Even though survey workers should always be trained to measure and record height to the nearest millimeter, many tend to round height to the nearest full centimeter or one- half centimeter. As a result, a disproportio- nate number of height measurements will end in .0 and/or .5, sometimes referred to as digit preference. An analysis of the deci- mal for the height measurements can tell survey managers if this error is common in the survey data. The distribution of the decimal for the Somalia survey is shown in Figure 3.8. Note the disproportionate num- ber of children whose height measurement ended in 0 or 5. Figure 3.9 shows a similar distribution for a survey done in Mongolia. There is also a preference for rounding the height measurement to the nearest centime- ter in this survey, but it is not so pronounced as in the Somalia data. Moreover, there seems to be no rounding to the nearest 0.5 cm. Such an analysis should be presented in the survey report in order to inform readers about the precision of the height measurements.
Figure 3.8 Distribution of Decimal of Height Measurements for Children Under 5 years of Age, Bardera, Somalia, January 1993 90 80 70 60 50 40 30 20 10 0 0 1 2 3 4 5 6 7 8 9 Number of childr en
Calculate the proportion of records with extreme Z-scores
Inevitably, mistakes will be made in either measuring children, transcribing the results to data collection forms or entering the data into a computer. Certain values for anthropometric indices are highly improbable or incompatible with life and should not be included in the analysis of Z-scores. For example, it is extreme- ly unlikely that a child would have a weight- for-height Z-score of -6.0. Many computer pro- grams which calculate anthropometric Z-sco- res or percents of median from anthropometric measurements will somehow “flag” those children whose Z-scores fall outside certain boundaries, indicating that the calculated index is probably the result of a mistake in the data and is not real. Epi Info™ creates a new field in the data set when calculating the Z-sco- res from measurements. This field is called
“flag,” and it contains a code number depen- ding on which anthropometric index is out of bounds. For Epi Info, the “flag” variable is based on an older definition of extreme values and should be used with caution.
A more acceptable method of flagging is to create acceptable boundaries for the Z-scores depending on the data set. For areas with a very high prevalence, an alternative appro- ach to defining Z-scores is any Z-score more than 4 standard deviations below or above the mean Z-score for that index. For exam- ple, if in a particular survey, the average weight-for-height Z-score is -2.0, then those children with Z-scores less than -6.0 and greater than +2.0 should be excluded from the analysis. Alternately, fixed boundaries can be used to flag records; Table 3.3 details boundaries as recommended by WHO.
DESIGNING A SURVEY CHAPTER
3
Figure 3.9 Distribution of Decimal of Height Measurements for Children 6-59 Months of Age, Mongolia Nationwide Nutrition Survey, 2001
180 160 140 120 100 80 60 40 20 0 0 1 2 3 4 5 6 7 8 9 Number of childr en
Decimal of height me asurement
Table 3.3 Boundaries for flagging outlying data
Index Weight-for-height Height-for-age Weight-for-age Lower boundary < -4.0 < -5.0 < -5.0 Upper boundary > +5.0 > +3.0 > +5.0
Source: Physical Status: The Use and Interpretation of Anthropometry - Report of a WHO Expert Committee. WHO, Geneva. 1995
Of course, each flagged record with an extreme Z-score should be investigated to correct errors in data recording or data entry, and obvious errors should be cor- rected in the computer data file. Those children whose flags are not the result of obvious errors must be excluded from the analysis of the Z-scores. The final sur- vey report should present the proportion of children with extreme Z-scores. 14. ANALYSE THE DATA AND PRESENT THE RESULTS
Once the survey data are entered into a computer program, checked, and the accuracy of measurements confirmed, the data can be analysed to describe chil- dren's nutritional status. In the final report, it is customary to present results in certain standardized ways.
• Primary results often will be the pre- valence of wasting, stunting and underweight, both <-2 SDs and <-3 SDs. These prevalences are merely the number of children with any mal-
nutrition and the number of children with severe malnutrition, respective- ly, divided by the number of children with a valid anthropometry index. For example, consider a survey in which 397 children have valid weight-for-height Z-scores, 24 have a Z-score between -3.0 and -2.0, and 7 have a weight-for-height Z-score < - 30. The prevalence of wasting would be 7.8 percent [(24 + 7) / 397]. The prevalence of severe wasting would be 1.8 percent (7 / 397).
• In addition, you should present the mean anthropometric Z-scores and their standard deviations for all chil- dren without oedema. This number gives an overall summary of the degree of acute malnutrition among children in the survey population. • You also should present a graph sho-
wing the distribution of anthropome- try Z-scores among children without oedema. Figure 3.10 shows one example of such a graph.
Figure 3.10 Distribution of Weight-for-Height Z-Scores, Children Under 5 Years of Age, Badghis Province, Afghanistan, March 2002 20 15 10 5 0 Sample -3.0 -2.0 -6,00 to -5 ,0 -4,99 to -4 ,5 -4,49 to -4 ,0 -3,99 to -3 ,5 -3,49 to -3 ,0 -2,99 to -2 ,5 -2,49 to -2 ,0 -1,99 to -1 ,5 -1,49 to -1 ,0 -0,99 to -0 ,5 -0,49 to 0 .0 0,01 to 0 ,5 0,51 to 1 ,0 1,01 to 1 ,5 1,51 to 2 ,0 2,01 to 2 ,5 2,51 to 3 ,0 3,01 to 3 ,5 3,5 to 6,0 NCHS/CDC/WHO reference Per cent Weight-for-height z-score
• The prevalence, mean Z-score and graph of the distribution of height- for-age Z-scores also can be presen- ted for chronic malnutrition. • If hemoglobin or serum retinol are
measured on survey subjects, the comparable results for these values can be presented (the prevalence of abnormally low values, the mean value, and a graph showing the distribution of all the values). • It is very important to calculate con-
fidence intervals for the major out- comes measured in a survey and to present them in the verbal presenta- tions and the final report. This will give the reader or listener an idea of the precision of the estimates. Most computer programs used to analyse data will give confidence intervals around prevalence rates; however, many such programs assume that your data come from a sample cho- sen by simple or systematic random sampling. If you have done cluster sampling, the confidence intervals calculated by these programs are not correct. For cluster samples, you must use a computer program which takes into account the cluster sam- ple and calculates the appropriate confidence intervals. Such computer programs will ask you the name of the variable which identifies the clu- ster to which each unit of analysis belongs. For example, the Analysis module of Epi Info will calculate confidence intervals, but they will not be correct for a cluster survey. In Epi Info for DOS (version 6), you must use the CSample module to calculate confidence intervals for cluster surveys. One of the boxes in CSample asks for the PSU, or prima-
ry sampling unit. You should indica- te in this window what variable con- tains the cluster number. In Epi Info for Windows, you must use the “Complex Sample Frequencies,” “Complex Sample Tables,” or “Complex Sample Means” com- mands. SAS and Stata have complex sample commands and SPSS has an option complex sampling module. Confidence intervals for mortality rates may be more difficult to obtain. If mortality data are collec- ted as recommended in this manual, you should be able to calculate the number of deaths and the popula- tion denominator in each house- hold. These numbers are then com- bined for all the households in the survey sample and analysed as a cluster sample. The total number of deaths divided by the total house- hold population gives an estimate of the proportion of the population which died during the recall period. Computer programs, such as CSample or Epi Info for Windows, should give a correct confidence interval for this proportion that takes into account the cluster sam- pling. The lower and upper ends of this confidence interval then can be converted to death rates expressed as the number of deaths per stan- dard population size per standard time period, such as the number of deaths per 1,000 per year.