Using Engel curves to measure CPI bias for Indonesia

Abstract

To measure real income growth over time, a price index is needed to adjust for changes in the cost of living. The consumer price index (CPI) is often used for this task; but several country studies show that it is a biased measure of such changes, leading to potentially inaccurate estimates of the rate of real income growth. This paper calculates CPI bias for urban Indonesia by estimating food Engel curves for households with the same level of CPI-deflated incomes at four different points in time between 1993 and 2008. The results suggest that CPI bias was negative during the 1997–98 crisis but has been positive since 2000. From 1993 to 2008, CPI bias averaged 4% annually, equivalent to almost one-third of the measured inflation rate.

Keywords:

• CPI bias
• Engel curve
• inflation
• measurement error

Acknowledgements

The authors are grateful to Iyut Ria Muttaqun and Yunita Rusanti for assistance, and to seminar audiences at the SMERU Research Institute, Jakarta. Olivia acknowledges support from the Monash University Category 1 Research Grant Incentive Scheme.

Notes

¹Substitution bias occurs when a Laspeyres-type index does not recognise that consumers make economising substitutions for items whose prices have risen since the base period (when the CPI basket of goods was formed). Outlet-substitution bias results from shoppers responding to lower prices by switching outlets while price surveyors do not. Quality-change bias occurs when improved goods have quality upgrades reflected in higher prices (wrongly treated as inflation). The delayed introduction of new goods into the basket creates a bias if a rapid fall in the quality-adjusted price early in the product life-cycle occurs before the goods are incorporated into the CPI. A fifth bias, due to the formula used to aggregate individual price observations into an index of price relativities, is easier for statistical authorities to control.

²This method only partially accounts for quality change and does not recover the consumer surplus from new goods. Even after correcting official statistics for CPI bias, living standards might be under-stated.

³Prior to 1996, the geographic coverage was 27 provincial capital cities in Indonesia, then it increased to the 45 cities described here, and since June 2008 coverage has further expanded to 33 provincial capital cities and 33 other big cities. Given that our sample spans from November 1993 to February 2008 (see the ‘Data and estimation framework’ section in this paper), we have based our description of the CPI on the number of cities surveyed by BPS within most of the studied period.

⁴See <http://stats.oecd.org/mei/default.asp?lang=e&subject=8&country=IDN>.

⁵See <http://www.bi.go.id/NR/rdonlyres/712BB89D-B49A-41AC-A7DC-F94C6F8D4F3D/10672/Boks2.pdf>.

⁶The CPI weights for 1993–97 relied on the consumption patterns of households from 27 provincial capital cities in the 1988 Cost of Living Survey. The number of cities surveyed expanded to 44 in the 1996 survey (with 249 to 353 commodities in each city's basket) and to 45 in the 2002 survey (with 283 to 397 commodities in each city's basket).

⁷Food prices rose faster during the 1997–98 crisis than during the 2005 inflationary episode, owing largely to the additional effects of a severe drought that damaged crops. For instance, rice prices increased by 50% in Java during the 1997–98 crisis (Tradeport 1999).

⁸Frankenberg, Thomas and Beegle (1999) suggest that actual inflation during the financial crisis may be as much as 15% higher than the BPS-derived inflation estimates, because food prices rose much faster than non-food prices and the BPS figures put a smaller weight on food than do those of other surveys.

⁹Diary surveys are believed to yield a more accurate level of consumption than recall surveys, but empirical evidence for this is not conclusive. Evidence from both urban Papua New Guinea (Gibson 1999) and Tanzania (Beegle et al. 2012) shows that recall surveys measure lower consumption than a carefully conducted diary, while Ahmed, Brzozowski and Crossley (2010) compared diaries and recall methods applied to the same households in the Canadian Food Expenditure Survey and found that there was considerable measurement error in the diary surveys. Furthermore, preliminary comparisons of consumption from the Bosnia and Herzegovina recall and diary surveys in 2004 showed no clear differences between the methods (World Bank 2006). Moreover, for recall error to interfere with the estimates of CPI bias made here, the error would need to have changed over time. We can see no reason for such a pattern.

¹⁰These provinces are North Sumatra, West Sumatra, South Sumatra, Lampung, DKI Jakarta, West Java, Central Java, DI Yogyakarta, East Java, Bali, West Nusa Tenggara, South Kalimantan and South Sulawesi.

¹¹Although the Engel-curve relationship should hold for any group after having controlled for taste variables, using a more homogeneous sample should produce a better estimate of CPI bias.

¹²Because the IFLS is a panel survey, the changing demographics do not reflect average changes occurring in Indonesia; the members of the population in the panel survey age, whereas those in refreshed samples drawn every year would not. This ageing is unlikely to affect the CPI bias estimates; Gibson, Stillman and Le (2008) obtained almost identical CPI bias estimates from panel and cross-sectional data for Russia.

¹³The hypothesis test results that we report are based on standard errors that are clustered at the enumeration area (n = 247). If we instead cluster at the household level (n = 3,769) – which is possible since there are repeated observations on households – the standard errors would be slightly smaller. The hypothesis tests, regression coefficients and descriptive statistics also rely on population sampling weights.

¹⁴Unlike in the standard situation, where random measurement error would attenuate the estimate of β, and therefore increase the estimate of the cumulative CPI bias (since β is in the denominator of equation (7) in appendix 1), expenditures in the current case are also in the denominator of the dependent variable, so it is not possible to determine a priori the direction of any bias.

¹⁵Even if the area fixed effects are ignored (which they should not be, based on the highly significant F-tests at the foot of table 2), the coefficients on the term measuring relative inflation rates for food and non-food become only weakly significant and still have a magnitude less than one-half of that of the coefficients on real income.

¹⁶The particular instrumental variables that we used are quadratics in annual income from wages, assets and business net profits, and dummy variables for households with nil wage income and for those with government workers.

¹⁷The n = 247 fixed effects for enumeration areas that are included (but unreported) in the results in table 2 are obliterated by the addition of these household fixed effects. The specification of the fixed-effects model is otherwise the same as that reported in table 2.

¹⁸In contrast, the annual CPI bias in 1997–2000 is less than one-third as large, and the 95% confidence interval (based on two standard errors) for the cumulative bias in this period includes zero. All of the other cumulative bias estimates in table 3 are precisely estimated and do not have confidence intervals that cover zero.