Créer une présentation
Télécharger la présentation

Télécharger la présentation
## 405 ECONOMETRICS

- - - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - - -

**405ECONOMETRICS**Chapter # 11: HETEROSCEDASTICITY: WHAT HAPPENS IF THE ERROR VARIANCE IS NONCONSTANT? By Domodar N. Gujarati Prof. M. El-Sakka Dept of Economics Kuwait University**As in Chapter 10, we seek answers to the following**questions: 1. What is the nature of heteroscedasticity? 2. What are its consequences? 3. How does one detect it? 4. What are the remedial measures?**THE NATURE OF HETEROSCEDASTICITY**One of the important assumptions of the classical linear regression model is that the variance of each disturbance term ui, conditional on the chosen values of the explanatory variables, is some constant number equal to σ2. This is the assumption of homoscedasticity, or equal (homo) spread (scedasticity), that is, equal variance. Symbolically, Eu2i = σ2i = 1, 2, . . . , n (11.1.1) Look at Figure 11.1. In contrast, consider Figure 11.2, the variances of Yiare not the same. Hence, there is heteroscedasticity. Symbolically, Eu2i = σ2i(11.1.2) Notice the subscript of σ2, which reminds us that the conditional variances of ui(= conditional variances of Yi) are no longer constant.**Assume that in the two-variable model Yi= β1 + β2Xi+ ui, Y**represents savings and X represents income. Figures 11.1 and 11.2 show that as income increases, savings on the average also increase. But in Figure 11.1 the variance of savings remains the same at all levels of income, whereas in Figure 11.2 it increases with income. It seems that in Figure 11.2 the higher income families on the average save more than the lower-income families, but there is also more variability in their savings. There are several reasons why the variances of uimay be variable, some of which are as follows. 1. Following the error-learning models, as people learn, their errors of behavior become smaller over time. In this case, σ2iis expected to decrease. As an example, consider Figure 11.3, which relates the number of typing errors made in a given time period on a test to the hours put in typing practice.**2. As incomes grow, people have more discretionary income**and hence more scope for choice about the disposition of their income. Hence, σ2iis likely to increase with income. Similarly, companies with larger profits are generally expected to show greater variability in their dividend policies than companies with lower profits. 3. As data collecting techniques improve, σ2iis likely to decrease. Thus, banks that have sophisticated data processing equipment are likely to commit fewer errors in the monthly or quarterly statements of their customers than banks without such facilities. 4. Heteroscedasticity can also arise as a result of the presence of outliers, (either very small or very large) in relation to the observations in the sample Figure 11.4. The inclusion or exclusion of such an observation, especially if the sample size is small, can substantially alter the results of regression analysis. Chile can be regarded as an outlier because the given Y and X values are much larger than for the rest of the countries. In situations such as this, it would be hard to maintain the assumption of homoscedasticity.**5. Another source of heteroscedasticity arises from**violating Assumption 9 of CLRM, namely, that the regression model is correctly specified, very often what looks like heteroscedasticity may be due to the fact that some important variables are omitted from the model. But if the omitted variables are included in the model, that impression may disappear. 6. Another source of heteroscedasticity is skewness in the distribution of one or more regressors included in the model. Examples are economic variables such as income, wealth, and education. It is well known that the distribution of income and wealth in most societies is uneven, with the bulk of the income and wealth being owned by a few at the top. 7. Other sources of heteroscedasticity: As David Hendry notes, heteroscedasticity can also arise because of: (1) incorrect data transformation (e.g., ratio or first difference transformations) (2) incorrect functional form (e.g., linear versus log–linear models).**Note that the problem of heteroscedasticity is likely to be**more common in cross-sectional than in time series data. In cross-sectional data, members may be of different sizes, such as small, medium, or large firms or low, medium, or high income. In time series data, on the other hand, the variables tend to be of similar orders of magnitude. Examples are GNP, consumption expenditure, savings.**THE METHOD OF GENERALIZED LEAST SQUARES (GLS)**If ui is hetroschedastic, βˆ2 is no longer best, although it is still unbiased. Intuitively, we can see the reason from Table 11.1. As the table shows, there is considerable variability in the earnings between employment classes. If we were to regress per-employee compensation on the size of employment, we would like to make use of the knowledge that there is considerable interclass variability in earnings. Ideally, we would like to devise the estimating scheme in such a manner that observations coming from populations with greater variability are given less weight than those coming from populations with smaller variability. Examining Table 11.1, we would like to weight observations coming from employment classes 10–19 and 20–49 more heavily than those coming from employment classes like 5–9 and 250–499, for the former are more closely clustered around their mean values than the latter, thereby enabling us to estimate the PRF more accurately.**Unfortunately, the usual OLS method does not follow this**strategy, but a method of estimation, known as generalized least squares (GLS), takes such information into account explicitly and is therefore capable of producing estimators that are BLUE. To see how this is accomplished, let us continue with the now-familiar two-variable model: Yi= β1 + β2Xi+ ui(11.3.1) which for ease of algebraic manipulation we write as Yi= β1X0i+ β2Xi+ ui(11.3.2) where X0i= 1 for each i. Now assume that the heteroscedastic variances σ2i are known. Divide through by σito obtain: which for ease of exposition we write as Y*i= β*1X*0i+ β*2X*i+ u*i(11.3.4)**We use the notation β∗1 and β∗2, the parameters of the**transformed model, to distinguish them from the usual OLS parameters β1 and β2. What is the purpose of transforming the original model? To see this, notice the following feature of the transformed error term: which is a constant. That is, the variance of the transformed disturbance term u*i is now homoscedastic.**The finding that it is u* that is homoscedastic suggests**that if we apply OLS to the transformed model (11.3.3) it will produce estimators that are BLUE. This procedure of transforming the original variables in such a way that the transformed variables satisfy the assumptions of the classical model and then applying OLS to them is known as the method of generalized least squares (GLS).In short, GLS is OLS on the transformed variables that satisfy the standard least-squares assumptions. The estimators thus obtained are known as GLS estimators, and it is these estimators that are BLUE.**The actual mechanics of estimating β*1 and β*2 are as**follows. First, we write down the SRF of (11.3.3) Yi/σi= βˆ*1X0i/σi+ βˆ*2Xi/σi + uˆi/σi or Y*i= βˆ*1X*0i+ βˆ*2X*i+ uˆ*i(11.3.6) Now, to obtain the GLS estimators, we minimize Σuˆ2*i=Σ(Y*i− βˆ*1X*0i− β*2X*i)2 that is, The actual mechanics of minimizing (11.3.7) follow the standard calculus techniques the GLS estimator of β*2 is**and its variance is given by**where wi= 1/σ2i .**The Difference between OLS and GLS**Recall from Chapter 3 that in OLS we minimize: Σuˆ2i=Σ(Yi− βˆ1 − βˆ2Xi)2 (11.3.10) but in GLS we minimize the expression (11.3.7), which can also be written as Σwiuˆ2i= Σwi(Yi− βˆ*1X0i− βˆ*2Xi)2 (11.3.11) where wi= 1/σ2i Thus, in GLS we minimize a weighted sum of residual squareswith wi= 1/σ2iacting as the weights, but in OLS we minimize an unweighted or (what amounts to the same thing) equally weighted RSS. As (11.3.7) shows, in GLS the weight assigned to each observation is inversely proportional to its σi, that is, observations coming from a population with largerσiwill get relatively smaller weight and those from a population with smaller σiwill get proportionately larger weight in minimizing the RSS (11.3.11). To see the difference between OLS and GLS clearly, consider the hypothetical scattergram given in Figure 11.7.**In the (unweighted) OLS, each uˆ2iassociated with points A,**B, and C will receive the same weight in minimizing the RSS. Obviously, in this case the uˆ2iassociated with point C will dominate the RSS. But in GLS the extreme observation C will get relatively smaller weight than the other two observations. Since (11.3.11) minimizes a weighted RSS, it is appropriately known as weighted least squares (WLS), and the estimators thus obtained and given in (11.3.8) and (11.3.9) are known as WLS estimators. But WLS is just a special case of the more general estimating technique, GLS. In the context of heteroscedasticity, one can treat the two terms WLS and GLS interchangeably.**CONSEQUENCES OF USING OLS IN THE PRESENCE OF**HETEROSCEDASTICITY As we have seen, both βˆ*2 and βˆ2 are (linear) unbiased estimators: In repeated sampling, on the average, βˆ*2 and βˆ2 will equal the true β2; that is, they are both unbiased estimators. But we know that it is βˆ*2 that is efficient, that is, has the smallest variance. What happens to our confidence interval, hypotheses testing, and other procedures if we continue to use the OLS estimator βˆ2? We distinguish two cases. OLS Estimation Allowing for Heteroscedasticity Suppose we use βˆ2 and use the variance formula given in (11.2.2), which takes into account heteroscedasticity explicitly. Using this variance, and assuming σ2iare known, can we establish confidence intervals and test hypotheses with the usual t and F tests? The answer generally is no because it can be shown that var (βˆ*2) ≤ var (βˆ2), which means that confidence intervals based on the latter will be unnecessarily larger. As a result, the t and F tests are likely to give us inaccurate results.**OLS Estimation Disregarding Heteroscedasticity**The situation can become serious if we not only use βˆ2 but also continue to use the usual (homoscedastic) variance formula given in (11.2.3) even if heteroscedasticity is present or suspected, whatever conclusions we draw or inferences we make may be very misleading.**DETECTION OF HETEROSCEDASTICITY**In most cases involving econometric investigations, heteroscedasticity may be a matter of intuition, educated guesswork, prior empirical experience, or sheer speculation. Some of the informal and formal methods are used for detecting heteroscedasticity. Most of these methods are based on the examination of the OLS residuals uˆisince they are the ones we observe, and not the disturbances ui. One hopes that they are good estimates of ui, a hope that may be fulfilled if the sample size is fairly large. Informal Methods Nature of the Problem. Very often the nature of the problem under consideration suggests whether heteroscedasticity is likely to be encountered. In cross-sectional data involving heterogeneous units, heteroscedasticity may be the rule rather than the exception. Thus, in a cross-sectional analysis involving the investment expenditure in relation to sales, rate of interest, etc., heteroscedasticity is generally expected if small-, medium-, and large-size firms are sampled together.**Graphical Method**If there is no a priori or empirical information about the nature of heteroscedasticity, in practice one can do the regression analysis on the assumption that there is no heteroscedasticity and then do an examination of the residual squared uˆ2ito see if they exhibit any systematic pattern. Although uˆ2iare not the same thing as u2i, they can be used as proxies especially if the sample size is sufficiently large. An examination of the uˆ2imay reveal patterns such as those shown in Figure 11.8. In Figure 11.8a we see that there is no systematic pattern between the two variables, suggesting that perhaps no heteroscedasticity is present in the data. Figure 11.8b to e, however, exhibits definite patterns. For instance, Figure 11.8c suggests a linear relationship, whereas Figure 11.8d and e indicates a quadratic relationship between uˆ2iand Yˆi. Using such knowledge, albeit informal, one may transform the data in such a manner that the transformed data do not exhibit heteroscedasticity.**Instead of plotting uˆ2iagainst Yˆi, one may plot them**against one of the explanatory variables Xi. A pattern such as that shown in Figure 11.9c, for instance, suggests that the variance of the disturbance term is linearly related to the X variable. Thus, if in the regression of savings on income one finds a pattern such as that shown in Figure 11.9c, it suggests that the heteroscedastic variance may be proportional to the value of the income variable. This knowledge may help us in transforming our data in such a manner that in the regression on the transformed data the variance of the disturbance is homoscedastic.**Formal Methods**Park Test.Park suggests that σ2iis some function of the explanatory variable Xi. The functional form he suggested was σ2i= σ2Xβi evi or ln σ2i= ln σ2 + β ln Xi+ vi(11.5.1) where viis the stochastic disturbance term. Since σ2iis generally not known, Park suggests using uˆ2ias a proxy and running the following regression: ln uˆ2i= ln σ2 + β ln Xi+ vi= α + β ln Xi+ vi(11.5.2) If βturns out to be statistically significant, it would suggest that heteroscedasticity is present in the data. If it turns out to be insignificant, we may accept the assumption of homoscedasticity. The Park test is a two stage procedure. In the first stage we run the OLS regression disregarding the heteroscedasticity question. We obtain uˆifrom this regression, and then in the second stage we run the regression (11.5.2).**Although empirically appealing, the Park test has some**problems. Goldfeld and Quandt have argued that the error term vientering into (11.5.2) may not satisfy the OLS assumptions and may itself be heteroscedastic. EXAMPLE 11.1 Relationship between compensation and productivity use the data given in Table 11.1 to run the following regression: Yi= β1 + β2Xi+ ui where Y = average compensation in thousands of dollars, X = average productivity in thousands of dollars, and i = ith employment size of the establishment. The results of the regression were as follows: Y ˆi= 1992.3452 + 0.2329Xi se = (936.4791) (0.0998) (11.5.3) t = (2.1275) (2.333) R2 = 0.4375**The residuals obtained from regression (11.5.3) were**regressed on Xi as suggested in Eq. (11.5.2), giving the following results: ln uˆ2i= 35.817 − 2.8099 ln Xi se = (38.319) (4.216) (11.5.4) t = (0.934) (−0.667) R2 = 0.0595 Obviously, there is no statistically significant relationship between the two variables. Following the Park test, one may conclude that there is no heteroscedasticity in the error variance**Glejser Test. Similar in spirit to the Park test. Glejser**suggests regressing |uˆi| on the X variable that is thought to be closely associated with σ2i. He used the following functional forms (viis the error term)**EXAMPLE 11.2Relationship Between Compensation and**Productivity: THE GLEJSER TEST Continuing with Example 11.1, the absolute value of the residuals obtained from regression(11.5.3) were regressed on average productivity (X), giving the following results: |uˆi| = 407.2783 − 0.0203Xi se = (633.1621) (0.0675) r2 = 0.0127 (11.5.5) t = (0.6432) (−0.3012) As you can see from this regression, there is no relationship between the absolute value ofthe residuals and the regressors, average productivity. This reinforces the conclusion basedon the Park test.**Goldfeld and Quandt point out that the error term vi has**some problems in that its expected value is nonzero, it is serially correlated and ironically it is heteroscedastic. An additional difficulty with the Glejser method is that models such as are nonlinear in the parameters and therefore cannot be estimated with the usual OLS procedure. Glejser has found that for large samples the first four of the preceding models give generally satisfactory results in detecting heteroscedasticity. As a practical matter, therefore, the Glejser technique may be used for large samples and may be used in the small samples strictly as a qualitative device to learn something about heteroscedasticity.**Spearman’s Rank Correlation Test.**rs= 1 − 6 (Σd2i/(n(n2 − 1))) (11.5.6) where di= difference in the ranks assigned to two different characteristics of the ith individual or phenomenon and n = number of individuals or phenomena ranked. Heteroscedasticity assuming Yi= β0 + β1Xi + ui , is Step 1. Regress the data on Y and X and obtain the residuals uˆi. Step 2. Rank both |uˆi| and Xi(or Yˆi) according to an ascending or descending order and compute rs given previously. Step 3. Assuming that the population rank correlation coefficient ρsis zero and n > 8, the significance of the samplerscan be tested by the t test as follows: with df = n − 2. If the computed t value exceeds the critical t value,we may accept the hypothesis of heteroscedasticity; otherwise we may reject it.**For models of more than one X variable,rscan be computed**between |uˆi| and each of the Xi and test for the significance by the t test given in (11.5.7). EXAMPLE 11.3**The capital market line (CML) of portfolio theory postulates**a linear relationship between expected return (Ei) and risk (as measured by the standard deviation, σ) of a portfolio as follows: Ei= βi+ β2σi Using the data in Table 11.2, the preceding model was estimated and the residuals from this model were computed. Since the data relate to 10 mutual funds of differing sizes and investment goals, a priori one might expect heteroscedasticity. To test this hypothesis, we apply therank correlation test. The necessary calculations are given in Table 11.2. Applying formula (11.5.6), we obtain rs= 1 − 6 (110 / (10(100 − 1))= 0.3333 (11.5.8) Applying the t test given in (11.5.7), we obtain t = (0.3333)(√8) / √1 − 0.1110 = 0.9998 (11.5.9) For 8 df this t value is not significant even at the 10% level of significance; the p value is 0.17. Thus, there is no evidence of a systematic relationship between the explanatory variable and the absolute values of the residuals, which might suggest that there is no heteroscedasticity.**Goldfeld-Quandt Test.**For simplicity, consider the usual two-variable model: Yi= β1 + β2Xi+ ui Suppose σ2iis positively related to Xias: σ2i= σ2X2i(11.5.10) where σ2 is a constant. Assumption (11.5.10) postulates that σ2iis proportional to the square of the X variable. If (11.5.10) is appropriate, it would mean σ2iwould be larger, the larger the values of Xi. If that turns out to be the case, heteroscedasticity is most likely to be present in the model. To test this explicitly, Goldfeld and Quandt suggest the following steps:**Step 1. Order or rank the observations according to the**values of Xi, beginning with the lowest X value. Step 2. Omit c central observations, where c is specified a priori, and divide the remaining (n − c) observations into two groups each of (n − c) / 2 observations. Step 3. Fit separate OLS regressions to the first (n − c) / 2 observations and the last (n − c) / 2 observations, and obtain the respective residual sums of squares RSS1and RSS2, RSS1 representing the RSS from the regression corresponding to the smaller Xivalues (the small variance group) and RSS2 that from the larger Xivalues (the large variance group). These RSS each have ((n− c) / 2) − k or ((n− c − 2k) / 2) df. where kis the number of parameters to be estimated, including the intercept. Step 4. Compute the ratio λ = RSS2/df / RSS1/df (11.5.11) λ of (11.5.10) follows theF distribution with numerator and denominator df each of (n− c − 2k)/2.**If in an application the computed λ (= F) is greater than**the critical F at the chosen level of significance, we can reject the hypothesis of homoscedasticity, that is, we can say that heteroscedasticity is very likely. EXAMPLE 11.4 To illustrate the Goldfeld–Quandt test, we present in Table 11.3 data on consumption expenditure in relation to income for a cross section of 30 families. Suppose we postulate that consumption expenditure is linearly related to income but that heteroscedasticity is present in the data. We further postulate that the nature of heteroscedasticity is as given in (11.5.10). The necessary reordering of the data for the application of the test is also presented in Table 11.3. Dropping the middle 4 observations, the OLS regressions based on the first 13 and the last 13 observations and their associated residual sums of squares are as shown next (standard errors in the parentheses). Regression based on the first 13 observations:**Yˆi= 3.4094 + 0.6968Xi**(8.7049) (0.0744) r2 = 0.8887 RSS1 = 377.17 df = 11 Regression based on the last 13 observations: Yˆi= − 28.0272 + 0.7941Xi (30.6421) (0.1319) r2 = 0.7681 RSS2 = 1536.8 df = 11 From these results we obtain λ = (RSS2/df)/(RSS1/df) = (1536.8/11)/(377.17/11) λ = 4.07 The critical F value for 11 numerator and 11 denominator df at the 5 percent level is 2.82. Since the estimated F( = λ) value exceeds the critical value, we may conclude that there is heteroscedasticity in the error variance. However, if the level of significance is fixed at 1 percent, we may not reject the assumption of homoscedasticity. Note that the p value of the observed λ is 0.014.**Breusch–Pagan–Godfrey Test. The success of the**Goldfeld–Quandt test depends not only on the value of c (the number of central observations to be omitted) but also on identifying the correct X variable with which to order the observations. This limitation of this test can be avoided if we consider the Breusch–Pagan–Godfrey (BPG) test. To illustrate this test, consider the k-variable linear regression model Yi= β1 + β2X2i+ ·· ·+βkXki+ ui(11.5.12) Assume that the error variance σ2iis described as σ2i= f (α1 + α2Z2i+ ·· ·+αmZmi) (11.5.13) that is, σ2iis some function of the non-stochastic variables Z’s; some or all of the X’s can serve as Z’s. Specifically, assume that: σ2i= α1 + α2Z2i +· · ·+αmZmi(11.5.14) that is, σ2iis a linear function of the Z’s. If α2 = α3 = · · · = αm = 0, σ2i = α1, which is a constant. Therefore, to test whether σ2i is homoscedastic, one can test the hypothesis that α2 = α3 = · · · = αm = 0. This is the basic idea behind the Breusch–Pagan test. The actual test procedure is as follows.**Step 1. Estimate (11.5.12) by OLS and obtain the residuals**uˆ1, uˆ2, . . , uˆn. Step 2. Obtain σ˜2 = Σuˆ2i /n. Step 3. Construct variables pidefined as pi= uˆ2i / σ˜2 which is simply each residual squared divided by σ˜2. Step 4. Regress pion the Z’s as pi= α1 + α2Z2i +· · ·+αmZmi+ vi (11.5.15) where viis the residual term of this regression. Step 5. Obtain the ESS (explained sum of squares) from (11.5.15) and defineΘ=1/2 (ESS) (11.5.16) Assuming uiare normally distributed, one can show that if there is homoscedasticity and if the sample size n increases indefinitely, then Θ∼asy χ2m−1 (11.5.17)**thatfollows the chi-square distribution with (m − 1)**degrees of freedom. Therefore, if the computed Θ(= χ2) exceeds the critical χ2 value at the chosen level of significance, one can reject the hypothesis of homoscedasticity; otherwise one does not reject it. EXAMPLE 11.5**THE BREUSCH–PAGAN–GODFREY (BPG) TEST**Revisit Table 11.3. Regressing Y on X, we obtain the following: Step 1. Yˆi= 9.2903 + 0.6378Xi se = (5.2314) (0.0286) (11.5.18) RSS = 2361.153 R2 = 0.9466 Step 2. calculate σ˜2 =Σuˆ2i /30 = 2361.153/30 = 78.7051 Step 3. Divide the squared residuals uˆi 2 obtained from regression (11.5.18) by 78.7051 to construct the variable pi. Step 4. Assuming that piare linearly related to Xi( = Zi) as per (11.5.14), we obtain the regression pˆi = −0.7426 + 0.0101Xi se = (0.7529) (0.0041) ESS = 10.4280 R2 = 0.18 (11.5.19) Step 5. Θ= ½ (ESS) = 5.2140 (11.5.20)**Under the assumptions of the BPG test in (11.5.20)**asymptotically follows the chisquare distribution with 1 df. Now from the chi-square table we find that for 1 df the 5 percent critical chi-square value is 3.8414 and the 1 percent critical χ2 value is 6.6349. Thus, the observed chi-square value of 5.2140 is significant at the 5 percent but not the 1 percent level of significance. Therefore, we reach the same conclusion as the Goldfeld–Quandt test. But keep in mind that, strictly speaking, the BPG test is an asymptotic, or large-sample test and in the present example 30 observations may not constitute a large sample. It should also be pointed out that in small samples the test is sensitive to the assumption that the disturbances uiare normally distributed. Of course, we can test the normality assumption by the tests discussed in Chapter 5.**White’s General Heteroscedasticity Test. White does not**rely on the normality assumption and is easy to implement. As an illustration of the basic idea, consider the following three-variable regression model (the generalization to the k-variable model is straightforward): Yi= β1 + β2X2i+ β3X3i + ui(11.5.21) The White test proceeds as follows: Step 1. Given the data, we estimate (11.5.21) and obtain the residuals, uˆi . Step 2. We then run the following (auxiliary) regression: uˆ2i= α1 + α2X2i + α3X3i + α4X2i2+ α5X3i2+ α6X2i X3i + vi(11.5.22) Note that there is a constant term in this equation even though the original regression may or may not contain it. Obtain the R2 from this (auxiliary) regression.**Step 3. Under the null hypothesis that there is no**heteroscedasticity, it can be shown that sample size (n) times the R2obtained from the auxiliary regression asymptotically follows the chi-square distribution with df equal to the number of regressors (excluding the constant term) in the auxiliary regression. That is, n · R2 ∼asy χ2 df (11.5.23) where df is as defined previously. In our example, there are 5 df since there are 5 regressors in the auxiliary regression. Step 4. If the chi-square value obtained in (11.5.23) exceeds the critical chi-square value at the chosen level of significance, the conclusion is that there is heteroscedasticity. If it does not exceed the critical chi-square value, there is no heteroscedasticity, which is to say that in the auxiliary regression (11.5.21), α2 = α3 = α4 = α5 = α6= 0. EXAMPLE 11.6