3. THE ESTIMATION OF PREVALENCE OF UNDERWEIGHT CHILDREN

3.1 The calculation of predicted prevalence of underweight
3.2 Clinic data - Figures 6B to 6E of the Report
3.3 Analysis of covariance - Figure 2 of the Report
3.4 Interactions in estimating underweight prevalence for Asia

3.1 The calculation of predicted prevalence of underweight

A brief description of the methods used to estimate prevalences of underweight by country groups can be found in section 4.1.6 of the Report (p. 48). These methods are discussed in more detail here, including additional statistical results and a more thorough explanation of the approach used.

Generally the procedure used for obtaining the prevalence of underweight by country group, involved the following steps.

1. Compilation of all available and nationally representative surveys with data on prevalences of underweight in 0 through 4 year old children in developing countries (n=45) from 1975 to present Underweight is defined as below -2 standard deviations weight-for-age (WA) using NCHS standards.

2. Using the prevalence of low weight-for-age data as the dependent variable (outcome), a regression equation was calculated by inserting estimates for a priori chosen independent factors. This is referred to here as the interpolation model.

3. The interpolation model is then applied to a set of developing countries (n=94), grouped by region of the world; by inserting values for the independent factors in the model, a predicted prevalence was calculated for each country.

4. The country group estimates of low WA prevalence were calculated by deriving the number of malnourished children by country, based on the predicted prevalence culated in step 3. totalling these values by country group, and dividing by total regional population of under five year olds.

Calculation of interpolation model - relationship of independent factors with prevalences: The dataset used for the calculation of the regression is listed in Table All. There are 36 countries with 45 surveys (9 countries had more than one survey) since 1975. The independent factors available for the analysis discussed above are all chosen from the same year as the prevalence data.

The bivariate relationships between the prevalence of malnutrition and the independent factors: kcals/caput/day, log of GNP, and infant mortality rate (IMR), are shown in Figures I-III. The simple regression line of prevalence on log GNP or IMR. respectively, is shown on these graphs, with the appropriate equation listed below the plot The letters represent points (i.e. countries) on the graph designated by different country groups.

Figure I: Plot of prevalence of low weight-for-age by Kcals/caput/day

KEY

F:	Sub Saharan Africa
C:	Central America
M:	South America
A:	South Asia
E:	Near East
S:	Southest Asia

Regression of PREV2SD (prevalence of under -2SD Weight-for-age) against KCALSY (Kcals/caput/day for survey year):

Correlation = -0.58, R Squared = 0.33, SE of Estimate = 17.2
Significance of regression p<0.0000

Intercept (SE): 126.001 (20.8); Slope (SE): -0.043 (0.009)
45 cases plotted

Figure II: Plot of prevalence of low weight-for-age by log GNP

KEY

F:	Sub Saharan Africa
C:	Central America
M:	South America
A:	South Asia
E:	Near East
S:	Southest Asia

Regression of PREV2SD (prevalence of under -2SD Weight-for-age) against LNGNP (log GNP):

Correlation = 0.69. R Squared = 0.48. SE of Estimate = 152 Significance of regression: p<0.0000

Intercept (SE): 129.575 (16.4); Slope (SE): -16.5,7 (2.65) 43 cases plotted

Figure III: Plot of prevalence of low weight-for-age by IMR

KEY

F:	Sub Saharan Africa
C:	Central America
M:	South America
A:	South Asia
E:	Near East
S:	Southest Asia

Regression results from prevalence <-2SD's wt-age by IMR, IMR * DASIA:

R squared = 0.83, significance of regression = 0.000, n = 45

Intercept (SE) = 15.19 (3.01) coefficient for IMR = 0.062 (P = 0.036)
coefficient for IMR * DASIA = 0.205 (P = 0.019)
coefficient for DASIA = 25.03 (P = 0.011)

The lines shown, calculated from this equation, are:

Asia: Prevalence = 40.218 + 0.267 (IMRSY)
Other regions: Prevalence.= 15.188 + 0.062 (IMRSY)

The relationship of both GNP and kcals with prevalence of low WA arc very similar (Figures I and II), and in the expected direction. As GNP increases, the level of malnutrition decreases; likewise as kcals per caput increases, prevalence decreases.

IMR is positively related to prevalence, as expected. The two lines on the graph (Figure III) represent separate relationships for Asia and the rest of the world. The line for Asia clearly shows that prevalence of malnutrition is much higher at a given IMR than the other countries in the sample. In. fact, in all three graphs the prevalence of low WA is higher than expected in Asia; an issue which is discussed in more detail in section 3.4.

Regression analysis - procedure and results: The regression method involved inserting the factors plotted in Figures I-III, and other dervised variables, to best explain the variation in the outcome, i.e. prevalence of low WA. The criteria used for determining the most appropriate model were based on the R-square statistic, the t and p- values for the independent factors, and an analysis for the residuals. All analyses were done using the SPSS/PC statistical package for the IBM PC/XT microcomputer.

Since GNP and kcals showed almost identical relationships with prevalence, two separate models were derived, and both are reported here. All factors in the two models were otherwise identical. The kcal model was used as the best predictive model because it showed the strongest association with prevalence.

Kcal model: Results of the interpolation model including kcals/caput/day are shown in Table I. The adjusted R-square is 0.927, indicating that 92.7% of the variation in prevalence of low WA is explained by the model. The analysis of variance (ANOVA) shows a highly statistically significant model with an F-value of 80.69 (p<0.000). All independent factors are listed in Table I.

Table I: Results of multiple regression analysis (OLS) for prevalence of low wt/age (KCAL model). See text for description of variables.
Multiple R	.96877	Analysis of Variance
R Square	.93852		DF	Sum of Squares		Mean Square
Adjusted R Square	.92689	Regression	7	18100.97065		2585.85295
Standard Error	5.66101	Residual	37	1185.74135		32.04706
		F - 80.68923		Signif F =.0000
Variables in the Equation
	Variable	B	SE B	Beta	T	Sig T
	DASIA	17.30734	6.56822	.31962	2.635	.0122
	IMRSY	-7.88043E-03	.02681	-.01895	-.294	.7705
	DSEASIA	13.63137	3.00127	.20693	4.542	.0001
	DSAMER	-10.85593	3.06321	-.16480	-3.544	.0011
	DCAMCA	-4.19829	2.79055	-.08112	-1.504	.1409
	KCALSY	-.01610 4.	21436E-03	-.21707	-3.821	.0005
	INTERI	.23531	.05482	.49299	4.292	.0001
	(Constant)	59.28480	11.66129		5.084	.0000

The regression equation can be written in the following form:

Predicted Prevalence = 59.285 - 0.0161(KCALSY) - 0.00788(IMRSY)+17.307(DASIA)"
- 4.198(DCAMCA) + 13.631 (DSEASIA) - 10.856(DSAMER)
+ 0.235(INTER1)

Where:

KCALSY is kcals/caput/day for survey year.
IMRSY is Infant Mortality Rate for survey year.
DASIA is a dummy variable for South Asia (1=yes, 0=no).
DCAMCA is a dummy variable for Middle America and Caribbean(0,l).
DSEASIA is a dummy variable for Southeast Asia (0,1).
DSAMER is a dummy variable for South America (0,1).
INTER1 is the interaction term for IMR in Asia (=DASIA x IMRSY).

The slope for KCALSY was statistically significant (p<0.000). The coefficient of - 0.0161 means that for every increase of 100 kcals/caput/day (for example), the prevalence estimate will decrease by 1.61%, after accounting for the influence of all other factors in the equation. It should be remembered that this is not assumed to be a causal relationship but simply means that the two factors are changing together by a measurable amount.

Although IMR showed no statistically independent relationship with prevalence it was included in the equation because of its highly significant interaction with the term for Asia.

The independent variables also include:

1. All variables beginning with "D" are dummy variables for country groups, as listed above. A value of "1" is applied for all countries in the country group, and a "0" for all else.

2. INTER1 is the interaction of IMR with DASIA. This term is included to differentiate this relationship from the rest of the country groups, as shown in Figure III.

The following example demonstrates the use of dummy variables and the interaction terms to calculate separate equations to interpolate national prevalences, in this case, one for Asia, and one for South America, From the regression equation above, a "1" is inserted for DASIA for Asia and the resulting equation is as follows, for Asia:

Prevalence = (59.285 + 17.307) - 0.0161 (KCALSY) + (- 0.00788 + 0.235)(IMRSY)
= 76.592 - 0.0161(KCALSY) + 0.227(IMRSY)

To do me same for a country in South America, a "1" is inserted for DSAMER for all countries in South America (note that there is no interaction term with IMR):

Prevalence = (59.285 - 10.856) - 0.161(KCALSY) - 0.00788(IMRSY)
= 48.429 - 0.0161(KCALSY) - 0.0078 8(IMRSY)

Notice that the relationship between kcals and prevalence of low WA is the same in both groups. However, both the intercept and the relationship of IMR with prevalence varies by country group.

Analysis of residuals for the kcal model: The residual⁴ values from the kcal model are shown in Figure IV. All cases arc listed by country group. The variables given are defined in the figure.

⁴. Residual = observed value residual minus (from model).

CCODE is the Country code (see Table All for full names, listed in the same order).
PREV2SD is the actual prevalence of below 2 SD's WA.
*PRED is the predicted value of prevalence of low WA from the model.
*RESID is the residual value. PREV2SD - *PRED.
*ZRESID is the standardized residual (residual/SD of residual) - plotted.
*SEPRED is the standard error of the predicted value.

The plot of residuals shows how well the model predicts the actual country/year value; points which are furthest away from the "0.0" line show the largest discrepancy between actual and predicted prevalences. It is important to note whether a majority of countries in a particular country group show a consistently high or low residual value (me standardized residual is more informative for this purpose). The residuals show an acceptable scatter with no discernible pattern. It should be borne in mind that only two countries in the Near East/North African region were available for interpolation. Further analysis of residuals (from an analogous model) is given in Haaga et al. 1985.

The plot of actual versus predicted prevalence is shown in Figure V. This plot, being an alternative presentation to Figure IV. provides a similar picture as the casewise plot of residuals (Figure IV), showing how far away the actual prevalence in a country is from the predicted value (the line designates a perfect agreement). Agreement is reasonably good, as expected from the high R-squared for the model.

Figure V: Plot of prevalence of low weight-for-age by predicted prevalence from KCAL model

KEY

F:	Sub Saharan Africa
C:	Central America
M:	South America
A:	South Asia
E:	Near East
S:	Southest Asia

Regression of PREV2SD (prevalence of under -2SD Weight-for-age) against PREDWA1 (Predicted WA from KCAL model):

Correlation = 0.96, R Squared = 0.93. SE of Estimate = 5.25
Significance of regression: p< 0.0000

Intercept (SE): -0.019 (1.38); Slope (SE): 1.000 (0.039)
45 cases plotted.

GNP model: Another model was calculated using the natural log of GNP as a predictor of prevalence of low WA, instead of kcals, in order to verify estimates from the kcal model. The log GNP model is expected to be very like the kcal model, given the similarity in their relationships with prevalence (Figures I and II). The natural log of GNP was used because it shows a linear relationship with prevalence (see Haaga, et al. 1985), whereas GNP alone shows an exponential relationship with prevalence. All other independent factors in the kcal model were included in this model (except kcals). The model was calculated for 1980 and 1984 only.

Regression results for the GNP model are shown in Table II. Both the adjusted R-square and the F-value for the regression are lower than the kcal model, (see Table I) albeit still very high. A plot of the actual versus predicted prevalence is shown in Figure VI.

Table II: Results of multiple regression analysis (OLS) for prevalence of low weight-for-age (log GNP model). See text for description of variables.

Multiple R	.96404	Analysis of Variance
R Square	.92937		DF	Sum of Squares		Mean Square
Adjusted R Square	.91524	Regression	7	17182.93442		2454.70492
Standard Error	6.10823	Residual	35	1305.86465		37.31042
		F = 65.79141		Signif F =.0000
Variables in the equation
	Variable	B	SE B	Beta	T	Sig T
	DASIA	17.95794	7.84966	.31972	2.288	.0283
	IMRSY	-9.04543E-03	.03311	-.02183	-.273	.7863
	DSAMER	-6.13377	3.56025	-.09482	-1.723	.0937
	DSEASIA	13.53705	3.72270	.18963	3.636	.0009
	DCAMCA	-.55302	3.07131	-.01085	-.180	.8581
	LNGNP	-6.08339	2.09687	-.25665	-2.901	.0064
	INTER1	.22708	.06183	.47489	3.673	.0008
	(Constant)	59.22568	15.33481		3.862	.0005

Figure VI: Plot of prevalence of low weight-for-age by predicted from GNP model

KEY

F:	Sub Saharan Africa
C:	Central America
M:	South America
A:	South Asia
E:	Near East
S:	Southest Asia

'Regression of PREV2SD (prevalence of under -2SD Weight-for-age) against PREDWA2 (Predicted WA from GNP model):

Correlation = 0.96, R Squared = 0.92, SE of Estimate = 5.64 Significance of regression: p<0.0000

Intercept (SE): -0.005 (1.49); Slope (SE): 1.000 (0.043) 43 cases plotted.

This model was not used for the interpolation of prevalence, but only to verity results from the kcal model. Comparisons of prevalence estimates from born models are shown in Table III.

Table III: Country group estimates of prevalences and numbers underweight calculated from the KCAL and GNP models (see Tables I and II)

Year	Model	1975		1980		1984
Country Group		Prev %	Numbers (millions)	Prev %	Numbers (million.)	Prev %	Numbers (million.)
Sub-saharan Africa
	kcal	24.7	14.78	23.6	16.07	25.3	19.45
	gnp			21.8	14.80	23.0	16.53.
Middle America/Caribbean
	kcal	12.7	1.93	10.2	1.75	8.9
	gnp			13.8	2.37	13.8	1.69 2.51
South America
	kcal	7.8	2.25	6.4	2.04	6.4	2.25
	gnp			6.9	2.19	7.5	2.62
South Asia
	kcal	73.9	89.33	70.0	93.80	66.7	97.70
	gnp			70.0	93.91	67.4	92.08
Near East/North Africa
	kcal	17.1	4.28	12.3	3.44	10.2	3.20
	gnp			14.5	4.02	14.9	4.17
Southeast Asia
	kcal	37.1	16.06	34.0	16.37	32.7	17.51
	gnp			34.8	14.19	33.6	14.09
China
	kcal	25.8	20.64	22.1	19.03	17.7	16.02
	gnp			24.4	20.98	24.0	21.68

Validity of model for interpolation of several time points: Because the purpose of calculating prevalence was to obtain an estimate of trends by country group, one concern was the validity of applying an equation calculated from a combined sample of countries and years (1975 -1985) to three points in time. The major concern was that the effects of time may be confounded by country group if a particular group (or groups) have surveys from only one point in time (e.g. 1982/4). In this case one could not tell if apparent differences over time were in fact due to differences between countries. The sample data was examined for this and was found to have an equal distribution in two time periods, 1975 - 1979. and 1980 - 1985, and most country groups were represented in both periods. Southeast Asia had more surveys in the latter period (N = 1 and 4 respectively). The Near East was only represented in the earlier period, but anyway only data on two countries were available from this region. No information on nutritional status in China existed for this analysis. Thus, because these two regions were under-represented in the original dataset, their predicted prevalence may be less reliable; however, the relationship between prevalence and kcals was so strong and consistent in all other countries, that this is unlikely to affect prediction to a large extent.

To examine whether kcals/caput/day could accurately predict prevalence of low weight-for-age in both time periods, the model was re-calculated separately for both samples. Results showed that kcal was significantly and negatively related to prevalence in both periods; it was a better predictor in the earlier years (as judged by the adjusted R²), but residuals indicated a poorer fit than in the later period. Thus, the relationship between kcals and prevalence held for both time periods, and the method was used to interpolate prevalences and aggregate them to country group for 1975, 1980 and 1984.

Interpolation of prevalences of low weight-for-age by country: When the calculation of a regression equation for the prediction of prevalence, was completed, the model was applied to a larger set of countries for the calculation of country group estimates of underweight. The intention was to include as many developing countries as possible in this larger set, divided into the following regions: Sub-Saharan Africa, Middle America and Caribbean, South America, South Asia, Near East and North Africa, Southeast Asia and China (not including Mongolia). Inclusion depended on me availability of data, particularly for kcals and population. Adequate data were available for 92 countries, as listed in Table AI (and Table 3 of the Report⁵). The countries not included were all either industrialized or of relatively low populations.

⁵ Two countries (Kuwait and United Arab Emirates) were eventually deleted from the list given in Table 3 of the Report due to unavailability of data.

The independent factors included in the kcal model were required for every country and for the three years chosen for interpolation of prevalence. Only kcals/caput/year and IMR were required for estimating prevalence: kcals was obtained as three year averages around the year of interest (e.g. the value for 1975 was the average of 1974, 1975, and 1976). Since IMR was only available in five year intervals, estimates for the model (as shown in Table AI) were interpolated from these values as follows:

"IMR 1975" = 172 ((IMR for 1970-75) + (IMR for 1975-80))
"IMR 1980" = 1/2 ((IMR for 1975-80) + (IMR for 1980-85))
"IMR 1984" = 3/5 (IMR for 1980-85) + 2/5 (IMR for 1985-90)

For every country and all three years of interest, a predicted prevalence of under -2 SD's weight-for-age was calculated by inserting the values for all factors into the model.

Calculation of country group estimates of underweight: Country group prevalences of underweight were derived by first calculating the numbers of 0 through 4 year old underweight children at the country level, and then totalling these by country group and dividing by me regional population of children of this age. The population estimates of 0 through 4 year old children for each country were derived by applying a known proportion of 0-5 years of age to the total population of the year of interest These proportions were calculated from 1985 population statistics taken from the UN Population Division database in AGROSTAT, as described in section 2.7 above. The total population estimates used to calculate prevalence over time were three year averages around the years 1975, 1980, and 1984. These were also extracted from data from the UN Population Division.

Country group estimates were obtained for both the kcal model and the GNP model. Prevalence estimates and numbers malnourished for both models arc shown in Table III; but only those from the kcal model were graphed in the panels in the Report.

3.2 Clinic data - Figures 6B to 6E of the Report

Trends in prevalence below 80% weight-for-age (WA) in Figures 6B to 6E are from clinic data compiled by the Catholic Relief Services (CRS) Growth Surveillance System (GSS). In many instances clinic data are the only data available on the level of malnutrition in a country, but the validity of using clinic data for estimating national levels of malnutrition is questionable.

Little work has been done to verify the representativeness of clinic data (see Test, et al, 1987 and Serdula, et al. 1987), i.e. how accurately clinic-based measures estimate the level of malnutrition in a population. This section discusses some of the issues concerned with the validity of clinic-based data for this purpose. It is important to distinguish between the use of clinic data for a cross- sectional estimate of prevalence (i.e. one point in time), and its use for comparing estimates over time. In the latter case. as long as the factor(s) which bias these estimates are not changing over time, then although the actual prevalence may be incorrect the trend is likely to be valid. These issues must be borne in mind when interpreting trends shown in Figures 6B to 6E.

Factors biasing clinic-based estimates of malnutrition: There are basically four major sources of bias in clinic data collected by the CRS-GSS, which must be considered when extrapolating to the general population. These are:

1. Self-selection by the mother.
2. Selection (at the clinic level) by age of the child.
3. Selection by nutritional status.
4. Graduation out of the programme (clinic level).

Selection at the clinic level by age of the child was practised in some clinics in the CRS programme. This could bias prevalence estimates if the ages selected had very different levels of malnutrition than the rest of the sample. For example, in malnourished populations children under the age of about one year tend to grow normally (i.e. similar to children from healthy populations) (Martorell, 1984). If these children are omitted from a sample of all children under five, then prevalence estimates will increase because the "best" group is being eliminated. The extent of this increase will depend on how large the under one year old population is relative to other age groups, and on how great a difference there is between the prevalence estimates of both groups. As far as could be ascertained, age distributions were unlikely to be changing greatly over time in the data used.

Selection of children at the clinic level based on their nutritional status would have the most drastic effect on estimating prevalence trends. For example, if only weights of children of less than 80% wt/age were recorded, the prevalence would consistently be reported as 100%. whatever the situation in the population. Even if this happened in only some clinics, trends would be obscured. Such selection (known as "medical selection" in some programmes) was looked for carefully, in the data and by enquiry, and suspect results eliminated. Nonetheless, more work on this aspect would be desirable.

Graduation out of the programme is usually based on the child's age; thus, its effects on prevalence estimates are similar to those of selection by age of the child into the programme. Information on prevalence by age of the child and the proportion of all children by age group would help determine to what extent this bias can alter prevalence estimates.

Validity of clinic estimates used in Figures 6B to 6E of the Report: All issues discussed above were addressed in the analysis which derived the estimates used. Details are discussed in Test, et al, (1987). One analytical approach used clinic coverage, calculated as the number of children attending the clinics in a given year divided by an estimate of under five year child population of the region, as a proxy variable reflecting possible changes in the biases discussed above. It was considered that the prevalence estimate based on clinic data may not be accurate at a given point in time, but it is the trend over time that was of concern in the study. Any biases affecting clinic-based estimates would only affect trends if the bias changed over time concomitantly with prevalence. To investigate this. change in coverage was examined. Statistical tests on the change in yearly and monthly prevalences were done after controlling for change in coverage; and results showed that although coverage was significantly related to prevalence in four of the five countries studied, it did not affect estimated trends in prevalence over time.

After examining other potential biases by country, it was concluded that the trends observed over time were likely to be real. This was supported by other available information including: 1) the consistent seasonal trend in prevalence over several years, 2) the impact of the drought in 1983/84 in many Sub-Saharan African countries, 3) associations with annual trends in agricultural production economic indicators.

3.3 Analysis of covariance - Figure 2 of the Report

Figure 2 in the Report shows global trends in mean prevalence (< -2 SD's) weight-for-age over time since 1975. Prevalence was calculated from the actual anthropometric survey data (Table AH) so may not accurately represent regional prevalences. For this reason a more appropriate title would be "Overall National Survey Prevalences of Low Weight-for-age in pre-school Children, 1975-1985, Adjusting for Country Group".

Mean prevalence over time was modeled using an analysis of covariance while controlling for the effect of country group, which if not accounted for could confound yearly trends. Analysis of covariance was chosen because it permits the testing of both categorical (country groups) and continuous (survey year) factors; the dependent variable being prevalence of low weight-for-age. In this case, the trends over time (survey year) was tested after accounting for differences between country groups.

Results of the analysis of covariance are shown in Table IV. There are highly significant differences between regions, as expected, with a highly significant F-value of 49.9 (p <0.000). Survey year as the covariate is significant at the 10% level (F = 2.85, p = 0.099), after controlling for region. The regression coefficient between prevalence of underweight and survey year was -0.739 indicating a negative relationship - in other words, prevalence decreased over time since 1975. Examination of the graph indicates that most of this downward trend was probably during 1975-1980. but this was not examined further by this method.

Table IV: Results of analysis of covariance testing prevalences of low weight-for-age by region and over time, using data shown in Table All

Source of Variation	Sum of squares	DF	Mean square	Sig.
				F Value	of F
Region	16574.72	5	3314.9	49.935	0.000
Year (covariate)*	189.37	1	189.4	2.853	0.099
Residual	2522.63	38	66.4
Total	19286.71	44	438.3

* The covariate, survey year, was tested after removing the effects of region. The regression coefficient for prevalence by survey year was - 0.739 (P <0.1).

The mean prevalence estimates graphed in Figure 2 of the Report were adjusted for regional effects obtained from an analysis of variance. This analysis was identical to the analysis of covariance except survey year was entered as a categorical variable and not a covariate (i.e. continuous), grouped by year as indicated on the graph. A Multiple Classification Analysis (MCA. from SPSS/PC for the IBM PC/XT) table gave the adjusted mean prevalence.

It must be emphasized that the downward trend in global prevalence observed over time. as shown in Figure 2 of the Report, must be interpreted with caution because mean prevalence was only based on available survey data, and the significance of the trend was fairly small after accounting for region.

3.4 Interactions in estimating underweight prevalence for Asia

It has been observed that prevalence of malnutrition, particularly wasting, is much higher than expected in South Asia when compared to populations in other developing countries. This is not thought to be an ethnic difference since well-nourished children from India for example, grow in height and weight like children from industrialized countries.

This phenomenon has been observed here, in the relationships of prevalence of low weight-forage with kcals/caput/day, IMR, and GNP, shown in Figures I - III. At a given level of the independent factor (X-axis), prevalences in the South Asian countries (designated by an 'A' on the graph) are much higher than all other countries in the sample. The most striking was with IMR (see Figure HI), where prevalence at a given IMR value in the South Asian countries was not only higher than all other country groups, but also increased much more rapidly as IMR increased. In this case, a separate relationship with prevalence was drawn for the Asian countries, as seen in Figure III.

This phenomenon has not been studied in sufficient detail to suggest a feasible explanation for the markedly higher prevalence of underweight in the Asian countries. For this reason, as mentioned in the Report, estimates of underweight prevalences in the South Asian region should be interpreted with caution in so far as the causes and consequences of malnutrition are concerned.