The Effect of a Cash Transfer Scheme on Children S Body Mass Index in Colombia: Prospective

The effect of a cash transfer scheme on children’s body mass index in Colombia: prospective cohort study

Ian Forde1,Research Training Fellow, Tarani Chandola2,Professor of Medical Sociology, Sandra Garcia3, Professor of Public Policy, Marcos Vera-Hernández4, Lecturer and Research Fellow, Orazio Attanasio4, Professor of Economics, Michael G. Marmot1, Professor of Epidemiology and Public Health.

1Department of Epidemiology and Public Health, University College London; 2Cathy Marsh Centre for Census and Survey Research, University of Manchester; 3School of Government, Universidad de los Andes, Bogotá; 4Department of Economics, University College London and Institute of Fiscal Studies, London.

October 2011

Corresponding author:

Dr Ian Forde, UCL Epidemiology and Public Health, 1-19 Torrington Place, LONDON, WC1E 6BT.

Word count:

Abstract:250

Main text:3117

References:21

Figures: 4

Tables: 4

Abstract:

Introduction:Familias en Acción is a Colombian government programme designed to improve life chances during children’s early years. Families receive around 60,000 pesos (US$20) per month if children comply with regular health checks and school attendance. Prior work found thatFamilias is associated with a 15% increase in household spending on food. We explored the relationship between exposure to Familias and standardised BMI-for-age (BMIZ), rates of overweight and obesity in children.

Methods: 2266 children aged 2-7 from treatment areas and 3022 from matched control areas were surveyed in 2002 and 2006.Follow-up was 62.5%. Height and weight was measured using standardized techniques. Overweight and obesity were defined using IOTF curves. The effect of Familias was estimated using multivariate regression, controlling for secular trends and several individual, household and area characteristics, using robust standard-errors clustered at area-level in an intention-to-treat analysis.

Results: At baseline, children’smean BMIZ was +0.01s.d.; 7.6% children were overweight or obese. BMIZ decreased over time in all children, but less in Familias participants. In linear regression with BMIZ as the dependent variable, participation in Familias was associated with a positive co-efficient in girls (β=0.16, 95% CI 0.02,0.29; p=0.02) and boys (β=0.17, 95% CI 0.04,0.29; p=0.01). In logistic regression on odds of overweight,Familias was associated with positive odds ratios in girls (O.R.=1.90; 95%CI 1.06,3.43; p=0.03) but not boys (O.R.=1.27; 95%CI 0.75,2.15; p=0.38).Odds of obesity were not increased.

Conclusion: Conditional cash transfers to poor householdsmay be independently associated with adverse nutritional outcomes in children.

KEY WORDS:body weight changes, child welfare, poverty, Colombia.

1. Introduction

The global epidemic of obesity is increasingly affecting children. Wang, in a review of the literature published between 1980 and 2005 covering more than 60 countries, reports that childhood obesity increased in all children in almost all countries for which data are available[1]. A recent study of 5 to 12 year olds in Bogota reported that over-nutrition had become a bigger problem than under-nutrition[2].

Familias en Acción is a Colombian government programme designed to improve life chances during children’s early years. Known as a “conditional cash transfer scheme (CCTS)”, eligible families receive 40,000 pesos (around US$15) per month if they have children aged 0 to 7, as long as they ensure up to date immunisation and regular health checks. Additional payments of 14,000 pesos (around US$5.5) are made for each child regularly attending primary school and 28,000 pesos (around US$11) for each child regularly attending secondary school (transfer amounts are those current at the time of this study). Familias began in 2002; as of 2010, some 2.6 million families were enrolled[3]. Prior workhas shown that participation in Familias is associated with approximately a 15% increase in household spending on food compared to unexposed controls[4]. Some modest effects on under-nutrition were also reported: children aged 0-2 were 6.9% less likely to suffer from stunting, although no effect was seen in older children[5]. We explored the effects of participation in Familiason body mass index in children.

2. Methods

2.1 Design, setting and participants

Households eligible for Familias are those from Colombia´s lowest income sextile (identified by routine government data) which include children under-17 and live rurally or in small towns (defined as <100,000 inhabitants).

Familias was implemented at area-level. An evaluation sample was constructed by randomly selecting 57 treatment areas from the 622 implementing the programme.In some areas, political pressure meant that the programme started before baseline data collection was completed - participants from these areas were discarded from the analysis.Treatment areas were matched with control areas within 25 strata based on region, health/education infrastructure, population, land area and quality of life score, based on routine government data. Approximately 100 eligible households were randomly sampled from each treatment and control area, generating an analytic sample of 7938 households.

Within each household, height and weight were measured in children aged under-7 and their biological mothers. We exclude children under-2 given the uncertain value of diagnosing overweight in this age group, whether in terms of current morbidity or prognostic significance[6,7].

Local institutional ethical approval was obtained. Adults signed consent to participate in the study and data were anonymised prior to analysis.

2.2 Explanatory, outcome and mediating variables

Three surveys took place between mid-2002 and early 2006. This analysis reports data from the first and third surveys. Treatment/control status was determined from programme administrative records.

Height and weight were measured by 18 trained fieldworkers using a protocol based on the Pan-American Health Organisation Manual on Anthropometrics[8], with standardised measuring boards (Shorr Productions, Olney, Maryland USA) and electronic scales (Seca 770, Vogel & Halke, Hamburg, Germany). Children were clothed but without shoes. BMI was calculated and converted into age and sex standardised Z-scores based on 2000 CDC Growth Reference curves and analysed as a continuously distributed outcome. “Overweight” and “obese” were defined using thresholds recommended by the Childhood Obesity Working Group of theInternational Obesity Taskforce[9] and analysed as dichotomous outcomes. Cases with missing anthropometrics at baseline or follow-up were deleted list-wise from the sample and patterns of missingness explored.

The conceptual framework established by Friel et al.[10] was used to inform the selection of further co-variates. As well as individual-biological factors such as food intake and energy expenditure, this model emphasizes the social determinants of obesity such as income, education, living conditions, social capital, remoteness and infrastructure. The model is particularly sensitive to health inequity and the shifting burden of obesity towards poorer households, hence is apt given Familias´ objectives.

At the individual level, age and sibling number were treated as continuous variables; previous participation in Hogares Comunitarios, a supplementary feeding program to which the same children were eligible, was dichotomised.At the household level, maternal age, BMI and household size (persons) and crowding (persons per room) were treated as continuous variables. Household wealth was proxied from total household spending in the past two weeks and log-transformed. Presence of piped water and rural/semi-urban location (referring to the presence of facilities such as a town hall, a school or health centre in contrast to more remote communities) were treated as dichotomous variables. Mother´s completed formal education was categorized as ‘primary education incomplete’, ‘primary education complete’, ‘secondary education complete’ or ‘higher education’. Individual and household-level covariates were self-reported. At the area-level, population (from 2000 census figures) and number of families eligible for Familias were log-transformed. Log-ratio of doctors to population and proportion of houses with piped water were included as markers of infrastructure, as was presence of a bank since this was known to be the most common criterion on which treatment and matched control areas differed. Average travel-time to the nearest medical centre was included as a proxy for remoteness. Quality of life score (taken from a routine national survey asking about local amenities) and average log-household wealth were included as additional area-level co-variates.

All co-variates were measured at baseline except travel time to medical centre (recorded at first follow-up)

Measures to ensure reliability of the data included extensive questionnaire piloting and fieldworker training, use of computer assisted personal interview (CAPI) technology, direct observation of ~10% of surveys by quality controllers and repeat measures on a subset of participants. Efforts to minimize attrition included widely publicized support for Familias from civic leaders and participants, regular contact with participants via newsletters and a website and efforts to trace households that had moved.

2.3 Statistical analysis

Data were cleaned by censoring children with BMIZ beyond +/- 4s.d. Baseline differences by exposure were explored using two-tailed t-test (for differences in means for continuously distributed variables) or chi-squared test (for categorical variables) and are reported in Table 2. Differences by attrition are reported in Table 1.

Given non-random programme implementation, a double-difference methodology was employed to estimate programme effects. This model allows for the outcome of interest to differ between treatment and control areas at baseline and for a secular trend in the outcome independent of the programme. Baseline difference in outcome between treatment and control areas is subtracted from the difference at follow-up, identifying the net effect of the programme as shown in Figures 2-4.

The main assumptions of the double-difference method are two-fold: that any secular trend is not substantially different between exposed and unexposed communities and that co-variate values do not change over time. Prior work examining trends in key co-variates immediately prior to inception of Familias suggests that the first assumption holds[5]. Regarding the second, co-variates were selected, as far as possible, to be fixed or net-neutral across exposure groups (such as age).

The regression estimated is:

where

Yi= outcome of interest for individual i

Time = 0 if baseline, 1 if follow-up

Comi= 1 if exposed area, 0 if control

Familiasi ,T=1 if Familias was in operation in the community where individual ilived at time T, 0 otherwise

Xi= all (observed) co-variates

i= error term

The co-efficient of interest is β3. This identifies the effect of the Familias independent of secular trends and unobserved area effects. Baseline BMIZ followed a normal distribution and OLS regression was used to estimate the effect of the programme on BMIZ and logistic regression for effects on odds of overweight or of obesity.

Intention-to-treat analyses are reported; that is, households are analysed according to the treatment/control status of their area at baseline, whether or not they took up the offer of Familiasor whether they changed status between baseline and follow-up. It is known that only around 85% of qualifying households took up the offer of Familias and that a small number changed status between baseline and follow-up.

Robust standard errors are reported, clustered at area-level. This is the coarsest and most conservative cluster-level hence it also accounts for non-independence at individual level. Stata SE11.2 was used for all analyses, including software written by Vidmar et al. to calculate BMIZ and apply IOTF overweight/obese categories[11].

3. Results

3.1 Participants and missing data

7492 children were included in the baseline survey and anthropometrics recorded in 6903 (92.1%) of these.Likelihood of recording anthropometrics varied by exposure (2=19.2, p<0.001), but not by other likely determinants of nutritional status such as maternal age, educational level,BMI, household size or wealth (all p>0.09). 5566 were aged between 2 and 7.

Maternal BMI was missing in 12.7% and replaced by multiple imputation, generating 5 imputations from a model containing treatment/control status, age, parity, educational attainment, household wealth and ruruality. After replacement all other co-variates were completely observed, except in 285 (5.0%) children who were deleted list-wise. Incomplete observation was not significantly associated with treatment/control status (2=0.71, p=0.40) or with other likely predictors of follow-up nutritional status including baseline BMIZ, age, gender, maternal BMI, household size or wealth (all p>0.09).

3.2 Sample attrition

Hence, 5281 children comprised the baseline cohort. 1847 (61.0%) children from control and 1427 (63.0%, 2=2.10, p=0.15) from treatment areas were re-observed at follow-up and form the analytic sample. A flow-diagram is given in Figure 1.

Children lost to follow-up had lower baseline BMIZ than the analytic sample (-0.06 vs 0.01, t=2.03, p=0.04). They were also older (62.7 vs 53.1 months, t=20.8, p0.001), with more siblings (5.3 vs. 5.1 live births, t=2.5, p=0.01), as shown in Table 1.

3.3 Baseline characteristics of the analytic sample

Mean baseline BMIZ was 0.04 in control and -0.04 (t=1.95, p=0.05) in treatment areas. Children from treatment areas were less likely to have attended Hogares Comunitarios (47.3 vs. 57.2%, t=9.14, p<0.001), had mothers with slightly lower educational level (2 =8.79, p=0.03), lived in less crowded (3.0 vs. 3.3 persons per room, t=5.3, p<0.01) and less semi-urban households (41.2 vs. 51.3%, t=7.31, p<0.001). Several differences were also apparent at area-level, as shown in Table 2, all of which were entered into the regression models.

3.4 Outcomes of interest

A crude indication of the effect of Familias can be obtained from double-difference tables. These data suggest that BMIZ and rates of overweight and obesity decreased in all areas with time, but less so in treatment areas than in control areas (Figures 2-4).

Multivariate analysis demonstrates a statistically significant positive association between exposure to Familias and BMIZin both girls (β=0.16, 95% CI 0.02, 0.29; p=0.02; Table 3) and boys (β=0.17, 95% CI 0.04, 0.29; p=0.01). Maternal BMI was also positively associated with BMIZ in each gender, and age and attendance at Hogares Comunitarios negatively associated after adjustment.

In girls, logistic regression on odds of overweight demonstrates a statistically significant positive relationship with exposure to Familias (O.R.=1.90; 95%CI 1.06,3.43; p=0.03). Age was negatively associated (O.R.=0.99; 95%CI 0.98, 1.00; p=0.01) and maternal BMIpositively associated (O.R.=1.08; 95%CI 1.04, 1.13; p<0.001). In boys, no relationship between odds of overweight and Familias was observed (O.R.=1.27; 95%CI 0.75, 2.15; p=0.38), although similar relationships were observed with age (O.R.=0.98; 95%CI 0.97, 0.99; p<0.01) and maternal BMI (O.R.=1.10; 95%CI 1.05, 1.14; p<0.001). Additionally, increasing sibling number was protective in boys (O.R.=0.89; 95%CI 0.82, 0.97; p=0.01).

Logistic regression on odds of obesity did not demonstrate statistically significant relationship in girls (O.R.=1.23; 95%CI 0.33, 4.61; p=0.76) or boys (O.R.=1.79; 95%CI 0.43, 7.34; p=0.42). In boys, increasing maternal BMI appeared as a risk factor for obesity (O.R.=1.17; 95%CI 1.08, 1.28; p<0.001) and increasing household crowding protective (O.R.=0.78; 95%CI 0.65, 0.94; p=0.01).

Across all outcomes, clear although non-significant trends were seen across levels of maternal education as a risk factor for abnormal weight gain in this population.

4. Discussion

4.1 Key results and interpretation

The study finds that BMIZ and rates of overweight or obesity decreased over time in all children, but less so in Familias participants: the programme appears positively associated with children´s BMIZ, after controlling for secular trends, several individual, household and community level co-variates and time-invariant unobserved co-variates. Increasing age and household wealth are also positively associated.These findings areconsistent with earlier reports that Familias households increase their spending on food [4].

Although the net difference in BMIZ between exposed and unexposed groupsis small (0.1-0.2), it is significantly associated with odds of overweight in girls. Energy expenditure in girls may be less than that of boys of the same age and female metabolism predisposes to fat deposition relative to males[12,13]. Girls may also consume a greater share of household food: at baseline stunting was significantly less common in girls (23.3% vs. 26.3%, p<0.05). The null finding in boys may also be artefactual: sample attrition was greater amongst older children and increasing age is associated with greater odds of overweight in boys, but not girls.

4.2 Strengths and limitations

The study benefits from a large sample size, information on a wide range of co-variates and a conservative statistical approach. Several quality control mechanisms were implemented, including independently commissioned data collection and analysis.

A primary weakness of the data is its non-randomized nature. Although the multivariate, double-difference statistical approach accounts for all observed baseline differences in co-variates and any unobserved time-invariant differences, the possibility of residual confounding remains. Most control areas lacked a bank, in contrast to treatment areas, and were therefore slightly poorer. Several mechanisms unrelated to Familias canbe imagined which would promote adverse weight gain in treatment areas, such as faster expansion of fast-food outlets, sedentary leisure or motorized public transport.

Follow-up rates were just over 60% and did not vary by exposure. Children lost to follow-up, however,were older, with more siblings and had smaller BMIZ than those retained in the sample. Although in multivariate regressions these factors are associated with less weight gain, they did not differ between treatment and control areas within those children lost to follow-up (all p>0.3), making any biasing effect unlikely.

4.3 Comparison with other studies

Comparison is complicated by lack of agreed norms of child growth and thresholds to define overweight. Nevertheless, Colombia’s 2005 Demographics and Health Survey reported comparable rates of excessive weight (defined as WHZ score > 2 s.d. on 1977 NCHS curves in children aged 0-5) of 1.5-8.1%[15]. A study of 3075 children aged 5-12 years old in Bogotá in 2006[2] estimated prevalence of overweight to be 11.1% and obesity 1.8% using IOTF definitions, as used in this study.

The negative trend in BMIZ over time is perhaps unexpected given global trends. It is not, however, unprecedented. Overweight (defined as WHZ score > 2 s.d.) dropped from 4.6% to 2.6% in pre-school children in Colombia, between 1986 and 1995 [16]. The phenomenon may occur if children gain height more rapidly than weight, perhaps due to improved dietary quality, which is bourne out by inspection of HAZ and WAZ trends in the Familias cohort (not shown). Alternatively, the decline may be artefactual and explained by use of a historic growth reference. Children’s BMI trajectory is sinusoidal, rising steeply in first year, falling until about age 5 and then rising again until plateauing at maturity. Modern cohorts (i.e. Familias children) typically plateau earlier than earlier cohorts, with prior phases of the growth curve commensurately compressed. The children used to construct IOTF curves were measured between 1963 and 1993 [9]. Hence, it may appear that Familias children’s standardised BMI falls over time because they follow a steeper downward gradient compared to historic IOTF peers, prior to rebound and plateau. Differential sample attrition would not account for the finding because children retained in the sample had higher mean BMIZ than children lost to follow-up (Table 1).