statistic should be large, and we would
reject a null hypothesis of equal variances.
Returning to the wage model, where metro F is 31.824 and rural F is 15.243. Open a new
worksheet and name it “GQ test”. Create the following template to use for any Goldfeld-Quandt
Test.
136 Chapter 8
Fill in the Input Data as T1=808 and T2=193, K=2, and Alpha should be .05, for testing at the
5% level. From the metro worksheet, highlight cell D13 (the MS residual from the ANOVA
table), right-click and choose Copy. Go back to GQ test worksheet and Paste the value into cell
B6. Remember to always place the larger of the estimated variances in the numerator of the
formula for the GQ test statistic. Copy cell D13 from the rural worksheet to cell B7 of GQ test.
The resulting template is
Let's test for heteroskedasticity in the food expenditures model. Formally, we test the null
hypothesis H0: σi2 = σ 2 against the alternative H1: σi2 = σ 2 xi .
Since the variance is an increasing function of income we have to sort the data in an
ascending order by income. Open the original data on food expenditures and income, food.xls.
• Highlight both columns of the data, including labels, choose Data>Sort from the menu
bar.
• Click on the down-arrow in the Sort By box. Choose income.
• Check the Ascending option;
• Check the Header Row option, since we included the labels and click OK.
Looking at the column containing income, the numbers should be increasing as you look down
the column.
Now we must estimate two regressions, using the first part of the data in the first one, and the
second part of the data in the second regression. Because there is no natural dividing point, we
will break the ordered sample into two equal halves. The only output we're interested in from
these regressions is the estimated variance of the model.
Estimate a regression on the data, using cells A2 through A21 for the Y-Range and cells B2
through B21 for the X-Range. Do NOT include labels. Place the output on a worksheet named
“regression 1.”
Repeat the procedures above, but now include cells A22 through A41 as the Y-Range and cells
B22 through B1 as the X-Range. Save output to a worksheet named “regression 2.”
138 Chapter 8
Fill in the Input data. T1 and T2 both equal 20, K=2 and Alpha should be .05, for testing at the
5% level.
• From the regression 1 worksheet, highlight cell D13 (the MS residual from the ANOVA
table), right-click and choose Copy.
• Go back to GQ test worksheet and Paste the value into cell B7. Remember to always
place the larger of the estimated variances in the numerator of the formula for the GQ test
statistic.
• Copy cell D13 from the regression 2 worksheet to cell B6 of GQ test.
The resulting template is
We reject the null hypothesis and conclude that heteroskedasticity IS present. If we assume
proportional heteroskedasticity, we would proceed as in section 11.3. If we couldn't assume any
particular form of the heteroskedasticity, then we should at least calculate White's standard errors
and report the corrected t-statistics and confidence intervals.
8.4.3 Testing the variance function

There are many other tests to test for the existence of heteroskedasticity. In this section, two more
of these tests will be discussed; Breusch-Pagan and White’s tests. In both of the tests the null is
homoskedastic errors and alternative is heteroskedastic errors. Both of the tests also require
auxiliary regressions.
First, we will illustrate the Breusch-Pagan test using the food expenditure data. Open the
original series, food.xls and estimate the regression, click the residual option to save the residuals.
Then, copy and paste the residuals to column C in the data sheet. Label D1 “resid2” and type
=C2^2 in D2.
Next, regress residuals on income.
The test statistic (LM) is calculated by multiplying the N*R2 and has a χ 2s−1 distribution.
140 Chapter 8
White’s test is just a minor modification to Breush-Pagan test. Add one more column to the data
and label it income2. This column will contain the squared income variable. After putting in the
formula and copying it down, regress resid2 on income and income2. Recall that you need to get
the income and income2 in adjacent columns to be able to run the regression. The test statistic is
again the product of N and R2 which has a χ 2s−1 distribution.
The results for both of the tests indicate a rejection of the null hypothesis of no heteroskedasticity.
CHAPTER 9
Dynamic Models, Autocorrelation,
and Forecasting
CHAPTER OUTLINE
9.1 Lags in the Error Term 9.4 Detecting Autocorrelation
9.2 Area Response for Sugar 9.4.1 The Durbin-Watson test
9.3 Estimating an AR(1) Model 9.4.2 An LM test
9.3.1 Least squares 9.5 Autoregressive Models
9.6 Finite Distributed Lags
9.7 ARDL Model
9.1 LAGS IN THE ERROR TERM
The multiple linear regression model of Chapters 5 and 6 assumed that the observations are not
correlated with one another. This assumption is not realistic if the observations are drawn
sequentially in time. With times-series data, where the observations follow a natural ordering
through time, there is a possibility that successive errors are correlated with each other. Shocks to
a model may take time to work out and effects may carry over to successive time periods. The
result is that the error term in period t can affect the error term in period t+1, or t+2, and so on.
Somehow, we must take these lasting effects into account.
In the first example the supply response for an agricultural crop is modeled as a log-log linear
model where area planted (acres) depends on price. The first dynamic model we will consider is
one with a lag in the error term.
ln( At ) = β1 + β2 ln( Pt ) + et
et = ρet −1 + vt
142
Dynamic Models, Autocorrelation and Forecasting 143
Where ρ (rho) is the parameter that describes the relation between et and et −1 , vt is the new
random error term. For stability (and stationarity) of the model, −1 < ρ < 1 .
9.2 AREA RESPONSE FOR SUGAR
First, we will estimate the above mentioned model using least squares. Open bangla.xls.
• Label columns C and D as LN(P) and LN(A) respectively.

• In cell C2, type =LN(A2), copy this formula to cell D2.
• Highlight cells C2 and D2, and copy the formulas down the columns.
Estimate a regression, using LN(A) as the Y-Range and LN(P) as the X-Range. Include labels.
Check the Residuals option so that the estimated errors are produced and click OK.
The least squares estimates are:

144 Chapter 9
Since we are using times-series data, we should explore the possibility of autocorrelation.
Visually, we can plot the residuals against time and see if we can detect any patterns.
Return to the worksheet containing the original data.
• Label cell E1 t, for time.

• Type “1” in cell E2.
• Type =E1+1 in cell E3.
• Highlight cells E2 and E3, place cursor on the lower right hand corner of the highlighted
area until it turns into a cross-hatch.
• Left-click and drag down the column to fill in the values in ascending order.
Return to the worksheet containing the regression output.
• Copy residuals to the worksheet containing the original data and Paste in column F.
• Create an XY Scatter graph with t on the horizontal axis and Residuals on the vertical
axis.
The results will look like

Looking at the values of the residuals (ehat) and at the graph above, there seems to be a tendency
for negative values to follow negative values, and positive values to follow positive values. This
is consistent with positive correlation between successive terms. While such a conclusion about
autocorrelation is subjective, we will later look at a more formal test. For now, however, it does
appear that there is a problem.
We can also check the correlation between the ehat and the lagged values. Rename the
residual column ehat, create a new column and call is ehat_1. Lag the ehat column by copying
F2-F35 and paste it to G3. Delete F2 and G36 to make the columns even. Go to Tools>Data
Analysis>Correlation. Put ehat and ehat_1 in the range and hit OK. The output indicates about
.40 correlation between the errors one period apart.
9.3 ESTIMATING AN AR(1) MODEL
When the errors follow and AR(1) model, et = ρet −1 + vt , the least squares assumption 4 is
violated. Least squares is unbiased and consistent, but no longer efficient. The reported standard
errors are no longer correct, leading to statistically invalid hypothesis tests and confidence
intervals.
9.3.1 Least Squares

In the previous chapter, we transformed our data so that we could move from a heteroskedastic
model to a homoskedastic. The same type of procedure can be used to correct for first order
autoregressive errors AR(1). Our objective is to transform the model
yt = β1 + β2 xt + et where et = ρet −1 + vt
such that the autocorrelated term et is replaced by the uncorrelated error term vt. After some
substitution and rearranging, the transformed model we obtain is
yt − ρyt −1 = β1 (1 − ρ) + β 2 ( xt − ρxt −1 ) + vt
All we need to do is to transform the dependent variable, intercept, and explanatory variable as
above and proceed with the generalized least squares estimation. The first problem is ρ is
unknown and must be estimated and observe that we now have T-1 observations since we lose the
"first" one.
Estimate the original model and store the residuals. Since
146 Chapter 9
T
∑ eˆt eˆt −1
ρˆ = t =2
T
∑ eˆt2−1
t =2
and we already have our residuals from the least squares estimation, we're ready to go!
Return to the worksheet containing the original data and the residuals. Label cells G1, H1,
and I1 as sum(et*et-1), ssq(et-1), and rhohat respectively. In cell G2, type
=SUMPRODUCT(F3:F35,F2:F34)
This corresponds to the numerator in the formula above. In cell H2, type
=SUMSQ(F2:F34)
This calculates the denominator. Finally, in cell I2, divide G2 by H2 by typing =G2/H2. The
result is 0.3992.
Now, to deal with the issue of transforming the first observation for the transformed model, we
have y1 = β1 + x1β2 + e1 with an error variance of var(e1)= σe2 = σv2 / (1 − ρ2 ) . The transformation
that gets to where we want (a variance of σv2 ) is multiplication of the terms in the model, for the
first observation, by 1− ρ2 .
Use the worksheet with columns P, A, LN(P), LN(A). Label columns E, F, and G as y*, int*
and x* respectively.
• In cell E2, type =SQRT(1-(0.3992^2))*D2.

• In cell F2, type =SQRT(1-(0.3992^2)).
• In cell G2, type =SQRT(1-(0.3992^2))*C2.
The first observation for the transformed model is now complete. For the remaining observations:
• In cell E3, type =D3-(0.3992*D2).

• In cell F3, type =1-0.3992.
• In cell G3, type =C3-(0.3992*C2).
• Highlight cells E3 through G3. Copy the formulas down the columns.
Run a regression, using y* as the Y-Range and int* and x* as the X-Range. Include labels as
usual AND suppress the intercept by checking the Constant is Zero box. Place output on a new
worksheet names GLS and click OK.
The generalized least squares results are:
Once again, interpretations of the estimates are as usual. The price elasticity of sugar cane area
response seems to be one.
148 Chapter 9
9.4 DETECTING AUTOCORRELATION
We will consider two formal tests to test for the existence and extent of autocorrelation; Durbin
Watson and LM (Lagrange Multiplier) serial correlation tests. Both tests test the hypothesis H0: ρ
= 0 versus the alternative H1: ρ > 0. If ρ is zero, then no transformation is necessary and ordinary
least squares estimation is BLUE.
9.4.1 The Durbin-Watson test

The Durbin-Watson test statistic uses the residuals from the least squares procedure and is
closely related to ρ̂ . The statistic is
T
∑ ( eˆt − eˆt −1 )
2
d= t =2
T
∑ eˆt2
t =1
It is approximately equal to 2(1 − ρˆ ) .

If the estimated value of ρ is zero, the Durbin-Watson test statistic equals 2. If the estimated
value of ρ is one, the Durbin-Watson test statistic equals 0. Therefore, a low value of the DW test
statistic suggests the null hypothesis should be rejected. The distribution of the DW test statistic
is difficult and Excel cannot compute the p-value associated with d, but tables are available for
performing the hypothesis test [see www.bus.lsu.edu/hill/poe], now called the Durbin-Watson
bounds test.
T
∑ ( eˆt − eˆt −1 )
2
= sum of squared differences =SUMXMY2(F3:F35,F2:F34) = 3.54386803
i =2
T
∑ eˆt 2 = sum of squared residuals: = SUMSQ(F2:F35) = 3.0316
i =1
Using the formula above, we calculate the Durbin-Watson statistic to be d ≅ 1.169 as follows:
The least squares residuals for bangle.xls are stored in cells F2:F35 in the spreadsheet. The
numerator and the denominator of the DW statistic can be obtained using the Excel functions
SUMXMY2, and SUMSQ, respectively.
The decision rule for the Durbin-Watson bounds test is
• if d > upper bound, fail to reject the null hypothesis of no serial correlation,
• if d < lower bound, reject the null hypothesis and conclude that positive autocorrelation is
present,
• if lower bound < d < upper bound, the test is inconclusive.
With T=34 and K=2, the lower bound is 1.393 and the upper bound is 1.514. Since d < the lower
bound, we reject the null hypothesis that ρ is zero, and find evidence of positive autocorrelation.
9.4.2 An LM test
Another way to test for autocorrelation is to use test whether residuals are correlated with one
another using the LM test. This test is based on the auxiliary regression where you regress least
square residuals on the original regressors and lagged residual(s). If the auxiliary regression
explains sufficient variation in eˆt , then we conclude there is autocorrelation.
Now, let’s see how it works in Excel. Return to the worksheet containing the original
data for bangla.xls, include the logs of the data, and the residuals.
• Insert a new column and label it lagged residuals.
• Write 0 into cell E2.
• In cell E3, type =G2, where G2 contains the first residual from the regression output and
copy the formula down the column.
Once we created the lagged values of the residuals, we can now run the regression, using
residuals as the Y-Range and LN(P) and lagged residuals as the X-Range. Since the explanatory
variables need to be next to each other, you will have to move some of the columns around. Do
NOT suppress the intercept and name the worksheet LM test. Click OK.
150 Chapter 9
Recall that Excel reports a p-value based on a two-tailed test. We would conclude that ρ > 0,
testing at the 5% level, but we cannot reject the null hypothesis H0: ρ = 0 at the 1% level. The LM
scalar is calculated by N*R2 where N and R2 are the number of observations and the coefficient of
determination in the auxiliary regression and has a χ(1)2
distribution if the null hypothesis is true.
The test statistic value in this case is 34* 0.16=5.44 . Wwe reject the null hypothesis and conclude
that there is significant autocorrelation.
9.5 AUTOREGRESSIVE MODELS
In addition to incorporating lagged values of error term into a regression equation, we can add
lagged values of dependent and/or independent variables. Autoregressive models include lags of
the dependent variable as regressors. The AR(p) model has p lags of yt as regressors.
yt = δ + θ1 yt −1 + θ2 yt − 2 + ... + θ p yt − p + vt
We will use the inflation.xls to estimate an AR(3) model of inflation rate. Open the data file and
add three columns and label them INFLAG1, INFLAG2 and INFLAG3. Copy the INFLN column
and paste it to INFLAG1 starting from cell G4, to INFLAG2 starting from cell H5 and finally to
INFLAG3 starting from cell I6 as shown below.
Then, estimate the model using INFLN as the Y-Range and INFLAG1 INFLAG2 and INFLAG3
as the X-Range variables. Since there area missing observations for each column, make sure to
start from row 6, the first row for which there are no missing observations.
The result is:
9.6 FINITE DISTRIBUTED LAG MODELS
Finite distributed lag models contain independent variables and their lags as regressors.
yt = α + β0 xt + β1 xt −1 + β2 xt − 2 + ... + β q xt − q + vt
Using the inflation.xls, let’s model the inflation rate as a function of the percentage change in
wages and three lags of wages. Open the original file and add three columns. Label them
pcwage_L1, pcwage_L2 and pcwage_L3. Copy the PCWAGE column and paste it to pcwage_L1
starting from cell F4, to pcwage_L2 starting from cell G5 and finally to pcwage_L3 starting from
cell H6 as shown below.
152 Chapter 9
Observe that we now have T − n complete observations since we lose observations when we
create the lags. Now run the regression with T − n observations.
The least squares results are:

9.6 AUTOREGRESSIVE DISTRIBUTED LAG MODELS (ARDL)
Finally, we consider a model that contains both finite distributed lags and is autoregressive.
yt = δ + δ0 xt + δ1 xt −1 + .. + δ q xt − q + θ1 yt −1 + θ2 yt − 2 + ... + θ p yt − p + vt
ARDL (p, q) model has p lags of the dependent variable, yt and q lags of the independent variable
xt. Let’s illustrate ARDL (2,3) model for the inflation data. Modify the inlation.xls to incorporate
the two lags of INFLN and three lags of PCWAGE
Observe the missing observations due to lags. Run the regression starting from the 6th observation
using the INFLN as Y-Range and PCWAGE, its three lags, and two lags of INFLN as the X-
Range
The least squares results are:

154 Chapter 9
CHAPTER 10
Random Regressors and Moment-
Based Estimation
CHAPTER OUTLINE
10.1 Least Squares with Simulated Data 10.6 Estimation using Mroz Data
10.2 Instrumental Variables Estimation with 10.6.1 Least squares regression
Simulated Data 10.6.2 Two-stage least squares
10.2.1 Correction of IV standard errors 10.7 Testing for Endogeneity of Education
10.2.2 Corrected standard errors for 10.8 Testing for Weak Instruments
simulated data 10.9 Testing for Validity of Surplus Instruments
10.3 The Hausman Test: Simulated Data
10.4 Testing for Weak Instruments: Simulated
Data
10.5 Testing for Validity of Surplus Instruments:
Simulated Data
10.1 LEAST SQUARES WITH SIMULATED DATA
When the explanatory variables are random, the relationship between x and the error term, e, is
crucial in deciding whether ordinary least squares estimation is appropriate. If x and e are
uncorrelated, then least squares can, and should, be used. However, if the cov(x, e) ≠ 0, then the
least squares estimator is biased and inconsistent. In this case, instrumental variables (IV) / two-
stage least squares estimation process gives us a consistent but inefficient estimator. IV is not
directly available in Excel as a built-in function. However, we will show that it is easy to perform
IV estimation using Excel functions you are already familiar with.
We will use ch10.xls to explore the properties of the least squares estimator when cov(x, e) ≠
0. The data set contains simulated data of sample size 100, ρ xe = 0.6 , where
E ( y ) = β1 + β2 x = 1 + 1x and xi , ei ~ N (0,1) .
First, get the summary statistics for the dataset where z1, z2, and z3 are the instrumental
variables we will consider later. Go to Tools>Data Analysis>Descriptive Statistics
155
156 Chapter 10
Below is a sample of the descriptive statistics provided by Excel.
Next estimate the simple regression of y on x using y as the Y-Range and x as the X-Range.
Choose the Residuals option for future reference and place the output in a worksheet named LSE
for random regress. Click OK.
Random Regressors and Moment-Based Estimation 157
The results are:
Notice that the estimated slope is 1.7, when in fact the true slope of the artificial data is 1. These
results are due to the correlation between the error term and explanatory variable that has been
built into the data.
10.2 INSTRUMENTAL VARIABLES ESTIMATION WITH SIMULATED DATA
Ch10.xls also contains two instrumental variables both of which are correlated to x yet
uncorrelated with the error term. Since our problem is that the explanatory variable we are using
is correlated with the error term, using these instrumental variables should help solve our
problem.
We will estimate a reduced form or 1st stage regression xi = α1 + α2z1i + ei to obtain the
predicted values of xi, xî . Then, we use xî as an instrumental variable in the equation yi = β1 +
β2 xî + ei. Estimation of this equation is the "2nd stage". Thus the name, 2-stage least squares.
Return to the worksheet containing the simulated data, ch10.xls. Run a regression, using x as
the Y-Range and z as the X-Range. Choose the Residuals option to obtain xî . Place output on a
worksheet names 1st stage.
158 Chapter 10
The "Predicted x" from the RESIDUAL OUTPUT is our instrumental variable.
Copy cells B24 through B124 containing Predicted x over to the worksheet containing the
original data. Now run another regression, using y as the Y-Range and Predicted x as the X-
Range. Place output on a worksheet named 2nd stage.
The results of this Two-Stage Least Squares Estimation are:

The IV estimate of the slope 1.19 is closer to the true value of 1 but the standard errors are
incorrect.
10.2.1 Correction of IV standard errors

In the simple linear regression model yi = β1 + β2 xi + ei the 2SLS estimator is the least squares
estimator applied to yi = β1 + β2 xî + ei where xî is the predicted value from a reduced form
equation. So, the 2SLS estimators are
βˆ 2 =
∑ ( xî − x )( yi − y )
∑ ( xî − x )
2
βˆ 1 = y − βˆ 2 x
In large samples the 2SLS estimators have approximate normal distributions. In the simple
regression model
⎛ σ2 ⎞
βˆ 2 ~ N ⎜ β2 , ⎟
⎜ ∑ ( xˆ − x ) ⎟
2
⎝ i ⎠
The error variance σ2 should be estimated using the estimator
∑( y )
2
i − βˆ 1 − βˆ 2 xi
σˆ 22 SLS =
N −2
160 Chapter 10
with the quantity in the numerator being the sum of squared 2SLS residuals, or SSE2SLS. The
problem with doing 2SLS with two least squares regressions is that in the second estimation the
estimated variance is
∑( y )
2
i − βˆ 1 − βˆ 2 xî
σˆ 2
=
N −2
wrong
The numerator is the SSE from the regression of yi on xî , which is SSEwrong.
Thus correct 2SLS standard error is
σˆ 22 SLS σˆ 22 SLS σˆ 2 SLS

( )
se βˆ 2 = = =
∑ ( xˆ − x) ∑ ( xˆ − x) ∑ ( xˆ − x)
2 2 2
i i i
and the “wrong” standard error, calculated in the 2nd least squares estimation, is
σˆ 2wrong σˆ 2wrong σˆ wrong

sewrong ( )
βˆ 2 = = =
∑ ( xˆ − x) ∑ ( xˆ − x) ∑ ( xˆ − x)
2 2 2
i i i
Given that we have the “wrong” standard error in the 2nd regression, we can adjust it using a
correction factor
σˆ 22 SLS σˆ
( )
se βˆ 2 =
σˆ wrong
2 ( )
sewrong βˆ 2 = 2 SLS sewrong βˆ 2
σˆ wrong
( )
10.2.2 Corrected standard errors for simulated data

To correct the inflated standard errors and modify the t-statistic and p-value with the correct
standard error, return to the worksheet containing the original data.
• Insert a new column and label it residuals=y−b1−b2x.

• Calculate residuals using the coefficients from the 2SLS estimation. In other words, in
cell G2, type =y−1.101101099−1.192445082*A2, where A2 contains the first observation
on x. Copy the formula down the column.
• Label cell H1 sighat_2sls, cell H2 sighat_wrong, and cell H3 correction factor.
• In cell I1, type =SQRT(SUMSQ(G2:G101)/98) where 98 is T-K. This cell calculates the
square root of correct mean of squared errors which is the sighat_2sls.
• In cell I2, type =SQRT(3.042), where 3.042 is the wrong MSE and its square root it the
sighat_wrong..
• In cell I3, type =I1/I2 which calculates the correction factor.
This is our correct model standard error.
Copy the correction factor, return to the worksheet 2nd stage and paste after the regression output.
Then, create the corrected standard errors by multiplying the current standard errors with the
correction factor. The correct t-stat will be obtained by dividing the coefficient estimates by the
corrected standard errors. And you can calculate the correct p-value by typing =TDIST(D20,98,2)
for a two-tailed test where D20 is the corrected t-stat.
The results will be:
The correct t-stat and p-value are quite different from those reported originally by Excel.
162 Chapter 10
10.3 THE HAUSMAN TEST: SIMULATED DATA
Since we don’t always know if the explanatory variables are endogenous or not, we can
empirically check it using the Hausman test. We want to test H0: cov(x, e) = 0 against the
alternative H1: cov(x, e) ≠ 0. The Hausman Test is a formal test of these hypotheses and can be
based on using the residuals from the 1st stage estimation of
xi = γ1 + θ1 zi1 + θ2 zi 2 + vi
Denote the least squares residuals from the reduced form as vî . Include them in an artificial
regression
yi = β1 + β2 xi + δvî + ei
Estimate this model by least squares and test the significance of vî using a standard t-test.
H0 : δ = 0 no correlation between x and e

H1 : δ ≠ 0 correlation between x and e
If the null hypothesis is true, the ordinary least squares estimators are more efficient and should
be used. If the null hypothesis is not true, the instrumental variables estimator is consistent and
should be used.
Return to the worksheet containing the data ch10.xls. Estimate a regression, using x as the Y-
Range and z1 and z2 as the X-Range. Make sure to click the residual box to get the residuals.
Place the results on a worksheet named Hausman step 1.
Copy cells C25 through C125 which contain the Residuals to the worksheet containing the
original data and Paste. Move the columns x and Residuals together. Now, run the 2nd regression,
using y as the Y-Range and x and residuals as the X-Range. Place the results on a worksheet
named Hausman step 2. No need to store residuals for this step.
The regression output is:
Based on the t-test of the coefficient on Residuals, we reject the null hypothesis of no correlation
between x and e, and conclude that instrumental variable estimation is the procedure in this case.
10.4 TESTING FOR WEAK INSTRUMENTS: SIMULATED DATA
Instrumental variables must be as strongly correlated with the endogenous variable as possible. A
standard rule of thumb is that they will at least have a t-statistic of 3.3 or F-statistic of 10 in the
reduced form regression.
164 Chapter 10
From the above output, t-statistic for the second instrumental variable is 2.28. Since the t-
statistics less than the rule of thumb of 3.3, usage of this instrumental variable may not yield
satisfactory results.
10.5 TESTING THE VALIDITY OF SURPLUS INSTRUMENTS: SIMULATED

DATA
In addition to being strongly correlated with the endogenous variable, a “good” instrument needs
to be uncorrelated with the error term. We can test the validity of the instrument by an LM test.
LM test uses 2SLS residual eˆ2 SLS , as the dependent variable, and all the available exogenous and
instrumental variables as independent variables. LM statistic is calculated from this regression by
LM = NR 2 which has χ(2k ) where k is the number of surplus instrumental variables. Estimate an
auxiliary regression using the eˆ2 SLS from the 2nd step of 2SLS estimation, yi = β1 + β2 xî + ei as the
dependent variables and use instruments as the independent variable. The output is:
10.6 ESTIMATION USING THE MROZ DATA
We will use the Mroz data to illustrate the endogenous variables with real data. Open file
mroz.xls. This is a log-linear wage model on working woman.
ln(WAGE ) = β1 + β2 EDUC + β3 EXPER + β4 EXPER 2 + e
10.6.1 Least squares regression
The data set includes observations for non-working women. To eliminate the non-working
woman, use the variable lfp (labor force participation). Delete the rows where lfp = 0. Next, create
lwage =ln(wage) and exper2=exper^2. Next, run the least squares regression model using the
lwage as Y-Range and educ, exper, exper2 as X-Range. Include the labels, name the worksheet
LSE and check the residual box. The results are:
10.6.2 Two-stage least squares

In this example, educ is considered the endogenous variable as it could be correlated with factors
in the regression error term such as ability. In the 1st stage of the 2SLS estimation, we will
estimate the first one by using the educ as Y-Range and exper, exper2 and mothereduc (the
instrumental variable) as the X-Range. Keep the predicted values from the first stage to be used
in the 2nd stage.
Observe that the mothereduc is a good instrument with a t-statistic of approximately 8.6 which is
greater than the rule of thumb of 3.3
166 Chapter 10
Estimate the 2nd stage regression by using the lwage as Y-Range and exper, exper2 and
predicted values of educ (from the previous regression) as X-Range.
Do not forget to modify the standard errors, t-statistics and the p-values are described in the
previous section. First we need to modify the model standard error.
Then, apply the new standard error to the coefficient standard errors, t-statistics and p-values.
10.7 TESTING THE ENDOGENEITY OF EDUCATION
We will use Hausman test to test whether educ is endogenous and correlated to the regression
error term. First run the 1st stage regression using educ as the Y-Range and all other explanatory
variables as the X-Range. We only need the residuals from the 1st step to use as an additional
explanatory variable in the 2nd stage. For the second stage, use the lwage as the Y-Range and
educ, exper, exper2, and residuals as X-Range.
Based on the t-test of the coefficient on Residuals, we fail to reject the null hypothesis at 5% level
of no correlation between x and e, and conclude that OLS estimation is appropriate. However,
educ is endogenous at 10% level of significance.
10.8 TESTING FOR WEAK INSTRUMENTS
To test whether the relationship between the instruments and educ is strong enough, estimate the
restricted and unrestricted reduced form equations and test the joint significance of the
instruments. The unrestricted and restricted models have the following ANOVE tables:
168 Chapter 10
Using the MSE from the above models, to test the instruments jointly, we can conduct an F-test
where the unrestricted model has all exogenous variables and the restricted model has exper and
exper2 as the explanatory variables. Recall that the rule-of-thumb threshold value for adequate
instruments is an F-value of 10.0.
10.9 TESTING THE VALIDITY OF SURPLUS INSTRUMENTS
We can check the validity of the surplus instruments using the LM test. For this purpose, we need
to run an auxiliary regression where the residuals are the Y-Range and all the exogenous
variables are the X-Range. We calculate the LM test statistic as N*R2 which is 0.3780714 for this
example.
CHAPTER 11
Simultaneous Equations Models
CHAPTER OUTLINE
11.1 Truffle Supply and Demand 11.4 Supply and Demand of Fish
11.2 Estimating the Reduced Form Equations 11.5 Reduced Forms for Fish Price and Quantity
11.3 2SLS Estimates of Truffle Demand and
Supply
11.3.1 Correction of 2SLS standard errors
11.3.2 Corrected standard errors in
truffle demand and supply
In this chapter, we estimate simultaneous equation models where there are two or more
dependent variables that need to be estimated jointly. Ordinary least squares estimation is not
possible when we are dealing with more than one equation. For example to explain both price and
quantity of a good, we need both supply and demand equations which work together to determine
price and quantity jointly.
11.1 TRUFFLE SUPPLY AND DEMAND
Consider the supply and demand for truffles:
Demand: Qi = α1 + α 2 Pi + α 3 PSi + α 4 DI i + eid
Supply: Qi = β1 + β2 Pi + β3 PFi + eis
In the demand equation, Q is the quantity of truffles traded at time (in ounces) P is the market
price for truffles in dollars per ounce, PS is the market price for substitutes for truffles in dollars
per pound, and DI is the per capita disposable income, in thousands of dollars. In the supply
equation, PF is the price of the factor of production. P and Q are endogenous variables meaning
their values are determined within the system of equations. The equilibrium levels of price and
169
170 Chapter 11
quantity, P* and Q* are determined by both of these equations. PS, DI, PF are exogenous
meaning that we take the values as given.
Open truffles.xls and obtain the summary statistics. Part of the print out is given below.
11.2 ESTIMATING THE REDUCED FORM EQUATIONS
The reduced form equation expresses each endogenous variable, P and Q, in terms of the
exogenous variables PS, DI, PF. This can be accomplished by setting the structural equation
equal to each other and solving for the endogenous variables.
Qi = π11 + π21 PSi + π31 DI i + π 41 PFi + vi1
Pi = π12 + π22 PSi + π32 DI i + π42 PFi + vi 2
These equations can be estimated by least squares since all independent variables are exogenous
and uncorrelated with the error terms. We will estimate two regressions for the reduced form
equations. Once we estimate these reduced form equations, we will obtain and store the predicted
values of price, P̂ , and then estimate the structural equations using P̂ and the other exogenous
variables using the 2SLS technique which was first introduced in Chapter 10.
First, estimate the quantity equation using Q as the Y-Range, and PS, DI and PF as the X-
Range. Include labels and place results on the worksheet named reduced eq quant and Click
OK.
Simultaneous Equations Models 171
Next, estimate the regression for the price equation using P as the Y-Range and PS, DI and PF as
the X-Range. Include labels and place results on a worksheet named reduced eq price. Make
sure to include Residuals option to obtain the P̂ and click OK.
The Quantity (Q) reduced form estimates are:
The Price (P) reduced form estimates are:
All explanatory variables in both models are significant.
11.3 2SLS ESTIMATES OF TRUFFLE DEMAND AND SUPPLY
To obtain 2SLS estimates, we replace P in the structural equation with P̂ from the reduced form
equation. To accomplish that, return to the worksheet containing the original data. Create an
empty column next to the other explanatory variables as Excel requires all explanatory variables
to be next to each other. Return to the reduced eq price worksheet and Copy cells B26 through
B56 (Predicted p) to the original data worksheet. Estimate a regression of the structural demand
equation using Q as the Y-Range and Predicted p, PS and PF as the X-Range. Include labels
and place results on the worksheet named stage 2 demand.
172 Chapter 11
Return to the worksheet containing the original data. Move the columns around so that phat and
PF are next to each other to estimate the structural supply equation since Excel requires the X-
Range data to be contiguous. Estimate the regression, using Q as the Y-Range and Predicted p,
and PF as the X-Range. Include labels and place results on a worksheet names stage 2 supply.
The 2nd stage regression results for demand equation are:
The results for supply equation are:

Since the model standard error is based on the LS residuals, the standard errors, t-statistics and
confidence intervals are incorrect. Since Excel doesn’t have a built in function for these errors,
we will need to do some calculations. The calculation is explained in the next section.
11.3.1 Correction of 2SLS standard errors

In the simple linear regression model yi = β1 + β2 xi + ei the 2SLS estimator is the least squares
estimator applied to yi = β1 + β2 xî + ei where xî is the predicted value from a reduced form
equation. So, the 2SLS estimators are
βˆ 2 =
∑ ( xî − x )( yi − y )
∑ ( xî − x )
2
βˆ 1 = y − βˆ 2 x
In large samples the 2SLS estimators have approximate normal distributions. In the simple
regression model
⎛ σ2 ⎞
βˆ 2 ~ N ⎜ β2 , ⎟
⎜ ∑ ( xˆ − x ) ⎟
2
⎝ i ⎠
The error variance σ2 should be estimated using the estimator
∑ ( yi − βˆ 1 − βˆ 2 xi )
2
σˆ 22 SLS =
N −2
with the quantity in the numerator being the sum of squared 2SLS residuals, or SSE2SLS. The
problem with doing 2SLS with two least squares regressions is that in the second estimation the
estimated variance is
∑ ( yi − βˆ 1 − βˆ 2 xî )
2
σˆ 2
=
N −2
wrong
The numerator is the SSE from the regression of yi on xî , which is SSEwrong.
Thus correct 2SLS standard error is
174 Chapter 11
σˆ 22 SLS σˆ 22 SLS σˆ 2 SLS

( )
se βˆ 2 = = =
∑ ( xî − x ) ∑ ( xî − x ) ∑ ( xî − x )
2 2 2
and the “wrong” standard error, calculated in the 2nd least squares estimation, is
σˆ 2wrong σˆ 2wrong σˆ wrong

( )
sewrong βˆ 2 = = =
∑ ( xî − x ) ∑ ( xî − x ) ∑ ( xî − x )
2 2 2
Given that we have the “wrong” standard error in the 2nd regression, we can adjust it using a
correction factor
σˆ 22 SLS σˆ
( )
se βˆ 2 = se ( )
βˆ = 2 SLS sewrong βˆ 2
ˆσ 2wrong wrong 2 σˆ wrong
( )
11.3.2 Corrected standard errors in truffle demand and supply
The first step to correcting the standard errors is to calculate the sigmahats of the 2SLS for both
the supply and demand equations. As in the previous chapter, go back to the original data and
create two columns.
• Label them residualdemand and residualsupply, respectively.

• Calculate the residuals for both columns using the 2SLS estimates.
• Label column I as shown below to calculate the correction factors for both the demand
and the supply functions.
• In cells J1 and J5, we calculate the correct sigmahat_2sls for the demand and supply
functions, respectively.
• In cells J2 and J6, we copy and paste the sighat_wrong from the demand and supply
outputs and
• calculate the correction factors in cells J3 and J7 for demand and supply, respectively.
The resulting sighat_2sls figures and the correction factor for demand are
Now, we can plug in the correction factor of 1.839479403 into the demand model coefficient
page and calculate the corrected standard errors, t-test, and the p-values. From the results, we
have a demand function is downward sloping, and shifts to the right as the price of substitutes
rises, and shifts to the right as income rises (truffles are a normal good). All of our coefficients
are significant at the 5% level.
11.4 SUPPLY AND DEMAND OF FISH
The second example is from the Fulton Fish market. The demand equation is:
ln(QUAN t ) = α1 + α 2 ln( PRICEt ) + α 3 MON t + α 4TUEt + α 5WEDt + α 6THU t + etd
Where ln(QUAN t ) the quantity is sold in pounds, and PRICEt is the average price per pound.
The remaining are the dummy variables for the days of the week which capture the demand
shifts. α 2 is the price elasticity of demand, which is should be negative. The supply equation is:
ln(QUAN t ) = β1 + β2 ln( PRICEt ) + β3 STORMYt + ets

176 Chapter 11
β2 is the price elasticity of supply. The variable STORMY is a dummy variable indicating stormy
weather during the previous three days. Below are the partial summary statistics on the data set
fultonfish.xls.
11.5 REDUCED FORMS FOR FISH PRICE AND QUANTITY
The reduced form equations are estimated using the LSE. We can get them by estimating the
price and quantity equations as a function of all the exogenous variables. In this case,
Qt = π11 + π21MONt + π31TUEt + π41WEDt + π51THU t + π61STORMYt + vt1
Pt = π12 + π22 MON t + π32TUEt + π 42WEDt + π52THU t + π62 STORMYt + vt 2
The output for the reduced form equations are:
For the quantity:
For the price:
Now, using the predicted price, phat, we can estimate the structural equations. The output for the
structural equations is:
Supply equation:
Demand equation:
Recall that, although the point estimates are correct, the standard errors are NOT. You need to
adjust the standard errors the way illustrated in Section 11.3.2 above. The discussion in the text
explains why the estimated supply equation is not credible.
CHAPTER 12
Nonstationary Time-Series Data
and Cointegration
CHAPTER OUTLINE
12.1 Stationary and Nonstationary Data 12.4 Integration and Cointegration
12.2 Spurious Regression 12.5 Engle-Granger Test
12.3 Unit Root Tests for Stationarity
12.1 STATIONARY AND NONSTATIONARY DATA
One of the fundamental principles of econometrics is that the statistical properties of estimators
depend on how the data behaves. When time series data is concerned, the data need to be
stationary for the usual econometric procedures to have desired statistical properties. Time series
data is stationary when the means, variances, and covariances are constant and don’t depend on
the period in which they are measured.
To illustrate stationarity, we will use usa.xls and look at the mean and variance of GDP in
different sub period using time series plots and summary statistics. Open your file usa.xls. Insert
a column to the left of GDP and name it t for time.
• Type 1 in cell A2.

• Type =A2+1 in cell A3 and copy the cell down the entire column.
• After creating a time counter, highlight the t and GDP columns and go to
Insert>Chart>Scatter Plot. After some formatting, your time series plot will look like:
178
Nonstationary Data and Cointegration 179
As many other macroeconomic variables, GDP is a non-stationary time series. It has an

increasing trend that changes the mean every period. The consequences of non-stationarity can be
quite severe, resulting in unreliable least squares estimates and test statistics.
We can also create summary statistics to compare the means in different periods. Below are
the summary statistics for the first half, second half and overall data.
Both plot and the descriptive statistics indicate the GDP data is non-stationary.
12.2 SPURIOUS REGRESSION
It is very possible to estimate a regression and find statistically significant relationship where no
such relationship exists. In time series analysis this is a common phenomenon when data is non-
stationary. Our example will use two time series, rw1 and rw2 that are generated by random
walks where
rw1 : yt +1 + v1t
rw2 : xt +1 + v2t
180 Chapter 12
The errors are independent standard normal generated by a pseudo random generator. By
construction, rw1 and rw2 are not related in any way. Now let’s look at the empirical relationship
by opening the file spurious.xls.
Insert a column at Column A and create the time counter as described in the previous section.
Then plot both of the series against the t variable.
Next, scatter plot rw1 against rw2 and observe the potential spurious relationship between the two
variables.
You can also estimate a linear regression that will confirm the suspicion. The coefficient on rw2
is positive and significant. However, these variables are not related in any way. The cause of the
spurious relationship is the nonstationarity of the two series.
Therefore, it is very important to check for stationarity first, whenever you are dealing with time
series data.
12.3 UNIT ROOT TEST FOR STATIONARITY
The Dickey Fuller (DF) and augmented Dickey-Fuller (ADF) are unit root tests that are used to
test stationarity. Before conducting the test, there are certain decisions that need to be made about
the nature of the data in order to implement the correct regression model. These choices can be
made by visual inspection of the data various plots. For example, if the data has nonzero mean, a
constant term in the regression is appropriate. Below are the different possible regression models
for series with different characteristics:
Δyt = γyt −1 + vt No constant and no trend
Δyt = α + γyt −1 + vt Constant but no trend
Δyt = α + γyt −1 + λt + vt Constant and trend
Recall the GDP plot, which is slightly quadratic in time, so you would choose the regression
model that included a constant and a trend to conduct the unit root test. The test is conducted by
estimating the regression and implementing a t-test for the following hypothesis:
H0 : γ = 0
H1 : γ < 0
182 Chapter 12
The augmented version of the DF test (ADF) adds lagged differences to the model and the models
become:
m
Δyt = γyt −1 + ∑ as Δyt − s + vt
s =1
m
Δyt = α + γyt −1 + ∑ as Δyt − s + vt
s =1
m
Δyt = α + γyt −1 + λt + ∑ as Δyt − s + vt
s =1
You have to pick a lag length to implement this test. The lag length should be enough to ensure
that the residuals are not autocorrelated.
Let’s consider Federal Funds rate (Ft) and the 3-year Bond rate (Bt) for this test. The plots
indicated both series are nonstationary.
Since the series fluctuated from a non-zero mean and didn’t seem to have trend, we will use the
model with a constant but no trend. We will use one lag for the ADF test. So we will be
estimating the following regression model for both the Ft and the Bt.
Δyt = α + γyt −1 + a1Δyt −1 + vt
Open the file usa.xls and estimate two regressions.
• Create 3 columns to the right of column F.

• Label them diffF, lagF and difflagF for differenced F, lagged F and differenced lagged F,
respectively.
• In cell E3, type =D3-D2.

• Copy this formula down the column.
• In cell F3, type =D2, and Copy the formula down the column.
• In cell G4, type =F4-F3, and Copy the formula down the column.
• Repeat the same for series F. The formulas have been highlighted in the above sheet.
Next, estimate two regressions; one for B and the other for F. Use the diffF as the Y-Range, and
lagF and difflagF as the X-Range. Notice that the first observations appear in Row 3. Place
results on a worksheet named ADF-f .
Repeat for the B series and store the output in ADF-b.
The results for the B series are:

184 Chapter 12
The results for the F series are:
The t-statistics in this case is also called the tau-statistics within the context of unit root testing.
These values are compared to the ADF critical values. For both of the series, when compared the
test statistics to the critical values for the Dickey-Fuller test (which are shown in Table 12.2 of
POE), we do not reject our null hypothesis, and conclude that the levels of the series are
nonstationary.
12.4 INTEGRATION AND COINTEGRATION
Two non-stationary series are said to be cointegrated if their differences are stationary. For
example, Fed Funds rate and the 3-year bond are non-stationary. Both series are “integrated of
order 1” or I(1). If the two series move together through time, they are said to be cointegrated.
We can test cointegration by running a regression of one of the I(1) series on the other and testing
the residuals for stationarity using the augmented Dickey-Fuller test where the null hypothesis is
that the residuals are non-stationary. Rejecting the hypothesis leads to the conclusion that the
residuals are stationary and therefore, the series are cointegrated.
12.5 ENGLE-GRANGER TEST
The test described above is commonly referred as the Engle-Granger test. We will illustrate the
test by regressing B on F, save the residuals then use these in an augmented Dickey-Fuller
regression on the residuals.
Return to usa.xls file. Estimate the regression B as the Y-Range, and F as the X-Range.
Include labels and choose the Residuals option. Place the results on a cointegration worksheet.
186 Chapter 12
Copy the residuals from the output section to the data section and paste it in the next available
column. Create differ, lage and difflage and write the relevant formulas as shown below.
Estimate a regression, using diffe as the Y-Range and lage and difflage as the X-Range.
The output is:

Compare the results to the 1%, 5%, 10% ADF critical values. If the values are less than the t-
value, reject the null hypothesis and conclude that the residuals are stationary. You can find the
critical values for this test in Table 12.3 of POE.
CHAPTER 13
An Introduction to
Macroeconometrics: VEC and
VAR Models
CHAPTER OUTLINE
13.1 VEC and VAR models 13.3 Estimating VAR
13.2 Estimating VECM
13.1 VEC AND VAR MODELS
The vector autoregressive (VAR) model is a general framework used to describe dynamic
interrelationship among stationary variables. So first step of any VAR analysis should be to
determine whether the levels of your data is stationary or not. If not, take the first difference and
try again. Usually, if the levels of the time series are not stationary, the differences will be. If the
time series is stationary, then we need to modify the VAR framework to allow consistent
estimation of the relationship among series. The vector error correction (VEC) model is a special
case of the VAR for variables that are I(1). The VEC can also take into account any cointegrating
relationship among the series.
13.2 ESTIMATING A VEC MODEL
We will use gdp.xls data on Gross Domestic Product of U.S. and Australia to estimate a VEC
model. We are using a VEC model because (1) both series are nonstationary in levels and
stationary in differences (2) the variables are integrated. The following is the time series plot for
the two series. Go to Insert>Chart> select line graph and highlight both columns of data.
188
An Introduction to Macroeconometrics 189
After some formatting your plot should look like
Neither of these series is stationary in its levels and they appear to have a common trend. Next,
we will take the difference of both series. To accomplish that, label cells D1 and E1, diffusa and
diffaus, for differenced U.S. GDP and differenced Australian GDP, respectively.
190 Chapter 13
• In cell C3, type =B3-B2 to difference the series usa and Copy this formula down the
column.
• Repeat it for aus; in cell D3, type =C3-C2, and Copy the formula down the column.
• Highlight both C and D columns and go to Insert>Chart> select line graph.
• After some formatting the plot for the differenced data will look like
The differences of both series seem to be stationary, and possible they may be cointegrated.
Based on the graphical information, we will estimate Dickey-Fuller regressions for levels
with intercept, time trend and 2 lags for US equation and 3 lags for Australian equation. Recall
that the model with intercept and trend is
m
Δyt = α + γyt −1 + λt + ∑ as Δyt − s + vt
s =1
Before we estimate the model, we need get the data ready. Return to gdp.xls.
• Insert 5 columns after usa and label those new columns diffusa, lag usa, diffl1usa,
diffl2usa, and trend.
• In cell B3, type =A3-A2 and Copy this formula down the column (You already have this
column from the previous section).
• In cell C3, type =A2 to get the lagged values of usa, and Copy the formula down the
column.
• In cell D4, type =C3 to get the lagged values of the differenced data for usa, and Copy
the formula down the column.
• In cell E5, type =C3 to get the double lagged values of the differenced data for usa, and
Copy the formula down the column.
• In cell F1, type 1, in cell F2 type 2. Highlight the two cells and drag it down to create a
time trend.
• Repeat the same process for series aus except add another lag for the difference since we
are using 3 lags for aus.. The formulas have been highlighted in the below sheet.
Next, estimate two regressions; one for usa and the other for aus. Use the diffusa as the Y-Range,
and lag usa, diffl1usa, diffl2usa, and trend as the X-Range. Notice that the first 4 rows will not
be used due to missing values. Place results on a worksheet named ADF for USA
Repeat for the aus series and store the output in ADF for AUS
192 Chapter 13
The ADF results for USA are:
and for Australia:

In each case, the null hypothesis of nonstationarity can not be rejected at any significance. Notice
that both lagged differences are significant in the US equation and the 3rd lag in the Australian
equation are significant. Next, we will estimate the cointegration equation. Notice that the
cointegration equation does not include a constant. Return to gdp.xls file. Estimate the regression
aus as the Y-Range, and usa as the X-Range. Include labels and choose the Residuals option.
Place the results on LS for cointegration worksheet.
The output is:

194 Chapter 13
We saved the residual to for the Engle-Granger test of cointegration as described in Section 12.5
of this manual. Copy the residuals from the output section to the data section and paste it in the
next available column and rename the column e. Create diffe, lag e and diffl1e and write the
relevant formulas as shown below.
We will estimate the relationship Δeˆt = φeˆt −1 + vt .
which produces the following result:

The t-ratio is −2.89. The 5% critical value for the model with no intercept is −2.76. The t-ratio
falls within the rejection region and the null hypothesis of no cointegration is rejected and we
conclude that the two real GDP series are cointegrated.
To get the VEC model results for Australia, we will estimate the following regression.
The results are:

196 Chapter 13
The significant negative coefficient on lag e indicates that the Australian GDP responds to
disequilibrium between the US and Australia.
If we estimate the VEC model for USA,
We obtain:
The US, does not appear to respond to the disequilibrium between the two economies. These
finding support the idea that the economic conditions in Australia depend on those in the US but
not vice verse. In this simple model of two economy trade, the US acts as a large closed economy
and Australia as a small open economy.
13.3 ESTIMATING A VAR
The VEC model incorporates cointegration equation. It is relevant when two variables are I(1)
and integrated. Vector autoregressive model (VAR) is used when there is no cointegration among
variables and is estimated using time series that have been transformed to their stationary values.
To illustrate VAR model, we will use growth.xls. As earlier, the first step is to determine whether
the variables are stationary. If they are not, we will difference them and make sure the difference
is stationary. Next, we will test for cointegration, if they are cointegrated, we will estimate the
VEC model. If the series are not cointegrated, then using the differences and the lagged
differences, we will estimate a VAR model.
Let’s first plot the series and the differenced series. Plotting will help us determine the
existence of intercept and/or trend components.
198 Chapter 13
The fitted least squares regression results of Gt on Pt are:
Observe the very large R2, and t-ratio. However, based on the plots, the series appear to be
nonstationary. To test for cointegration, we estimate the Dickey-Fuller regression. The following
are the results from that regression:
Since the tau (unit root t-value) of −0.977 is greater than the 5% critical value of the test for
stationarity of −3.37, we conclude the residuals are nonstationary. Hence the relationship between
G and P is spurious. Therefore instead of a VEC model, we need to apply a VAR model to these
series.
For illustrative purposed, the order of lag in this example will be restricted to 1. In general,
one should test for the significance of lag terms greater than 1. The models we will be running
are:
ΔPt = β10 + β11ΔPt −1 + β12 ΔGt −1
ΔGt = β20 + β21ΔPt −1 + β22 ΔGt −1
We need to modify the growth.xls worksheet to include the above variables as shown below:
The least square results for CPI equation are:
The results indicate that the growth in price ( ΔPt ) is significantly related to its own past value
( ΔPt −1 ) but insignificantly related to growth rate in GDP in the previous period ( ΔGt −1 ).
The least square results for G are:
These results indicate that ΔGt is significantly and positively related to its own past value and
significantly and negatively related to last period’s inflation ( ΔPt −1 ).
CHAPTER 14
An Introduction to Financial
Econometrics: Time-Varying
Volatility and ARCH Models
CHAPTER OUTLINE
14.1 ARCH Model and Time Varying Volatility
14.2 Testing, Estimating, and Forecasting
14.1 ARCH MODEL AND TIME VARYING VOLATILITY
In this chapter, we will estimate several models in which the variance of the dependent variable
changes over time. These types of models are broadly referred to as autoregressive conditional
heteroskedasticity (ARCH) models. These models have become popular because the variance
specification can capture commonly observed features of the time series of financial variables.
They are especially useful in modeling volatility and changes in volatility over time. To
understand volatility and time-varying volatility, we will look at the returns.xls which shows the
time series of the monthly returns to four different stock market indices.
Open returns.xls and plot each series using Insert>Chart and choosing Line plot. After some
editing, the plots look like this:
200
Financial Econometrics: ARCH 201
The series are characterized by random, rapid changes and are said to be volatile. The volatility
seems to change over time for all 4 series. Next, to generate histograms of the series, go to
Tools>Data Analysis>Histogram. After some editing the histograms look like:
Based on the histograms, we can say the series are leptokurtic. In other words, they have a lot of
observations around the average and a relatively large number of observations that are far from
average.
202 Chapter 14
14.2 TESTING, ESTIMATING, AND FORECASTING
The basic ARCH model has two equations; one equation to describe the behavior of the mean,
and another to describe the behavior of the variance.
The mean equation is a linear regression function that contains a constant and possibly some
explanatory variables. The example below contains only an intercept.
yt = β0 + et where et ~ N (0, σt2 )
In this case, the series is expected to vary random about its mean. The error of the regression is
normally distributed and heteroskedastic. The variance of the current period depends on the
information revealed in the proceeding period.
ht = α + α1et2−1
where the variance of et is given by ht . Notice that ht depends on the squared error in the
preceding period. The parameters in this equation have to be positive to ensure a positive
variance.
To test for the presence of the ARCH effects, we can use a Lagrange Multiplier (LM) test. To
conduct this test, we first estimate the mean equation, store least squares residuals from the mean
equation, square them and run an auxiliary regression to observe the impact of previous error
term for a 1st order ARCH process. In other words, testing for ARCH(1), requires to regress eˆt2 on
the lagged residual eˆt2−1 .
eˆt2 = γ 0 + γ1eˆt2−1 + vt
where vt is the random error of the auxiliary relationship. The null and alternative hypotheses to
be tested are:
H 0 : γ1 = 0
H 1 : γ1 ≠ 0
The LM test statistic is NR2 with a χ(1) distribution under null hypothesis, where N is the number
of observations in the auxiliary regression. The general version of the LM test is to conduct an
ARCH(q) test and include q lags of eˆt2 as regressor. This version will have χ ( q ) distribution if the
null hypothesis is true and N in this case will be T − q.
We will use byd.xls data to as the first ARCH example. Open the Excel file named byd.xls
and go to Insert>Chart and plot the line plot for visual inspection of the time series.
Financial Econometrics: ARCH 203
As can be seen from the plot, there is visual evidence of volatility in the series. We will propose
and test and ARCH(1) model against the null hypothesis of no ARCH using the LM test
discussed earlier. First, estimate the mean equation. Recall that the mean equation we are running
does not have any explanatory variables. Excel will not estimate the model without defining the
X-Range so we will have to trick Excel. In cell B1, type intercept and create a column of ones.
Next, estimate the regression using the r as the Y-Range, and intercept as the X-Range. Make
sure to suppress the intercept, include the labels and keep the residuals.
In order to test the ARCH(1) model, we now need to estimate the auxiliary regression. Copy the
residuals from the output section to the data section and paste it in the next available column.
Rename it e, and create e square , and lag e square and write the relevant formulas as shown
below.
204 Chapter 14
Now, to test for first-order ARCH, regress eˆt2 on eˆt2−1 by estimating the regression using the e
square as the Y-Range, and lag e square as the X-Range.
Given the LM=62.16, we reject the null hypothesis, there is evidence of presence of ARCH(1)
process.
Unfortunately, ARCH models and all its extensions are estimated using by the maximum
likelihood method and beyond the capabilities of Excel.
CHAPTER 15
Panel Data Models
CHAPTER OUTLINE
15.1 Sets of Regression Equations 15.3 The Fixed Effects Model
15.2 Seemingly Unrelated Regressions 15.3.1 A dummy variable model
15.2.1 Breusch-Pagan test of independence 15.4 Random Effects Estimation
15.1 SETS OF REGRESSION EQUATIONS
Times series data are observations on the same unit taken over time. For example, annual GDP
over a ten-year period would be time-series data. Cross-sectional data are observations at one
point in time, over different units, such as 1990 per capita income for each of the 50 U.S. states.
In this section we will examine investment data from two firms (cross section) and for 20 periods
(time series). Open grunfeld2.xls. The descriptive statistics for the two firms can be obtained by
choosing Tools>Data Analysis>Descriptive Statistics. Below are some of the descriptive
statistics after some editing:
The equations we consider first are the individual investment models.
INVGE ,t = β1 + β2VGE ,t + β3 K GE ,t + eGE ,t t=1,…,20
INVWE ,t = β1 + β2VWE ,t + β3 KWE ,t + eWE ,t t=1,…,20
205
206 Chapter 15
If the models have the same parameters, we can estimate a pooled regression model using all 40
observations. However, if the parameters are not identical the models will be:
INVGE ,t = β1,GE + β 2,GEVGE ,t + β3,GE K GE ,t + eGE ,t t=1,…,20
INVWE ,t = β1,WE + β2,WEVWE ,t + β3,WE KWE ,t + eWE ,t t=1,…,20
We will first estimate two separate regressions using Tools>Data Analysis>Regression. Check
the Residuals option for later use. Click OK.
Estimate another regression for WE using D1 through D21 as the Y-Range and E1 through F21
as the X-Range, include labels and check the Residuals option for later use. Click OK.
The least squares results for GE are:

Panel Data Models 207
The least squares results for WE are:
To test whether we should pool the data or not, create the below dummy variable model:
INVi ,t = β1,GE + δ1 Di + β2,GEVi ,t + δ2 Di + β3,GE K i ,t + δ3 Di + ei ,t
where D = 1 for Westinghouse observations and 0 otherwise.

To estimate the dummy variable model in Excel, create three new columns and name them D,
DV and DK respectively. Column D will take the value 1 for WE and 0 for GE, Column DV will
take the respective v values for Westinghouse and 0 for General Electric, and Column DK will
take the respective k values for Westinghouse and 0 for General Electric. The worksheet will look
like:
208 Chapter 15
Estimate a regression, using A1 through A41 for the Y-Range, and cells B1 through F41 as the
X-Range. Place the results on a worksheet named OLS with dummies.
The results are
From these results, if we plug in 1 for D, we obtain two estimated regressions, one for GE and
one for Westinghouse. Since D = 1 for WE and D = 0 for GE, the respective equations are:
GE (D = 0): INVi ,t = −9.9563 + 0.02655Vt + 0.15169 K i ,t
Westinghouse (D = 0): INVi ,t = −0.5094 + 0.0529Vt + 0.0924 K t
The parameter estimates here are exactly as those from the separate regression equations above,
but the standard errors are different. This new model treats the coefficients in the same way as
before, but now assumes constant variance across all the observations.
To test for the significant differences in the variances, we use the Goldfeld-Quandt test,
where -
GQ = SSE _ GE / SSE _ WE
We can get the SSEs for the separate regressions from the output and calculate GQ as:
GQ = 13216.5871/1773.23405 = 7.45338
The F(17,17,.95) critical value is 2.2718 with a p-value of 0.3284. We reject the null hypothesis
of equal variance and find that there is strong evidence that the error variance for the two
equations are different.
15.2 SEEMINGLY UNRELATED REGRESSIONS
We will now assume that the error terms in the two equations, at the same point in time, are
correlated. This is called contemporaneous correlation.. Adding this assumption of
contemporaneous correlation provides additional information to our model and it should be
incorporated. Seemingly unrelated regression (SUR) permits equation coefficients and variances
to differ and also allows for contemporaneous correlation between the errors,
cov(eGE ,t , eWE ,t ) = σGE ,WE
While most statistical packages perform SUR estimation automatically, Excel does not. The basic
procedures used in SUR are (1) estimate the two equations separately via least squares, (2) use
the residuals in step (1) to obtain estimates of the variances and covariances in order to transform
the data, and (3) estimate the equations jointly via generalized least squares. Unfortunately, the
transformation necessary to perform the generalized least squares is beyond the scope of text.
15.2.1 Breusch-Pagan test of independence

In order to determine when to pool the data and use SUR or when to estimate the equations
separately, we need to test for the independence of errors. If the errors are not correlated, separate
estimation is fine. The test for correlation between the errors is called the Breusch-Pagan test. If
the null hypothesis of zero correlation is not rejected, again, separate estimation is fine.
BP = TrGE
2
,WE
σGE
2
1 20
2
where rGE ,WE =
,WE
and σGE ,WE = = ∑ eˆGE ,t eˆWE ,t .
σ σ
2
GE
2
WE T i =1
The test statistic is distributed as χ(21) under the null hypothesis of no contemporaneous
correlation. To calculate the estimated correlation between the equations, we will use the
residuals from the separate OLS results.
• Return to the worksheet GE and highlight cells C25 to C45.

• Choose Edit>Copy or click the Copy icon.
• Go to the WE worksheet and Paste in cells D25 through D45.
• In cell E26, type =C26*D26 and copy this formula down to cell E45.
210 Chapter 15
• In cell E46, type =SUM(E26:E45).

• In cell E47, type =E46/20. The result is 176.4490614 and represents σGE ,WE .
Next, we need a degrees of freedom adjustment for the calculation. Excel calculates the estimated
model variance by dividing by T−K. Because the Breusch-Pagan test is only asymptotically
justified, we divide only by T. Since the estimated model variance is the SSE divided by T−K, we
can simply recalculate using information from the ANOVA tables of our separate regressions.
• In cell D9 of the GE worksheet, type =C13/20. The result is 660.8293885.

• In cell D9 of the WE worksheet, type =C13/20. The result is 88.66169652.
• In a blank cell, say E7 of WE, type =(176.4490614^2)/(660.8293885*88.66169652).
The result is 0.531389929.
• In cell F7, type =E7*20 and
• in cell G7, type =CHIDIST(F7,1).
Based on the BP test, we reject the null hypothesis and find evidence of contemporaneous
correlation between the error terms. SUR is the appropriate analysis.
15.3 THE FIXED EFFECTS MODEL
The fixed effects model is a model for pooling data where the intercept is allowed to change
across firms but not across time and slope coefficients are assumed to be the same across firms.
Since all behavioral differences between firms and over time will be captured by the intercept,
this model is called a fixed effects model.
15.3.1 A Dummy Variable Specification

The fixed effects model permits cross-section heterogeneity by allowing only the intercept to vary
across individuals using dummy variables. We will also extend our model to include data on all
ten firms in grunfeld.xls. This data set has the investment data for ten companies for 20 periods
each. Below is an edited portion of the descriptive statistics of the dataset.
We will need nine dummy variables for ten forms, one firm will be the base firm and captured by
the intercept term. And the coefficient for each firm will be the difference between the intercept
for its firm and the intercept for the "base" firm (the "variable of omission"). Recall that if we
include all ten dummy variables, we will have perfect multicollinearity and estimation would not
be possible.
• First, we will create the dummy variables. Open the file named grunfeld.xls.
• Label cells F1 through N1, D1 through D9.
• In cells F2 through F21, type "1" and type "0" in the remaining cells of the column.
• In cells G22 through G41, type "1" and type "0" in all other cells of the column.
• Continue in this fashion, typing "1" appropriately for each firm, and "0" otherwise using
the Copy and Paste functions.
• Estimate a regression, using column C as the Y-Range, and columns D through N as the
X-Range.
• Include labels and place results on a worksheet named Unrestricted.
212 Chapter 15
The results are
To obtain the results in Table 17.2, remember that we did not include a dummy variable for the
tenth firm, so its intercept is −6.546. And for all the other firms, the intercept is the difference
between the coefficient of the firm dummy and the intercept. For example, for Firm 1, the
intercept is the difference −6.546−62.60 = −69.146. Based on the p-values for the dummy
coefficients, it appears that Firms 1, 4, and 6 do not have intercepts that differ significantly from
−6.546.
To test the equality of the intercepts, the null and alternative hypotheses are:
H 0 : β1, D1 = β1, D 2 = β1, D 3 = β1, D 4 = β1, D 5 = β1, D 6 = β1, D 7 = β1, D8 = β1, D 9
H1 : β1,i are not all equal
This is a regular F-test where we can use the template we previously created. The model we
estimated above is the unrestricted model, so now we just need to run the restricted model, where
we force all nine intercept coefficients to be equal. We can run the restricted model by returning
to the worksheet containing the data and running a regression, including only V and K as the X-
Range. Place the results on a worksheet named Restricted.
The results are:
• Open the F-template created earlier.

• Copy the SSE from cell C13 of the Unrestricted model and
• Paste it to cell B7 of the template.
• From the Restricted model, Copy the SSE in cell C13 and
• Paste it to cell B6 of template.
• We are testing for 9 restrictions and sample size is 200, so set the J = 9, T = 200, and K =
12.
214 Chapter 15
We reject the null hypothesis and find evidence that at least one firm has an intercept different
from the rest.
15.4 RANDOM EFFECTS ESTIMATION
We only make inferences about the firms on which we have data. The error components model
assumes the intercepts are random variables, drawn from a population distribution of firm
intercepts. One result of this is that the error terms from the same firm in different time periods
are correlated. The error components model, therefore, is sometimes called a random effects
model.
We know that generalized least squares estimation is appropriate when we have correlated
error terms. While it is possible to carry out the calculations with Excel, you are best advised to
use an econometric software package with a specialized routine for estimation of the random
effects model.
CHAPTER 16
Qualitative and Limited Dependent
Variable Models
CHAPTER OUTLINE
16.1 Models with Binary Dependent Variables 16.1.2 Least squares estimation of the linear
16.1.1 The linear probablility model probablility model
16.1 MODELS WITH BINARY DEPENDENT VARIABLES
So far, our focus has been on models in which the dependent variables is continuous; prices,
revenues, quantities etc. However, as a general theory of choice, microeconomics deals with
many choices where the outcomes are qualitative. In this chapter, we will consider choices that
are of the "either-or" type. That is, we choose to buy a particular car or not, we choose one job
over another, we vote either for or against a particular issue. In trying to explain these types of
choices, the dependent variable is dichotomous, or binary, since we quantify the choices by
assigning values zero or one. We then construct a statistical model that explains why particular
choices are made and what factors influence those choices.
To illustrate a model with a dichotomous dependent variable, consider a problem from
transportation economics. Workers can either drive to work (private transportation) or take a bus
(public transportation). For simplicity, we will assume these are the only two alternatives. The
individual’s choice will be represented by a dummy variable y, where
⎧1 if the individual drives to work

y=⎨
⎩0 if the individual takes a bus to work
If we collect random sample of workers who commute to work, then the outcome y will be
unknown to us until the sample is drawn. Thus, y is a random variable. If the probability that an
individual drives to work is p, then P [ y = 1] = p . It follows that the probability that a person uses
public transportation is P [ y = 0] = 1 − p . The probability function for this random variable is
215
216 Chapter 16
f ( y ) = p y (1 − p )
1− y
, y = 0, 1
where p is the probability that y takes the value 1. This discrete random variable has expected
value E [ y ] = p and variance var ( y ) = p(1 − p) .
If we assume that the only factor that determines the probability that an individual chooses
one mode of transportation over the other is the difference in time to get to work between the two
modes, then we define the explanatory variable x as
x = (commuting time by bus – commuting time by car)
While there are other factors that affect this choice, we will focus on this single explanatory
variable. A priori we expect a positive relationship between x and p that is as x increases, the
individual will be more inclined to drive.
16.1.1 The linear probability model

In regression analysis, the dependent variable is broken into two parts; fixed (systematic) and
random (stochastic). If we apply this to random variable y, we have
y = E ( y) + e = p + e
We then relate the systematic portion of y to the explanatory variables that we believe will help
explain the expected value. We are assuming that the probability of driving is positively related to
the difference in driving times, x, in this example. If we assume a linear relationship, then we will
have the following linear probability model.
E ( y ) = p = β1 + β2 x
In this chapter, we will examine the problems with least squares estimation in the context of
binary choice models. However, for these models least squares estimation methods are not the
best choice. Instead, maximum likelihood estimation (see Appendix C.8 of your book) is the
method to use. Excel does not have the capabilities to perform maximum likelihood estimation.
Other statistical packages such as EViews or SAS should be used when dealing with binary
choice models.
16.2 Least squares estimation of the linear probability model

The linear regression model for explaining the choice variable y is called the linear probability
model and is given by
y = E ( y ) + e = β1 + β2 x + e
To see how to apply the linear probability model in Excel, open the file transport.xls. Estimate
the regression, using auto as the Y-Range and dtime as the X-Range. Include labels and check
the Residuals option. label the worksheet “linear probability model” and click OK.
Qualitative and Limited Dependent Variable Models 217
The least squares results are
The explanatory variable is significant, suggesting that an increase of one minute in the difference
between the time it takes to get to work by bus versus by car increases the probability of driving
to work. However, the linear probability model has a very serious problem. Let’s look at the fitted
model, using least squares estimation:
E ( y ) = pˆ = 0.4848 + 0.007 dtime
For certain values of dtime, the estimated probability might turn out less than zero or greater than
one which is NOT possible for any valid probability function. If we look at the residual output we
can observe multiple occurrences of this problem.
218 Chapter 16
The problem arises because, the linear probability model is an increases function in x and the
increase is constant. However, given the requirement for a valid probability function of 0 ≤ p ≤ 1,
a constant rate of increase is not possible. Unfortunately, more appropriate models such as the
Logit and Probit model can not be estimated using the standard version of Excel.
CHAPTER 17
Importing Internet Data
Up to now, we have taken you through various econometric methodologies and applications
using already prepared Excel workfiles. In this chapter, we show you how to import data into an
Excel spreadsheet.
Getting data for economic research is much easier today than it was years ago. Before the
Internet, hours would be spent in libraries, looking for and copying data by hand. Now we have
access to rich data sources which are a few clicks away.
Suppose you are interested in analyzing the GDP of the United States. As suggested in POE
Chapter 17, the website Resources for Economists contains a wide variety of data, and in
particular the macro data we seek.
Websites are continually updated and improved. We shall guide you through an example, but
be prepared for differences from what we show here.
First, open up the website: www.rfe.org :
219
220 Chapter 17
Select the Data option and then select U.S. Macro and Regional Data.
This will open up a range of sub-data categories. For the example discussed here, select the
National Income and Produce Accounts to get data on GDP.
From the screen below, select the Gross Domestic Product (GDP) option.
Importing Internet Data 221
Most websites allow you to download data convenietly in an Excel format.
Be sure to save the file which is called gdplev.xls.

222 Chapter 17
Once the file has been downloaded (in this example, to C:\gdplev.xls), we can open the file and a
sample of the data in Excel format is shown below.
APPENDIX B
Review of Probability Concepts
CHAPTER OUTLINE
B.1 Binomial Probabilities B.2 The Normal Distribution
B.1.1 Computing binomial probabiliies directly
B.1.2 Computing binomial probabilities using
BINOMDIST
Excel has a number of functions for computing probabilities. In this chapter we will show you
how to work with the probability function of a binomial random variable, how to compute
probabilities involving normal random variables.
B.1 BINOMIAL PROBABILITIES
A binomial experiment consists of a fixed number of trials, n. On each independent trial the
outcome is success or failure, with the probability of success, p, being the same for each trial.
The random variable X is the number of successes in n trials, so x = 0, 1,…., n. For this discrete
random variable, the probability that X = x is given by the probability function
⎛ n! ⎞ x
P ( X = x ) = f ( x ) = ⎜⎜ ⎟⎟ p (1 − p ) ,
n− x
x = 0,1,… , n
⎝ x !( n − x )! ⎠
We can compute these probabilities two ways: the hard way and the easy way.
B.1.1 Computing binomial probabilities directly

Excel has a number of mathematical functions that make computation of formulas
straightforward. Assume there are n = 5 trials, that the probability of success is p = 0.3, and that
we want the probability of x = 3 successes. What we must compute is
223
224 Appendix B
⎛ 5! ⎞ 3
P ( X = 3) = f ( 3) = ⎜⎜ ⎟⎟ .3 (1 − .3 )
5−3
⎝ (
3! 5 − 3 )⎠
!
Eventually you will learn many shortcuts in Excel, but should you forget how to compute some
mathematical or statistical quantity, there is a Paste Function (f*) button on the Excel toolbar,
Click on the Paste Function button, select Math & Trig in the first column, and scroll down the
list of functions in the right-hand column. When you reach Fact you see that this function returns
the factorial of a number.
Click OK. In the resulting dialog box, enter 5 and Excel determines that 5! = 120.
Review of Probability Concepts 225
Alternatively, click on Help. In the resulting dialog box, enter factorial and click Search
Click on FACT.
You are presented with an Excel function, FACT(number), a definition and some examples.
The other mathematical operations we need to compute the binomial probability are
multiplication (*), division (/) and power (^).
In cell A1 type “f(3)”, and in B1 type the formula
=(FACT(5)/(FACT(3)*FACT(2)))*(0.3^3)*(0.7^2)
It will look like
Note that we have used parentheses to group operations.

Hit <enter>, and the result is 0.1323.
B.1.2 Computing binomial probabilities using BINOMDIST

The Excel function BINOMDIST can be used to find either cumulative probability, P ( X ≤ x )
or the probability function, P ( X = x ) for a Binomial random variable. Syntax for the function is
BINOMDIST(number_s, trials, probability_s, cumulative)
where number_s is the number of successes in n trials

trials is the number of independent trials (n)
probability is p, the probability of success on any one trial
cumulative is a logical value. If set equal to 1 (true), the cumulative probability is
returned; if set to 0 (false), the probability mass function is returned.
Access this function by clicking the Paste Function button. Select Statistical in the Function
category and BINOMDIST in the Function name.
226 Appendix B
Using the values n = 5, p = .3, and x = 3 we obtain the probability 0.1323, as above.
Alternatively, we can type the function equation directly into a cell. For example, if p = .2 and n
= 10, to find the probability that X = 4 and X ≤ 4, the worksheet would appears as follows:
=BINOMDIST(4,10,0.2,0) 0.08808
=BINOMDIST(4,10,0.2,1) 0.967207
The formulas in the first column produce the results reported in the second column.
B.2 THE NORMAL DISTRIBUTION
Excel provides several functions related to the Normal and Standard Normal Distributions.
1. The STANDARDIZE function computes the Z value for given values of X, μ, and σ. The
format of this function is
STANDARDIZE(X, μ, σ)
Referring to the example in POE Section B.5.1 in which μ = 3 and σ = 3, if we wanted to find the
Z value corresponding to X = 6, we would enter =STANDARDIZE(6,3,3) in a cell, and the value
computed would be 1.0.
2. The NORMSDIST function computes the area, or cumulative probability, less than a given Z
value. Geometrically, the cumulative probability is
The format of this function is
NORMSDIST(Z)
If we wanted to find the area below a Z value of 1.0, we would enter =NORMSDIST( 1.0) in a
cell, and the value computed would be .8413.
3. The NORMSINV function computes the Z value corresponding to a given cumulative area
under the normal curve. The format of this function is
NORMSINV(prob)
228 Appendix B
where prob is the area under the standard normal curve less than z. That is, prob = P(Z < z). If we
wanted to find the z value corresponding to a cumulative area of .10, we would enter
=NORMSINV(.10) in a cell and the value computed would be −1.2815.
4. The NORMDIST function computes the area or probability less than a given X value. The
format of this function is
NORMDIST(X, μ, σ, TRUE)
TRUE is a logical value, which can be replaced by 1. If we wanted to find the area below an X
value of 6, we would enter =NORMDIST(6,3,3,1) in a cell, and the value computed would be
.8413.
5. The NORMINV function computes the x value corresponding to a cumulative area under the
normal curve. The format of this function is
NORMINV(prob, μ, σ)
where prob is the area under the normal curve less than x. That is, prob = P(X < x). To compute
the value of x such that .10 of the probability is to the left, enter =NORMINV(.10,3,3) in a cell,
yielding −0.8446.
For the example in Section 2.6, a template can be built in Excel to compute probabilities and
values of X corresponding to particular probabilities. The highlighted cells require user input.
The formulas in the other cells do the computations. Set up a spreadsheet that looks like the
following
A B
1 Normal Probabilities
2 mean
3 standard_dev
4
5 Left-tail Probability
6 a
7 P(X<=a) =NORMDIST(B6,B2,B3,1)
8
9 Right-tail Probability
10 a
11 P(X>=a) =1-NORMDIST(B10,B2,B3,1)
12
13 Interval Probability
14 a
15 b
16 P(a<=X<=b) =NORMDIST(B15,B2,B3,1)-NORMDIST(B14,B2,B3,1)
17
18 Inverse cumulative
19 Left-tail probability
20 Quantile =NORMINV(B19,B2,B3)
230 Appendix B
Using X∼N(3,9), the above template would produce the following results:
Normal Probabilities
mean 3
standard_dev 3
Left-tail Probability
a 6
P(X<=a) 0.84134474
Right-tail Probabiity
a 6
P(X>=a) 0.15865526
Interval Probability
a 4
b 6
P(a<=X<=b) 0.210786144
Inverse cumulative
Left-tail probability 0.95
Quantile 7.934559001
Note that the Quantile = 7.93 gives the top 5% "cut off" value.
Once again, if you forget these formulas, use the Paste Function (f*) button on the Menu Bar.

Using Excel For Eco No Metric. Book

Uploaded by

Copyright:

Available Formats

Using Excel For Eco No Metric. Book

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Using Excel For Eco No Metric. Book

Uploaded by

Copyright:

Available Formats

Using Excel

For Principles of Econometrics, Third Edition

JOHN WILEY & SONS, INC

Carter Hill dedicates this work to Todd and Peter

ACQUISITIONS EDITOR Xxxxxx Xxxxxxxx

This book is printed on acid-free paper. ∞

Copyright © John Wiley & Sons, Inc. All rights reserved.

We welcome comments on this book, and suggestions for improvement. *

CHAPTER 2 The Simple Linear Regression CHAPTER 5 Multiple Linear Regression 68

CHAPTER 6 Further Inference in the

CHAPTER 14 An Introduction to Financial

CHAPTER 15 Panel Data Models 205

1.1 STARTING EXCEL

Excel opens. Click on the New Workbook icon.

The worksheet looks like this

Perhaps the most important of all these is Help.

1.2 ENTERING DATA

• To highlight a column—place cursor in A1. Hold down Ctrl-key and Shift-key

After selecting A1:A6, click Center

1.3 USING EXCEL FOR CALCULATIONS

The result its

The column sum (X + Y) is now in column C.

Add a header in column D.

1.3.1 Arithmetic operations

Arithmetic operators: To perform basic mathematical operations such as

Arithmetic operator Meaning (Example)

1.3.2 Mathematical functions

Math and trigonometry functions

ABS Returns the absolute value of a number

COMBIN Returns the number of combinations for a given number of objects

EVEN Rounds a number up to the nearest even integer

EXP Returns e raised to the power of a given number

FACT Returns the factorial of a number

FACTDOUBLE Returns the double factorial of a number

FLOOR Rounds a number down, toward zero

GCD Returns the greatest common divisor

INT Rounds a number down to the nearest integer

LCM Returns the least common multiple

LN Returns the natural logarithm of a number

LOG10 Returns the base-10 logarithm of a number

ODD Rounds a number up to the nearest odd integer

PI Returns the value of pi

POWER Returns the result of a number raised to a power

PRODUCT Multiplies its arguments

QUOTIENT Returns the integer portion of a division

RAND Returns a random number between 0 and 1

RANDBETWEEN Returns a random number between the numbers you specify

ROUND Rounds a number to a specified number of digits

ROUNDDOWN Rounds a number down, toward zero

ROUNDUP Rounds a number up, away from zero

SERIESSUM Returns the sum of a power series based on the formula

SIGN Returns the sign of a number

SQRT Returns a positive square root

SQRTPI Returns the square root of (number * pi)

SUBTOTAL Returns a subtotal in a list or database

SUM Adds its arguments

SUMIF Adds the cells specified by a given criteria

SUMPRODUCT Returns the sum of the products of corresponding array components

SUMSQ Returns the sum of the squares of the arguments