Difference between revisions of "Randomization in SurveyCTO"

Jump to: navigation, search
(33 intermediate revisions by 3 users not shown)
Line 1: Line 1:
<onlyinclude>During surveys, you might often need to randomize various aspects of the questionnaire. While SurveyCTO has a random number generator, is is usually not recommended that you use it. This article will argue for doing the randomization in Stata, R or similar software, before the start of the survey, and preload the results of the randomization as dummies or categorical variables.</onlyinclude>


== Read first ==


Unless your survey falls into very rare exceptions, do not randomize in SurveyCTO, do it in Stata, R or similar and preload the result into SurveyCTO.


== Why randomization is better to do before the Survey ==


== Read First ==
During surveys, we often need to randomize various aspects of the questionnaire. For example – sometimes we need to randomize which household members to interview, and sometimes - which set of questions to ask. While most CAPI software have random number generators, it is not the preferred option. Using, for example, Stata to randomize and then preloading the generated data file into the survey software is in almost all cases the better option among the two. The main advantages of using Stata over CAPI software during randomization are as follows:


* Randomization in Stata is transparent and reproducible which is important for publishing research.
* Randomization results in Stata can be dependent, so that we are guaranteed that no disproportional large share of the results falls into any group. Randomization is always independent in SurveyCTO which means that no groups could be assigned observations if the number of observation per groups is low.
* Randomization in Stata provides the option of ensuring that the result of the randomization is balanced over other variables, i.e. stratas. This means that we can guarantee that, for example, not all female respondents end up in a certain group.
* Randomization in Stata is done before the survey takes place. This provides an opportunity to double check the result of a randomization and fix bugs and typos in the randomization code before it is used in the field, as it then would be too late to fix.


== Guidelines ==
== How to randomize in Stata for a survey in SurveyCTO ==
===Subsection 1===
 
===Subsection 2===
This is a basic example on how to do randomize in Stata and preload the results to SurveyCTO. See [[randomization in Stata]] more details for how to implement more advanced randomization, but the procedure for how to preload the result will still be the same as described below.
===Subsection 3===
 
In the example below we have a survey in which we want a random 30% to answer to a long survey and the rest to take a shorter survey. We also have a gender variable and we want the ration of 30/70 to be as exact as possible within the both genders.
 
=== Step by step - known respondents ===
 
* Let's say we a data set with all the respondents and replacement respondents, and that we have the variables, ''unique_id'' and ''gender''. Make sure that all observations have [[ID_Variable_Properties|uniquely and fully identifying]] ID in ''unique_id''. 
** If you do not have ID for your respondents you should create it.
** If you do not have a list of your respondents (for example, if you randomize the sample in the field by drawing lots) see the section of [[Randomization_in_SurveyCTO#Preload_Randomization_without_IDs| prelaoding randomization without IDs]].
* Set version, set seed and sort the data to guarantee a replicable sort. See [[Randomization_in_Stata#Steps_needed_for_replicability_when_randomizing_in_Stata| replicable randomization in Stata]] why all these three variables are needed.
* Generate a random number and create a dummy variable (''long_survey'' in this example) that indicates for each observation if that observation should answer the long or the short survey.
** Note that it does not have to be a dummy, we could just as well have had randomized into three groups and saved the result into a categorical variable with the value 1, 2, 3.
*In the end we should have a data set with the ''unique_id'' and the categorical variables with the result of the randomization, in this example only ''long_survey''.
*In your SurveyCTO survey, have a question where the enumerator enters the ID for the respondent currently interviewed. For this to be possible the enumerators needs a list with both the name and the ID.
**It is very important that these lists are not publicly disclosed as that would allow anyone to identify observations even in data sets that have names and other identifying information removed. If the data collected is extra sensitive, consider using a different ID than the main ID for this purpose.
* Preload the data set you generated in Stata into your SurveyCTO survey using the ID entered by the enumerator.
* Restrict the relevant section of the survey in SurveyCTO using the value just preloaded.
 
=== Code example - known respondents ===
<pre>
*Set version
ieboilstart , version(12.1)
`r(version)'
 
*Open the data set after ieboilstart
use sample.dta
 
*Set seed
set seed 123456 //this is an example seed, replace this with another number
 
*Sort data set
sort unique_id
 
*Generate random number, rank that random number per gender, and assign
* long survey if the rank is less than or equal the total number of observations
* in that gender
gen rand = uniform()
bys gender : egen rank = rank(rand)
bys gender : gen long_survey = (rank/_N <= .3)
</pre>
 
== Preload randomization without IDs ==
It is still possible to preload randomized categories even if there is no way of knowing the respondents in the survey.  For example, you might only have a list of villages to survey and intend to do a lottery to decide which respondents to interview. Or you might have randomized school classes and you will interview all students in those classes but you do not know exactly how many students each class has. In both these cases we can not assign a randomized category to each respondent as we do not know who they are yet. Instead we will randomize a category to the first respondent in the village/class, randomize a category for the second respondent, etc. And in the field the enumerators have lists for each village/class where they cross off each ID being used.
 
The following list explains the steps needed for this. We will keep using the example where we new which school classes we wanted to interview but we do not know the size of those class. Let's say that we want to collect basic demographic information on all students, we want to do a short survey on 40% of the students and a long survey on 10% of the students.
 
=== Step by step - unknown respondents ===
* Create a data set with one observation per class if you do not already have that. Make sure that all classes have unique IDs.
* Expand the data set (code example below) with the maximum number of students you want to interview per class, or the maximum number of students there could be in a class if you always want to interview all students no matter how long they are. If you expect that the highest of number of students in a class is 50, expand the number of observations to 100 so that there is no risk your expectation was too low. We pick the number 99 so that no ID requires three digits.
* Create a unique ID for each student. It is easy to do in combination with the class ID.
* Set version, set seed and sort the data to guarantee a replicable sort. See [[Randomization_in_Stata#Steps_needed_for_replicability_when_randomizing_in_Stata| replicable randomization in Stata]] why all these three variables are needed.
* Since we are generating 99 IDs but most classes will be considerably smaller, there is a risk that all or a large share of the 10% of the students ID that will take the long surveys are among the higher IDs 50-99 that we never expect to be used. To guarantee that that will never happen, we can create stratas for each group of 10 IDs, 1-10, 11-20, 21-30 etc. and randomize within each group. See the code example for how to do so.
** There are many different ways to do this. We can make sure that any number of categories shows up among the first ten or 20 numbers. Just make sure that the students are selected in random order so that the order they are assigned IDs is random. The way they sit in a class room might not be random as it might be differences in motivation or socioeconomic status between the students in the front or in the back.
* The example code below generates a variable ''surveyType'' with a randomized category for the first respondents, for the second, third etc. up to 99. This variable is preloaded the same way as explained above where we have IDs for all respondents.
* The final but equally important step is to create a system where the enumerators use each ID in correct order and only once. The best way to do that is to print out lists for each village where the enumerator crosses of the IDs used. **If there are multiple enumerators per class, then the lists can be split into IDs with odd or even numbers and the list is crossed of by the enumerators.
* Make sure that the randomization information is not on the lists available to the enumerators so that the enumerators has a change to influence the assignment of ID.
** For the same reason we need to randomize differently for each class as enumerators quickly learn the randomization order if it would have been the same for each class.
 
=== Code example - unknown respondents ===
<pre>
ieboilstart , version(12.1)
`r(version)'
 
*Open the data set with all classes you have randomized after ieboilstart
use class.dta
 
* Create one observation per each student ID. We use 99 as the highest number of students
expand 99
 
* Create unique ID
bys class_id : gen student_id = class_id * 100 + _n
 
* Generate strata for each group of 10 id, 1-10, 11-20, 21-30 etc.
bys class_id : gen id10strata = floor(_n/10)
 
*Set seed and sort
set seed 123456 //this is an example seed, replace this with another number
sort student_id
 
*Generate the result variable and the random number, rank that random number
* per class_id and ID strata, and assign survey type to 2 if the rank is less than or
* equal the total number of observations in that class and ID strata, and assign survey
* type 3 if the rank is larger than 90% of the total number of observations in that
* class and ID strata. The rest of the observations keep the survey type 1.
gen surveyType = 1
gen rand = uniform()
bys class_id id10strata : egen rank = rank(rand)
bys class_id id10strata : replace surveyType  = 2 if (rank/_N <= .5)
bys class_id id10strata : replace surveyType  = 3 if (rank/_N > .9)
</pre>
 
=== Other things to consider- unknown respondents ===
 
'''Test field system''. The most obvious step where this could go wrong is the step where enumerators use the correct ID and do not use IDs more than once. There is a risk that the enumerators do not do this properly, or that the system they are using is not designed well enough, but we can easily check that by looking into which IDs are used by each enumerators each day. We should already in the pilot have a good understanding if our system is good.
 
'''Stratification'''. If you want to stratify over, for example, gender you can have two lists. One for male respondents and one for female respondents. However, this increases the complexity in the field and more than one layer of strate organized in this way immediately becomes too complicated to be preferred. This is one of the few cases where randomization in SurveyCTO is preferred. However, make sure that there are enough observations in each strata so that even the independent randomization will end up having a balanced result.


== Back to Parent ==
== Back to Parent ==
This article is part of the topic [[Randomized Controlled Trials]]
This article is part of the topic [[Randomized Control Trials]]
 


== Additional Resources ==
== Additional Resources ==
* list here other articles related to this topic, with a brief description and link


[[Category: *category name* ]]
[[Category: Impact Evaluation Design ]]

Revision as of 11:20, 5 April 2018

During surveys, you might often need to randomize various aspects of the questionnaire. While SurveyCTO has a random number generator, is is usually not recommended that you use it. This article will argue for doing the randomization in Stata, R or similar software, before the start of the survey, and preload the results of the randomization as dummies or categorical variables.

Read first

Unless your survey falls into very rare exceptions, do not randomize in SurveyCTO, do it in Stata, R or similar and preload the result into SurveyCTO.

Why randomization is better to do before the Survey

During surveys, we often need to randomize various aspects of the questionnaire. For example – sometimes we need to randomize which household members to interview, and sometimes - which set of questions to ask. While most CAPI software have random number generators, it is not the preferred option. Using, for example, Stata to randomize and then preloading the generated data file into the survey software is in almost all cases the better option among the two. The main advantages of using Stata over CAPI software during randomization are as follows:

  • Randomization in Stata is transparent and reproducible which is important for publishing research.
  • Randomization results in Stata can be dependent, so that we are guaranteed that no disproportional large share of the results falls into any group. Randomization is always independent in SurveyCTO which means that no groups could be assigned observations if the number of observation per groups is low.
  • Randomization in Stata provides the option of ensuring that the result of the randomization is balanced over other variables, i.e. stratas. This means that we can guarantee that, for example, not all female respondents end up in a certain group.
  • Randomization in Stata is done before the survey takes place. This provides an opportunity to double check the result of a randomization and fix bugs and typos in the randomization code before it is used in the field, as it then would be too late to fix.

How to randomize in Stata for a survey in SurveyCTO

This is a basic example on how to do randomize in Stata and preload the results to SurveyCTO. See randomization in Stata more details for how to implement more advanced randomization, but the procedure for how to preload the result will still be the same as described below.

In the example below we have a survey in which we want a random 30% to answer to a long survey and the rest to take a shorter survey. We also have a gender variable and we want the ration of 30/70 to be as exact as possible within the both genders.

Step by step - known respondents

  • Let's say we a data set with all the respondents and replacement respondents, and that we have the variables, unique_id and gender. Make sure that all observations have uniquely and fully identifying ID in unique_id.
    • If you do not have ID for your respondents you should create it.
    • If you do not have a list of your respondents (for example, if you randomize the sample in the field by drawing lots) see the section of prelaoding randomization without IDs.
  • Set version, set seed and sort the data to guarantee a replicable sort. See replicable randomization in Stata why all these three variables are needed.
  • Generate a random number and create a dummy variable (long_survey in this example) that indicates for each observation if that observation should answer the long or the short survey.
    • Note that it does not have to be a dummy, we could just as well have had randomized into three groups and saved the result into a categorical variable with the value 1, 2, 3.
  • In the end we should have a data set with the unique_id and the categorical variables with the result of the randomization, in this example only long_survey.
  • In your SurveyCTO survey, have a question where the enumerator enters the ID for the respondent currently interviewed. For this to be possible the enumerators needs a list with both the name and the ID.
    • It is very important that these lists are not publicly disclosed as that would allow anyone to identify observations even in data sets that have names and other identifying information removed. If the data collected is extra sensitive, consider using a different ID than the main ID for this purpose.
  • Preload the data set you generated in Stata into your SurveyCTO survey using the ID entered by the enumerator.
  • Restrict the relevant section of the survey in SurveyCTO using the value just preloaded.

Code example - known respondents

*Set version 
ieboilstart , version(12.1)
`r(version)'

*Open the data set after ieboilstart
use sample.dta

*Set seed
set seed 123456 //this is an example seed, replace this with another number

*Sort data set
sort unique_id

*Generate random number, rank that random number per gender, and assign 
* long survey if the rank is less than or equal the total number of observations 
* in that gender
gen rand = uniform()
bys gender : egen rank = rank(rand)
bys gender : gen long_survey = (rank/_N <= .3)

Preload randomization without IDs

It is still possible to preload randomized categories even if there is no way of knowing the respondents in the survey. For example, you might only have a list of villages to survey and intend to do a lottery to decide which respondents to interview. Or you might have randomized school classes and you will interview all students in those classes but you do not know exactly how many students each class has. In both these cases we can not assign a randomized category to each respondent as we do not know who they are yet. Instead we will randomize a category to the first respondent in the village/class, randomize a category for the second respondent, etc. And in the field the enumerators have lists for each village/class where they cross off each ID being used.

The following list explains the steps needed for this. We will keep using the example where we new which school classes we wanted to interview but we do not know the size of those class. Let's say that we want to collect basic demographic information on all students, we want to do a short survey on 40% of the students and a long survey on 10% of the students.

Step by step - unknown respondents

  • Create a data set with one observation per class if you do not already have that. Make sure that all classes have unique IDs.
  • Expand the data set (code example below) with the maximum number of students you want to interview per class, or the maximum number of students there could be in a class if you always want to interview all students no matter how long they are. If you expect that the highest of number of students in a class is 50, expand the number of observations to 100 so that there is no risk your expectation was too low. We pick the number 99 so that no ID requires three digits.
  • Create a unique ID for each student. It is easy to do in combination with the class ID.
  • Set version, set seed and sort the data to guarantee a replicable sort. See replicable randomization in Stata why all these three variables are needed.
  • Since we are generating 99 IDs but most classes will be considerably smaller, there is a risk that all or a large share of the 10% of the students ID that will take the long surveys are among the higher IDs 50-99 that we never expect to be used. To guarantee that that will never happen, we can create stratas for each group of 10 IDs, 1-10, 11-20, 21-30 etc. and randomize within each group. See the code example for how to do so.
    • There are many different ways to do this. We can make sure that any number of categories shows up among the first ten or 20 numbers. Just make sure that the students are selected in random order so that the order they are assigned IDs is random. The way they sit in a class room might not be random as it might be differences in motivation or socioeconomic status between the students in the front or in the back.
  • The example code below generates a variable surveyType with a randomized category for the first respondents, for the second, third etc. up to 99. This variable is preloaded the same way as explained above where we have IDs for all respondents.
  • The final but equally important step is to create a system where the enumerators use each ID in correct order and only once. The best way to do that is to print out lists for each village where the enumerator crosses of the IDs used. **If there are multiple enumerators per class, then the lists can be split into IDs with odd or even numbers and the list is crossed of by the enumerators.
  • Make sure that the randomization information is not on the lists available to the enumerators so that the enumerators has a change to influence the assignment of ID.
    • For the same reason we need to randomize differently for each class as enumerators quickly learn the randomization order if it would have been the same for each class.

Code example - unknown respondents

ieboilstart , version(12.1)
`r(version)'

*Open the data set with all classes you have randomized after ieboilstart
use class.dta

* Create one observation per each student ID. We use 99 as the highest number of students
expand 99

* Create unique ID
bys class_id : gen student_id = class_id * 100 + _n

* Generate strata for each group of 10 id, 1-10, 11-20, 21-30 etc.
bys class_id : gen id10strata = floor(_n/10)

*Set seed and sort
set seed 123456 //this is an example seed, replace this with another number
sort student_id

*Generate the result variable and the random number, rank that random number 
* per class_id and ID strata, and assign survey type to 2 if the rank is less than or 
* equal the total number of observations in that class and ID strata, and assign survey 
* type 3 if the rank is larger than 90% of the total number of observations in that 
* class and ID strata. The rest of the observations keep the survey type 1.
gen surveyType = 1
gen rand = uniform()
bys class_id id10strata : egen rank = rank(rand)
bys class_id id10strata : replace surveyType  = 2 if (rank/_N <= .5)
bys class_id id10strata : replace surveyType  = 3 if (rank/_N > .9)

Other things to consider- unknown respondents

'Test field system. The most obvious step where this could go wrong is the step where enumerators use the correct ID and do not use IDs more than once. There is a risk that the enumerators do not do this properly, or that the system they are using is not designed well enough, but we can easily check that by looking into which IDs are used by each enumerators each day. We should already in the pilot have a good understanding if our system is good.

Stratification. If you want to stratify over, for example, gender you can have two lists. One for male respondents and one for female respondents. However, this increases the complexity in the field and more than one layer of strate organized in this way immediately becomes too complicated to be preferred. This is one of the few cases where randomization in SurveyCTO is preferred. However, make sure that there are enough observations in each strata so that even the independent randomization will end up having a balanced result.

Back to Parent

This article is part of the topic Randomized Control Trials

Additional Resources