Additional information on statistical approaches used in the analysis [As supplied by authors]

The effectiveness of the albendazole offered at the Child Health Days is likely to be related to several factors which were controlled for using multivariate regression models in STATA software in which weight gain was the dependent variable. For each child in the sample one observation on weight gain was used, calculated as the difference in recorded weight between the first and last time the child was recorded as attending a health day. Body weights more than six standard deviations above or below reference values for age were excluded from subsequent analyses. Z-scores of weight-for-age were calculated for each child using EpiInfo software, which uses the reference values recommended by the World Health Organization.

Possible influences on weight gain are controlled as follows.

- The number of health days visited at which the treatment with albendazole was given (hence the number of treatments) varies across the sample. Hence the number of visits for both the treatment and the control group are included in the initial estimates.

-The age of the child at recruitment, which varied between 12 and 83 months, can influence subsequent weight gain. Thus, all regressions include the age of the child when first measured as well as its square to account for any age specific patterns in the velocity of weight gain. A cubic term for age proved not to be significant as an explanatory variable, and was not included in the models.

- As the expected weight gain may vary between genders, a bivariate dummy variable for gender is included in all regressions.

- Local environmental factors within each parish may affect food production or disease transmission and thus the nutritional status of the children within each parish. This possibility was addressed by including dummy variables in each regression for district and round effects. That is, for each of the six districts there were four bivariate dummy variables for the round the child entered the sample, giving a total of 24 fixed effects. Every observation was in one, and only one, of these cells. Each cell contained observations from both the treatment as well as the control groups. As these dummy variables also correspond to the field teams organizing the Child Health Days they also control for differences in the quality of administration of Child Days.

- Since the sample was based on cluster sampling effects at the parish level the standard errors are corrected for cluster effects. This correction accounts for any correlation of errors in measurement within a cluster but assumes independence between clusters. This is a Huber-White or sandwich estimator in which a matrix based on the sum of observation level likelihood scores summed over each cluster is pre- and post-multiplied by the usual model based variance matrix.^{1, 2}

- Since the length of time a child was observed will affect total weight gain, the interval between the first and last visit was included in most of the regression models.

- Given that this was a large scale intervention assessing a government program, no attempt was made to visit children in their home. Thus, some children attended the program more than others. We, therefore, need to take the frequency of attendance into account and in so doing test whether frequency of visits influence the outcome. For a portion of the analysis the treatment group was divided into three sub-samples: those who had attended Child Health Days at intervals of 7.5 months or less, those for whom the average time between visits was 7.5 to 13 months, and those for whom the average time was longer. While this is not an equal division of the sample (it represents 18, 67 and 15 percent of the treatment group respectively) it corresponds to practical targets of biannual, annual or less frequent visits, with some allowance for delays in service delivery.

- Because the subsequent weight gain of each child may be influenced by the initial nutritional status we include regression models that have initial weight as a right hand side regressor. Moreover, as inclusion of initial weight as an independent variable can introduce a correlation of measurement error with the dependent variable, initial height is also used to control for measurement error using an instrumental variables regression model.

However, given the potential for measurement error due to the design of the study as a monitored large scale program rather than a more fully controlled clinical research we needed to consider the potential for errors in initial measurement. While the measurement error is unlikely to be correlated with the treatment itself and, thus, would not bias the result of interest, additional regressions are presented to verify whether the key result is robust to alternative specification. First, initial weight was included in the regression model. No biological interpretation can be assigned to the coefficient of initial weight in this model since any measurement error in the initial weight is also in the dependent variable, total weight gain, leading to a bias towards minus one for that coefficient. However, under the assumption that there was no correlation between the measurement error of weight and the treatment effect, this initial weight variable picks up unexplained variance without biasing the variable of interest. In an additional model, predicted initial weight was substituted for the actual observed initial weight. Thus, measurement error for weight was removed from the right hand side regressor while information which correlates with other observed variables was utilized. If measurement error in height was uncorrelated with that in weight, the coefficient of predicted initial weight should no longer have a bias towards minus one and can be interpreted as regression towards the mean.

- Finally, in order to facilitate a discussion of the expected impact in terms of incremental weight gain per month or per year, we specified one regression in terms of weight gain per month rather than total weight gain.

To summarize, in addition to a comparison of mean total weight gain, we use regression analysis to compare weight gain per visit to the Child Health Days, and as a function of frequency of visits. The latter is also expressed in terms of weight gain per month in the program.

1. Williams, R. A note on robust variance estimators. Biometrics 56: 645-646. 2000.

2. Rogers W. Regression standard errors in clustered samples. Stata Technical Bulletin 13: 19-23. 1993.

 

 

 

Table 1. Dates of Child Health Days, the Proportion of Children Classified As Underweight (-2 S.D. below reference vales) and the Mean Z-Scores of Weight-For-Age

		Round
		1	2	3	4	5
Start date		Nov 2, 2000	August 14, 2001	February 18, 2002	Sept. 2, 2002	March 18, 2003
End date		Dec. 8, 2000	Nov.30, 2001	June 29, 2002	Oct. 19, 2002	June 26, 2003
Proportion underweight	Treatment	0.26	0.24	0.23	0.23	0.24
	Control	0.26	0.25	0.25	0.23	0.24
Average Z- score of weight-for-age	Treatment	-1.14 (1.48)	-1.06 (1.52)	-1.07 (1.45)	-1.14 (1.32)	-1.23 (1.25)
	Control	-1.17 (1.45)	-1.11 (1.52)	-1.14 (1.44)	-1.17 (1.29)	-1.20 (1.25)
Number of children weighed		37,165	33,711	21,124	20,787	20,443
Average age in years		3.69 (1.66)	3.64 (1.80)	3.69 (1.77)	3.63 (1.63)	3.54 (1.54)
Percent female		50.0	50.3	50.6	51.1	50.7
Percent in treatment		50.7	51.2	51.9	53.8	56.8

(Standard deviations of z-scores and age in parentheses).

 

 

Table 2. Average Weight Gain, Months in Program and Number of Visits To Child Health Days For Children With Two or More Measurements.

		Treatment	Control	All
Weight Gain (g)	Average	2413	2259	2341
	Standard Deviation	2536	2474	2508
Number of Months in Program	Average	16.9	16.2	16.6
	Standard Deviation	7.7	7.5	7.6
Number of Visits to Child Health Days	Average	2.7	2.6	2.7
	Standard Deviation	0.9	0.8	0.9
Number of Children with Repeated Measurements		14940	13055	27995

 

 

Table 3. The Parameters Generated By Five Regression Models Indicating Impact of Attendance on Weight Gain (in grams) In Sites Providing Albendazole

Variable	1. Total Weight Gain	2. Total Weight Gain	3 – Total Weight Gain	4 -– Total Weight Gain	5– Total Weight Gain	6 – Weight Gain per Month
Visits to Treatment Site	705.0** [648-762]	143.7** [87-201]		---	---	---
Visits to Control Site	649.6** [592-707]	92.5** [25-160]
Initial age (months)	-689.4** [-790 - -589]	-693.9** [-791 - -597]	-694.2** [-797 - -597]	-519.9** [-603 - -510]	-838.23** [-955 - -721]	-41.0** [-48.9 - -33.0]
Age squared	56.5** [43-70]	66.3** [54-78]	66.3** [54.2-78.4]	41.7** [29.8 – 53.6]	78.5** [64.0 – 93.0]	3.8** [2.8-4.8]
Interval between 1st and final measurement		99.0** [89-109]	111.6** [104.6-118.7]	102.8** [95.6 – 109.9]	102.0** [93.3 – 110.7]	-2.0 * [-2.4 - -1.5]
Female	49.6* [1-98].	48.5 [-0.8 – 98]	49.1 [-0.7-100]	125.2** [73.0 – 177.3]	101.4** [50 - 153]	1.9 [-2.5 – 6.3]
Treatment if time between visits <=7.5 months			158.2** [59.0-262.3]	169.5** [52.4 – 286.6]	178.0* [37.3 –318.6]	13.8 * [1.3-26.3]
Treatment if time between visits >7.5 and <= 13 months			106.6* [23.2-190.0]	127.7* [30.0 – 225.5]	128.8* [14.8 – 242.8]	6.6 * [0.2 – 12.9]
Treatment if time between visits >= 13 months			-125.4 [-304.7-53.9]	-127.7 -307.5 – 52.1]	-150.6 [-142.5 – 41.0]	-4.3 [-13.4 – 4.8]
Initial weight			-	-555.2** [-600.1 - -510.3]	---	---
Initial weight (Predicted)					-273.3** [-333.8 - -212.8] ---
Constant	2684** [2452-2915]	1853** [1581-2124]	1801** [1524-2078]	911** [668-1154]	1917** [1625-2208]	257** [235-280
N	27,995	27,995	27,995	27,995	21,134	27,995
R-square	0.146	0.178	0.178	0.283	0.183	0.018

95% Confidence Intervals corrected for cluster sampling effects in brackets. The constant is the average of the fixed effects.

*Coefficient differs from zero (two tailed test) with p<=.05.

**Coefficient differs from zero (two tailed test) with p<=.01.

  

 

Table 4. Proportion of Children 1 Year or Older Reported To Be Dewormed in Samples of Households Between Two Surveys Three Years Apart in Rural Uganda

	Deworming Sites		Control Sites
Round	1st Q 2000	1st Q 2003	1st Q 2000	1st Q 2003
No of children	1392	1296	1393	1314
% of children dewormed from any source in last 6 months	21.7	65.8	23.9	34.6
% getting treatment from private or NGO sources	NA	11.4	NA	15.2
Cost (USH) at private providers	NA	748	NA	776
Attend Child Health Day in last two years y/n	NA	0.735	NA	0.730
Number of Child Health Days attended	NA	1.738	NA	1.764

NA indicates that the question was not asked in the first round. Child Health Days were not organized in the region at the time of the initial survey.

Q = quarter

NGO = non-governmental organization

USH = Ugandan shillings

Online poll

Find out more >>