Cover Image
close this bookPositive Deviance in Child Nutrition - with Emphasis on Psychosocial and Behavioural Aspects and Implications for Development (UNU, 1990, 153 pages)
close this folderResearch considerations
View the documentIntroduction and purpose
View the documentUnderlying assumptions or hypotheses for research in positive deviance
View the documentRelationship of positive-deviance research to epidemiological methods
View the documentDefinition of terms and specification of research goals
View the documentThree-stage research and pilot-project model
View the documentResearch design for stage 1
View the documentA conceptual framework for the design of positive-deviance studies
View the documentImportant variables: results of the positive-deviance mail survey
View the documentMicro-level variables measuring caretaker-child interactions
View the documentVariables measuring maternal characteristics and socio-cultural support
View the documentMeasuring growth
View the documentControlling for socio-economic status
View the documentLimiting the number of covariables: restriction by age and topic
View the documentRationale for existing behaviours and social structures
View the documentTimeline for change
View the documentNutrition and infection
View the documentManagement of multidisciplinary teams

Measuring growth

Measuring growth is particularly important to positive-deviance studies because the growth variables identify the children who are positive deviants. As of 1987, growth should be measured and assessed using the WHO methods and NCHS standards presented by Lavoipierre and colleagues (1983). Well-nourished (W), average-nourished (A), and malnourished (M) may be categorized using methods that differ according to the nature of the study design, as indicated in the following sections.

Cross-sectional Designs

It is legitimate to compare infants measured at one point in time, contrasting those whose length or weight falls below given cut-off points with those who are larger. On a probability basis, the group of larger children will undoubtedly be better nourished than the smaller ones. If, for example, we pick-2 SD in height-for-age z (HAZ), based on the NCHS reference population, as our cut-off point, the probability that a truly well-nourished pre-school child will fall into our malnourished group is only about 2.5 per cent. This cut-off point has excellent specificity for identifying malnourished children. Children in our well-nourished group have a higher probability of being misclassified because some may have experienced recent growth failure without dropping below our cut-off point.

If children are classified into groups on the basis of a single measurement, comparison between the caretakers of the two groups would be expected to yield significant differences, since the caretakers' long-term behaviour patterns would have contributed to the cumulative status of the children. A comparison between the short-term eating behaviours of the children themselves might be more confounded because some of the big children may be losing weight and some of the small ones may be undergoing catch-up growth at the time of the study.

Age-matching in Cross-sectional Designs

In cross-sectional designs and other studies where sophisticated methods are not used for the classification of positive deviance, well-nourished (W), average-nourished (A), and malnourished (M) children must be matched for age. In most developing countries, the entire growth distribution shifts downwards in comparison to international reference standards at about six months of age. The most accurate simple method of dividing children into W! A, and M groups is to sort out children within each month of age separately during the period when nutritional status is falling off rapidly: i.e. the top third of the seven-month-old children are defined as W, the middle third as A, and the bottom third, M. This means that the W children in the 12-month age-group may actually be more poorly nourished than the A group of sevenmonth-olds, according to international reference standards. Over age periods when nutritional status is relatively stable, e.g. 12 to 21 months, it is possible to pool all children across the agegroup for sorting purposes, so long as the classification procedure produces the same age distribution in the W, A, and M groups.

The alternative, of sticking to a strict definition of good nutrition according to international standards, classifies more of the younger children as well-nourished, and more of the older ones as malnourished. Thus the W, A, and M groups are noncomparable in average age, and the percentage of children falling into W, A, and M categories changes as the children get older. A consistent definition of malnutrition across age-groups was maintained in the Burmese study cited earlier (Nutrition Research Division, 1985). This analysis defined Ws as > 1 SD in weightfor-age z (WAZ), As from -1 SD to-2 SD, and Ms as c-2 SD. For 3,298 children aged between 0 and 36 months, the percentages falling into the W, A, and M categories by agegroup are shown in table 10.

Because the overall sample size was so large in the Burmese study the small percentages falling into the W category in the older age-groups still left sufficiently large groups of children for analysis.

Table 10

Age-group (months) Percentages
  W A M
0-3 36 45 18
4-6 29 44 22
7-12 17 67


1 3-24 5 69 26
25-36 4 77 20

Classifications and Analyses on the Basis of Different Anthropometric Indicators

The different anthropometric indicators change differently over time. Average HAZ in a population often drops rapidly while weight-for-height Z (WHZ) remains more nearly normal. Ideally, the well-nourished should be near the top of their distributions on all three indicators: WAZ, HAZ, and WHZ.

In fact, this may not be possible. Where classifications diverge, the indicator most affected by malnutrition should be the main criterion indicator. Where children are stunted but chubby, this tends to be HAZ; where they are stunted and thin, WAZ.

The way in which a preliminary subset of the Burma study data was analysed with technical assistance from Zeitlin (1983) provides an illustration of the manner in which indicators can be combined in classification criteria, as well as demonstrating how the criteria for nutritionalstatus groups can shift downward with age if the W, A, and M groups are to be age-matched. Infants and young children were classified into W. A, and M categories according to the following criteria, applied to three measurements WAZ, HAZ, and WHZ, calculated according to the NCHS/WHO standards, where W³ -ISD, A< -1, and ³ -2SD and M< -2SD.

W = W W W (applied to all three indicators).

A = (1) any combination of W and A or AAA in children below 7 months;
(2) any combination with A in final place above 7 months.

M = (1) any combination with M below 7 months;
(2) any combination with M in final place above 7 months.

Because true Ws were scarce, the As were further divided into high As and low As. Two matching procedures were used by hand to form triplets consisting of: (1) a true W, and A, and M child; (2) any child in the top third of the distribution (Ws plus high As), matched with a low A child from the middle third and an M child from the bottom.

In Burma, weight-for-age was the main criterion used for matching, yet problems arose in classifying children who were normal in height but very thin or very short but chubby. To avoid problems encountered in applying a single classification system, such children were excluded from the pairing procedure when pairmates would have been very different in WAZ, HAZ, and WHZ.

If multivariate methods are used, e.g. regression or analysis of covariance, using age as a covariate, elaborate pairing may be avoided and separate analyses can be conducted for positive-deviance classifications defined by WA, HA, and WH. Attempting to use statistical methods to control for the effects of age is not fully satisfactory when many phenomena change qualitatively, not simultaneously, with age. In adjusting for age effects, the procedures described below for longitudinal analyses can also be applied to cross-sectional data.

Longitudinal Designs

Longitudinal designs are desirable in the interests of accuracy but tend to require advanced computer capabilities for their analysis. It is not necessary to use longitudinal methods. The complexities described in this section can be very time- and resource consuming. Therefore, longitudinal positive-deviance studies probably should not be undertaken by researchers lacking a computer with a statistical package and an accessible statistician to provide ongoing guidance.

The Value of Longitudinal Growth Measures

Rate of growth is a matter of concern in the definition of positive deviance. Infants who grow well during one period may falter and grow poorly subsequently. The behaviours and circumstances that promoted their growth in the good period may disappear or prove maladaptive during the poor growth phase. When resources are abundant, it is desirable (though still not necessary) to measure growth longitudinally and to define positive deviance versus poor growth over a time period of six months or more. Cross-sectional studies that measure the child at one moment in time cannot tell whether the child's condition has recently improved or deteriorated.

At Least Six Months of Longitudinal Growth Data

Rate of growth can be different or impossible to measure over the short term. Between one year and two years of age, the reference growth-rate for weight is only about 200 g per month. Short-term variability in weight due mainly to differences in stomach, bladder and bowel content has been reported to be 290 g at 30 months (Habicht, 1983). Therefore, if the child is weighed once monthly, it is difficult to tell with certainty whether she has gained weight from one month to the next. At least six months' worth of longitudinal data should be collected in order to assess growth in length (height) or weight.

Adjusting for Age and Season in Longitudinal Data

An adjustment has to be made for changes in growth with age and sex. A useful first stage of adjustment frequently consists of transforming anthropometric raw scores to Z-scores according to the NCHS Standards.

A second stage of adjustment is then necessary. In many developing-country populations, almost all infants are well-nourished between about 1 and 4 months and malnourished by the age of 18 months. If no further adjustments are made, the youngest children will appear to be the positive deviants, as noted earlier. Similarly, where there are seasonal changes, infants at given ages will be consistently betternourished in some seasons than in others. In order to compare the growth status of children it is important to subtract from each child's growth measurement (in Z-scores if a Z-score adjustment has been used) at each month of age a value that represents the average growth measurement of the other children in the sample at the same age in the same season. The subtracted value left over is a residual Z-score that measures how well each particular child is doing compared to the others at each month in time. To adjust adequately for age and season requires a sufficient sample size of children at each given age in each given season.

Developing Summary Scores from Longitudinal Anthropometry

Two summary scores should be constructed representing the child's absolute size and his growth rate. The absolute size must be considered because a normal rate of growth at a very low Z-score (or percentile) may be maintained on a diet that would not support the same growth-rate at a higher Z-score (or percentile). Therefore, it is not possible to assume that a child who maintains a normal growth rate at -2.5 SD in HAZ is as well nourished as a child with the same growth rate at-1.5 SD.

For the first summary score, the average overall measurement points of each child's residual Z-score, as described above, provide an adequate ranking of the child's size relative to others in the group.

The second summary score for growth-rate over the measurement period must adjust for regression to the mean. The term regression to the mean describes the fact that the largest children tend to grow more slowly and the smallest more rapidly over a longitudinal measurement period. Causes of regression to the mean include measurement error, temporary illness, differing maturation rates, and environmental influences. The simplest way to make the second summary score with this adjustment is to construct a "value-added" score, using a method first introduced into the literature by Heimendinger and Laird (1983). This procedure involves the following steps:

  1. Construct a correlation matrix of the residual Z-scores at each age against all other ages.
  2. Find the r value of the correlation of the residual Z at age of first measurement and residual Z at final age of measurement for each child.
  3. Multiply this r value with the child's initial residual Z-score to get his final expected residual Z-score.
  4. Subtract this final expected residual Z-score from the child's actual final residual Z-score.
  5. This is the raw value-added score. lt should be divided by an age-specific constant to correct for the different rates of growth of the children at different ages. This constant is calculated by dividing the expected growth of the child (in cm or kg) at the age at which the final measurement is taken. The raw value-added score divided by this constant is the summary score measuring the child's growth-rate relative to the growth-rate of the others in the group.
  6. If the standard deviations of the study population differ significantly from those of the reference population, which is not usually the case, other procedure may have to be applied to the residuals before undertaking step 1 of this process.

If a large sample of children have been measured monthly from the starting age to the same final age the raw value-added score may be used without adjustment, or a differ ent second summary score can be obtained using some variation of the approach of Johnston and colleagues (1980). This approach pools all the anthropometric data points of all the children into a file in such a manner that they are arranged as crosssectional data, as if each monthly measurement represented a separate child. It may then divide the children on the basis of their starting Z-scores into 4 quartiles for HAZ, WAZ, and WHZ. Within each quartile, it regresses HAZ = a + b age; WAZ = a + B age; and WHZ = a + b age. Squared or cubic or log terms and seasonal dummy variable can be put into these regressions if they describe the data. Within each quartile, each individual's residuals can be taken from these regression lines. The slope of the linear regression through each child's residual scores can serve as the summary measure of growth-rate. Rather than using quartiles, if computer facilities permit, a separate regression may be calculated for each child, including in his regression equation the 60 children whose measurements were closest to his at the first measurement date.

Yet another approach, principal components analysis of the children's difference scores from month to month of the age- and season-adjusted residual Z-score variable, may also yield summary scores describing growth. However. these scores would be more likely to capture different patterns of growth spurts rather than growth-rates.

Statistics for Longitudinal Analysis

The statistical procedures currently available for longitudinal analyses are far more limited than those for cross-sectional approaches. This is the main reason for deriving summary scores of the longitudinal growth measurements, so that these summaries can be used cross-sectionally with summaries of other variables.

In theory, longitudinal methods such as repeat-measures multivariate analysis of covariance (MANCOVA) should be able to handle the covariates that are of interest to nutritional epidemiology. In practice, as of 1987, the existing statistical packages cannot accept as many covariates as one would wish, the manipulation of the covariates by the computer programs is difficult to control according to the needs of the analyst, and the results tend to be difficult to interpret. Time-series analysis cannot handle many individual cases.

Avoiding Shifts in Classifications

Prospective longitudinal designs in which children are classified as W and M at the beginning of the study will run into problems because some of the children will change in category over time. Therefore, prospective studies are advised to take children of all nutritional status categories and classify them according to their final measurements, or to sort them retrospectively into growth categories during the analysis.

Household versus Dyad-level Status

Innate child characteristics are confounded variables for household-level analyses, where the research goal is to compare mothers and families who produce wellnourished children versus those who do not. Some children are born survivors who thrive despite unfavourable environments.

For household-level studies, only families in which all children show satisfactory nutritional status and in which none have died should be classified as positive-deviant. This restriction minimizes the likelihood that the individual child, rather than the mother or the environment, is responsible for the favourable outcome.

Positive-deviant interaction patterns between caretaker and child may still occur regardless of which member of the dyed is more responsible for initiating them. For some research purposes, for example to determine the child characteristics associated with positive deviance, well-nourished children should be selected from homes in which another sibling is malnourished or deceased.

Because of the extreme immaturity of the human infant compared to the mother, it is only reasonable to expect that the mother's characteristics are more important than those of the child in determining the quality of their interaction. She has a far greater repertoire of responses as well as complex reasoning ability at her disposal. A study of cognitive development that did attempt to separate out the relative importance of the mother's versus the child's role (Ruddy and Bornstein, 1982) found the mother's contribution to be more significant than the child's.

Genetic Differences in Child Size and Growth-rate in Malnourished Populations

The issue of the degree to which malnourished children are genetically influenced by the short stature of their parents always comes up in positive-deviance studies. The best evidence currently available indicates that stunting below -2 SD of the NCHS standards cannot be considered to be genetic in origin. If variability in length and weight of young children in malnourished populations were predominantly determined by genetic growth potential, it would be very difficult to classify some as wellnourished and some as malnourished. Given the potential importance of this problem, this section discusses the heritability of growth in some detail.

Let us first examine whether uniform cross-generational stunting could be created in laboratory rats, for example, by making sure that the rat parents and rat pups in sequential generations all received exactly 60 per cent of their nutrient requirements from identical lab chow. In this case, one might assume simplistically that parents and pups would both be 75 per cent (or some consistent proportion) of their potential genetic lengths for their ages. In this case the parent-pup length correlations would be identical to those of well-nourished rats. If this were true, it would be possible to say that some of the malnourished rats were worse nourished than others. All would be equally malnourished' compared to their genetic potential.

In actuality, however, some of the rats would be more metabolically efficient than others, so that some would find the diminished ration adequate and would grow at or close to their genetic potential while some would experience severe growth retardation because their higher nutrient requirements were not met. Therefore, the small ones would in fact be less well nourished than the large ones.

Moreover, if they lived freely in colonies with a limited food supply, some would establish dominance over others and get more of the food. Some would experience more growth failure caused by illness than others. Some rat dams would manage their newborn pups less stressfully than others and have bigger pups with lower mortality rates. Each of these metabolic or behavioural sources of variability could contribute to significant positive parent-pup correlations in length (e.g. Iess stressed dams would be likely to be bigger and to have less stressed, bigger pups). However, these correlations might imply little or nothing about the genetic length potential of the rat, had they been raised on diets adequate for all members of the colony.

Fig. 13. Parent-child correlations for stature in well-nourished population (after Tanner and Israelsohn, 1963).

There is a body of research concerning parent-child height correlations that should be reviewed before drawing conclusions concerning the genetic component of child size in deprived communities.

Height is known to be highly heritable according to a primarily additive polygenic model (Carter and Marshall, 1978). Numerous empirical studies have confirmed predicted correlation coefficients of about r = 0.5 for stature between siblings and between parent and child, and about 0.7 between child's stature and midparent height (the average of the two parents' heights). Figure 13 (Tanner and Israelsohn, 1963) shows that parent-child correlations in wellnourished populations are low at birth, but are well-established by one year of age and fairly stable after two years. Paediatriclans in industrialized countries have been advised to use parentadjusted growth standards to assess the growth of young children (Tanner et al., 1970).

A number of studies found parent-child correlations in stature to be low in developing countries where environmental variables prevent the full expression of genetic growth potential. Two studies in 1977 (Martorell et al., 1977; Mueller and Titcomb, 1977), however, reported that parent-child correlations for stature (and other physical dimensions) remained high in endemically malnourished populations in which diet and health-related environmental variables had remained stable from one generation to the next.

The study populations in the latter studies may have differed from those that preceded them in the amount of intergenerational change that had occurred and in the homogeneity of the environments in which they lived. In the later studies showing high correlations, environmental influences appeared to exert very similar effects on the sets of families included in the analyses.

Mansour (1985) approached this issue using national level data from Tunisia collected from diverse regions of the country. The sample was divided into two groups of 2- to 6-year-old children whose heights fell above and below the regression line of height-for-age (HAZ) on age (between 2 and 6 years this line was nearly horizontal at about -1. 1 SD according to the NCHS standards). Because the child HAZ distribution was somewhat bimodal this division into halfdistributions did not make each half too narrow for further analysis. Multiple-regression analyses regressing child's HAZ against mother's height, socio-economic factor score, and sex of the child within each half showed mother's height to be the only significant correlate of child's height in the children of normal stature, and socio-economic score and child's sex the only significant correlates among the stunted children. Yet other analyses by Mansour showed that mother's height and child's height were correlated within more homogeneous subgroups of stunted children.

These findings strongly suggest that parent-child height correlations between stunted preschool children and their parents are due not to the biological expression of the children's genetic height potential but rather to cross-generational similarities in socio-economic status, metabolic responses to given diets, and other variables. When families entered into correlational analyses were taken from a homogeneous community, the correlations between parent and child height were inflated by local environmental, dietary, and behavioural/cultural features and morbidity patterns that affect both parents and children consistently. When families entered into the analysis were taken from many disparate regions within a country' the divergent effects of microenvironmental factors from different communities cancelled each other out.

These findings should not be taken to imply that genetics do not operate in stunted populations with high morbidity rates. Rather, the genetically regulated responses to multiple environmental insults are very complicated. Therefore, the height of stunted parents and children may be highly correlated but may not reflect their genetic height potential.

As an example of the conclusions that may reasonably be drawn from parent-child size correlations in malnourished groups, we cite Johnston and co-workers ( 1980) who found that parents' heights, and shoulder and hip widths, were highly correlated with the growth-rate of malnourished Mexican children. They concluded "that the etiology of chronic malnutrition, indicated here by growth failure, involves a significant generational aspect. These parents apparently replicated the conditions which led to their own malnutrition, so that their children are significantly more likely to display the failing growth which is characteristic of chronic malnutrition."

Mansour (1985) found that within the stunted group of children, parent-child height correlations began to become significant after age five, versus age three within the taller groups. This and other findings from his analysis support a model for older children and adults in which quantity of food is calorically sufficient for all, hut come position of the diet is systematically low in protein and micro-nutrients needed to promote optimal growth. Under these conditions growth of individuals might indeed be significantly correlated to genetic potential but at a lower level than would occur if genetic height were fully expressed.

In children below five, and below three years of age particularly, irregularity in growth-rate caused by frequent infections with erratic catch-up growth, and by faulty weaning diets, would appear to obliterate any consistent relationship between actual size and genetic size potential among the malnourished.

Given the evidence referred to above, we conclude that genetic heritability of size should not be a predominant consideration in positive-deviance research in nutrition on infants aged 0 to 3 years. We concur with Johnston and colleagues (1980), however, that tall versus short stature among parents may be considered as rough screening factors for identifying households having historically good versus poor crossgenerational adaptations to poverty and resource scarcity.

Genetic Differences in Growth in Well-nourished Populations

Variables affecting growth operate in a dose-response relationship. Above a specific threshold, further increases in a given variable will not increase body size. In nutritionally normal populations, correlation coefficients of close to 0.9 between the heights of identical twins imply that about 80 per cent of the variance in height between normally nourished children is genetically determined. Therefore, positive-deviance research in normal populations may yield relatively few psychosocial or dietary differences between large and small infants.

The twin studies indicate, however, that environmental influences still play some role in determining growth achievement even with identical twins within the same household, whose heights are correlated at an r value of about .95 (Newman et al., 1937). Identical twins reared in separate homes in presumably well-nourished environments show height correlations with r values closer to 0.8 (Shields, 1962). This suggests that certain maternal-child interaction characteristics still potentiate the exe pression of genetic size among well-nourished groups. Such growth-promoting characteristics would be expected to be more frequent among families whose children ranked high on the growth distributions of both developing-country and industrialized-country populations.