chapter 10: multicollinearity - hem | karlstads universitet consequences of heteroskedasticity ols...

56
Chapter 10: Multicollinearity Chapter 10: Multicollinearity Iris Wang [email protected]

Upload: hoangkhanh

Post on 25-May-2018

225 views

Category:

Documents


6 download

TRANSCRIPT

Page 1: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Chapter 10: MulticollinearityChapter 10: Multicollinearity

Iris Wang

[email protected]

Page 2: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Econometric problemsEconometric problems

Page 3: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

MulticollinearityWhat does it mean? A high degree of correlation amongst theexplanatory variablesWhat are its consequences? It may be difficult to separate outWhat are its consequences? It may be difficult to separate outthe effects of the individual regressors. Standard errors maybe overestimated and t‐values depressed. Note a symptom may be high R2 but low t valuesNote: a symptom may be high R2 but low t‐valuesHow can you detect the problem? Examine the correlationmatrix of regressors ‐ also carry out auxiliary regressions

t thamongst the regressors. Look at the Variance‐inflating factor (VIF)

NOTENOTE: be careful not to apply t tests mechanically without checking for multicollinearitymulticollinearity is a data problem, not a misspecification problemmulticollinearity is a data problem, not a misspecification problem

Page 4: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Variance‐inflating factor (VIF)Variance inflating factor (VIF)

Multicollinearity inflates the variance of anMulticollinearity inflates the variance of an estimator 

VIF = 1/(1 R 2)VIFJ = 1/(1‐RJ2)

where RJ2 measures the R2 from a regression of X h h X i bl /Xj on the other X variable/s

⇒serious multicollinearity problem  if VIFJ>5

Page 5: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Econometric problemsEconometric problems

Page 6: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

HeteroskedasticityHeteroskedasticity

What does it mean? The variance of the error term is not t tconstant

What are its consequences? The least squares results l ffi i d d lare no longer efficient and t tests and F tests results may 

be misleading

H d t t th bl ? Pl t th id l i tHow can you detect the problem? Plot the residuals against each of the regressors or use one of the more formal tests

Ho can e remed the problem? Respecif the modelHow can we remedy the problem? Respecify the model –look for other missing variables; perhaps take logs or choose some other appropriate functional form; or make sure relevant variables are expressed “per capita”p p p

Page 7: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

The Homoskedastic CaseThe Homoskedastic Case

Page 8: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

The Heteroskedastic CaseThe Heteroskedastic Case

Page 9: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

The consequences of heteroskedasticity

OLS estimators are still unbiased (unless there are also i d i bl )omitted variables)

However OLS estimators are no longer efficient or minimum varianceminimum varianceThe formulae used to estimate the coefficient standard errors are no longer correct• so the t-tests will be misleading• confidence intervals based on these standard errors will be

wrong

Page 10: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Detecting heteroskedasticityDetecting heteroskedasticity

Visual inspection of scatter diagram or the residuals

Goldfeld‐Quandt test suitable for a simple form of heteroskedasticitysuitable for a simple form of heteroskedasticity

Page 11: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Goldfeld‐Quandt test (JASA, 1965)Goldfeld Quandt test (JASA, 1965)

P. 382, Suppose it looks as if σ i = σ XiP. 382, Suppose it looks as if σui   σuXii.e. the error variance is proportional to the square of one of the X’sqRank the data according to the variable and conduct an F test using RSS2/RSS12 1

where these RSS are based on regressions using the first and last [n‐c]/2 observations [c is a 

l f d ll b f ]central section of data usually about 25% of n]Reject H0 of homoskedasticity if Fcal > Ftables

Page 12: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

RemediesRemedies

Respecification of the modelRespecification of the modelInclude relevant omitted variable(s)Express model in log-linear form or some other p gappropriate functional formExpress variables in per capita form

Where respecification won’t solve the problem use robust Heteroskedastic Consistent Standard Errors (due to Hal White Econometrica 1980)Errors (due to Hal White, Econometrica 1980)

Page 13: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Basic Econometrics, Spring 2012

Chapter 11: Heteroskedasticity

Iris Wangg

[email protected]

1

Page 14: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Chapter 11: Heteroskedasticity

Definition:Heteroskedasticity occurs when the constant variance assumption, i.e. Var(ui|Xi)= σ2, fails. This happens when variance of the error term (ui) changes across different values of Xi. Heteroskedasticity is present if theof Xi.

Example: Savingsi=α0+α1income+ui

Heteroskedasticity is present if thevariance of unobserved factorsaffecting savings (ui) increases with income

- Higher variance of ui for higherHigher variance of ui for higherincome

2

Page 15: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Ch t 11 H t k d ti itChapter 11: Heteroskedasticity

Outline1. Consequences of Heteroskedasticity2. Testing for Heteroskedasticity

3

Page 16: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

1. Consequences of Heteroskedasticity

• OLS is unbiased and consistent under the following 4 assumptions:p– Linear in parameters

– Random sampling

No perfect collinearity– No perfect collinearity

– Zero conditional mean (E(u|X)=0)

• Homoskedasticity assumption (MLR.4) stating  constant error variance (Var(u|X)= σ2) plays no role in showing that OLS is unbiased & consistentunbiased & consistent

– Heteroskedasticity doesn’t cause bias or inconsistency in OLS iestimators

4

Page 17: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

1. Consequences of Heteroskedasticity ‘cntd• However, estimators of variances, Var(βj) are biased without 

homoskedasticityOLS standard errors are biased– OLS standard errors are biased

– Standard confidence interval, t, and F statistics which are based on standard errors are no longer valid.

& l h & d b– t & F statistics no longer have t & F distribution  resp.

– And this is not resolved in large samples

• OLS is no longer BLUE and asymptotically efficientg y p y– It is possible to find estimates that are more efficient than OLS (e.g. 

GLS, Generalized Least Squares)

• Solutions involve using:• Solutions  involve using:i. Generalized least squares (GLS)

ii. Weighted least squares (WLS) is a special case of GLS, p.373

5

Page 18: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Weighted Least Squares (WLS)g q

• Aim: to specify the form of heteroskedasticity detected and use weighted least squares estimatorleast squares estimator. 

– If we have correctly specified the form of the variance, then WLS is more efficient than OLS

If d f f i WLS ill b bi d b t it i– If we used wrong form of variance, WLS will be biased but it is generally consistent as long as E(u|X)=0. 

– But, efficiency of WLS is not guaranteed when using wrong form of ivariance.

• We use this to transform the original regression equation with homoskedastic error term

i.e. the bias will improvewith large N

homoskedastic error term

6

Page 19: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

2. Testing for Heteroskedasticity• Why test for heteroskedasticity?

– First, unless there is evidence of heteroskedasticity, many prefer to use the usual t under OLSthe usual t under OLS

– This is because the usual t statistics have exact t distribution under the assumptions of homoskedasticity & normally distributed errors. 

Second if heteroskedasticity is present it is possible to obtain better– Second, if heteroskedasticity is present, it is possible to obtain better estimator than OLS when the form of heteroskedasticity is known. 

• In the regression model:Y= β0+β1x1+…+βkxk +u

• We assume that E(u| x1, …xk )=0  OLS is unbiased and consistentconsistent.

• In order to test for violation of the homoskedasticityassumption, we want to test the null hypothesis:

Ho: Var(u| x1, …, xk )=σ27

Page 20: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

2. Testing for Heteroskedasticity ‘cntd

• To test the null hypothesis above, we test whether expected value of u2 is related to one or more of the explanatory variablesvalue of u is related to one or more of the explanatory variables.

• If we reject Ho, then heteroskedasticity is a problem & needs to be solved.

• Two types heteroskedasticity tests:

A. Goldfeld‐Quandt Test for heteroskedasticity, p.382 

hi ’ l k d i iB. White’s General Heteroskedasticity Test, p.386

• Once we reject Ho of homoskedasticity, we should treat the heteroskedasticity problemheteroskedasticity problem

8

Page 21: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

B. White heteroskedasticity test

• homoskedasticity assumption, Var(u| X)=σ2 , can be replacedwith weaker assumption that u2 is uncorrelated with:

All the independent variables (x )– All the independent variables (xj)

– Their squared terms (x2j) and

– Their cross‐products (xj xh for all h≠j)

• Under this weaker assumption, OLS standard errors and test statistics are asymptotically valid

• White heteroskedasticity test is motivated by this assumption• White heteroskedasticity test is motivated by this assumption. For e.g. for k=3,û2= δ0+ δ1x1+ δ2x2+ δ3x3+δ4x21  +δ5x22 + δ6x23+δ7x1x2 + δ8x1x3+ δ9x2x3 +v

• White test is F statistics for testing all δj, except δ0, are zero.

• Limitation: it consumes degrees of freedom (for k=3, we needed 9 gvariables)

9

Page 22: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Basic EconometricsBasic Econometrics

AutocorrelationAutocorrelation

Iris Wang

[email protected]

Page 23: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Econometric problemsEconometric problems

Page 24: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Topics to be coveredTopics to be covered

Overview of autocorrelationOverview of autocorrelation

First‐order autocorrelation and the Durbin‐Watson testWatson test

Higher‐order autocorrelation and the Breusch‐G dfGodfrey test

Dealing with autocorrelation

Examples and practical illustrations

Page 25: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Autocorrelated series and autocorrelatedAutocorrelated series and autocorrelateddisturbances

Page 26: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Overview of autocorrelationWhat is meant by autocorrelation ?The error terms are not independent from observation to observation – utdepends on one or more past values of udepe ds o o e o o e pas a ues o uWhat are its consequences? The least squares estimators are no longer “efficient” (i.e. they don’t have the lowest variance). More seriously autocorrelation may be a symptom of 

d l i ifi timodel misspecificationHow can you detect the problem? Plot the residuals against time or their own lagged values, calculate the Durbin‐Watson statistic or use some other tests of autocorrelation such as the Breusch‐Godfrey (BG) testHow can you remedy the problem?Consider possible model re‐specification of the model: a different functional form,missing variables lags etc If all else fails you could correct for autocorrelationmissing variables, lags etc.  If all else fails you could correct for autocorrelation by using the Cochrane‐Orcutt procedure or Autoregressive Least Squares

Page 27: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

First‐order autocorrelation

Page 28: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

The sources of autocorrelation

Page 29: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

The consequences of autocorrelationThe consequences of autocorrelation

Page 30: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Detecting autocorrelationDetecting autocorrelation

Page 31: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

The Durbin‐Watson testThe Durbin Watson test

Page 32: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

More on the Durbin‐Watson statisticMore on the Durbin Watson statistic

Page 33: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Using the Durbin‐Watson statisticUsing the Durbin Watson statistic

Page 34: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Durbin‐Watson critical valuesDurbin Watson critical values

Page 35: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

The Breusch‐Godfrey (BG) testThe Breusch Godfrey (BG) test

Page 36: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

The Breusch‐Godfrey test continuedThe Breusch Godfrey test continued

Page 37: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Dealing with autocorrelationDealing with autocorrelation

How should you deal with a problem of autocorrelation?y p

Consider possible re‐specification of the model:a different functional form,the inclusion of additional explanatory variables,the inclusion of lagged variables (independent and dependent) 

If all else fails you can correct for autocorrelation by using the Autoregressive Least Squaresg g q

Page 38: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Quick questions and answers

Page 39: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Question 1: 

What is the problem of autocorrelation?

Page 40: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Answer: 

Autocorrelation is the problem whereAutocorrelation is the problem where 

the disturbances in a regression model are 

not independent of one another 

from observation to observation 

(it is mainly a problem for models 

estimated using time series data)estimated using time series data)

Page 41: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Question 2: Is serial correlation the same as autocorrelation?autocorrelation?

Page 42: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Answer:Answer: Yes. Serially correlated disturbances or errors are the same as autocorrelatedonesones.

Page 43: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Question 3: Wh i b AR(1) ?What is meant by AR(1) errors?

Page 44: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Answer: This means that the errors or disturbances follow a first‐order d stu ba ces o o a st o deautoregressive pattern

+ut = ρut‐1 + εt

Page 45: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Question 4:Question 4: What is the best known test for AR(1) disturbances?

Page 46: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Answer: The Durbin‐Watson test. The nullThe Durbin Watson test. The null hypothesis of no autocorrelation ( i l i d d ) i H 0(serial independence) is H0 ρ=0

Page 47: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Question 5: What is the range of possibleWhat is the range of possible values for the DW statistic?

Page 48: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Answer:0≤ DW ≤ 4.   If there is no autocorrelation youIf there is no autocorrelation you would expect to get a DW stat of around 2.

Page 49: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Question 6: What are the three main limitations of the DW test?limitations of the DW test?

Page 50: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Answer: 1. It only tests for AR(1) errors2 It has regions where the test is2. It has regions where the test is inconclusive (between dL and dU) 3. The DW statistic is biased towards 2 in models with a lagged dependentin models with a lagged dependent variable.

Page 51: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Question 7: How do you test for higher order a tocorrelated errors?autocorrelated errors?

Page 52: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Answer: Using the Breusch‐Godfrey (BG) testtest

Page 53: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Question 9: How do I know what order of autocorrelation to test for?autocorrelation to test for?

Page 54: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Answer: With annual data a first order test isWith annual data a first order test is probably enough, with quarterly or 

thl d t h k f AR(4)monthly data check for AR(4) or AR(12) errors if you have enough data. If in doubt repeat the test for a number of different maximum lagsnumber of different maximum lags.

Page 55: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Question 10: What should I do if my model exhibitsWhat should I do if my model exhibits autocorrelation?

Page 56: Chapter 10: Multicollinearity - Hem | Karlstads universitet consequences of heteroskedasticity OLS estimators are still unbiased (unless there are also omiditted varibl )iables) However

Answer: On the first instance try model re‐On the first instance try model respecification (additional lagged values f i bl l t f ti fof variables or a log transformation of 

some series). If this doesn’t deal with the problem use Autoregressive Least Squares rather than OLS estimationSquares rather than OLS estimation.