Print this page
Themes > Features
19.09.2000

Messing about with Poverty Statistics

Throughout the 1990s, the results of the NSS "thin samples" on household consumption expenditure generated much interest in both academic and policy circles. These results, which suggested at an all-India level that rural poverty did not show any declining trend over the 1990s despite higher rates of aggregate GDP growth, were seen as important inputs into the ongoing policy debate on the effects of the liberalising economic policies instituted by successive governments over the 1990s.
 
Those who had been questioning the economic reform package have pointed out that these policies have led to a neglect of rural and agricultural investment, resulting in reduced productive employment opportunities in rural areas and to higher food prices, developments likely to be associated with persistent or even increasing rural poverty. They pointed out that the evidence of stagnation or even decline in rural non-agricultural employment was also in conformity with the evidence of persisting rural poverty.
 
By contrast, proponents of the economic policies of the 1990s suggested that the "thin samples" simply did not allow for such conclusions, and that nothing could be said about rural consumption or poverty until the next large sample results were available. It was also argued by proponents of the official economic reform strategy, that the NSS consumption expenditure estimates from the thin samples were out of sync with the CSO estimates of GDP. Thus they maintained that the association of higher GDP growth rates with the persistence of rural poverty in particular was not a real fact which needed to be understood and addressed, but more of a failure of the statistical system to capture the actual consumption in rural areas.
 
With this background, obviously the results of the 55th Round (1999-2000) of the NSS, which is the first "large sample" since 1993-94, have been eagerly awaited. Now that the round is complete and the data due to be released later this year, it has also become evident that this NSS Round is important and interesting for another reason. This is that in the 55th Round, a different technique has been used to establish household consumption levels. The basic change is in terms of a change in the reference period.
 
Since the 1950s, the NSS consumption surveys have been using an uniform reference period of one month. Since the interviews are evenly spread out over the year, problems of seasonality were ironed out. However, in recent years the NSS has revived the issue of whether a one week reference period is more suitable for determining non-durable consumption than the one month reference period currently used.
 
This is not a new issue : indeed, it is a question which has been of concern to the NSS since the very inception of the surveys in the early 1950s. In fact, in the formative years of the NSS, considerable attention was paid to the length of reference period suitable for ascertaining the correct level of consumption of different items of goods and services, and a special report was brought out covering the period April 1951 to March 1954 on the suitability of reference period.
 
Most interestingly, the NSS under guidance of P.C. Mahalanobis had carried out a special investigation into this very issue in March-April 1952, based on 1254 households of 76 villages of West Bengal. The households were divided into two groups. For one group, consumption details were procured by actual weighing of food items (clean rice, pulses, sugar and salt) by field staff. For the other group, data collection was by questioning, and here again the group was divided equally between those for whom the questions pertained to a reference period of one week and those for whom the reference period was one month.
 
The results were quite interesting. It was found that the two sets of data obtained by questioning differed quite sharply, with the consumption estimates obtained on the basis of the one week reference period being much higher than those obtained on the basis of one month recall. It was also found that the one month reference period generated information that corresponded much more closely to the data on the basis of actual weighing of food items. This led Mahalanobis and others to conclude that the one month reference period was better suited to the purposes of the estimates especially of food consumption.
 
Since then, it has been standard for the National Sample Survey to use the one month reference period for food items, although both the one month and one year reference period have been used for some non-food items. In the five quinquennial surveys of household consumption expenditure between 1972-73 and 1993-94, information for clothing, footwear and durable goods was collected from each sample household for two reference periods - "last 30 days" and "last 365 days". In the 50th round, "educational" and "institutional medical" expenses were also added to the list of items for which data were collected by these two reference periods.
 
However, during the 1990s, the question of the most suitable reference period has resurfaced, and the NSS has in its thin samples experimented on the basis of alternative schedules (based on one week and one month recall) in independent samples during the course of the same survey. This was done for the 51st Round (1994-95), the 52nd Round (1995-96), the 53rd Round (1997) and the 54th Round (Jan-Jun 1998).
 
In all of these surveys, half the sample was canvassed using a Type 1 schedule which had a 30 day reference period for all items, but the other half was canvassed using a Type 2 schedule which had a one week reference period for food, pan, tobacco and intoxicants and a one year reference period for clothing, footwear, durable goods and educational and medical (institutional) services. However, since the Type 2 schedule was not comparable to earlier NSS surveys, the results by this schedule were not tabulated in the NSS Reports of the relevant rounds, so that all the available analysis of consumer expenditure and of poverty during the nineties are based on the Type 1 schedule.
 
It is now reported that for the large sample of 1999-2000, this experiment has been carried one step further, to the point where the two different schedules of Type 1 (one month) and Type 2 (one week) have been canvassed from every sample household. To the extent that the two schedules give varying results, incorporating them so that all the households respond to both schedules is obviously problematic since it would bias the results of both schedules. It also means that there are definite problems of comparability with data from past surveys, for which the reference period was essentially one month and one week was not used at all.
 
Fortunately, the NSS has now released the preliminary results on the effect of the choice of reference period for the 51st, 52nd, 53rd and 54th Rounds in its Report No. 447: Choice of Reference Period for Consumption Data. It is, therefore, possible to examine the effect of choosing one reference period over another. As mentioned, above, in all of these Rounds, one half of the sample (Type 1) had a reference period of one month (30 days) throughout; for the other half (Type 2) the reference periods were as follows : one week for all food, pan, tobacco and intoxicants; one month for fuel and light and miscellaneous goods and services; and one year for clothing, footwear and durable goods as well as education and institutional medical expenses.
 
The results of these samples based on alternative schedules are extremely interesting. It emerges that Type 2 samples give higher overall food consumption, exactly as Mahalanobis had predicted and as was also confirmed by the pilot investigation in the West Bengal villages in 1952. Thus, the Type 2 schedule-based samples also suggest that poverty is much lower.
 
In Charts 1A and 1B, the cumulative distribution of population below specified per capita total consumption levels as obtained from the Type 1 and Type 2 schedules are plotted for the rural and urban sectors, using data from the 52nd round. It may be seen that the proportion of population below any expenditure level is always higher by the Type 1 schedule than by Type 2, and the difference is extremely large. Thus, using the Planning Commission poverty line, about 39 per cent of the rural population would be below the poverty line in 1995-96 by the Type 1 schedule but this percentage would be only around 20 per cent by the Type 2 schedule. The corresponding percentages for urban areas are 30 and 15. Similarly large differences are obtained for the 51st, 53rd and 54th rounds.



This of course raises the question about which of these two schedules gives a better measure of the actual incidence of poverty in India. But an even more important problem is the implication that the 55th round may end up giving a totally biased picture of poverty trends during the 1990s if the results from this are compared to earlier rounds which did not use the "one week" reference period.
 
Thus, suppose, for example, the 55th Round comes up with an estimate of 25 per cent of rural population below the poverty line in 1999-2000. This would be compared to the corresponding 37 percent rural poverty obtained from the 50th Round in 1993-94, and the implied large reduction in poverty would be greeted by the reformers as vindication of their policies. But, since this does not compare like with like, such a conclusion would obviously be erroneous and the debate about trends in poverty would enter a statistical minefield which the 55th Round results will not be able to resolve.
 
The problem is that since the 55th Round has canvassed both the Type 1 and Type 2 schedules from all households, there would have been a pressure for consistency between the answers to the "one week" and "one month" reference periods on the part of both respondents and investigators. It is obvious that when the household is questioned using both the one week and one month reference period, the answers are likely to be tested by simple multiplication of the one week reply for the monthly response as well. This means that the two schedules can no longer be seen as independent and may give misleading results depending upon how the conflation of the referenced periods affects the responses.
 
If such pressure for consistency has led to primacy being given to the "one week" response, a 25 per cent rural poverty incidence in the 55th round would correspond to a 45 per cent poverty incidence by the procedure followed in 1993-94 so that the proponents of liberalising reforms could end up claiming massive poverty reduction while in fact poverty might have increased massively.
 
Even if the pressure for consistency has worked evenly across the schedules, so that the 55th Round results are an average of the two, a 25 per cent rural poverty incidence in this Round would correspond to a 36 per cent poverty by the method used in earlier rounds, so that a large poverty reduction could be claimed without there having been any significant reduction from the actual poverty level in 1993-94.
 
In view of this, the 55th Round stands severely compromised in its ability to give estimates of mean consumer expenditure and poverty comparable to that in earlier Rounds. This is particularly the case because, as Charts 2A and 2B show, the errors associated with measuring the population in lower expenditure classes is much higher by the Type 2 schedule. These errors are computed by the NSSO on the basis of the variance obtained across sub-samples in the same survey and for the same schedule type, and show the much greater inherent unreliability of poverty estimates calculated with data from the Type 2 schedule.



The National Sample Survey Organisation needs, therefore, to conduct another large sample survey on the basis of the earlier schedule exclusively, to give comparable estimates. In the meantime, the 55th Round should be treated as an experimental survey whose comparability with past surveys is poor, but which may further elucidate the intriguing differences revealed by the results of the 51st to 54th Round surveys. This is important because much mud has been slung at the NSS consumer expenditure data in the course of the recent debate on poverty, and the air needs to be cleared so that questions about the reliability of data does not continue to cloud assessment of such an important issue.
 
Much of the criticism of NSS data in the recent past has concentrated on the fact that these give lower estimates of total consumer expenditure than the alternative estimates from the National Accounts Statistics (NAS), and some have even claimed that the difference between these two estimates have grown alarmingly during the nineties. In an earlier edition of Macroscan (February 22nd, 2000), it had been pointed out that although the NSS does indeed give a lower estimate of overall consumption than the NAS, and although the difference did increase during the seventies and eighties, the ratio between these two estimates have remained fairly stable during the nineties so that the 1990s trends in poverty obtained from NSS data cannot be challenged on this ground.
 
Nonetheless, since this has been the main thrust of the attack on the NSS data so far, we present in Table 1 the ratios of the NSS to NAS estimates of consumption by broad items and over successive full-year NSS rounds beginning 1977-78. Since only the NAS estimates with base 1980-81 cover this entire period, these ratios have been calculated with these NAS estimates rather than the new estimates with 1993-94 as base. As mentioned earlier, the ratio of total NSS consumption to NAS consumption is seen to decline from 0.81 in 1977-78 to 0.69 in 1990-91 but remains almost constant thereafter. However, the more important differences between these two estimates relate to the item-wise results.

As may be seen there are certain persistent differences between these two data sets at the level of individual items. Thus, for sugar, edible oils, fruits & vegetables, milk & products, pan, tobacco & intoxicants, and other goods & services, the NSS has consistently measured lower consumption but with no obvious time trend in these ratios. In the case of meat,fish & eggs and clothing, NSS has lower consumption and the ratio has fallen over time. For cereals, the ratio has always been close to unity but with some tendency to decline over time. But, for pulses, other food and fuel & light, the NSS has consistently measured higher consumption than the NAS.
 
These persistent differences between NSS and NAS data have been analysed thoroughly by resarchers, notably B.S. Minhas and his associates, who have noted that it is normal all over the world for items like intoxicants to be under-reported by respondents and that something similar is probably true also for non-vegetarian items in a country such as India.
 
Also, they note that much of sugar, edible oils, milk & products and fruit & vegetables are consumed not directly by households but are purchased after processing either by hotels and restaurants or by other manufacturers. In such cases, these would appear differently in the NSS and NAS data, with the former including these under "other foods" while the latter would include them directly under the item concerned.
 
This explains also why the NSS has tended to measure higher expenditure under "other food". The relative over-estimation by the NSS of fuel & light has likewise been explained by failure of the NAS to adequately capture fuel wood and twigs collected directly by households. Thus, for most of the above items, the differences are not particularly surprising or unexpected, especially given that the NSS does not capture all consumer expenditure since it leaves out institutional consumption such as in hostels, prisons and ceremonials.
 
However, for certain items such as clothing and "other non-food" the differences are large and have been attributed in past analysis both to a failure of NAS to measure household consumption correctly and to a failure of the NSS to adequately capture the consumption of the relatively richer household who consume relatively more of these.
 
Indeed, it is on the basis of the these observations, that the Expert Group on Poverty Estimates had decided in 1993 that the differences between the NSS and NAS were unlikely to cause any serious bias in poverty measures estimated directly from NSS estimates, and had accordingly decided to end the practice till then of adjusting these estimates to conform to the NAS.
 
However, the issue of alternative reference periods has now again opened up this issue since apparently much of the difference in food consumption between the NAS and the NSS can be resolved if the one week rather than the one month schedule is used in the latter. Indeed, in its Report No. 477, the NSSO reports higher errors for week-based estimates as compared to estimates based on 30 days, but on noting that "the substantial and systematic differences between the week and month based estimates indicate that one or both methods are not depicting the real life situation", goes on to claim some support for the Type 2 schedule since the total of the one week estimates is closer to the NAS than the totals of the 30 day estimates.
 
Although the NSSO itself is careful on the matter, suggesting that further methodological surveys would be advisable, others may not be so careful. They may not only ignore the fact that the relative standard error of each item canvassed on the one week basis is higher than by the 30 day schedule, but also fail to notice that the recent experiments merely reproduce what Mahalanobis and his associates had found almost 50 years ago.
 
As mentioned earlier, it was found in the 1950s that the one month estimate of most food items was lower than the one week estimate but was in greater conformity with the physical (weight) measures of actual household consumption of foodstuffs, than the one week estimate. Thus, the bias between results from the two reference periods continues to remain in the same direction, and with no experiment through physical weighment repeated, there is no further evidence for judging the relative plausibility of the two estimates. Comparison with NAS is a poor substitute given the past judgement of researchers such as Minhas and the assessment of the Expert Group on Poverty Estimates, both of which found strong reason not to accept the NAS as necessarily giving reliable estimates.
 
Table 2 gives the perentage differences between the NSS and NAS (1980-81) using both the schedules in the former. It may be noticed that although the Type 2 NSS estimates are fairly close to the NAS for all food items taken together, while the Type 1 NSS estimates are about 20 per cent lower, this result comes about because the Type 2 estimate gives higher estimates for all food items including for those where the Type 1 estimate is higher than the NAS estimate. As a consequence, the apparent concordance between the Type 2 NSS estimates and the NAS is something of statistical artifact because these week-based results continue to show large shortfalls from the corresponding NAS estimates for sugar, edible oils, milk & products and meat, fish & eggs but almost double the estimate for "other food". If both positive and negative divergence are given equal weight to measure the difference between the NSS and NAS, the results from the Type 1 schedule turn out to be closer to the NAS for food items. Also the large gap between the NAS and NSS for non-food is further widened when the Type 2 rather than the Type 1 results are considered.

Table 3 gives the absolute value of consumption estimates for 1995-96 from both the Type 1 and Type 2 NSS schedules and from both the NAS estimates with base 1980-81 and base 1993-94. This not only shows the patterns discussed above but also certain inherent infirmities in the NAS data.

Thus, for two items, "pan, tobacco and intoxicants" and "clothing", the NAS has substantially revised downwards its consumption estimates, between the 1980-81 and 1993-94 series, bringing these closer to the NSS estimates. But for two others, "fruits and vegetables" and "other non-food" the NAS has revised upwards its estimates and thus increased the gap with the NSS. For "other non-food" there is at least the likelihood that new goods and services were being underestimated earlier and may not be captured in the NSS which does miss out on the rich who consume these more, but the doubling of fruit and vegetable consumption is intriguing and highly suspicious.
 
As discussed in an earlier Macroscan, not only does this not correspond with the known area under horticultural crops, it has the effect of making the NAS estimate three times the NSS Type 1 estimate and more than double even the NSS Type 2 estimate. In view of these large and sometimes inexplicable revisions, the NAS can hardly be said to represent the type of benchmark that Mahalanobis had set himself when he actually carried out physical weighment to check the validity of reference periods.

Finally, as Charts 3A and 3B show, there is the intriguing fact that the Type 2 schedule sample results show richer households consuming relatively more food and less non-food as compared to the Type 1 schedule, thus overturning much of what is now accepted wisdom regarding changing consumption habits along the Engels Curve. This too should suggest a more careful look at the estimates emanating from the Type 2 schedules, and by implication, of the overall consumption results of the 55th Round.



 

© MACROSCAN 2000