RS Chapter 6

Lecture 6
Specification and Model
Selection Strategies

Model Selection Strategies

So far, we have implicitly used a simple strategy:
(1) We started with a DGP, which we assumed to be true.
(2) We tested some H0 (from economic theory).
(3) We used the model (restricted, if needed) for
prediction/forecasting.
Under CLM assumptions (A1) to (A5), t-tests, F-tests and predictions
have desirable properties. But if assumptions do not hold, then,
- Tests can be weak with unknown distributions.
- Tests may be biased i.e., more likely to reject H0 when it is
true than when it is false. Same for predictions.
- Tests may be inconsistent i.e., power does not approach 1
for every alternative for a given significance level

RS Chapter 6

Model Selection Strategies

In this lecture we will address assumption (A1). How do we propose
and select a model (a DGP)?
Potentially, we have a huge number of possible models (different
functional form, f(.), and explanatory variables, X). Say, we have
Model 1
Y = X +
Model 2
Y = Z +
Model 3
Y = (W) +
We want to select the best model, the one that is closest to the DGP.
In practice, we aim for a good model.

Model Selection Strategies

According to David Hendry, a good model should be:
- Data admissible
-i.e., modeled and observed y should have the
same properties.
- Theory consistent -our model should make sense
- Predictive valid
-we should expect out-of-sample validation
- Data coherent
-all information should be in the model.
Nothing left in the errors (white noise errors).
- Encompassing
-our model should explain earlier models.

RS Chapter 6

Model Selection Strategies

FAQ in practice:
- Should I include all the variables in the database in my model?
- How many explanatory variables do I need in my model?
- How many models do I need to estimate?
- What functional form should I be using?
- What are interactive dummies do I need them?
- Which regression model will work best and how do I arrive at it?

Model Selection Strategies: Some Concepts

The term diagnostic testing describes how we test assumptions behind
the model. In our case, assumptions (A1)-(A5) in the CLM
The term parameter testing describes how we test economic H0s.
There are several model-selection methods. We will consider two:
- Specific to General
- General to Specific

RS Chapter 6

Model Selection Strategies : Specific to General

Begin with a small theoretical model
for example, the CAPM
y = X + .
Estimate the model
say, using OLS
Do some diagnostic testing
If the assumptions do not hold, then use:
- More advanced econometrics
GLS instead of OLS?
- A more general model
APT?
Test economic hypotheses on the parameters Is size significant?
This strategy is known as specific to general, Average Economic Regression
(AER).

Model Selection Strategies: General to Specific

Begin with a general unrestricted model (GUM), which nests restricted
models and, thus, allows any restrictions to be tested. Say:
y = X + Z + .
Then, reduction of the GUM starts. Mainly using t-tests, and F-tests, we
move from the GUM to a smaller, more parsimonious, specific model.
If competing models are selected, encompassing tests or information
criteria (AIC, BIC) can be used to select a final model. This is the
discovery stage. After this reduction, we move to:
y = X + .
Creativity is needed for the specification of a GUM. Theory and
empirical evidence play a role in designing a GUM. Estimation of the
GUM should be feasible from the available data.

RS Chapter 6

Model Selection Strategies: General to Specific

General-to-Specific Method:
Step 1 - First ensure that the GUM does not suffer from any diagnostic
problems. Examine the residuals in the general model to ensure that
they possess acceptable properties. (That is, test for heteroskedasticity,
non-normality, incorrect functional form, etc.).
Step 2 - Test the restrictions implied by the specific model against the
general model either by exclusion tests or other tests of linear
restrictions.
Step 3 - If the restricted model is accepted, test its residuals to ensure
that this more specific model is still acceptable on diagnostic grounds.
This strategy is called general to specifics (gets), LSE, TTT (Test, test,
test). It was pioneered by Sargan (1964).

Model Selection Strategies: General to Specific

The role of diagnostic testing is two-fold.
- In the discovery steps, the tests are being used as design criteria.
Testing plays the role of checking that the original GUM was a good
starting point after the GUM has been simplified.
- In the context of model evaluation, the role of testing is clear cut.
Suppose you use the model to produce forecasts. These forecasts can
be evaluated with a test. This is the critical evaluation of the model.

John Dennis Sargan (1924 1996, England)

RS Chapter 6

Modeling Strategies: Pre-testing

A special case of omitted variables.
- First, a researcher starts with an unrestricted model (U):
y = X + .
(U)
- Then, based on (preliminary) tests say, an F-test- a researcher
decides to use restricted estimator. That is,
y = X + .
s.t. R = q
(R)
- We can think of the estimator we get from estimating R as:
bPT = I{0,c}(F) b* + I{c,}(F) b,
where I{0,c} is an indicator function:
I{0,c}(F) =1 if F-stat not in the rejection region say, F<c-,
I{c,}(F) =0, otherwise.
c : critical value chosen for testing the H0:R = q , using the F-stat.

Modeling Strategies: Pre-testing

The pre-test estimator is a rule which chooses between the restricted
estimator, b*, or the OLS estimator, b:
bPT = I{0,c}(F) b* + I{c,}(F) b.
where b* = b - (XX)-1R[R(XX)-1R]-1(Rb - q)
Two situations:
(1) H0:R = q is true. The F-test will incorrectly reject H0 % of the
time. That is, in % of the repeated samples, OLS b => No bias,
inefficient estimator.
(2) H0:R = q is false. The F-test will correctly reject H0 a % of times
equal to the power of the test. That is, (100- )% of the time, R=q
will be incorrectly imposed, b* will be used: => bias!

RS Chapter 6

Modeling Strategies: Pre-testing

The failure of the OLS estimator to have the properties under
correct specification is called pre-test bias.
Pre-testing (also called sequential estimation, data mining) is common in
practice. In general, it is ignored and not even acknowledged.
Main argument to ignore pre-testing: We need some assumptions to
decide which variables are included in a model. Is the probability that
pre-testing yields an incorrect set of X greater than the probability of
selecting the correct assumption?
The LSE methodology does not see pre-testing in the discovery
stage as a problem. For the LSE method, pre-testing at that stage is
part of the process of discovery.

Modeling Strategies: Pre-testing

Checking the MSE of bPT, b* and b helps to evaluate the problem
Practical advise: Be aware of the problem. Do not rely solely on stats
to select a model use economic theory as well.
Do not use same sample evidence to generate an H0 and to test it!
Example: The Fama-French factors have been discovered using the
CRSP/Compustast database for a long, long time. Thus, testing the
Fama-French factors using the CRSP/Compustat is not advisable!
(You can test them with another dataset, for example, get
international data.)

RS Chapter 6

Modeling Strategies: Pre-testing

Taken from Wallace (1977, AJAE) --non-centrality parameter .

MSE (b1) = MSE (b) = OLS MSE
MSE (^1) = MSE (b*) = Restricted OLS MSE
MSE (*1) = MSE (bPT) = Pretest estimator MSE

Modeling Strategies: Properties

A modeling strategy is consistent if its probability of finding the true
model tends to 1 as T -the sample size- increases.
Properties for strategies
(1) Specific to General
- It is not consistent if the original model is incorrect.
- It need not be predictive valid, data coherent, & encompassing.
- No clear stopping point for an unordered search.
(2) General to Specific
- It is consistent under some circumstances. But, it needs a large T.
- It uses data mining, which can lead to incorrect models for small T.
- The significance levels are incorrect. This is the problem of mass
significance.

RS Chapter 6

Modeling Strategies: Mass significance

We perform k different tests each with a nominal significance level of :
= Prob (Rejecting for a given test |H0 for this test is true)
The overall significance of the test procedure is, however, given by
* = Prob (Rejecting at least one test | all H0 are true).
The probability of rejecting at least one H0 is obviously greater than
of rejecting a specific test. This is the problem of mass significance.
Two cases
(1) Independent tests
(1 *) = (1 )k

* = 1 (1 )k
= 1 (1 *)1/k

Modeling Strategies: Mass significance

Two cases
(1) Independent tests
= 1 (1 *)1/k
(2) Dependent tests (Bonferroni bounds)
* k
*/k
These are close to the independent values for small , but can differ
substantially for large .
Example: =0.05 and k=5
* =0.05 and k=5
* =0.20 and k=5

=> *(Indep)=.23 and *(Dep)=.25

=> (Indep) =.0102 and (Dep) =.01
=> (Indep) =.044 and (Dep) =.04

RS Chapter 6

Modeling Strategies: Mass significance

In repeated parametric testing (overall level 5%):
- Only accept variables as important when their p-values are less
than 0.001, preferably smaller
- Maybe look for other ways of choosing variables, say IC.
In repeated diagnostic testing (overall level 20%), we should only accept
there is no misspecification if
- All p-values are greater than 0.05, or
- Most p-values are greater than 0.10 with a few in the range
0.02 to 0.10

Modeling Strategies: Information Criteria

ICs are equal to the estimated variance or the log-likelihood
function plus a penalty factor, that depends on k. Many ICs:
- Theil Information Criterion (Adjusted R2)
Adj. R2 = 1 - [(T-1)/(T-k)](1 R2) = 1- [(T-1)/(T-k)] RSS/TSS
=> maximizing Adjusted R2 <=> minimizing s2
- Akaike Information Criterion (AIC)
AIC = -2/T(ln L k) = -2 ln L/T + 2 k/T
=> if normality AIC = ln(ee/T) + (2/T) k

(+constants)

- Bayes-Schwarz Information Criterion (BIC)

BIC = -(2/T ln L [ln(T)/T] k)
=> if normality AIC = ln(ee/T) + [ln(T)/T] k (+constants)

10

RS Chapter 6

Modeling Strategies: Information Criteria

The goal of these criteria is to provide us with an easy way of
comparing alternative model specifications, by ranking them.
General Rule: The lower the IC, the better the model. For the
previous ICs, then choose model to minimize s2J, AICJ,, or BICJ.
Some remarks about ICs:
- They are used for ranking. The raw value tends to be ignored.
- They have two components: a goodness of fit component based on
lnL and a model complexity component the penalty based on k.
- Different penalties, different ICs.
- Some authors do not scale the ICs by T, like we do above. If raw
values are irrelevant, this is not an issue.

Modeling Strategies: Information Criteria

ICs are not test statistics. They do not test a model. But, they are
statistics i.e., they are functions of RVs- with sampling distributions.
We would like these statistics i.e., the ICs- to have good properties.
For example, if the true model is being considered among many, wed
want the information criteria to select it. This can be done on average
(unbiased) or as T increases (consistent).
Usually, inconsistency is a fatal flaw for a statistics. But, in model
selection, it is very likely that the true DGP is not among the models
considered. That is, inconsistency may not matter in these cases.
Information? It refers to Kullback and Leiblers (1951) information
discrepancy measure, used in information theory (in telecom literature).

11

RS Chapter 6

Modeling Strategies: IC - K-L divergence

Kullback and Leiblers (1951) information discrepancy measure is also
called information divergence.
Information divergence measures the difference between two
probability distributions P and Q; where P represents the true DGP.
Here, we look at the difference between the expected values of Y
when Y is determined by: (i) P and (ii) some Q model.
Minimizing the K-L divergence, when considering several Q models,
gets us close to the true DGP.
But, expected values are unobservable, they need to be estimated.
The information associated with Y is given by L i.e., the joint pdf.
The AIC uses ln L evaluated at the estimated parameter values.

Modeling Strategies: IC AIC and BIC

Some results regarding AIC and BIC.
- AIC and Adjusted R2 are not consistent.
- AIC is conservative i.e., it tends to over-fit; that is, choose too large
models.
- AIC selects the model that minimizes the leave-one-out crossvalidation MSE for cross-sectional data. In time series, it selects the
model that minimizes the out-of-sample one-step ahead forecast
MSE.
- BIC is more parsimonious than AIC. It penalizes the inclusion of
parameters more (kBIC kAIC).
- BIC is consistent in hierarchical (gets) autoregressive models.

12

RS Chapter 6

Modeling Strategies: IC AIC and BIC

There are several small sample corrections of ICs. But, asymptotically
they have no impact. Because of this feature, using corrected ICs is
not a bad choice.
Comparing models based on ICs can be expensive
In unstructured problems (natural order to the hypotheses to be
tested), there is a huge number of potential combinations to
investigate: 2m possible models for m candidate variables.
For the Lovell (1983) database, that would be 240 1012 models.
Even at a USD 0.001 per model, that would cost USD 1 billion.

Non-nested Models and Tests

Sometimes, we have two rival models to choose between, where
neither can be nested within the other -i.e., neither is a restricted
version of the other.
Example:
Model 1
Model 2

Y = X + W +
Y = X + Z +

If the dependent variable is the same in both models (as is the case
here), we can simply use Adjusted-R2 to rank the models.
We can also use AIC and/or BIC.

13

RS Chapter 6

Non-nested Models and Tests

Alternative approach: Encompassing --Mizon and Richard (1986).
(1) Form a composite or encompassing model that nests both rival
models --say, Model 1 and Model 2.
(2) Test the relevant restrictions of each rival model against it.
Assuming the restrictions cannot be rejected, we would prefer the
model with the lower F statistic for the test of restrictions.
Example: From Model 1 and Model 2, the encompassing model is:
Encompassing Model:
Y = X + W + Z +
We now test, separately, the hypotheses (1) =0 and (2) =0
Note: We test a hybrid model. Also, multicollinearity may appear.

Non-nested Models: J-test

Davidson-MacKinnon (1981)s J-test .
We start with two non-nested models. Say,
Model 1
Y = X +
Model 2
Y = Z +
Idea: If Model 2 is true, then the fitted values from the Model 1, when
added to the first equation, should be insignificant.
Steps:
(1) Estimate Model 1 => obtain fitted values: Xb.
(2) Add Xb to the list of regressors in Model 2. => Y = Z + Xb +
(3) Do a t-test on . A significant t-value would be evidence against
Model 2 and in favour of Model 1.
(4) Repeat the procedure for the models the other way round.
(5) Rank the models on the basis of this test.

14

RS Chapter 6

Non-nested Models: J-test

Note: As some of the regressors in step (3) are stochastic, the t-test is
not strictly valid. Davidson and MacKinnon (1981) show that it is
asymptotically valid.
Of course neither of these methods may not reject both models.
This is possible in small samples, even if one model, say Model 2, is
true.
Rejecting both models is not unusual. McAleers (1995), in a survey,
reports that out of 120 applications all models were rejected 43 times.
One would also want to examine the diagnostic test results when
choosing between two models.

Non-nested Models: J-test

J-test does not work very well when we compare 3 or more models.
Encompassing interpretation of the J-test .
Lets encompass both models:
Y = (1-) Z + X +
Under H0 (Model 2 is true): =0.
Under H1 (Model 1 is true): =1.
Nice model, but unfortunately, this model is not intrinsic linear!

15

RS Chapter 6

Non-nested Models: J-test - Application

We want to test
H0: y = X + 0
(additive)
vs
H1: ln y = (ln X) + 1
(multiplicative)
We look at the Davidson-MacKinnon J-test
Step 1: OLS on H1: get c
OLS y = 0 exp{ln(X) c}+ X +
=> t-test on 0
Step 2: OLS on H0: get b
=> t-test on 1
OLS ln y = ln(Z) + 1 Xb +
Situations:
(1) Both OK: 0=0 and 1=0. => get more data
(2) Only 1 is OK: 00 and 1=0 (multiplicative is OK);
1 0 and 0=0 (additive is OK)
(3) Both rejected: 00 and 1 0 => new model is needed.

Non-nested Models: J-test - Considerations

The J-test was designed to test non-nested models (one model is the
true model, the other is the false model), not for choosing competing
models the usual use of the test.
The J-test is likely to over reject the true (model) hypothesis when
one or more of the following features is present:
- a poor fit of the true model
- a low or moderate correlation between the regressors of the 2 models
- the false model includes more regressors than the correct model.
Davidson and MacKinnon (2004) state that the J-test will over-reject,
often quite severely in finite samples when the sample size is small or
where conditions (i) or (iii) above are obtained.

16

RS Chapter 6

Modeling Strategies: Significance level,

So far, we have assumed that the distribution of the test statistic say
the F-statistic-- under H0 is known exactly, so that we have what is
called an exact test.
Technically, the size of a test is is the supremum of the rejection
probability over all DGPs that satisfy H0. For an exact test, the size
equals the nominal level, i.e., the Prob[Type I error] = .
Usually, the distribution of a test is known only approximately
(asymptotically). In this case, we need to draw a distinction between the
nominal level (nominal size) of the test and the actual rejection probability
(empirical size), which may differ greatly from the nominal level.
Simulations are needed to gauge the empirical size of tests.

Modeling Strategies: A word about

Ronald Fisher, before computers, tabulated distributions. He used a
.10, .05, and .01 percentiles. These tables were easy to use and, thus,
those percentile became the de-facto standard for testing H0.
It is usual and convenient for experimenters to take 5% as a
standard level of significance. --Fisher (1934).
Given that computers are powerful and common, why is p = 0.051
unacceptable, but p = 0.049 is great? There is no published work that
provides a theoretical basis for the standard thresholds.
Rosnow and Rosenthal (1989): ... surely God loves .06 nearly as
much as .05.

17

RS Chapter 6

Modeling Strategies: A word about

Practical advise: In the usual Fishers null hypothesis (significance)
testing, significance levels, , are arbitrary. Make sure you pick one,
say 5%, and stick to it throughout your analysis or paper.
Report p-values, along with CIs. Search for economic significance.
Q: .10, .05, or .01 significance?
Many tables will show *, **, and *** to show .10, .05, and .01
significance levels. Throughout the paper, the authors will point out
the different significance levels. In these papers, it is not clear what
is the paper using for inference.
In a Neyman-Pearson world, we can think of these stars (or p-values)
as ways of giving weights to H0 relative to H1.

Modeling Strategies: A word about H0

In applied work, we only learn when we reject H0. Failing to reject
H0 provides almost no information about the state of the world.
Thus, failing to reject H0 does not rule out an infinite number of
other competing research hypotheses.
Null hypothesis significance testing is asymmetric: if the test statistic
is too large for a given H0 then the H0 is rejected; but if the test
statistic is not too large then the H0 is not automatically accepted.
It is dangerous to accept the conclusion from a non-rejected H0.
But, it is common. Eight of the twenty (40%) articles in the American
Political Science Review Volume 91 (1997), that used a H0, drew
substantive conclusions from a fail to reject decision.

18

RS Chapter 6

Modeling Strategies: A word about H0

In applied work, we only learn when we reject H0; say, when the pvalue<. But, rejections are of two types:
- Correct ones, driven by the power of the test,
- Incorrect ones, driven by Type I Error (statistical accident, luck).
It is important to realize that, however small the p-value, there is
always a finite chance that the result is a pure accident. At the 5%
level, there is 1 in 20 chances that the rejection of H0 is just luck.
Since negative results are very difficult to publish (publication bias) this
means that an unknown but possibly large number of false claims are
taken as truths.
Example from Lecture 4: If 0.05, proportion of false H0=10%,
and power=.50, 47.4% of rejections are true H0 -i.e., false positives.

Model Selection Methods: Summary

Eight literature strands can be delineated:
(1) Simple-to-general: Anderson (1962), Hendry and Mizon (1978), and
Hendry (1979), for critiques;
(2) Retaining the general model: Yancey and Judge (1976), and Judge and
Bock (1978);
(3) Testing Theory-based models: Hall (1978), criticized by Davidson and
Hendry (1981), and Hendry and Mizon (2000); Stigum (1990) proposes
a formal approach;
(4) Other rules for model selection, such as:
- step-wise regression: Leamer (1983a), for a critical appraisal
- optimal regression: algorithm to maximize the Adj-R2 with a
specified set of regressors. See Coen, Gomme and Kendall (1969);

19

RS Chapter 6

Model Selection Methods: Summary

Eight literature strands can be delineated (continuation):
(5) Model comparisons, often based on non-nested hypothesis tests or
encompassing: Cox (1961, 1962), Pesaran (1974), and the survey in
Hendry and Richard (1989);
(6) Model selection by information criteria: Schwarz (1978), Hannan and
Quinn (1979), Amemiya (1980);
(7) Bayesian model comparisons: Leamer (1978) and Clayton, Geisser and
Jennings (1986);
(8) Gets: Anderson (1962), Sargan (1973, 1981), Hendry (1979), and
White (1990).

Criteria for Model Selection: Judgement Call

In the end, judgment must be used in weighing up various criteria:
- The Economic Criterion are the estimated parameters plausible?
(Economic Significance)
- The First Order Statistical Criterion does the model provide a
good fit (in-sample) with statistically significant parameter estimates?
- The Second Order Statistical Criterion is the model generally free
of misspecification problems as evidenced in the diagnostic tests?
- The Out of Sample Predictive Criterion does the model provide
good out of sample predictions?

20