Targeting nutrition intervention using machine learning prediction of future malnutrition in India

Last registered on December 29, 2024

Pre-Trial

Trial Information

General Information

Title
Targeting nutrition intervention using machine learning prediction of future malnutrition in India
RCT ID
AEARCTR-0014343
Initial registration date
December 29, 2024

Initial registration date is when the trial was registered.

It corresponds to when the registration was submitted to the Registry to be reviewed for publication.

First published
December 29, 2024, 11:23 PM EST

First published corresponds to when the trial was first made public on the Registry after being reviewed.

Locations

There is information in this trial unavailable to the public. Use the button below to request access.

Request Information

Primary Investigator

Affiliation
IDinsight

Other Primary Investigator(s)

Additional Trial Information

Status
On going
Start date
2024-06-01
End date
2025-04-01
Secondary IDs
Prior work
This trial does not extend or rely on any prior RCTs.
Abstract
This study evaluates an innovative approach to preventing child malnutrition in rural India. The intervention combines machine learning predictions with targeted nutritional counseling to identify and support children at risk of malnutrition. The program operates in two states (Madhya Pradesh and Odisha) where the American India Foundation (AIF) already provides nutritional support services through local Anganwadi centers.
In this randomized controlled trial, we will evaluate whether using machine learning to identify high-risk children and providing their families with additional home-based counseling improves nutritional outcomes. The study includes villages randomly assigned to treatment or control groups. In treatment villages, families of children identified as high-risk will receive additional home visits and nutritional counseling. In control villages, standard nutritional services will continue as usual.
The study will measure impacts on children's nutritional status (using standard anthropometric measures) and on caregivers' knowledge and practices related to child nutrition. This will help us understand whether combining predictive analytics with targeted counseling can effectively prevent malnutrition among young children in rural India.
External Link(s)

Registration Citation

Citation
von Grafenstein, Liza. 2024. "Targeting nutrition intervention using machine learning prediction of future malnutrition in India." AEA RCT Registry. December 29. https://doi.org/10.1257/rct.14343-1.0
Sponsors & Partners

There is information in this trial unavailable to the public. Use the button below to request access.

Request Information
Experimental Details

Interventions

Intervention(s)
Dimagi and American India Foundation (AIF) are partnering to first identify children at risk of falling into malnutrition and then provide targeted counselling services to their caregivers.
Dimagi has developed a predictive algorithm using existing data and machine learning to estimate "risk scores" on the nutritional status of children, estimating their risk of falling into undernutrition or malnourishment. The top k children with the highest risk scores will be identified as high-risk to receive further counselling. Facilitators from American India Foundation (AIF) will then deliver counselling to the caregivers of high-risk children to improve their knowledge, attitudes and practices regarding childhood nutrition. This counselling will be delivered through weekly home visits by facilitators.
The intervention has two arms:
Treatment group: Frontline workers are told about risk scores from Dimagi's risk assessment. Households of the top-k high-risk children receive home visits and counselling on childhood nutrition.
Control group: Frontline workers are not told about risk scores. AIF continues to provide care to children based on the status quo.
AIF will provide personalised counselling to children who are non-malnourished but predicted to worsen and children who are moderately underweight and predicted to worsen. The counselling is delivered through AIF's Integrated Community Facilitators (ICFs), who conduct home visits. This intervention supplements AIF's existing SNEH (Skilling, Nutrition, Education and Health) Programme.
Intervention Start Date
2024-06-01
Intervention End Date
2025-02-01

Primary Outcomes

Primary Outcomes (end points)
Indicators P.1: Weight-for-Height (WFH), Weight-for-Age (WFA) and Height-for-Age (HFA) z-scores
The z-scores for each of the categories of wasting (WFH), underweight (WFA) and stunting (HFA) are standard deviations of weight or height from the median of the same height or age in a reference population defined by the World Health Organisation (WHO).
Indicator P.2: Average number of lapses per child into a worse category of malnourishment at two points of time
For each child, we will calculate the number of times between baseline, a predefined midpoint, and the endline that the child is in a worse category of malnourishment than in the previous period for stunting, underweight, and wasting separately.
Indicators P.3: Malnutrition/nutrition status
The malnutrition status is calculated as a binary indicator which records whether the child falls into the Yellow and Red category in any of the three types of malnutrition, i.e. underweight, wasting, and stunting. In contrast, nutrition status is calculated as a binary indicator that records whether the child falls exclusively into the Green category for any type of malnutrition or the Yellow category for underweight.
Indicators P.4: Severity of lapses into a worse category of malnourishment
For each child, we will measure the "distance" between the starting, middle and ending nutritional status, defined as the number of categories a child moves through between baseline, midline and endline nutritional status for stunting, underweight, and wasting separately.
Primary Outcomes (explanation)
"For each child, we will calculate the number of times between baseline, a predefined midpoint, and the endline that the child is in a worse category of malnourishment than in the previous period for stunting, underweight, and wasting separately. Dimagi has categorised children into three categories based on the z-scores (green, for non-malnourished; yellow, for moderately malnourished; and red, for severely malnourished for each of the three types of malnutrition). For each child, we will:

Measure status at baseline
Measure status at midpoint (during a 30-day period starting 45 days after the baseline). If a child has moved to a worse or the worst category of malnutrition than at baseline, they will have a lapse value of 1.
Measure status at endline. If a child has moved to a worse or the worst category of malnutrition than at midpoint, they will have a lapse value of 1.
The maximum lapse value will be 2."

For the severity measure:
"For each child, we will measure the "distance" between the starting, middle and ending nutritional status, defined as the number of categories a child moves through between baseline, midline and endline nutritional status for stunting, underweight, and wasting separately. For example, if a child is of normal weight at baseline, moderate weight at midline and severely underweight at endline, severity of lapse will be 2 (the child fell through two categories of malnutrition). Similarly if a child is at normal weight at baseline, severely underweight at baseline and normal weight at endline, the severity of lapse will be 2 ( the child fell through two categories of malnutrition)."

Secondary Outcomes

Secondary Outcomes (end points)
Indicator group S.1: Multiple indicators on caregivers' knowledge of childhood nutrition
Definition: We will look at either an aggregate score or single domains listed below that capture knowledge and attitudes toward childhood nutrition as binary variables. These indicators are aligned with AIF's counselling material to the maximum extent possible. Domains include: Caregivers' understanding of children's nutritional requirements, exclusive breastfeeding knowledge, complementary feeding knowledge, food preparation knowledge, malnutrition knowledge, awareness of growth monitoring.
Indicator S.2: Child morbidity
Definition: Child morbidity is defined as a binary variable whether a child has been reported to have any one of three symptoms in the past two weeks: fever, diarrhoea, or respiratory illnesses.
Indicator S.3: Minimum dietary diversity
Definition: Minimum dietary diversity is a binary indicator of whether a child consumed more than 5 out of 8 food groups for children 6 to 23 months, or more than 4 out of 10 food groups for children 24 to 59 months over the past 24 hours.
Indicator S.4: Minimum meal frequency
Definition: Meal frequency is a binary indicator of whether a child aged 6 – 23 months has consumed solid, semi-solid, and soft foods (including milk feeds for non-breastfed children) at least a) 2 times for breastfed infants 6–8 months b) 3 times for breastfed children 9–23 months c) 4 times for non-breastfed children 6-24 months.
Indicator S.5: Unhealthy food consumption
Definition: Unhealthy food consumption is a binary indicator of whether a child has consumed any products in three food groups: i) savoury and fried snacks, ii) sweets, and iii) sugar-sweetened beverages over the past 24 hours.
Indicator group S.6: Caregivers' attitudes and beliefs about childhood nutrition
Definition: We will measure multiple indicators to assess caregivers' attitudes toward nutritional practices, and will look at either an aggregate score or single indicators listed below. These include perceived barriers to meal frequency, perceived barriers to continued breastfeeding, trust in external counsellors, and perceived benefits of frequent feeding.
Indicator group S.7: Caregivers' implementation of best practices related to feeding and care of children
Definition: We will measure multiple indicators to measure the frequency and consistency of implementing recommended practices. We will look at either an aggregate score or single indicators listed below. Indicators S.2, S.3 and S.4 listed above will also be included in this category. These include continued breastfeeding, complementary feeding, and zero fruit or vegetable consumption.
Indicator S.8: Other household practices regarding childhood nutrition
Definition: We will look at either an aggregate score or single domains listed below that capture household practices regarding childhood nutrition as binary variables. These include frequency of visits to AWCs, utilization of Services, growth monitoring practices, and handwashing practices.
Secondary Outcomes (explanation)
For knowledge indicators (S.1): "For knowledge indicators, we will use a score-based index, where every respondent will be given a score equivalent to the number of knowledge-related questions they have answered correctly. We will construct a binary variable indicating whether the respondent has answered all questions correctly, and a continuous variable indicating the score the respondent has received. Based on these, we will calculate two indicators:
Proportion of respondents that have answered all questions correctly
Average number of questions answered correctly across respondents"
For practice related indicators: "For the practice related indicators, we will construct binary variables indicating whether the practice is being implemented correctly. We will calculate the proportion of respondents that are following each practice. We will use binary variables for p-value correction.

Experimental Design

Experimental Design
We will use a clustered Randomised Controlled Trial (RCT), wherein we will randomly assign clusters (i.e. villages) to treatment or control. We define villages as the cluster level because targeting children requires a group. We will include only one AWC per village to ensure no village is over-represented.
We aim to have 150 villages in the control group and 150 villages in the treatment group in total across Madhya Pradesh and Odisha. Based on routinely collected monitoring data, we will randomly assign each village in the study area to treatment or control status stratified by state, AWC size, and the prevalence of not-malnourished and moderately underweight children.
The two arms are:
Treatment group: Frontline workers are told about risk scores from Dimagi's risk assessment. Households of the top-k high-risk children receive home visits and counselling on childhood nutrition.
Control group: Frontline workers are not told about risk scores. AIF continues to provide care to children based on the status quo.
Target population consists of non-malnourished and moderately underweight children between 6 and 55 months of age at baseline who could be at risk of malnutrition. AIF will provide personalised counselling to children who are non-malnourished but predicted to worsen and children who are moderately underweight and predicted to worsen.
Experimental Design Details
Not available
Randomization Method
We will use stratified random assignment based on ,State, AWC size (based on median split), Prevalence of non-malnourished and moderately underweight children (based on median split) and Gender (based on proportion of females)"
Randomization Unit
We will use a clustered Randomised Controlled Trial (RCT), wherein we will randomly assign clusters (i.e. villages) to treatment or control. We define villages as the cluster level because targeting children requires a group. We will include only one AWC per village to ensure no village is over-represented.
Was the treatment clustered?
Yes

Experiment Characteristics

Sample size: planned number of clusters
We aim to have 150 villages in the control group and 150 villages in the treatment group in total across Madhya Pradesh and Odisha.
The planned number of clusters is 300 villages.
Sample size: planned number of observations
Overall, our sample will consist of 300 villages total (150 villages in Madhya Pradesh and 150 in Odisha). Of these 150 villages in each state, 75 will be randomly assigned to treatment and 75 to control. For the outcomes, we will select 4 children in each village, with a total sample size of 1,200 children The planned sample size is 1,200 children.
Sample size (or number of clusters) by treatment arms
Treatment arm: 150 villages (with 600 children - 4 per village)
Control arm: 150 villages (with 600 children - 4 per village)
Total: 300 villages with 1,200 children
Minimum detectable effect size for main outcomes (accounting for sample design and clustering)
For the entire sample with children belonging to top-k as well as bottom-k risk scores, i.e. 4 children per village, we estimate the MDES of 0.003 s.d. for the anthropometric z-score outcomes. For the intermediate outcomes, we estimate MDES to be in the range of 0.09 percentage points (pp) to 0.12 pp depending on the outcome measure and respective study parameters." Specifically for the primary outcomes: WFH z-score: 0.00291 s.d. WFA z-score: 0.00258 s.d. HFA z-score: 0.00321 s.d. For secondary outcomes: Minimum dietary diversity: 0.109 pp Prevalence of child morbidities: 0.086 pp Knowledge, attitudes and practices: 0.117 pp Consumption of Vitamin A Rich Foods: 0.108 pp Consumption of Iron-Rich Foods: 0.106 pp
IRB

Institutional Review Boards (IRBs)

IRB Name
Monk Praygoshala
IRB Approval Date
2024-11-05
IRB Approval Number
#157-024
Analysis Plan

There is information in this trial unavailable to the public. Use the button below to request access.

Request Information