Chapter 6

Further Inference in the Multiple Regression Model

Econometrics
Handong Global University

Seon Tae Kim

Chapter Outline
 6.1 Testing Joint Hypotheses: The F-test

 6.2 The Use of Nonsample Information

 6.3 Model Specification

 6.4 Prediction

 6.5 Poor Data, Collinearity, and Insignificance

 6.6 Nonlinear Least Squares

2
6.1 Testing Joint Hypotheses: The F-Test (1 of 3)
 A null hypothesis with multiple conjectures, expressed with more than one

equal sign, is called a joint hypothesis.

1. Example: Should a group of explanatory variables be included in a

particular model?

2. Example: Does the quantity demanded of a product depend on the prices

of substitute goods, or only on its own price?

3
6.1 Testing Joint Hypotheses: The F-Test (2 of 3)
 Both examples are of the form:

 (6.1) H 0 : β 4  0, β 5  0, β 6  0

 The joint null hypothesis in (6.1) contains three conjectures (three equal signs): β4

= 0, β5 = 0, and β6 = 0

 A test of H0 is a joint test for whether all three conjectures hold simultaneously.

4
 6.1 Testing Joint Hypotheses: The F-Test (3 of 3)
 The F-statistic determines what constitutes a large reduction or a small reduction in the sum of squared
errors.
F 
 SSER  SSEU  J
(6.4)
SSEU  N  K 
where J is the number of restrictions, N is the number of observations and K is the number of
coefficients in the unrestricted model.

 If the null hypothesis is true, then the statistic F has what is called an F-distribution with J numerator
degrees of freedom and N - K denominator degrees of freedom.

 If F is bigger than the critical value, then reject the null hypothesis.
 If the null hypothesis is not true, then the difference between SSER and SSEU becomes large.

5
6
Example 6.1 Testing the Effect of Advertising (1 of 4)
 Consider the model:

 (6.2) SALES  β1  β 2 PRICE  β 3 ADVERT  β 4 ADVERT 2  e

 Test whether or not advertising has an effect on sales – but advertising is in the model as two

variables.

 Advertising will have no effect on sales if β3 = 0 and β4 = 0.

 Advertising will have an effect if β3 ≠ 0 or β4 ≠ 0 or if both β3 and β4 are nonzero.

7
Example 6.1 Testing the Effect of Advertising (2 of 4)
 The null hypotheses are:
H 0 : β 3  0, β 4  0
H1 : β 3  0 or β 4  0 or both are nonzero
 Relative to the null hypothesis H0 : β3 = 0, β4 = 0 the model in (6.2) is called the unrestricted

model.

 The restrictions in the null hypothesis have not been imposed on the model.

 It contrasts with the restricted model, which is obtained by assuming the parameter

restrictions in H0 are true.

8
Example 6.1 Testing the Effect of Advertising (3 of 4)
 When H0 is true, β3 = 0 and β4 = 0, and ADVERT and ADVERT2 drop out of the model.

(6.3) SALES  β1  β 2 PRICE  e

 The F-test for the hypothesis H0 : β3 = 0, β4 = 0 is based on a comparison of the sums of

squared errors (sums of squared least squares residuals) from the unrestricted model in

(6.2) and the restricted model in (6.3).

 Shorthand notation for these two quantities is SSEU and SSER, respectively.

9
Example 6.1 Testing the Effect of Advertising (4 of 4)
 Adding variables to a regression reduces the sum of squared error.

 More of the variation in the dependent variable becomes attributable to the variables in the

regression and less of its variation becomes attributable to the error.

 In terms of our notation, SSER – SSEU ≥ 0.

 We find that SSEU = 1532.084 and SSER = 1896.391. Adding ADVERT and ADVERT2 to the equation

reduces the sum of squared errors from 1896.391 to 1532.084.

10
Example 6.2 The F-Test Procedure (1 of 3)
1. Specify the null and alternative hypotheses:

 The joint null hypothesis is H0 : β3 = 0, β4 = 0. The alternative hypothesis is H0 : β3 ≠ 0 or

β4 ≠ 0 both are nonzero.

2. Specify the test statistic and its distribution if the null hypothesis is true:

 Having two restrictions in H0 means J = 2.

 Also, recall that N = 75, K = 4. F 

 SSER  SSEU  2
SSEU  75  4 

 𝛼 = 0.05. Critical value: 𝐹𝐶 = 𝐹(0.95,2,71) = 3.126.

11
 Example 6.2 The F-Test Procedure (2 of 3)

3. Set the significance level and determine the rejection region.

4. Calculate the sample value of the test statistic and, if desired, the p-value.

F 
 SSER  SSEU  J

1896.391  1532.084  2
 8.44
SSEU  N  K  1532.084  75  4 

 The corresponding p-value is p = P(F(2, 71) > 8.44) = 0.0005.

12
Example 6.2 The F-Test Procedure (3 of 3)
5. State your conclusion.

 Because F = 8.44 > Fc = 3.126, we reject the null hypothesis that both β3 = 0 and

β4 = 0, and conclude that at least one of them is not zero.

 Conclusion: advertising has a significant effect upon sales revenue.

 The same conclusion is reached by noting that p-value = 0.0005 < 0.05.

13
6.1.1 Testing the Significance of the Model (1 of 3)
 Consider again the general multiple regression model with (K - 1) explanatory

variables and K unknown coefficients.

 (6.5) y  β1  β 2 x2  β 3 x3  L  β K xK  e

 To examine whether we have a viable explanatory model, we set up the

following null and alternative hypotheses:

H 0 : β 2  0, β 3  0,K , β K  0
 (6.6)
H1 : At least one of the β k is nonzero for k  2, 3,K K
 Under the null, the model is useless; J = K-1.
14
6.1.1 Testing the Significance of the Model (2 of 3)
 Because we are testing whether or not we have a viable explanatory model, the test for (6.6) is
sometimes referred to as a test of the overall significance of the regression model.

 Given that the t-distribution can only be used to test a single null hypothesis, we use the
F-test for testing the joint null hypothesis in (6.6).
 The restricted model, assuming the null hypothesis is true, becomes:

 The least squares estimator of β1 in this restricted model is: yi  1  ei

b   i 1 yi N  y
* N
1

15
6.1.1 Testing the Significance of the Model (3 of 3)
 The restricted sum of squared errors from the hypothesis (6.6) is:

 
N N
SSER   yi  b1*    yi  y   SST
2 2

i 1 i 1

 Thus, to test the overall significance of a model, but not in general, the F-test

statistic can be modified and written as:

 SST  SSE   K  1
(6.8) F 
SSE  N  K 

Recall that in this case, J = K-1; SSE means SSEU.

16
6.1.2 The Relationship Between t- and F-Tests (1 of 2)

 What happens if we have a null hypothesis that is not a joint hypothesis? it only has one equality

in H0.

 When testing a single “equality” null hypothesis (a single restriction) against a “not equal to”

alternative hypothesis, either a t-test or an F-test can be used.

 Two-tail t-tests are equivalent to F-tests when there is a single hypothesis H0.

17
6.1.2 The Relationship Between t- and F-Tests (2 of 2)
 The elements of an F-test:
1. The null hypothesis H0 consists of one or more equality restrictions on the model
parameters βk. J = Number of restrictions.
2. The alternative hypothesis states that one or more of the equalities in the null hypothesis is
not true.
3. The test statistic is the F-statistic in (6.4).
4. If the null hypothesis is true, F has the F-distribution with J numerator degrees of freedom
and N - K denominator degrees of freedom.
5. When testing a single equality null hypothesis, it is perfectly correct to use either the t- or F-
test procedure: they are equivalent.

18
6.1.3 More General F-Tests
 The conjectures made in the null hypothesis were that particular coefficients

are equal to zero.

 The F-test can also be used for much more general hypotheses.

 Any number of conjectures (≤ K) involving linear hypotheses with equal signs

can be tested.

19
 Example 6.5 Testing Optimal Advertising (1 of 3)
 The hypothesis (theory): the optimal level of advertising 𝐴𝐷𝑉𝐸𝑅𝑇0 should satisfy

 (6.11) β 3  2β 4 ADVERT0  1
 ADVERT0 = $1,900 per month in the data. We want to test if it is really optimal, as:

H 0 : β 3  2  β 4  1.9  1 H1: β 3 +2  β 4  1.9  1

Or H 0 : β 3  3.8β 4  1 H1: β 3 +3.8β 4  1

 Note that when H0 is true, β3 = 1 – 3.8β4 so that:

SALES  β1  β 2 PRICE  1  3.8β 4  ADVERT  β 4 ADVERT 2  e

20
Example 6.5 Testing Optimal Advertising (2 of 3)
 (6.12)  SALES  ADVERT   β1  β 2 PRICE  β 4  ADVERT 2  3.8 ADVERT   e
 The regression equation above is the restricted model.

 Dependent variable: SALES – ADVERT.

 X1 = 1.

 X2 = PRICE.

 X3 = 𝐴𝐷𝑉𝐸𝑅𝑇 2 − 3.8 ∗ 𝐴𝐷𝑉𝐸𝑅𝑇.

 Estimate this restricted model. Results:

21
Example 6.5 Testing Optimal Advertising (3 of 3)
 Compare to . J = 1. N = 75. K = 4.

 The calculated value of the F-statistic is:

F 
1552.286 - 1532.084  1
 0.9362
1532.084 71
 For α = 0.05, the critical value is Fc = 3.976. Because F = 0.9362 < Fc = 3.976, we do not

reject H0.

 We conclude that an advertising expenditure of $1,900 per month is optimal and compatible

with the data.

22
 6.1.5 Large Sample Tests (1 of 3)
 There are two key requirements for the F-statistic to have the F-distribution in samples of all
sizes.

1. Assumptions MR1–MR6 must hold.

2. The restrictions in H0 must be linear functions of the parameters β1, β2, …, βk a.

 We are concerned with what test statistics are valid in large samples when the errors are not
normally distributed or when the strict exogeneity assumption is weakened to E (ei) = 0
and cov(ei , xjk) = 0(i ≠ j).

23
 6.1.5 Large Sample Tests (2 of 3)
 An F random variable is defined as the ratio of two independent chi-square (χ2 ) random
variables, each divided by their degrees of freedom.
SSE R  SSE U
2
V1 / J J (SSE R  SSE U ) / J
(6.13) F   ~ F( J , N  K )
V2 /( N  K ) ( N  K ) 2 2
2
( N  K)

 Note that 𝜎 2 = 𝑆𝑆𝐸𝑈 /(N − K), and so the result in (6.13) is identical to the F-statistic first
introduced in (6.4).

 When we drop the normality assumption or weaken the strict exogeneity assumption, the
argument becomes slightly different.

24
6.1.5 Large Sample Tests (3 of 3)
 We can go one step further and say that replacing σ2 by its consistent estimator 𝜎 2
does not change the asymptotic distribution of V1. That is,

𝑉1 ‘s distribution is approximated well by the chi-square distribution with d.f. = J.

 This statistic is a valid alternative for testing joint linear hypotheses in large samples
under less restrictive assumptions.

25
26
6.2 The Use of Nonsample Information (1 of 5)
 In many estimation problems we have information over and above the information

contained in the sample observations.

 This nonsample information may come from many places, such as economic

principles or experience.

 When it is available, we should find a way to use it.

27
6.2 The Use of Nonsample Information (2 of 5)
 Consider the log-log functional form for a demand model for beer:

 (6.17) ln 𝑄 = β1 + β2 ln 𝑃𝐵 + β3 ln 𝑃𝐿 + β4 ln 𝑃𝑅 + β5 ln 𝐼 + 𝑒

 This model is a convenient one because it precludes infeasible negative prices, quantities, and
income, and because the coefficients β2, β3, β4, and β5 are elasticities.

 A relevant piece of nonsample information can be derived by noting that if all prices and
income go up by the same proportion, we would expect there to be no change in quantity
demanded.

28
 6.2 The Use of Nonsample Information (3 of 5)
 Having all prices and income change by the same proportion is equivalent to

multiplying each price and income by a constant, say λ:

(6.18) ln 𝑄 = β1 + β2 ln 𝜆𝑃𝐵 + β3 ln 𝜆𝑃𝐿 + β4 ln 𝜆𝑃𝑅 + β5 ln 𝜆𝐼

=β1 + β2 ln 𝑃𝐵 + β3 ln 𝑃𝐿 + β4 ln 𝑃𝑅 + β5 ln 𝐼 + β𝟐 + β𝟑 + β𝟒 + β𝟓 𝐥𝐧(𝝀) + 𝑒

29
6.2 The Use of Nonsample Information (4 of 5)
 To have no change in ln(Q) when all prices and income go up by the same proportion, it must
be true that: β 2  β3  β 4  β5  0

 Start with ln 𝑄 = β1 + β2 ln 𝑃𝐵 + β3 ln 𝑃𝐿 + β4 ln 𝑃𝑅 + β5 ln 𝐼 + 𝑒

 Solve the restriction for one of the parameters, say β4: β𝟒 = −β𝟐 − β𝟑 − β𝟓

ln  Q   β1  β 2 ln  PB   β 3 ln  PL    β 2  β 3  β 5  ln  PR   β 5 ln  I   e

 Substituting gives:  β1  β 2 ln  PB   ln  PR    β 3 ln  PL   ln  PR  

β5 ln  I   ln  PR    e
 PB   PL   I 
 β1  β 2 ln    β 3 ln    β 5 ln  e
 PR   PR   PR 
30
6.2 The Use of Nonsample Information (5 of 5)
 Properties of this restricted least squares estimation procedure:

1. The restricted least squares estimator is biased, unless the constraints we

impose are exactly true.

2. The restricted least squares estimator: its variance is smaller than the variance

of the (unrestricted) least squares estimator, whether the constraints imposed are

true or not.

Reduction in variance: due to the additional information used.

31
 6.3 Model Specification
 In any econometric investigation, choice of the model is one of the first steps.

 What are the important considerations when choosing a model?

 What are the consequences of choosing the wrong model?

 Are there ways of assessing whether a model is adequate?

 Three essential features of model choice are

1. Choice of functional form

2. Choice of explanatory variables (regressors) to be included in the model

3. Whether the multiple regression assumptions MR1–MR6, listed in Chapter 5, hold.

32
6.3.1 Causality versus Prediction (1 of 2)
 With causal inference we are primarily interested in the effect of a change in a regressor on

the conditional mean of the dependent variable.

 We wish to be able to say that a one-unit change in an explanatory variable will cause a

particular change in the mean of the dependent variable with other factors held constant.

 This type of analysis is important for policy work.

33
6.3.1 Causality versus Prediction (2 of 2)
 On the other hand, if the purpose of a model is to predict the value of a dependent

variable, then, for regressor choice, it is important to choose variables that are highly

correlated with the dependent variable and that lead to a high R2.

 Predictive analysis using variables from the increasingly popular field of big data is an

example of where variables are chosen for their predictive ability rather than to examine

causal relationships.

34
 6.3.2 Omitted Variables
 It is possible that a chosen model may have important variables omitted that are correlated
with independent variables included in the model.

 Our economic principles may have overlooked a variable, or lack of data may lead us to
drop a variable even when it is prescribed by economic theory.

 Omitted variable bias:

Omitting a relevant variable is a special case of using a restricted least squares estimator,
where the restriction is that the coefficient on the omitted variable is zero, and the restriction
is not true. It leads to a biased estimator for coefficients on other explanatory variables, but
with a smaller variance.

35
36
37
6.3.3 Irrelevant Variables
 You think that a good strategy is to include as many variables as possible in your model.

 Doing so will not only complicate your model unnecessarily, but may also inflate the variances

of your estimates because of the presence of irrelevant variables.

 Those whose coefficients are zero because they have no direct effect on the dependent

variable.

38
6.3.4 Control Variables (1 of 3)
 Variables included in the equation to avoid omitted variable bias in the coefficient of interest
are called control variables.

 For a control variable to serve its purpose and act as a proxy for an omitted variable, it
needs to satisfy a conditional mean independence assumption.

 Actual regression equation: 𝑦𝑖 = 𝛽1 + 𝛽2 𝑥𝑖 + 𝛽3 𝑧𝑖 + 𝑒𝑖

 Want to estimate 𝛽2 , coefficient on 𝑥𝑖 . Suppose 𝑧𝑖 is unobservable and omitted.  𝑦𝑖 = 𝛽1 +

𝛽2 𝑥𝑖 + 𝑣𝑖 𝑤ℎ𝑒𝑟𝑒 𝑣𝑖 = 𝛽3 𝑧𝑖 + 𝑒𝑖 .

 If 𝑐𝑜𝑟(𝑥𝑖 , 𝑧𝑖 ) = 0: E[𝑣𝑖 |𝑥𝑖 ] = 0  OLS unbiased.

 If 𝑐𝑜𝑟(𝑥𝑖 , 𝑧𝑖 ) ≠ 0: E[𝑣𝑖 |𝑥𝑖 ] ≠ 0  OLS biased.

39
6.3.4 Control Variables (2 of 3)
 What to do?

 Find and include a control variable 𝑞𝑖 as a proxy for 𝑧𝑖 : it needs to satisfy a conditional
mean independence assumption.
E[𝑧𝑖 𝑥𝑖 , 𝑞𝑖 = E[𝑧𝑖 𝑞𝑖

Meaning: conditional on 𝑞𝑖 , 𝑧𝑖 is independent of 𝑥𝑖 .

 Part of 𝑧𝑖 𝑢𝑛𝑟𝑒𝑙𝑎𝑡𝑒𝑑 𝑡𝑜 𝑞𝑖 is unrelated to 𝑥𝑖 , too.

 Include in the regression equation 𝑞𝑖 as a proxy for 𝑧𝑖 .

 OLS unbiased.

40
6.3.4 Control Variables (3 of 3)
Example: regression of wages on education.
 Omitted variable: ability, clearly relevant, but unobserved.

 Include IQ as a proxy for ability, under the assumption:

 𝐸 𝐴𝐵𝐼𝐿𝐼𝑇𝑌 𝐸𝐷𝑈𝐶𝐴𝑇𝐼𝑂𝑁, 𝐼𝑄 = 𝐸 𝐴𝐵𝐼𝐿𝐼𝑇𝑌 𝐼𝑄 .

 IQ is correlated with ability.

 Once we know somebody’s IQ, knowing their level of education does not add
any extra information about their ability.

41
 6.3.5 Choosing a Model (1 of 3)
1. Is the purpose of the model to identify causal effects or prediction? If the goal is to
identify causal effects, then careful selection of control variables is necessary

2. If the goal is to predict, then use variables that have high predictive power (highly
correlated to the dependent variable).

3. Having theoretical knowledge and understanding of the relationship are important for
choosing the variables and functional form.

4. If an estimated equation has coefficients with unexpected signs, or unrealistic

magnitudes, they could be caused by a misspecification.

42
6.3.5 Choosing a Model (2 of 3)
4. Patterns in least squares residuals can be helpful for uncovering problems caused by
an incorrect functional form.

5. One method for assessing whether a variable or a group of variables should be

included in an equation is to perform significance tests (either t-test or F-test).

6. Are the estimated coefficients robust with respect to alternative specifications?

43
6.3.5 Choosing a Model (3 of 3)
8. A test known as RESET (Regression Specification Error Test) can be useful for
detecting omitted variables or an incorrect functional form.

9. Various model selection criteria, based on maximizing R2, or minimizing the sum of
squared errors (SSE), subject to a penalty for too many variables, have been
suggested.

10. A more stringent assessment of a model’s predictive ability is to use a “hold-out”

sample.

11. Following these guidelines can almost inevitably lead to revisions of originally
proposed model.
44
 6.3.6 RESET (1 of 3)
 RESET (REgression Specification Error Test) is designed to detect omitted variables
and incorrect functional form.

 Suppose we have the model: y  β1  β 2 x2  β 3 x3  e

 Let the predicted values of y be: (6.32) y  b1  b2 x2  b3 x3

 Now consider the following two artificial models:

 (6.33) 𝑦 = β1 + β2 𝑥2 + β3 𝑥3 + 𝛾1 𝑦 2 + 𝑒

 (6.34) 𝑦 = β1 + β2 𝑥2 + β3 𝑥3 + 𝛾1 𝑦 2 + 𝛾2 𝑦 3 + e

45
6.3.6 RESET (2 of 3)
 In (6.33) a test for misspecification is a test of H0:γ1 = 0 against the alternative H1:γ1 ≠ 0.

 In (6.34), testing H0:γ1 = γ2 = 0 against H1: γ1 ≠ 0 and/or γ2 ≠ 0 is a test for misspecification.

 Rejecting H0 implies that the original model is inadequate and can be improved. A failure to

reject H0 says that the test has not been able to detect any misspecification.

46
6.3.6 RESET (2 of 3)
 𝑦 2 𝑎𝑛𝑑 𝑦 3 are polynomial functions of 𝑥2 𝑎𝑛𝑑 𝑥3 .

 If the original functional form of 𝑥2 𝑎𝑛𝑑 𝑥3 are not correct, then polynomials of 𝑥2 𝑎𝑛𝑑 𝑥3 ,

captured by 𝑦 2 𝑎𝑛𝑑 𝑦 3 , would improve the fit of the model.

 If there were omitted variables that are correlated with 𝑥2 𝑎𝑛𝑑 𝑥3 , then such omitted variables

will be correlated with polynomials of 𝑥2 𝑎𝑛𝑑 𝑥3 , captured by 𝑦 2 𝑎𝑛𝑑 𝑦 3 . Thus, including those

omitted variables via 𝑦 2 𝑎𝑛𝑑 𝑦 3 would improve the fit of the model.

47
 6.4.1 Predictive Model Selection Criteria (1 of 2)
 In this section, we consider three model selection criteria:

1. 𝑅2 and 𝑅2

2. AIC

3. SC

 These criteria should be treated as devices that provide additional information about
the relative merits of alternative models.

48
6.4.1 Predictive Model Selection Criteria (2 of 2)
 An alternative measure of goodness of fit called the adjusted-𝑹𝟐 , denoted 𝑹𝟐 .

𝑆𝑆𝐸/(𝑁−𝐾)
 Computed as 𝑅 2= 1−
𝑆𝑆𝑇/(𝑁−1)

 (N-K)in numerator: including more variables does not increase 𝑅 2 .

 Selecting variables to maximize 𝑅2 can be viewed as selecting variables to minimize

SSE.
𝑆𝑆𝐸 2𝐾 𝑆𝑆𝐸 𝐾𝑙𝑛 𝑁
𝐴𝐼𝐶= ln + ( ) and SC = ln +( )
𝑁 𝑁 𝑁 𝑁

 including more variables reduces SSE but increases K. Minimize AIC (or SC).

49
6.5 Poor Data, Collinearity, and Insignificance
 When data are the result of an uncontrolled experiment, many of the economic (explanatory)
variables may move together in systematic ways.

 Such variables are said to be collinear, and the problem is labeled collinearity.

 It is not just relationships between variables in a sample of data that make it difficult to isolate
the separate effects of individual explanatory variables.

 If the values of an explanatory variable do not vary much within a sample of data, then it is
difficult to use that data to estimate a coefficient that describes the effect of change in that
variable.

50
6.5.1 The Consequences of Collinearity (1 of 3)

 Exact or extreme collinearity exists when x2 and x3 are perfectly correlated, in which case

r23 = 1 and var(b2) goes to infinity.

 Similarly, if x2 exhibits no variation   xi  x2 

2
equals zero and var(b2) again goes to

infinity.

 In this case x2 is collinear with the constant term.

51
 6.5.1 The Consequences of Collinearity (2 of 3)
 In general, whenever there are one or more exact linear relationships among the explanatory

variables, then the condition of exact collinearity exists.

 In this case the least squares estimator is not defined.

 We cannot obtain estimates of βk’s using the least squares principle.

 The effects of this imprecise information are:

52
6.5.1 The Consequences of Collinearity (3 of 3)
1. When estimator standard errors are large, it is likely that the usual t-tests will show that

parameter estimates are not significantly different from zero.

2. Estimators may be very sensitive to the addition or deletion of a few observations, or to the

deletion of an apparently insignificant variable.

53
 6.6 Nonlinear Least Squares (1 of 2)
 Estimation of models that are nonlinear in the parameters.

 Consider the model (6.50) 𝒚𝒊 = 𝜷𝒙𝒊𝟏 + 𝜷𝟐 𝒙𝒊𝟐 + 𝒆𝒊 .

 This example differs from the conventional linear model because the coefficient of xi,2

is equal to the square of the coefficient of xi,1, and the number of parameters is not

equal to the number of variables.

54
 6.6 Nonlinear Least Squares (2 of 2)
 To estimate β we can use (6.51) 𝑆 𝛽 = 𝑁
𝑖=1(𝑦𝑖 − 𝛽𝑥𝑖1 − 𝛽2 𝑥𝑖2 )2 .

 When we have models that are nonlinear in the parameters, we cannot in general derive
formulas for the parameter values that minimize the sum of squared errors function.

 However, for a given set of data, we can ask the computer to search for the parameter values
that take us to the bottom of the bowl.

 Those minimizing values are known as the nonlinear least squares estimates.

55
 Key Words
 χ2-test  influential observations  RESET
 AIC  irrelevant variables  restricted least squares
 auxiliary regressions  nonlinear least squares  restricted model
 BIC  nonsample information  restricted SSE
 causal model  omitted variable bias  SC
 collinearity  overall significance  single and joint null
 control variables  prediction hypotheses

 F-test  predictive model  unrestricted model

 unrestricted SSE

56