Australasian Agribusiness Review - Vol. 13 - 2005

Paper 12
ISSN 1442-6951

Unit Value Biases in Price Elasticities of Demand for Meat in Indonesia

Susan Olivia and John Gibson^[1]

Susan Olivia, University of California, Davis, USA and John Gibson, University of Waikato, New Zealand

Abstract

Indonesia is the largest market for Australian live cattle exports so estimates of income and price elasticities of meat demand in Indonesia may help exporters to set appropriate pricing strategies and to model future demands. In contrast to developed countries, where meat demand studies often use aggregate data, Indonesian studies rely on household surveys, with unit values (ratios of expenditures to quantities) used instead of market prices. When price elasticities of demand are estimated from unit values, various quality and measurement error biases can result. These biases may cause inappropriate pricing and marketing strategies to be adopted by Australian beef and cattle producers and exporters.

In this paper, data from 29,000 households on Java are used to estimate a demand system for beef, chicken and other meat groups. Java contains almost two-thirds of the Indonesian population and meat consumption is similarly concentrated there. When estimation procedures are used that correct for the biases caused by unit values, the own-price elasticities of both beef (-0.46) and chicken (-0.42) are smaller than in previous studies. This difference is consistent with the theoretical literature, which suggests that using unit values instead of prices makes own-price elasticities too large in absolute terms. The own-price elasticities for beef are much more sensitive to the choice of procedure for dealing with unit values than are the estimates for chicken. Hence, pricing strategies for beef producers that are based on estimated demand elasticities from Indonesia, may prove to be inappropriate if the wrong method for dealing with unit value biases is used.

I. Introduction

Indonesia is the major market for Australian live cattle exports, taking over 400,000 head in 2002. These exports are worth over $250 million, which is more than 40 percent of the total export cattle market. While beef exports from Australia to Indonesia are much smaller, they are still worth over $50 million in some years (Hutasuhut, et. al. 2002). Australia is the dominant supplier of beef and cattle to Indonesia, with 51.4 percent of the imported beef in 2000 originating from Australia, compared with 21.6 percent came from New Zealand (MLA, 2001). As a result, accurate estimates of income and price elasticities of meat demand in Indonesia may be useful for Australian agribusiness when deciding on pricing and marketing strategies and when predicting future demands.

In contrast to developed countries, where meat demand studies often use aggregate data, Indonesian studies rely on household surveys, with unit values (ratios of expenditures to quantities) used instead of market prices. Unlike market prices, unit values reflect household’s quality choices, are affected by reporting errors, and are unavailable for non-purchasing households. Each of these three factors can result in biased estimates of price elasticities coming from unit value data.

The quality effects matter because household surveys typically aggregate different varieties, so even if consumers faced the same prices, as the mix of varieties changes with variations in household income and other characteristics, the unit values change. In fact, unit values will tend to vary less than prices if consumers react to high prices by choosing lower quality, and this is likely to create a systematic overstatement in the absolute value of estimated price elasticities (Deaton, 1988). Reporting errors in either expenditures or quantities will also matter because they are reflected in unit values, and these errors are likely to cause spurious correlations with the demands that are being ‘explained’ by the unit values, causing estimated price elasticities to be biased. The lack of information for non-purchasers matters because excluding these households may result in sample selection bias.

Although methods for correcting the biases in demand elasticity estimates from unit value data have been developed, most notably by Deaton (1987, 1988, 1990), researchers rarely apply them. To see whether such corrections have any practical impact, this paper estimates demand systems for beef, chicken and other meat groups, using data from 28,964 households on the island of Java in Indonesia. Java contains almost two-thirds of the Indonesian population and meat consumption is similarly concentrated there. We compare price elasticities of meat demand calculated from procedures which attempt to correct for the various biases caused by unit values with the results from simpler procedures which do not make these corrections.

The rest of the paper is organised as follows. The next section reviews previous studies of meat demand in Indonesia. In Section III, we discuss the biases that can result when unit values are used in demand studies. Section IV discusses the demand specification and the different unit values procedure used, paying particular attention to the method proposed by Deaton (1990). Section V describes the data used in this paper. The estimation results and elasticity comparisons are presented in Section VI and Section VII presents our conclusions.

II. Previous Meat Demand Studies in Indonesia

Meat demand has been studied intensively in the United States (e.g. Hahn, 1994; Alston and Chalfant, 1993; Eales and Unnevehr, 1993; Yong and Hayes, 1993; Eales and Wessels, 1999). By contrast, interest in this field is relatively recent in Indonesia although there are many estimates of demand parameters for food crops in either Indonesia or Java. The emphasis on food crops and neglect of meat demand studies reflects the structure of consumption and is also partly due to the fact that Indonesia has a history of intervening in the supply, pricing and trade of key staple food crops, and especially rice, in order to maintain food security. It is therefore not surprising that food crop consumption analyses have been done more often than meat demand analyses. Most of these demand studies in Indonesia use unit values (expenditures divided by quantities) as proxies for ‘prices’ due to the unavailability of market prices in the household survey data.

Examples of previous Indonesian studies of income and price elasticities of food demand include Timmer and Alderman (1979) and Dixon (1982) both of whom estimated food crop (rice, corn, cassava) elasticities. A broader set of foods were included by Chernichovsky and Meesok (1984). Unlike most studies, Chernichovsky and Meesok’s (1984) estimated demand elasticities for both Java and off-Java regions. They found that the income elasticity of demand for meat (1.68) was the 3^rd highest of the 12 foods they studied, while the own price elasticity (-1.86) was the fifth highest for Java.  Income elasticity of meat demand for Outer Islands residents was the 2^nd highest of the 12 foods they studied. They also found that the own-price elasticity of demand for meat is higher in the Outer Islands than in Java; specifically, a 10 percent increase in the price of meat brought a 23 percent decline in the quantities consumed in the Outer Islands, but only an 18 percent decline for households in Java. All three of these studies used data from the SUSENAS household survey, and used unit values as measure of prices. While Deaton (1990) also uses unit values, he reports a lower estimate of the own price elasticity for meat (-1.09 for rural Java), due to the impact of his correction method on unit values (see below for details).

More recent, and also more disaggregated meat demand elasticities come from a study by Hutasuhut (2000), using data from the 1990, 1993 and 1996 SUSENAS surveys for Jakarta and West Java. A double truncated Linear Approximate version of the Almost Ideal Demand System (LA/AIDS) model was used to account for the fact that budget shares lie between zero and one.  In terms of data aggregation, Hutasuhut (2000) applied the ‘Stochastic Hicksian Aggregates’ technique (where goods which are highly correlated in prices are put together in one group) and formed 4 meat groups (MG1-4) from the 16 disaggregated meat types.^[2] As with previous studies, Hutasuhut (2000) used unit values as implicit prices in his analysis.

On average, the estimated expenditure elasticities for MG-1 (dominated by beef) and MG-3 (comprises of other poultry meat, canned meat, other processed meat and untrimmed bone) were less than unity while those for MG-2 (dominated by commercial and native chicken) and MG-4 (made up of shredded fried meat and beef liver) were slightly greater than unity.  With regards to the estimated own-price elasticities, Hutasuhut (2000) reports that the own-price elasticity for MG-1 is inelastic (ranges from -0.91 to -0.93) whereas MG-2, 3, and 4 have elastic own price elasticities (for example, the own price elasticity for MG-2 range from -1.08 to -1.09). A point to note is that the estimated own-price elasticities of all meat groups for Jakarta and West Java reveals that they are near unitary elasticities. To make comparisons with previous meat demand studies, Hutasuhut (2000) reports estimates done by Kesavan et al. (1992) who used data from 1987 SUSENAS. Kesavan et al. reported that the estimated expenditure elasticities for the demand for beef and poultry meat are 0.88 and 0.84 respectively. In addition, the compensated own-price elasticities for beef and poultry meat were estimated to be -0.515 and -0.647 respectively. Another study by Puslitbangnak (1992) found that expenditure elasticities for beef and preserved meat ranges from 0.64 in Yogya to 1.15 in Jakarta. The own price elasticity ranges from -0.69 for Yogya and -1.69 for Central Java while it was 0.30 for Jakarta. In addition, the expenditure elasticity for poultry ranges from 0.88 in East Java to 1.04 in Central Java, while the estimated own-price elasticity for chicken range from -0.25 for East Java to -2.16 for Jakarta. In other words, Puslitbangnak’s results indicate that, on average, beef is a more of a luxury than chicken, which is the opposite to the pattern found by Hutasuhut (2000) and Hutasuhut et. al. (2002).

Hutasuhut (2000) points out that one of the limitations in his study is that the quality of individual meat commodities is assumed to be homogenous, and if the ignored quality effect is substantial, the estimates might give biased results because they rely on unit values rather than prices. The next section discusses the biases that can result when unit values are used in demand studies.

III. Econometric Problems Created by Using Unit Values

The problems that may occur when unit values are used as proxies for market prices have been discussed previously in the literature by Cox and Wohlgenant (1986) and Deaton (1988, 1990, 1997). However, these problems remain widely ignored so in this section we summarize some of the potential biases that can occur from the use of unit values. In some cases, the direction of the bias depends on the particular demand specification used, so the discussion considers the double log and the share-log specifications. These two specifications are the ones used by Deaton (1987, 1990) in his unit value correction procedures, and the share-log is the one used in the empirical section of this paper.^[3] 

The models, written in terms of an arbitrary good i (but not indexing the individual household) are:

where Q_i is the quantity of food i, w_i is the budget share of food i, x is household total expenditure, p_i is the own-price, the p_j are cross-prices (i¹j ), the z_k are other relevant household characteristics, and the u_i is a random error. The use of unit values involves replacing ln p_i and ln p_j with ln v_i º ln E_i - ln Q_i, and ln v_j º ln E_j - ln Q_j. In the double log model, the own-price elasticity of quantity demand is directly estimated as γ_ii. In the share-log model, the elasticity is (γ_ii / w_i ) -1. The nature of the elasticity formulas, as it turns out, is important in evaluating the direction of bias.

Bias due to quality variation

The first problem that arises from using unit values as proxies for prices is quality variation. In markets in which prices are high, consumers may react by choosing goods that are lower quality.  In contrast, in markets in which prices are low, consumers may choose to consume items that are higher quality (Deaton, 1988). Hence, unit values, which reflect both price and quality, will tend to vary by less than prices, i.e., (¶lnv_i / ¶lnp_i )<1.  As a result, the absolute value of the γ_ii coefficient in (1) and (2) will be bigger when unit values are used than when market prices are used because the same movement in the left-hand side variables is attributed to smaller movements in the right-hand side variable. 

The direction of the bias depends on the sign of the γ_ii coefficient.  In model (1), γ_ii normally would be expected to be negative, so the bias makes demand appear more elastic, overstating the response of quantity to price (that is, the elasticity will tend to be further from zero--Deaton, 1988). In model (2), if the demand for food is own-price inelastic, then γ_ii >0, and the exaggerated size of γ_ii will make it appear as if the commodity demand is even more inelastic (that is, the elasticity will tend to be closer to zero).  Conversely, if the demand is own-price elastic, then γ_ii <0, and the use of unit values will make it appear that demand is even more price elastic than it truly is. Hence, for budget share models, when unit values are used, the effect of bias due to quality variation always moves estimates of the own-price elasticity away from minus one (-1).

Bias due to measurement errors

Two types of measurement error bias are relevant.  The first is “attenuation bias” due to the fact that unit values are noisy measures of prices.  Attenuation bias is expected to force the estimated γ_ii toward zero in both models (1) and (2).  As a consequence, the elasticities are biased toward 0 and -1, respectively.  Thus, attenuation bias due to random measurement error generally operates in the opposite direction to the bias due to quality variation in the unit values.

The second type of bias is due to “correlated errors” in expenditures and/or quantities that appear on both the left-hand and right-hand sides of equations (1) and (2). For example, consumers may not correctly recall the quantity of food consumed, Q_i, and instead over- or under‑estimate it as Q_i±ε_Q.  In this case, the denominator of the unit value expression contains an error.  The error, however, is not simply passed to the random error term of the regression (as it would if quantity only appeared on the left hand side of a demand regression).  Instead, because the unit value can be written as ln v_i º ln E_i - ln Q_i  there is a common component on the left-hand (ε_Q), and right-hand side (-ε_Q) of equation (1).  Thus, no matter what the true relationship between price and quantity, the estimated relationship will be more negative, due to the spurious negative correlation between quantity and unit value. Thus, correlated errors bias potentially counteracts the effect of the attenuation bias due to random errors and causes the response of quantity to price to be overstated.^[4]

IV. Specification of the Demand Models

To illustrate the possible biases from the uncritical use of unit values as proxies for market prices, we compare meat demand elasticities calculated from procedures which attempt to correct for the various biases with the results from simpler procedures which do not make these corrections. Ideally, we would have elasticities estimated from market prices to use as our reference point, but such estimates are unavailable because there is not sufficiently disaggregated market price information available to match with the SUSENAS household survey data.^[5] The procedure that we take as a reference point is the one developed by Deaton (1990), which was also illustrated using data from rural Java. Although this is the most sophisticated econometric procedure for dealing with the biases caused by unit values, it has rarely been applied.^[6] The other unit value procedures that we use are:

1.  replacing missing unit values with the mean unit value calculated across other households in the same province.
2. replacing missing unit values with the mean unit value calculated across other households in the same district. This procedure, and the replacement with provincial means are similar to what Minot (1998) has used, noting that there is no seasonal variation in SUSENAS because all households are observed in the same month.
3. replacing missing unit values with the cluster mean of the unit value (Sahn, 1988);
4. replacing missing unit values with the predictions from a regression of observed unit values on regional dummies and household total expenditures (Jensen and Manrique, 1998; Heien and Pompelli, 1989);^[7]
5. using cluster mean unit values, in place of both household-specific and missing unit values (Case, 1991; Rae, 1999).

The Deaton Procedure

In a series of papers, Deaton (1987, 1988, 1990) developed a methodology for making use of spatial variation in unit values from household survey data to estimate own- and cross price elasticities. The procedure starts with a two-equation system of budget shares (w_Gic ) and unit values (v_Gic) that are both functions of the unobserved prices, (p_Hc ):

the G indicates goods, i indicates households and the c indexes clusters of surveyed households in close geographical proximity. Amongst the explanatory variables, x_i is total expenditure of household i, p_H are the unobserved prices, z_i is a vector of other household characteristics,  f_Gc is a cluster fixed-effect in the budget share for good G  and  and  are idiosyncratic errors.

In other words, the budget share equation is assumed to be a linear function of the logarithm of total expenditure, of the prices and of a vector of household characteristics. Deaton’s method recognises that data are collected on clusters of household that live together in the same village and are surveyed at the same time. As a result, households within each cluster can be assumed to face the same market prices and there should be no variation in market prices within each cluster. The logarithm of the unit value is a function of the same variables that appear in the budget share equation, with the exception of the cluster fixed-effects.^[8]

The calculations take place in three stages. In the first stage, the within cluster variation in budget shares and unit values is used to estimate. As prices are constant within a cluster, both the cluster fixed effects and market prices disappear and the  parameters can be estimated consistently in their absence.   Any residual variation in unit values (and covariance with budget share residuals) is assumed to reflect measurement error, and the residuals from the first-stage regression give an empirical estimate of these errors. In the second stage of the estimation, a between-cluster errors in variables regressions is applied to the (adjusted) average budget shares and unit values, which have been purged of household characteristics at the first stage. At the third and final stage, a separability theory of quality, developed in Deaton (1988), has to be used to isolate the effects of price on the quantity demanded.^[9]

V. Data

The data used in this paper come from the detailed consumption module of the 1999 SUSENAS. The survey covered a random sample of 65,664 households, residing in 1,864 rural and urban communities.  Data for the present study were limited to 28,964 households located on Java. This island consists of the capital, Jakarta, and other fast growing cities including Surabaya (East Java), Bandung (West Java), Semarang (Central Java) and Yogyakarta. Java contains approximately 60 percent of the Indonesian population and is likely to account for a larger share of meat demand. The households in our sample were the ones for whom it was possible to merge the consumption module (used only every third year) with the core questionnaire, which is administered to a larger sample. It is the core questionnaire (used every year) that provides the information on the demographic characteristics and economic activity of the household.  The survey’s sampling procedure involved two-stage selection. At the first stage, sub districts are selected from Java’s provinces, districts and sectors (urban and rural); and at the second stage 16 households are selected per cluster. This spatial clustering encourages the assumption that households within each cluster face the same prices.

The consumption module includes 17 disaggregated meat items and for each of these, households were asked to recall the quantity and value of each of these meat items purchased from the market during the last week, or given to them as gifts or consumed out of own production. The latter quantities are valued by local interviewers using the imputed prices. The survey also collects non-food expenditure over the past month but expenditure on durables are excluded from the aggregate of household consumption. From this detail we formed 3 meat groups, which consist of beef, chicken, and other meat types. Most of the aggregation was for the other meat group, which is comprised of 14 individual products.

An important feature of the survey is that it does not collect market prices. Also no other survey gives sufficiently disaggregated prices to match with the spatial distribution of the clusters in our sample. Therefore, we used unit values (expenditures divided by quantities) as proxies for market prices. These were constructed by first converting all purchase quantities into kilograms, and then the unit values for each of the 17 meat items were aggregated, using a weighted geometric index, to give a unit value index for each of the 3 meat groups.  The weights used are the average budget shares for each component meat in the group, calculated over all households in the survey.  For example, for household j the unit value index for meat group k, ln V_jk depends on the unit values, v_ij, and weights for the i individual meat types that make up group k:

^[10]

The characteristics of the unit values for all Java, and for urban and rural Java are presented in Tables 1a-1c. Overall, chicken (consists of purebred and free-range chicken) is the most widely consumed type of meat in Java. Means of the unit values are also shown in the tables. These are computed from those households who make market purchases of the commodity under consideration. In general, beef prices are higher compared to other meat types, especially in urban areas. On average, consumers paid Rp. 22,157 ( ≈ A$4.14)  per kilo for beef and Rp. 12,390 (≈ A$2.32) per kilo for chicken.

The coefficient of variation (the standard deviation divided by the mean) can be used to indicate the degree of heterogeneity within each group, which is greatest for the other meat category, consisting as it does of 14 different types of meat.  Column (d) of the tables shows how many clusters record at least one household making a purchase for each of the meat categories. Overall, the number of clusters with unit values is the largest for chicken, followed by other meat and beef. It is also apparent that a substantial number of clusters have only one household that provides the unit values information, which in many procedures will provide the ‘price’ information used for all of the households in the cluster.

The shares of expenditures on each meat category are shown in the last columns of Tables 1a, 1b and 1c. A point to note is that these budget shares are not the ratios of aggregate consumption of each food to aggregate total consumption (so called ‘plutocratic budget shares’, see Deaton 1997), but the average of the budget shares for each household. These meat items contribute almost 2 percent of the average budget, ranging from chicken at 1.04 percent to other meat at 0.27 percent. The average budget share for chicken is almost three times as high as the budget share for beef. As can be seen from the tables, the urban population spends relatively more of their consumption budget on meat (especially for beef) than does the rural population, an indication of the relative affluence of the urban dwellers.

Table 1a. Commodities, Sample Sizes and Budget Shares for Java, 1999

Table 1b. Commodities, Sample Sizes and Budget Shares for Urban Java, 1999

Table 1c. Commodities, Sample Sizes and Budget Shares for Rural Java, 1999

VI. Empirical Results of the Elasticity Estimation

Deaton Procedure

Tables 2a-2c contain results from the first stage (within-cluster) estimation of the budget share and unit value equations. The first stage estimates explain much of the variation in unit values (between 51 to 68 percent for all Java), but somewhat less of the variation in budget shares. All of the expenditure effects on budget shares are statistically significant, as are all of those on unit values except for chicken and other meat in rural Java. The other (unreported) variables used at the first stage include (log) household size, a set of demographic variables (the number of household members in each of thirteen age and sex categories as a ratio of household size), and nine educational dummies. These variables are based on those used by Deaton (1990).

The positively signed coefficients in Tables 2a-2c indicate goods whose budget shares rise as household expenditures rise (i.e. luxury goods with expenditure elasticities greater than one) and this is the case for all three meat groups considered in the study. On average, beef has the highest expenditure elasticity (2.7 for Java) compared to other meat types(2.1 for both chicken and other meat). Tables 3a-3c also show that rural households tend to have larger expenditure elasticities than urban households, indicating that meat is even more of a luxury good in rural areas of Java. For instance, rural households have expenditure elasticities of 3.44 and 2.73 for beef and chicken compared with urban households whose expenditure elasticities are 2.17 and 1.67. These results contradict  Hutasuhut (2000), who found that the estimated expenditure elasticity is greater for urban households. The results also confirm the earlier evidence of Puslitbangnak (1992) that beef is more of a luxury than chicken, in contrast to the results of Hutasuhut et. al. (2002).

The quality elasticities,  show the rate at which unit values rise as households become better off, reflecting the purchase of higher quality meats within a group. In general, the quality components of expenditure elasticities are the highest for the other meat category (ranging from 12 percent for Java to 14 for rural Java), suggesting that this group is fairly heterogeneous in quality. It is apparent from Tables 2a-2c that the quality elasticity for beef is almost twice the size of the quality elasticity for chicken (the difference is even larger for rural Java). The fact that unit values vary systematically with household income cautions against treating unit values as if they were prices. When prices rise, households are effectively made poorer, so they will downgrade the quality of their purchases and the reported unit value will not rise by as much as the (unreported) price level. Hence, price elasticities calculated as the percentage decrease in quantity divided by the percentage increase in unit value will tend to be too big, and this effect may be especially apparent for beef compared with chicken.

Table 2a. First Stage Estimates: Effect of Total Expenditures on Quantity and Quality for Java 

Table 2b. First Stage Estimates: Effect of Total Expenditures on Quantity and Quality for Urban Java 

Table 2c. First Stage Estimates: Effect of Total Expenditures on Quantity and Quality for Rural Java 

Tables 3a-3c contain the estimated own- and cross-price elasticities for Java, as calculated by the Deaton procedure. In addition to the three meat groups, there is an extra row and column for “all other goods”, the estimates for which are obtained from the homogeneity and adding-up restrictions. The elasticities are conditional not only on household size and the dummy variables for household characteristics mentioned above, but also on a set of province and urban dummy variables. These dummy variables are used at the second stage (between-clusters) to control for any longer-term interregional price differences.^[11] In addition to the price elasticities, the tables also include bootstrapped estimates of “standard errors”. To calculate these standard errors, 1000 random draws are taken from the second stage data (i.e., the cluster average budget shares and unit values, after the effect of household total expenditures and other characteristics have been controlled for). For each of these random draws, all of the elasticities are recalculated. The length of the interval around the mean of each bootstrapped elasticity that contains 63.8 percent of the bootstrap replications is calculated and one-half of this interval is used as the estimate of the standard error. The rationale is that if the distribution of the elasticity estimates was normal, 0.638 is the fraction of a normal random variable within two standard deviations of the mean (Deaton, 1997).

All of the own price elasticities of demand are negative (except for the other meat category), as would be expected.^[12] Overall, the consumption of beef is more sensitive to changes in its own price than is the consumption of chicken. For example, a 10 percent increase in the price of beef (chicken) would reduce the amount consumed by 5.6 percent (2.5 percent).    Demand for beef is considerably more price elastic in rural areas than in urban areas, possibly reflecting the lower incomes of the rural population.  However, even with this greater elasticity, the results in Table 3c are considerably less elastic than almost all estimates in the literature either for Java or Indonesia, many of which also use SUSENAS data (see for example Chernichovsky and Meesok, 1984; Hutasuhut, 2000). However, Deaton (1990) points out that when unit values from SUSENAS are used as proxies for market prices and no correction is made for quality and measurement error effects, the expected bias is towards making elasticities too large in absolute terms.^[13]

To check whether the results of interest, for beef and chicken are being affected by the unexpectedly positive own-price elasticity for other meats, we also estimated a system where other meats are included with the non-meat aggregate. This forces the elasticities for other meats (along with for other, non-meat consumption) to be obtained from the homogeneity and adding up restrictions. The results for this new, two-meat system are reported in Tables 3d-3f, and it is apparent that there are similar patterns to those found previously. Specifically, the demand for beef is more price elastic than is the demand for chicken (with the exception for urban Java). However, the exclusion of other meat as a separately specified item in the demand system does change the point estimates of the own-price elasticities for chicken (from -0.251 to -0.421 for all Java), making the distinction between the own-price elasticity of demand for beef and chicken less apparent. It is also apparent from Tables 3e and 3f that the demand for chicken in urban Java is more own-price elastic than it is in rural Java, which is the opposite to the pattern for beef. This could be due to many rural households in Java raising chicken for their own consumption.

Table 3a.   Estimates of Own and Cross Price Elasticities for Java, 1999

       Note: Standard error in ( ); Results for “All Other Cons” derived from homogeneity and adding  up restriction.

Table 3b.   Estimates of Own and Cross Price Elasticities for Urban Java, 1999

Table 3c.   Estimates of Own and Cross Price Elasticities for Rural  Java, 1999

Table 3d.   Estimates of Own and Cross Price Elasticities for Java, 1999

Table 3e.   Estimates of Own and Cross Price Elasticities for Urban Java, 1999

Table 3f.   Estimates of Own and Cross Price Elasticities for Rural Java, 1999

                  Note: Standard error in ( ); Results for “All Other Cons” derived from homogeneity and adding up restriction.

Results Using Other Estimation Procedures

Table 4 contains the estimated own-price and cross-price elasticities using five different estimation procedures which do not make the corrections for unit value quality biases and measurement errors that the Deaton procedure makes. According to these simpler procedures, and in contrast to the Deaton method results in Tables 3d-3f, beef demand is much more own-price elastic than is the demand for chicken. The own-price elasticity of demand for beef appears largest when missing unit values are replaced with province-specific mean unit values. While this procedure also inflates the own-price elasticity for chicken, the effect on the beef elasticity is much larger. Thus, when province mean unit values are used for those households without a unit value, the demand for beef appears five-times as own-price elastic as is the demand for chicken, whereas when the Deaton method is used there is almost no difference in the beef and chicken own-price elasticities (based on the results for all Java).

Table 4. Comparisons of Price Estimates Using Different Unit Values Procedures^[14]

Replacing missing unit values with either the district mean unit value or with the predicted unit value from a regression on regional dummy variables and household total expenditures causes some small improvement in the estimated own-price elasticities – in the sense of bringing them closer to the estimates from the Deaton procedure. However, there is still no overlap with the confidence interval for the beef own-price elasticity that comes from applying the Deaton procedure (Figure 1). The fact that the own-price elasticities are especially large for beef when estimation methods make no correction for unit value biases is not surprising. Uncorrected quality variation is expected to make own-price elasticities larger in absolute value (Deaton, 1988). Thus, the fact that the quality elasticity for beef is twice the size of the quality elasticity for chicken (see Table 2) suggests that the neglect of quality effects will exaggerate own-price elasticities more for beef than for chicken.

The methods of either replacing missing unit values with cluster mean unit values or using cluster mean unit values in place of both household–specific and missing unit values have a closer correspondence of point estimates and a high overlap of confidence intervals with the elasticities from the Deaton method.^[15] The similarities of these two methods may reflect the fact that in the sample there are a large proportion of clusters that only have one consumer, so replacing missing values with the cluster mean is effectively the same as using the cluster mean for all households. Because of this feature of the current sample, these two methods may not agree as closely on other samples where there are more households with unit values in each cluster.

The issue of clusters with only one, or even no consumers is also relevant to choices about the appropriate aggregation level for the meats in the demand system. On the one hand, less aggregation makes it more likely that households report no purchase of the finely defined product, so there are more clusters having no unit value which can create modelling problems. On the other hand, more aggregated products have potentially greater bias in estimated elasticities due to quality variation. The results in Table 4 are from a reasonably disaggregated demand system, compared with previous applications of the Deaton method, so measurement error rather than quality variation may be the major cause of the differences in the elasticities coming from each method.

Figure 1: Comparison of Own-Price Elasticities from Different Methods of Using Unit Values

Statistical Comparisons of the Elasticities

The results in Tables 3 and 4 suggest that the different methods of calculating price elasticities from unit value data produce substantially different results, especially for beef. However, we would like to know which results are ‘better’ and whether the differences are statistically significant. Although we do not know that the Deaton model is the correct one, it is reasonable to presume that its results are more likely to be correct because this is the model that pays more attention to the problems created by unit values.^[16]

The second question of whether the differences are statistically significant is also difficult to answer. The difficulty comes from the two different methods used to calculate standard errors: bootstrapping for the Deaton method and the “delta method” approach to getting analytic standard errors of non-linear transformations of regression coefficients for the other methods. Therefore, for the purpose of making comparisons, we run bootstrapping experiments for all of the estimation approaches, even though they are not needed for the non-Deaton methods. Within each bootstrap replication we calculate the sum of squared deviations (SSD), where if  is the vector of elasticities calculated from the Deaton method and  is the corresponding elasticity vector from one of the other unit value methods, the sum of squared deviations is calculated as  The standard deviation of the SSD across each of the bootstrap replications gives an empirical estimate of the sampling variability surrounding the SSD, which then helps to assess the statistical significance of the differences.

Table 5 contains the results of these bootstrapping comparisons. The calculations are restricted to the own-price and cross-price elasticities for chicken and beef and are also calculated just for Java, rather than for the urban and rural sectors separately. We find that for all five of the non-Deaton procedures whose results were reported in Table 4, the sum of squared deviations (SSD) is more than two times the associated standard errors. The smallest SSD is when missing unit values are replaced with the cluster mean unit value (SSD=1.69±0.66), while the largest come from replacement with province means and the use of predictions from a regression of unit values on expenditures and other covariates. Overall, the comparisons in Table 5 suggest that the differences in the elasticity estimates are statistically significant and are unlikely to just result from the particular idiosyncratic characteristics of the sample studied here.

VII. Conclusions

In Indonesia, as in many developing countries, demand studies often rely on household surveys, with unit values used instead of market prices. Because Indonesia is an emerging market for beef and cattle exports, accurate estimates of the price responsiveness of beef and competing product demand may be useful to producers (Hutasuhut et al., 2002). In this paper we attempt to assess the possible role that unit value biases may play in interfering with the accurate estimation of meat demand elasticities. Specifically, meat demand elasticities calculated from procedures which attempt to correct for the various biases that may result from the use of unit values are compared with the elasticities that result from simpler procedures which do not make these corrections.

The results suggest that estimated meat demand elasticities are sensitive to the choice of procedure for dealing with unit values. The simplest procedures make beef demand appear much more own-price elastic than do the more sophisticated estimation methods. While the estimated own-price elasticity of demand for chicken also varies according to the method used to deal with unit values, the variation in the price elasticities is less severe than for beef. One consequence of these results is that pricing strategies for beef producers, which are based on estimated demand elasticities from Indonesia, may prove to be inappropriate if the wrong method for dealing with unit value biases is used.

References

Alston, J. and Chalfant, J. (1993) The silence of the lambdas: a test for the AIDS and the Rotterdam models. American Journal of Agricultural Economics 75(4): 304-313.

Case, A. (1991) Spatial patterns in household demand. Econometrica 59(4): 953-965.

Chernichovsky D, and Meesook A. (1984) “Patterns of food and nutrition consumption in Indonesia” World Bank Working Papers Series No. 670. Washington, DC: World Bank.

Cox, T. and Wohlgenant, M. (1986). Prices and quality effects in cross-sectional demand analysis. American Journal of Agricultural Economics 68(4): 908-919.

Deaton, A. (1987). Estimation of own and cross-price elasticities from household survey data. Journal of Econometrics 36(1): 7-30.

Deaton, A. (1988). Quality, quantity, and spatial variation of price. American Economic Review 78(3): 418-430.

Deaton, A. (1990). Price elasticities from survey data: extensions and Indonesian results. Journal of Econometrics 44(3): 281-309.

Deaton, A. (1997). The Analysis of Household Surveys: A Microeconometric Approach to Development Policy Johns Hopkins, Baltimore.

Dixon, J.A. (1982) “Use of expenditure survey data in staple food consumption analysis: Examples from Indonesia” in Chisholm, H. and Tyers, R. (eds), Food Security: Theory, Policy and Perspectives from Asia and the Pacific Rim. Lexington, Massachusetts: Lexington Books.

Eales, J. and Unnevehr (1993). Simultaneity and structural change in U.S. meat demand. American Journal of Agricultural Economics 75(2): 259-268.

Eales, J. and Wessells, C.R. (1999). Testing separability of Japanese demand for meat and fish with differential demand systems. Journal of Agricultural and Resource Economics 24(1): 114-126.

Gracia, A., and Albisu, L. (1998).  The demand for meat and fish in Spain: urban and rural areas. Agricultural Economics 19(3): 359-366.

Hahn, W.F. (1994). A random coefficient meat demand model. The Journal of Agricultural Economics Research 45(3): 21-30.

Heien, D. and Pompelli, G. (1989). The demand for alcoholic beverages: economic and demographic effects. Southern Economic Journal 55(3): 759-770.

Hutasuhut, M. (2000). The demand for meats in Indonesia: a censored regression approach. Unpublished doctoral thesis, University of New England, Armidale.

Hutasuhut, M., Chang, H., Griffith,G., O’Donell, C. and Doran H. (2002). The demand for beef in Indonesia: implications for Australian agribusiness. Australasian Agribusiness Review 10(4): 1-10.

Jensen, H., and Manrique, J. (1998). Demand for food commodities by income groups in Indonesia. Applied Economics 30(4): 491-501.

Kesavan, T., Altemeier, K., Rake, C., Alirahmand and Adinugroho, B. (1992). An Analytical Model of Indonesian Agriculture: Design and Structure, Winrock International and Bappenas Cited in Hutasuhut (2000).

Meat & Livestock Australia (2001). Other Markets & Competitors: South Asia, MLA Market Information Services, January.

Minot, N. (1998). Distributional and nutritional impact of devaluation in Rwanda. Economic Development and Cultural Change 46(2): 379-402.

Puslitbangnak (1992). Estimasi Parameter Sistem Permintaan Komoditas Ternak dan Hasil Ternak di Jawa, Pusat Penelitian dan Pengembangan Peternakan, Badan Penelitian dan Pengembangan Pertanian, Departemen Pertanian, Jakarta Cited in Hutasuhut (2000).

Rae, A. (1999). Food consumption patterns and nutrition in urban Java households: the discriminatory power of some socio-economic variables. Australian Journal of Agricultural and Resource Economics 43(3): 359-383.

Sahn, D. (1988). The effect of price and income changes on food-energy intake in Sri Lanka. Economic Development and Cultural Change 36(2): 315-340.

Timmer, C.P. and Alderman, H. (1979) “Estimating consumption parameters for food policy analysis” American Journal of Agricultural Economics 61(5): 982-987.

Yong, S. and Hayes, D. (1993). Testing the stability of preferences: a non parametric approach. American Journal of Agricultural Economics 75(2): 269-277.

^[1]Address for correspondence: Department of Agricultural and Resource Economics, University of California, Davis, One Shields Avenue, Davis, CA 95616. E-mail: olivia@primal.ucdavis.edu. We are grateful for helpful comments received from two anonymous referees and from participants at the Australian Agricultural and Resource Economics Society conference in Perth, 2003. Financial support from the Waikato Management School is gratefully acknowledged.

^[2] There are initially 17 meat types in the SUSENAS data set. However, Hutasuhut  (2000) discarded households with pork consumption from the data set because the majority of respondents in the chosen study areas do not consume pork due to religious reasons.

^[3] The share-log model is also closely related to the linear approximate Almost Ideal Demand System, with budget shares treated as a linear function of log income and log food prices.

^[4] See Deaton (1997) for a more detailed treatment of the effects of measurement errors in unit values.

^[5] The Indonesia Family Life Survey (IFLS) collects data on community-level prices, but there is insufficient overlap of the commodities with the SUSENAS budget items, and the survey was also carried out in a different year.

^[6] Gracia and Albisu (1998) are one of the few examples in the agricultural economics literature. Sample code for the Deaton procedure is available in Appendix A.5 of Deaton (1997), although this needs considerable adaptation to each particular sample. The version of the program used here is available from the authors upon request.

^[7] A related procedure is to regress the deviation of household-specific unit values from the mean for each region in each quarter on a set of household characteristics and use this equation to predict adjusted unit values for the non-consuming households (Cox and Wohlgenant, 1986).

^[8] Cluster fixed effects are not allowed in the unit value equation because they would obscure the link between unit values and the unobserved cluster prices.

^[9] If it was not for the effect of prices on cluster-wide quality variation, the parameters estimated at the second stage would be sufficient for calculating price elasticities.

^[10] Although the same weights are applied for all households, they are rescaled in each case to account for goods not purchased to ensure that the weights add to one for each household (Deaton, 1997).

^[11] It is not possible to add them at the first (within-cluster) stage because the cluster fixed effects obliterate them.

^[12] The unexpected result for other meat could be because this item is an aggregation of 14 different types of meat. This aggregation also means that there is less intrinsic interest in this category than there is for beef and chicken. To keep the focus on those two products, the results reported below constrain the own-price elasticity for other meat to be the same as for all other, non-meat products.

^[13] Deaton (1990) used SUSENAS data to estimate price elasticity for meat. He does not however disaggregate the meat category.  So we cannot compare with his estimates.

^[14] The estimated own-price elasticities using unadjusted unit values on the subset of households recording consumption of each good do not confirm with the economic theory as it gives positive own-price elasticity for beef and chicken (with the exception of own-price elasticity for chicken in urban Java). The elasticity matrices using unadjusted unit values are available from the authors.

^[15] Some of the previous food demand elasticities calculated from SUSENAS data have in fact used the method of replacing household-specific unit values with cluster means (Case, 1991; Rae, 1999), so the results here may provide some support for this procedure.

^[16] Moreover, there is some Monte Carlo simulation evidence in favour of the model reported in Deaton (1990).