Effects of Extreme Values on Price Indexes: The Case of the Air Travel Price Index

Effects of Extreme Values on Price Indexes: The Case of the Air Travel Price Index

JANICE LENT*

ABSTRACT

This paper examines the effects of extreme price values on the Fisher and Trnqvist index formulas. Using a simple model, we first consider the impact of outliers on the unweighted arithmetic, harmonic, and geometric means of a collection of values. Then, under the same model, we investigate the effect of a single extremely high or low price on the price index formulas (weighted means). Further investigation using Taylor series approximations leads to some general conclusions regarding the relative robustness of the Fisher and Trnqvist indexes. These are illustrated with empirical results based on airfare data from the U.S. Department of Transportation's Origin and Destination Survey.

KEYWORDS: Price index, extreme value, Taylor series. JEL Categories: C43, C13, E31.

INTRODUCTION

Many economists have come to favor the "superlative" Fisher and Trnqvist price indexes over the more traditional Laspeyres formula (see, e.g., Diewert 1976; Aizcorbe and Jackman 1993). The U.S. Bureau of Labor Statistics recently began publishing a new price index series targeting the Trnqvist formula. The choice between the Fisher and Trnqvist formulas may be based on a variety of factors, including other price index formulas currently in use by the organization producing the index and the relative sensitivity of the two formulas to extreme values. In this study, we compare the Fisher and Trnqvist formulas with respect to the latter criterion-sensitivity to extreme values.

Extreme-valued price ratios often occur as a result of deep discounts or "free" promotional goods or services. Such extreme-valued ratios can be either large or small, depending on whether the discounted price appears in the numerator or denominator of the price ratio. Less often, extremely high prices appear with converse effects. The Laspeyres formula is sometimes criticized as sensitive to extreme values, because it is based on an arithmetic mean of the price ratios. We will see, however, that such sensitivity depends on the direction of the outlying value (high or low), as well as on the weights used in the selected mean.

In the next section, we consider the effect of an extreme value on the unweighted arithmetic, harmonic, and geometric means. The third section contains a discussion of the corresponding effects on the Fisher and Trnqvist index formulas under differing assumptions regarding the correlation between the expenditure-share weights and the prices. This correlation is related to the elasticity of substitution (i.e., the extent to which consumers shift their purchases toward lower priced items when relative prices change).

The fourth section presents an empirical example: the case of air travel index estimates computed using data from the Passenger Origin and Destination Survey collected by the Bureau of Transportation Statistics. The extreme-valued price ratios in this application resulted from a change in data-collection procedures and are in this sense artificial. They do, however, provide an opportunity to compare the performances of the different index formulas under the conditions represented by the application. We summarize our conclusions in the final section.

EFFECTS OF EXTREME VALUES ON UNWEIGHTED MEANS

The following simple model shows the effects of an extreme value on three types of unweighted means. Let x1, ..., xn be a collection of nonnegative values, where xi = for i = 1, ..., n - 1, while xn = y for some factor y > 0; that is, xn is an outlier in the collection. We define the unweighted arithmetic, harmonic, and geometric means, respectively, as follows:

uppercase a = 1 divided by lowercase n summation from lowercase i = 1 to n lowercase x subscript {lowercase i},

uppercase h = (1 divided by lowercase n summation from lowercase i = 1 to n 1 divided by lowercase x subscript {lowercase i}) superscript {-1},

and

uppercase g = product from lowercase i = 1 to n lowercase x subscript {lowercase i} superscript {1 divided by lowercase n} .

For M ∈ {A, H, G}, let

lowercase f subscript {uppercase m} = uppercase m divided by lowercase mu.

Then

lowercase f subscript {uppercase a} = (lowercase n - 1 + lowercase y) divided by lowercase n, lowercase f subscript {uppercase h} = lowercase n divided by (lowercase n - 1 - lowercase y superscript {-1}), and

lowercase f subscript {uppercase g} = lowercase y superscript {1 divided by lowercase n}.

We first consider the rate at which the various means approach as n approaches infinity. For fixed y, we have

| 1 - lowercase f subscript {uppercase a) | = | (1 - lowercase y) divided by lowercase n |
                          = uppercase o (1 divided by lowercase n)

as n → ∞. For the harmonic mean also,

| 1 - lowercase f subscript {uppercase h} | = | (1 - lowercase y superscript {-1}) divided by (lowercase n - (1 - lowercase y superscript {-1})) |
                  = uppercase o (1 divided by lowercase n)

as n → ∞, and, similarly,

| 1 - lowercase f subscript {lowercase g} | = | 1 - lowercase e superscript {1 divided by n ln lowercase y} |
                                 = | -1 divided by n ln lowercase y - uppercase o (1 divided by (lowercase n superscript {2})) |
                                 = uppercase o (1 divided by lowercase n)

as n → ∞. Thus, as n becomes large, all three of the means approach at approximately the same rate. Their behavior in the presence of an outlier differs, however, under various assumptions about the outlier itself. If we suppose that n is fixed, we may follow derivations similar to those above to arrive at the results, which are summarized in table 1.

The results shown in table 1 for fG(y) may lead us to conclude that price index formulas based on the geometric mean are, overall, the most robust formulas available; at the very least, they represent a sensible choice when both high and low outliers are expected to occur. By contrast, while A is robust to low outliers, it is sensitive to high outliers; similarly, H is robust to high outliers but sensitive to low ones. In most applications, however, price indexes are not computed as unweighted means. In the next section, we examine the effect of expenditure-share weights on the Laspeyres, Paasche, Fisher, and Trnqvist indexes, with special emphasis on the latter two.

EFFECTS OF EXTREME VALUES ON PRICE INDEXES

Price Index Formulas

We begin by presenting several population index formulas. The Laspeyres index measuring price change between time periods 1 and 2 is defined as

uppercase l = (summation from lowercase j = 1 to uppercase n lowercase q subscript {lowercase j 1} lowercase p subscript {lowercase j 2}) divided by (summation from lowercase j = 1 to uppercase n lowercase q subscript {lowercase j 1} lowercase p subscript {lowercsae j 1}) = summation from lowercase j = 1 to uppercase n lowercase w subscript {lowercase j 1} (lowercase p subscript {lowercase j 2} divided by lowercase p subscript {lowercase j 1}) ,

where pjt denotes the price of item j at time t ∈ {1,2}, qjt denotes the quantity of item j purchased at time t,

lowercase w subscript {lowercase j lowercase t} = lowercase p subscript {lowercase j lowercase t} lowercase q subscript {lowercase j lowercase t} divided by summation from lowercase k = 1 to uppercase n lowercase p subscript {lowercase k lowercase t} lowercase q subscript {lowercase k lowercase t},

and N denotes the number of items in the target population. The weight wjt is the expenditure share for item j in period t; the price ratios pj2/pj1 are often called price relatives. Clearly L is the arithmetic mean of the price relatives with weights representing first period expenditure shares. The Paasche index is a harmonic mean of the price ratios, with second period expenditure-share weights:

uppercase p = summation from lowercase j = 1 to uppercase n lowercase q subscript {lowercase j 2} lowercase p subscript {lowercase j 2} divided by summation from lowercase j = 1 to n lowercase q subscript {lowercase j 2} lowercase p subscript {lowercase j 1} = 1 divided by summation from lowercase j = 1 to uppercase n lowercase w subscript {lowercase j 2} (lowercase p subscript {lowercase j 2} divided by lowercase p subscript {lowercase j 1}) subscript {-1}.

The Fisher index is simply defined as uppercase f = square root uppercase l uppercase p, while the Trnqvist is a geometric mean of the price ratios with weights representing the averages of the period 1 and period 2 expenditure shares, shown as

uppercase t = product from lowercase j = 1 to uppercase n (lowercase p subscript {lowercase j 2} divided by lowercase p subscript {lowercase j 1}) superscript {lowercase w lowercase j, 1, 2},

where w j,1,2 = (w j1 + w j2 ) /2.

Extreme Values and the Elasticity of Substitution

To examine the effects of an outlier on the indexes described above, suppose we have a collection of n items priced in time periods 1 and 2. Suppose further that for j = 1, ..., n we have pj1 = qj1 = 1 and that for j = 1, ..., n - 1 we also have pj2 = 1, while pn2 = y (i.e., we assume for simplicity that the above is 1.) For t ∈ {1,2} , let xjt = pjtqjt, the expenditure level for item j in period t. We wish to allow the quantity of an item purchased to vary in response to price change and an assumed elasticity level. When pj2 = pj1, we assume that qj2 = qj1. Otherwise, let

lowercase q subscript {lowercase j 2} = lowercase q subscript {lowercase j 1} (lowercase p subscript {lowercase j 2} divided by lowercase p subscript {lowercase j 1}) superscript {- lowercase tau}

where ≤ τ ≤ 1, and τ is assumed constant. Then

lowercase x subscript {lowercase j 2} = lowercase x subscript {lowercase j 1} (lowercase p subscript {lowercase j 2} divided by lowercase p subscript {lowercase j 1}) superscript {1 - lowercase tau}.

We define the elasticity τ in this way, because τ provides a convenient means of examining the effects of extreme values under conditions of high and low elasticity, defined relatively. Note that higher values of τ indicate less impact of price change (represented by the price ratios) on second period item-level expenditure levels.

For j = 1, ..., n - 1, we have qj2 = qj1 = 1; and

qn2. = y-τ

The resulting first and second period expenditure-share weights are as follows:

lowercase w subscript {lowercase j 1} = 1 divided by lowercase n, lowercase j = 1, ..., lowercase n;

lowercase w subscript {lowercase j 2} = 1 divided by (lowercase n - 1 + lowercase y superscript {1 - lowercase tau}), lowercase j = 1, ..., lowercase n - 1;

and

lowercase w subscript {lowercase n 2} = lowercase y superscript {1 - lowercase tau} divided by (lowercase n - 1 + lowercase y superscript {1 - lowercase tau}).

The "average weights" used in the Trnqvist index are

lowercase w subscript {lowercase j, 1, 2} = 1 divided by 2 (1 divided by lowercase n + 1 divded by (lowercase n - 1 + lowercase y superscript {1 - lowercase tau})), lowercase j = 1, ..., lowercase n - 1,

and

lowercase w subscript {lowercase n, 1, 2} = 1 divided by 2 (1 divided by lowercase n + lowercase y superscript {1 - lowercase tau} diviided by (lowercase n - 1 + lowercase y superscript {1 - lowercase tau})).

Note that when τ is small (low or zero elasticity) and y is large,

wn 1 < wn, 1, 2 ,       (1)

so the Laspeyres index gives less weight to high outliers than does the Trnqvist index. Similarly, when τ and y are both small,

wn 2 < wn, 1, 2 ,       ,(2)

indicating that the Paasche index gives less weight to low outliers than the Trnqvist. Under conditions of low elasticity, we therefore observe the following phenomena: although the Laspeyres index, based on the arithmetic mean, is sensitive to high outliers, it assigns them weights that are low relative to the Trnqvist weights, while the Paasche index, a harmonic mean, assigns lower weights to low outliers. The weights in the Laspeyres and Paasche indexes can therefore be expected to compensate, at least partially, for the sensitivity of the arithmetic and harmonic means to high and low outliers, respectively.

Under this simple model, the values of the Laspeyres, Paasche, Fisher, and Trnqvist indexes are as follows:

uppercase l (lowercase n, lowercase y) = (lowercase n - 1 + lowercase y) divided by lowercase n;

uppercase p (lowercase n, lowercase y, lowercase tau) = (lowercase n - 1 + lowercase y superscript {1 - lowercase tau}) divided by (lowercase n - 1 + lowercase y superscript {- lowercase tau});

uppercase f (lowercase n, lowercase y, lowercase tau) = [((lowercase n - 1 + lowercase y) divided by lowercase n)((lowercase n - 1 + lowercase y superscript {1 - lowercase tau}) divided by (lowercase n - 1 + lowercase y superscript {- lowercase tau}))] superscript {1 divided by 2};

and

uppercase tau (lowercase n, lowercase y, lowercase tau) = exp [1 divided by 2 (1 divided by lowercase n + lowercase y superscript {1 - lowercase tau} divided by (lowercase n - 1 + lowercase y superscript {1 - lowercase tau})) ln lowercase y].

Both the Fisher and Trnqvist indexes are known as superlative indexes, because economic theory suggests that they approximate a true cost of living index under relatively weak assumptions regarding economic conditions (Diewert 1987). (In the application considered in the next section, these indexes should be viewed as cost of flying indexes rather than as cost of living indexes.) We, therefore, focus on the relative robustness of F (n, y, τ) and T (n, y, τ) under the assumptions τ = 1 and τ = 0. The value τ = 1 indicates that consumers shift their purchases toward items (or item categories) whose relative prices have decreased between periods 1 and 2, while τ close to zero represents the case of little or no change in buying behavior in response to price change.

First consider the case τ = 1, where a value of τ represents the assumption that consumers alter the quantities of the items they purchase so as to maintain the same level of expenditure on each item-a situation corresponding to a fairly high level of elasticity. In this case, we have, for fixed n and large y,

uppercase f (lowercase n, lowercase y, 1) = [((lowercase n - 1 + lowercase y) divided by lowercase n)(n divided by lowercase n - 1 + lowercase y superscript {-1})] superscript {1 divided by 2}
                           = ((lowercase n - 1 + lowercase y) divided by (lowercase n - 1 + lowercase y superscript {-1})) superscript {1 divided by 2}
                           = (1 + lowercase y divided by lowercase n) superscript {1 divided by 2},       (3)

while

uppercase t (lowercase n, lowercase y, 1) = lowercase y superscript {1 divided by lowercase n}.       (4)

So, for reasonably large n, T is more robust than F in the presence of high outliers. For the case of low outliers, we have

uppercase f (lowercase n, lowercase y, 1) = ((lowercase n - 1 + lowercase y) divided by (lowercase n - 1 + lowercase y superscript {-1})) superscript {1 divided by 2}
                          uppercase o (lowercase y superscript {1 divded by 2})       (5)

and

uppercase t (lowercase n, lowercase y, 1) = uppercase omega (lowercase y superscript {1 divided by lowercase n})       (6)

for fixed n as y approaches 0. Under the simple model, we may therefore conclude that, with regard to robustness, conditions of high elasticity favor the Trnqvist index over the Fisher.

With τ = 0, we have

uppercase f (lowercase n, lowercase y, 0) = ((lowercase n - 1 + lowercase y) divided by lowercase n),

and

uppercase t (lowercase n, lowercase y, 0) = exp[1 divided by 2 (1 divided by lowercase n + lowercase y divided by (lowercase n - 1 + lowercase y) ln lowercase y].

Note that F (n, y, 0) = L (n, y, 0) = P (n, y, 0). For fixed n and large y,

uppercase f (lowercase n, lowercase y, 0) is almost equal to 1 + lowercase y over lowercase n,       (7)

while

uppercase t (lowercase n, lowercase y, 0) is almost equal to lowercase y superscript {(lowercase n + 1) over 2 lowercase n}.       (8)

As a rough rule of thumb, the above approximations suggest that T is likely to outperform F whenever outliers are as large as n2. The relative robustness of T and F thus depends on the relative values of y and n, which may, in turn, depend on the aggregation level being considered. Equations (4) and (8) also indicate that, for large values of n, T is much more robust to high outliers under high elasticity than it is under low elasticity. For low outliers, however, the elasticity assumption has less impact on T. With n fixed and y small, we have

uppercase f (lowercase n, lowercase y, 0) = uppercase omega ((lowercase n - 1) divided by lowercase n),       (9)

and

uppercase t (lowercase n, lowercase y, 0) = uppercase o (lowercase y superscript {1 divided by (2 lowercase n)}),       (10)

revealing that, under conditions of low elasticity, T is more sensitive to low outliers than F. Equations (5), (6), (9), and (10) suggest that T is somewhat more robust to low outliers for τ = 0 than for τ = 1, while F is much more robust.

The above results lead us to conclude that, under conditions of low elasticity, the Fisher index may often be more robust to outliers than the Trnqvist: the Fisher is more robust to low outliers and, when n is sufficiently large relative to any price ratios in the dataset, the Fisher is also more robust to high outliers. Conditions of higher elasticity (τ close to 1) render both indexes more robust to extremely high values. Under conditions of high elasticity, the Trnqvist is preferable to the Fisher, because it is less sensitive to both high and low outliers.

The numerical examples shown in appendix A illustrate these conclusions. Tables A1 and A2 give values of the Fisher and Trnqvist indexes under the single outlier scenario described above. (Note that these are not random values produced by a Monte Carlo simulation but simply the values of the functions F (n, y, τ) and T (n, y, τ) for the given parameters.) Table A1 gives index values under the assumption that τ = 1 (high elasticity). Under this assumption, the Trnqvist is clearly more robust than the Fisher to both high and low outliers.

Table A2 shows values of the indexes under the assumption that τ = 0. The bold numbers in this table highlight points at which y becomes large enough, relative to n, to render the Trnqvist index better than the Fisher for approximating the mean = 1 in the presence of a high outlier. As expected, the turning points occur as y approaches n2. The examples in table A2 also illustrate that, under low elasticity, both indexes are more sensitive to high outliers and less sensitive to low outliers than they are under high elasticity.

Taylor Series Results

The single outlier model employed in the previous subsections does not, of course, account for the data complexity often encountered in practical applications. Here, we look beyond the single outlier model to examine Taylor series expansions that shed further light on the relative robustness of the Fisher and Trnqvist indexes under the general assumption of low elasticity. Following and expanding on the development of Lent and Dorfman (2004a), we assume that the price indexes are computed from a collection of expenditure share weights and sub-indexes Ig, which here take the place of the price ratios pj2/pj1 in the previous subsection. Each g is an aggregate of the ratios pj2/pj1 for all items j in a particular item category g. In practice, the standard formulas are often applied in this two-stp fashion. The categories into which we divide the items may be defined according to item characteristics, geographic area of purchase, or both.

We begin by defining some notation. For t ∈ {1,2}and for each item category g, let

lowercase w subscript {lowercase g lowercase t} = summation over lowercase j element of lowercase g lowercase p subscript {lowercase j lowercase t} lowercase q subscript {lowercase j lowercase t} divided by summation over lowercase g summation over lowercase j element of lowercase g lowercase p subscript {lowercase j lowercase t} lowercase q subscript {lowercase j lowercase t},

and let

lowercase mu subscript {lowercase t} = summation over lowercase g lowercase w subscript {lowercase g lowercase t} uppercase i subscript {lowercase g}, lowercase sigma superscript {2} subscript {lowercase t} = summation over lowercase g lowercase w subscript {lowercase g lowercase t} (uppercase i subscript {lowercase g} - lowercase mu) superscript {2},

and

lowercase gamma subscript {lowercase t} = summation over lowercase g lowercase w subscript {lowercase g lowercase t} (uppercase i subscript {lowercase g} - lowercase mu) superscript {3}.

Next, with wg defined as the Trnqvist weights,

lowercase w subscript {lowercase g} = (lowercase w subscript {lowercase g 1} + lowercase w subscript {lowercase g 2}) divided by 2,

let

lowercase mu = summation over lowercase g lowercase w subscript {lowercase g} uppercase i subscript {lowercase g}, lowercase sigma superscript {2} = summation over lowercase g lowercase w subscript {lowercase g} (uppercase i subscript {lowercase g} - lowercase mu) superscript {2},

and

lowercase sigma = summation over g lowercase w subscript {lowercase g lowercase t} (uppercase i subscript {g} - lowercase mu) superscript {2}.

We expand each of the superlative indexes about the point at which all of the sub-indexes Ig equal the mean . The relevant partial derivatives are given in appendix B. From the general form of the third-order approximation given by Lent and Dorfman (2004a) for a geometric mean, we have the following approximation of the Trnqvist index:

uppercase t subscript {uppercase i} = product over lowercase g uppercase i superscript {lowercase w subscript {lowercase g}} subscript {lowercase g}
     is almost equal to lowercase mu - lowercase sigma superscript {2} divided by (2 lowercase mu) + lowercase gamma divided by (3 lowercase mu superscript {2}).       (11)

The second-order approximation for the Fisher index is

uppercase f subscript {1} = (summation over lowercase g lowercase w subscript {lowercase g 1} uppercase i subscript {lowercase g} divided by summation over lowercase g (lowercase w subscript {lowercase g 2} divided by uppercase i subscript {lowercase g})) superscript {1 divided by 2}
    is almost equal to lowercase mu - lowercase sigma superscript {2} divided by (lowercase sigma superscript {2} subscript {2} divided by (2 lowercase mu))       (12)

Thus, to the second order, we have

uppercase t subscript {lowercase t} - uppercase f subscript {lowercase t} is almost equal to (lowercase sigma superscript {2} - lowercase sigma superscript {2} subscript {2}) divided by (2 lowercase mu)
                    = 1 divided by 4 lowercase mu summation over lowercase g (lowercase w subscript {lowercase g 2} - lowercase w subscript {lowercase g 1})(uppercase i subscript g - lowercase mu) superscript {2}.       (13)

Consider the relative values of wg1 and wg2 in the presence of high outliers among the Ig and high correlation between the Ig and the wg2 (the case of low elasticity). Under these conditions, we are likely to have wg2 > wg1 for large values of (Ig - )2 and thus TI - FI > 0. Similarly, in the presence of low outliers, we are likely to have wg2 < wg1 for large values of (Ig - )2, resulting in negative values of TI - FI. Thus, the approximation shown in equation (13) indicates that, under conditions of low elasticity, the Fisher index may be more robust than the Trnqvist to both high and low outliers.

AN EMPIRICAL EXAMPLE

For the air travel price index series, the apparent elasticity of substitution is low-in some cases, even negative. The series, therefore, exemplify only the behavior of the different indexes under conditions of low elasticity (τ close to 0). Note that the elasticity reflected in the data, rather than the actual elasticity, is the quantity that affects the performance of the indexes; Dorfman et al. (1999) showed that the elasticity reflected in sample survey data need not always equal the true population elasticity.

The air travel price index series shown in figures 1, 2, 3, 4, 5, and 6 are based on data from the Bureau of Transportation Statistics' quarterly Origin and Destination (O&D) Survey. The sample for the O&D Survey comprises about 10% of all passenger itineraries having some U.S. component (i.e., itineraries that include at least one flight arriving at or departing from a U.S. airport) and includes about 6 to 7 million itineraries per quarter. Data items collected include trip route, class of service (e.g., coach, first class), and transaction fare including taxes. Note that the scales differ across figures, so comparisons across figures are distorted in some cases.

When goods and services are sampled for the purpose of estimating a price index, the sample items generally remain in the sample over an extended time period (e.g., two years) unless they are taken off the market by the retailer. The stable sample allows comparison of prices across time for identical items. Ratios of prices in different time periods for individual items are the building blocks of the traditional price index estimators.

In the O&D Survey, however, the sampling is performed independently for each reference quarter. Since the itineraries selected in a given quarter may not match those selected for a previous or subsequent quarter, we developed and tested a two-stage process for matching categories of itineraries across quarters and comparing average prices within categories across time. The ratio of average prices for different time periods is called a unit value index. These sub-indexes are then aggregated by the Fisher, Trnqvist, and other index formulas. The index series are based only on data from sample itineraries flown on domestic carriers and are chained quarterly.1 Lent and Dorfman (2004b) provide a more detailed description of the index estimation methodology.

The figures show the Laspeyres, Paasche, Fisher, and Trnqvist index series for various classes of service and for all classes combined. Note that, in all of the figures, the Paasche series runs either slightly below the Laspeyres series or even (for business class service) above the Laspeyres, indicating low or negative elasticity of substitution. Lent and Dorfman (2004a) describe a method of estimating the elasticity of substitution; elasticity estimates computed by their method, measuring elasticity of substitution between unit value categories as described above, run close to 0 for these data. Although air travel passengers readily substitute one carrier for another in response to fare changes, little substitution across trip origin/destination pairs occurs. Since origin/destination pairs far outnumber carriers, this substitution behavior leads to low overall estimates of elasticity of substitution between the unit value categories.

In examining figures 1 through 6, it is important to note that the Class of Service variable in the O&D Survey was redefined and standardized in 1997-98. (Formerly, air carriers had used a variety of service classifications in reporting this information, so the variable values had to be recoded by the Bureau of Transportation Statistics.) We therefore expect some unusual data values to affect the index series during this period; indeed, many of the series display a visible break between the fourth quarter of 1997 and the first quarter of 1998. These breaks may be exacerbated, because a lower percentage of the O&D Survey observations were "matched" across time during 1997-98 (see Lent and Dorfman 2004b for a description of the across-time matching method), resulting in lower than usual effective sample sizes.

Figures 1 and 2 show the series for all classes of service combined and for restricted coach class (by far the largest class), respectively. The series in figure 2 behave in typical fashion: the Laspeyres series runs just above the others, displaying a slight upward drift, while the Paasche shows a similar downward drift, and the two superlative series run between them, closely tracking each other. This type of behavior results from the large number of observations and because the 1997-98 break has relatively little impact on these series. Figure 1 is similar to figure 2, except for the noticeably larger effect of the 1997-98 change, which lifts the Trnqvist series slightly above the others. Recall that, under conditions of low elasticity, the Trnqvist index is often more sensitive to outliers than the Fisher.

Index series for other classes of service (categories comprising fewer observations) are shown in figures 3, 4, 5, and 6. For the unrestricted first and restricted first class indexes (figures 3 and 4), the Laspeyres series runs very slightly above the Paasche, indicating low but positive elasticity. For the unrestricted first class series, the 1997-98 break sends the Trnqvist above the other series, while the Trnqvist for restricted first class is "bumped down" and runs well below the others for 1998 and subsequent years. In both cases, the Trnqvist continues to roughly parallel the Fisher after the break, indicating that unusual data values generated the level shifts. Note also that the Trnqvist's upward shift for unrestricted first class is noticeably less severe than its downward shift for restricted first class, perhaps due to its greater robustness to high outliers than to low ones.

The business class index series (figures 5 and 6) display the relatively rare phenomenon of negative elasticity. The Paasche series runs above the Laspeyres, indicating that consumers are shifting their purchases toward higher priced services as relative prices change. It is important to note that sample survey data may not always reflect true population elasticity; in this case, the class-of-service categories are coarsely defined, and many different types of restrictions may apply to tickets in the restricted categories. (Restrictions may include, for example, a requirement of advance ticket purchase or, in the case of roundtrip itineraries, a Friday or Saturday night stay at the destination.) Elasticity estimates based on these data reflect substitution between these categories but not within them (for the same route and carrier) and may therefore suffer a downward bias.

On the other hand, since business class service is typically paid for by a third party (i.e., the passenger's employer), very low elasticity is expected. Some business class passengers may even choose higher priced tickets assuming that "you get what you pay for," and such behavior could also explain the negative elasticity indicated. Under negative elasticity, quantities purchased are positively correlated with price change, and this correlation may cause expenditure shares to increase dramatically when prices increase. The Trnqvist index, whose weights are average expenditure shares, therefore assigns large weights to some high price ratios. Apart from the negative elasticity, the movements of the business class series appear similar to that of the first class series, that is, the Trnqvist index is shifted up or down during the 1997-98 period, while the other series are less affected by the unusual values. Table A3 in appendix A shows unweighted percentiles of the distributions of the unit value indexes for the first class and business class categories over the crucial period. The outliers are clearly sufficient in number and severity to impact the tails of the sample distributions.

CONCLUSIONS

Practitioners may often consider robustness to outliers an important criterion in selecting a price index formula, especially for item categories such as airfares, in which extreme prices may regularly result from frequent flyer awards and other price discriminatory discounts. Although price index formulas based on different types of means inherit the relative robustness of these means, the weights applied in price index calculation also play a crucial role. This paper shows that, under conditions of low elasticity of substitution, the high correlation between the weights and the price ratios may offset the sensitivity of the Laspeyres and Paasche indexes, making the Fisher a more attractive option than the Trnqvist. The choice between index formulas is therefore more complex than the mere selection of an arithmetic, harmonic, or geometric mean. It requires information on the elasticity of substitution reflected in the data as well as an estimate of the magnitude of outliers (high or low) that can be expected.

ACKNOWLEDGMENT

Comments from Alan Dorfman of the Bureau of Labor Statistics resulted in significant improvements to this paper.

REFERENCES

Aizcorbe, A. and P. Jackman. 1993. The Commodity Substitution Effect in CPI Data, 1982-1991: Anatomy of a Price Change. Monthly Labor Review , December 1993:25-33. Washington, DC: U.S. Government Printing Office.

Diewert, W.E. 1976. Exact and Superlative Index Numbers. Journal of Econometrics 4:115-145.

_____. W.E. 1987. Index Numbers. The New Palgrave: A Dictionary of Economics. Edited by J. Eatwell, M. Milate, and P. Newman. London, England: MacMillan.

Diewert, W.E., R. Feenstra, and W. Alterman. 1999. International Trade Price Indexes and Seasonal Commodities. Washington, DC: U.S. Department of Labor, Bureau of Labor Statistics.

Dorfman, A., S. Leaver, and J. Lent. 1999. Some Observations on Price Index Estimators. Proceedings of the Federal Conference on Survey Methodology, Arlington, VA, November 1999.

Lent, J. 2003. Chain Drift in Experimental Price Index Series for Air Travel, Proceedings of the 2003 Joint Statistical Meetings, Section on Survey Research Methods, CD. Alexandria, VA: American Statistical Association.

Lent, J. and A. Dorfman. 2004a. Using a Weighted Average of the Jevons and Laspeyres Indexes to Approximate a Superlative Index. Working paper.

_____. 2004b. A Transaction Price Index for Air Travel. Working paper.

APPENDIX B

Partial Derivatives

To derive equations (11) and (12), the function FI is expanded around the point I = μ = (μ, ., μ). The general formula for the third-order Taylor expansion is

lowercase f (uppercase i underscore) = lowercase f (lowercase mu underscore) = summation lowercase f superscript {'} subscript {lowercase g} (lowercase mu underscore)(uppercase i subscript {lowercase g} - lowercase mu)

+ 1 divided by 2 summation summation lowercase f superscript {

+ 1 divided by 6 summation summation summation lowercase f superscript{'''} subscript {lowercase g 1, lowercase g 2, lowercase g 3} (lowercase mu) dot

      (uppercase i subscript {lowercase g 1} - lowercase mu) dot (uppercase i subscript {lowercase g 2} - lowercase mu) dot (uppercase i subscript {lowercase g 4} - lowercase mu).

For a derivation of the third-order expansion of TI (equation (11)), see Lent and Dorfman (2004a). The first- and second-order partial derivatives of FI evaluated at I = μ (used in the derivation of equation (12)) are as follows:

lowercase delta uppercase f subscript {uppercase i} divided by lowercase delta uppercase i subscript {lowercase g} long vertical bar over uppercase i underscore = lowercase mu underscore = lowercase w subscript {lowercase g}

lowercase delta uppercase f superscript {2} uppercase f subscript {uppercase i} divided by lowercase delta uppercase i superscript {2} subscript {lowercase g} long vertical bar over uppercase i underscore = lowercase mu underscore = lowercase mu superscript {-1} (lowercase w superscript {2} subscript {lowercase g} - lowercase w subscript {lowercase g 2} + (lowercase w superscript {2} subscript {lowercase g 2} - lowercase w superscript {2} subscript {lowercase g 1}) divided by 2)

lowercase delta superscript {2} uppercase f subscript {uppercase i} divided by lowercase delta uppercase i subscript {lowercase g 1} lowercase delta uppercase i subscript {lowercase g 2}) long vertical bar over uppercase i underscore = lowercase mu underscore = lowercase mu superscript {-1} ((lowercase w subscript {lowercase g 1 2} lowercase w subscript {lowercase g 2 2} + lowercase w subscript {lowercase g 1 1} lowercase w subscript {lowercase g 2 1}) divided by 2 + lowercase w subscript {lowercase g 1} lowercase w subscript {lowercase g 2})

END NOTES

1Price index chaining is done by estimating long-term price changes as products of shorter term changes (links). Quarterly chaining can lead to "chain drift," as seen in the Laspeyres and Paasche series in the figures in this section. For more information on chain drift in the airfare indexes, see Lent 2003.

ADDRESS FOR CORRESPONDENCE

*J. Lent, Bureau of Transportation Statistics, Research and Innovative Technology Administration, U.S. Department of Transportation, 400 7th Street SW, Washington, DC 20590. Email: janice.lent@dot.gov