Menu ENN Search
Language: English Français

Stratified randomization

This question was posted the Assessment forum area and has 5 replies. You can also reply via email – be sure to leave the subject unchanged.

» Post a reply

Nick

Normal user

26 Dec 2015, 18:41

Hi,
Can someone (especially: epidemiologist) check my though on Stratified Randomization for one study? Thanks in advance.
Our team is going to implement a one nutrition sensitive conditional (soft) cash transfer program and have idea to include one learning pieces on how our intervention effect on “Incidence of Stunting in beneficiaries’ children”. Idea is to provide cash transfer since 3 months gestation of pregnancy to child up to 2 years – cover most part of the early 1,000 days.
This study will compose with 3 arms: 1 – Only Cash, 2 – Cash + BCC and 3 – Control. In term of sample size, we estimate 2,300 (with additional 1,500 observation) under 2 children to detect the 5% reduction of stunting prevalence (estimated).
So, for stratification of three arms in our area, we are thinking to apply the following factors (including causal factors of malnutrition based on the Unicef conceptual framework) for randomization;
1. Population size of villages (rural community),
2. Public Health condition (health services, disease prevalence – ARI and Diarrhea),
3. Food Availability and Accessibility (distance to market and Food Group Availability per season) and
4. Water Source and Environmental Sanitation condition.
Does it enough to consider in randomization or missing any points which can effect on our outcome (stunting)? So, please feel free to provide your comments on suggestion on that and have any resource or similar study on this, please do share for better planning on our study.
Thanks again on your kind contribution.
Sincerely,
Nick

Tamsin Walters

en-net moderator

Forum moderator

28 Dec 2015, 12:56

From Monira Parveen at WFP:

Similar study was conducted by IFPRI in Bangladesh on WFP’s 5 armed transfer modalities including cash, food and intensive Nutrition BCC. Dr. Akhter Ahmed, Chief of Party, IFPRI, Bangladesh can be contacted for the methodolody.

Bradley A. Woodruff

Self-employed

Technical expert

29 Dec 2015, 00:29

Dear Nick:

You have posed a good question. Stratified sampling is frequently done using geographic strata in order to minimize the sample size necessary to make reasonably precise estimates of outcomes in each geographic subdivision; for example, in each province or region. However, another reason to do stratified sampling is to increase the precision obtained in your sample. If the sample is stratified on a variable which is associated with the outcome, taking into account the stratification during data analysis will result in better precision than would be obtained with a non-stratified sample of the same size.

My question is: Why are you stratifying your sample? If you are measuring the effectiveness of the two different interventions (cash and cash plus BCC) against the control. you are not attempting to estimate outcomes in each stratum. If you are stratifying to get better precision, then you need stratify only on factors you know will be associated with the outcome you are measuring. Stratifying on factors not associated with your outcome is a waste of time and resources.

But there are costs and problems with stratification. First, to account for stratification when comparing the outcome variable in each intervention group, you will need to do stratified or multi-variate analysis which greatly complicates data analysis. Second, you must have data on all stratification variables for each sampling unit. Any sampling unit without complete data on all stratification variables cannot be included in the sample because you will not be able to place it in a stratum. Third, if you stratify on too many variables, the number of sampling units in each stratum can become vary small. The stratum is defined as the set of values for all stratification variables. For example, if you stratify on wealth quintile (5 possible values), safe water supply (2 possible values), adequate sanitation (2 possible values), and 3 different levels of disease prevalence (3 possible values) you will have 60 strata (5 x 2 x 2 x 3). This will lead to very small numbers in each stratum which is not a good idea.

So, in short, if you wish to improve the precision of your study, I would select 1 or 2 variables which you think are most strongly related to stunting and stratify on them, if you have the necessary data.

Nick

Normal user

29 Dec 2015, 11:10

Thanks Tamsin Walters and Monira Parveen.

Could you please guide me the contact address of Dr. Akhter Ahmed. You can sent it to my gmail (nicholustintzaw@gmail.com).

best regards,
Nick

Nick

Normal user

29 Dec 2015, 12:00

thanks a lot for your very informative comment.

Yes, the objective of stratifying is to measure the effectiveness of two different intervention against the control and also comparing between each arms (cash only vs cash + BCC). We also considering as estimation to reduce 5% stunting prevalence to estimate sample size for each arm and hope that the sample size mentioned in my question have enough power to detect this 5% variance.

There will be also census data collection in all sample area before performing stratification (we are going to allocate three stratum based on three arms). I do agree with your point as considering one or two main key variables which have strong relationship to our outcome. In this points, what will be your recommendation on those key variables.

best regards,
Nick

Bradley A. Woodruff

Self-employed

Technical expert

29 Dec 2015, 15:39

Dear Nick:
I'm not sure what you mean by "we are going to allocate three stratum based on three arms". It sounds like you may be confusing allocation to treatment arms and stratification of a random sample. If you like, you can e-mail me at bradleyawoodruff@gmail.com for further discussion.

Back to top

» Post a reply