Contents - Previous - Next

Correlation between low BMI and morbidity

The association between low levels of BMI and illness is examined in Figs 3A and 3B. There are two ways by which the data are tabulated. Figure 3A draws attention to the probability of falling below 18.5 BMI for those reporting illness in the recent 2, 4, or 8 weeks against those who were not ill. Figure 3B on the other hand looks at the morbidity status for two groups of individuals: those falling below 18.5 BMI and those termed as having a normal BMI (18.5-23). Clearly, those individuals with BMI >24 needed to be excluded to control for confounding by morbidity risks due to obesity. Results of this bivariate analysis indicate that the probability of being sick does not vary substantially between those with low BMI and those who are considered to be 'normal'. The differences are not statistically significant in any of the countries examined except in males in the Philippines.

Fig. 4. Probability of low body mass index (BMI) in adults by illness days in last 2 weeks.

Fig. 5. Probability of low body mass index (BMI) in adults by illness days in last 8 weeks.

The relationship between low BMI and duration of illness is examined in Figs 4 and 5. The results indicate that the probability of an individual falling below the 18.5 cut-off BMI does not vary significantly with duration of illness except for the males in Pakistan where the prevalence of low BMI tends to rise for men who were sick for more than 12 days in the preceding 8 week period.

It is likely that the bivariate results are confounded by many factors, including income, age, energy expenditure and activity levels. It is for this reason that a multiple regression analysis is more useful in drawing the direction and size of the relationships between morbidity and low BMI.

To test the probability of being malnourished (below the 18.5 BMI cut-off), a set of regressions were run. The dependent variable in this equation is the dummy variable which takes the value of one if BMI is <18.5 and zero otherwise. The model is run excluding those with BMI >24 to control for the confounding by disease risk factors associated with obesity. As noted earlier, the problem of circularity in the BMI-illness causation should be considered in the model. This implies that ordinary least squares (OLS) is not suitable for the estimation of the illness coefficients. In this paper, we will attempt to mitigate any possible covariances of low BMI and illness by using instrumental variables (two-stage least squares) estimation procedures. We also modelled using OLS estimates the unadjusted illness effect on BMI.

The instrumental variables used to predict illness in the first stage of the regressions include: income measured by per capita expenditures, household size, potable water and a community illness variable. Morbidity is the result of many factors including exposure to pathogens and parasites and by availability and quality of health services. A community level variable, called community illness, is used as an explanatory instrumental variable in order to control for the fact that community pathogens are one of the potential causes of illness and infections. This is calculated as the mean illness prevalence of the village where the individual resides, and is estimated by excluding the individual in question. The procedures for this have been found to be useful in the analysis of the nutrition of children (Alderman & Garcia, 1992). The results of the instrumenting equations to predict illness are given in Annex 1. The most important instrument is that of community illness which explains a large proportion of the variance in illness patterns of the individual women. The other notable instrument is that of age which is in line with the expectation about health and ageing.

The main results using probit regressions (maximum likelihood probit estimates) are found in Table 3. After controlling for simultaneity, the model predicts moderate statistical significance for the morbidity variables in Pakistan and Kenya but not in the Philippines and Ghana. This implies that at least in the environment in Pakistan and Kenya, that illness predisposes individuals to low BMI. The size of the coefficients in Pakistan is much larger than in Kenya. It must be noted that earlier analysis by Kennedy (1989) using OLS on the same Kenyan data shows the strong impact of low BMI on illness. The recalculation of such effects using the instrumental variable approach here indicates that the strength of such relationships diminishes, although they are still clearly statistically significant. This result implies that after controlling for the two-way causation, the probability of low BMI is caused by a high morbidity.

No such relationships are found in the Philippines and Ghana. The findings in the Philippines confirm results by Haddad & Bouis (1991) on the same Philippine households where BMI is not a good predictor of agricultural productivity in that environment. On theo other hand, height was found to be more strongly associated (positively) with wage rates. The results for Ghana show a much lower level of significance compared with those estimated by Higgins & Alderman (1992) using a different estimation technique. The direction of causality in the present study is, however, in line with their study.

Table 3. Probit regressions: probability of low body mass index (< 18.5) in adults 15-59 years old

Maximum likelihood probit estimates

Pakistan (male)

Kenya (female)

Ghana (female)

Philippines (female)










Expenditure per capitaa

- 0.0019

- 0.00027

- 0.00023

- 0.0054







- 0.0018

- 0.0250

- 0.0371





Age, squared









Days ill in recent 2 weeksa,b









Household size

- 0.0062

- 0.0020

- 0.014



(- 1.76)*

Energy expenditure



Sex of household head



Percent pre-school children




(- 0.14)

No. of observations





*P <0.1; **P <0.01.
t-values are in parentheses.
a Endogenous right-hand-side variables.
b Pakistan, 8 weeks; Kenya, Philippines, 2 weeks; Ghana, 4 weeks. Pregnant and lactating women are excluded from sample.

The regressions in Table 3 also provide some explanations for the probability of a low BMI. The age variable is significantly negative for all the countries under study, and it is also clear that the effect is not linear as indicated by the quadratic term for age. This implies that the prevalence of low BMI declines as one advances to ~35 years, then rises thereafter. One of the surprising results relates to the effect of household size. Smaller size households tend to have more women with a low BMI, which is not consistent with the expectations about scarcity of resources as families get bigger. In Kenya, however, the larger households tend on average to be better off. One of the explanations could be the effect of household composition. The presence of more preschool children in Kenya also predisposes the Kenyan women to low BMI. Higher incomes also reduce the probability of low BMI, and it is strongly associated in Pakistan and Ghana.

Conclusions and summary

The analysis from four countries has shown mixed results with respect to the issue of the impact of low BMI on morbidity of individuals. It is likely that many unobserved factors affect thinness in adults other than morbidity. One of the important pathways that is difficult to model is the impact of activity level, as well as the role of community pathogens and infectious disease patterns which are outside the control of households and the individuals. It is clear from the modelling used here that circularity of causation is recognized.

One of the main concerns from this empirical investigation is the utility of the 18.5 cut-off BMI for diagnosing CED. It seems reasonably clear that the threshold at which morbidity begins to rise is not generally consistent with this cut-off at least in the types of households depicted in the present analysis. The cut-off appears meaningful only in the case of Pakistan, but even here the sizes of the relationships are quite small although statistically significant.

Acknowledgements - We thank Elizabeth Jacinto and Ellen Payongayong for their excellent research assistance.


ACC-SCN (1992): Second report on the world nutrition situation. Geneva, Switzerland, October.

Alderman H & Garcia M (1992): Food security or health security? Explaining the levels of nutrition in Pakistan. Washington DC: WB Working Papers Series.

Garcia M & Alderman H (1989): Patterns and determinants of malnutrition in children in Pakistan: impact of community health. Pakistan Dev. Rev. 28, no. 4.

Gurney M & Gorstein J (1988): The global prevalence of obesity. An initial overview of available data. Wld Hlth Stat. Q. 41, 251-254.

Haddad L & Bouis H (1991): The impact of nutritional status on agricultural productivity: wage evidence from the Philippines. Oxford Bull Econ. Stat 53, 45-68.

Higgins P & Alderman H (1992): Labor and women's nutrition: a study of energy expenditures, fertility, and nutritional status in Ghana. Ithaca, NY: Cornell Food Nutr. Policy Prog.

James WPT, Ferro-Luzzi A & Waterlow JC (1988): Definition of chronic energy deficiency in adults. A report of a working party of the IDECG. Eur J. Clin. Nutr 42, 969-981.

Kennedy E (1989): Effects of sugarcane production on food security, health and nutrition in Kenya: a longitudinal analysis. Washington, DC: IFPRI Res. Rep. 79.

Kennedy E & Garcia M (1992): BMI and economic productivity. Washington, DC: IFPRI.

King E, Rosenzweig M & Yan Wang (1991): Assessing the economic burden of illness: evidence from 8 countries. Washington, DC: World Bank.

Lopez AD (1990): Causes of death: an assessment of global patterns of mortality around 1985. Wld Hlth Stat. O. 43, 91-104.

Annex I. Illness instrumenting equations (dependent variable: days ill in recent weeks)b

Pakistan (male)

Kenya (female)

Ghana (female)

Philippines (female)










Household expenditure per capita









Age (years)









Average village illnessa









Household size









Potable water supply (I = clean)








F value










*P <0.1; **P < 0.01 (t values in parentheses).
aAverage days ill in recent weeks of sample individuals in the village. The individual's own illness is excluded in the calculation of the village average.
bPakistan (8 weeks); Kenya, Philippines (2 weeks); Ghana (4 weeks).


Durnin: How did you estimate the income of the females?

Kennedy: We had data on sources of income within the household for off-farm income. For farm income we had information on who controls the plots, crop production and distribution, whether it was marketed or for home consumption and if it was for home consumption it was valued at what they would have to pay in the open market.

Durnin: In the Philippines, males were ill for 7 days out of 14. Is that representative of that population?

Kennedy: Yes, there was a lot of malaria. Increased income improved energy consumption and anthropometry but did not change the morbidity patterns of children or adults. It is probably because it is such a squalid environment with so much disease about that the extra money for food does not make up for this.

Naidu: With such high morbidity, poor environments, and lack of distinction between kinds of severity of illness, do you think any relationship with BMI is being masked?

Kennedy: The methodological problems are indeed great and this self-reporting method is not ideal. The 2 week period of recall is too long: 1 week would be better. The reported numbers are high but we know there is underreporting of illness.

Ferro-Luzzi: I would like to ask Eileen Kennedy, are there any more data on LSMS World Bank on Ghana or from elsewhere to use to address the issue of mortality and disease?

Kennedy: Yes, there is a series of LSMS studies that are nationwide representative samples. But their method for collecting morbidity was a 4-week recall and validation studies suggest that recall beyond 1 week is unreliable; it underestimates illness and people only remember their most severe disease.

Scrimshaw: The period of recall is situation specific. In Guatemala where the same worker went to the same family every 2 weeks for 3 years, we did get useful data from a 2-week recall. In another situation, validation of 2-week recall showed underestimation, and picked up only severe illness. If you have a large enough sample, a point prevalence, in which you combine the self-reporting of the individual, with probing and direct observation, can be useful for that period only.

Waterlow: In many countries most children seem to be ill about 15-20% of the time. If you do a 1-day point prevalence, out of every 100 children you would get 15 ill, and that is a good number. In many situations you could easily visit 200 children a day, and from a practical point of view it seems little work, so I like the idea of point prevalence.

Ferro-Luzzi: A point prevalence for recognized disease is OK but it would be difficult for nonspecific disease, not easily identified.

Scrimshaw: There are problems with point prevalence because of fluctuating levels of endemicity and hyperendemicity with superimposed epidemicity and seasonal fluctuations with diarrhoea, respiratory disease etc. and, of course, no field method will pick up unsymptomatic disease like HIV. With frequent rectal swabs and a good laboratory you could pick up very high levels of non-symptomatic disease.

Waterlow: I don't think we can expect to diagnose diseases like HIV, tuberculosis, malaria, but we can record symptoms and that is useful.

Allen: There is a case for detecting illnesses such as malaria that are clearly caused by environmental pathogens.

Contents - Previous - Next