How To Apply Different Weighting Structures To Sub-Samples

Displayr's weighting tool is quite robust and has the ability to recompute a weighting scheme for each sub-sample within a grouping variable (such as a date, segment, wave, etc.) using the Recompute weights for field. However, sometimes you may need to apply a different weighting scheme (either different targets or variables) to each sub-sample. For example:

1. You need to create weights based on regions within a country, but there are different regions for each country
2. Some waves in your tracker have different target percentages or use different adjustment variables
3. A category in an adjustment variable has 0 respondents within a wave or other sub-sample
4. You want to weight just part of your data set. 

You can manually figure out the population counts or overall weighting percentages and use those in a grid within the weighting tool, but many times that's a little complicated and tedious. The steps below use the more easily understood and commonly used target percentages for each adjustment variable. The general steps are:

Step 1. Create weights for each sub-sample

Step 2. Combine the weights

Step 3. Check the combined weight

See a Worked Example

Requirements

• Familiarity with the process in How to Configure a Weight from Variable(s), along with the requirements to configure a weight.
• A variable used to group respondents into the sub-samples (this can be a Nominal variable like Wave, Country, Region, or a Date variable that has been aggregated to the appropriate periods).

Method

Step 1. Create weights for each sub-sample

First, we will need to create a weight variable for each weighting scheme for our sub-samples. This ensures that the correct weighting scheme will be constructed for each sub-sample. Note that if you only want to weight part of your data set (so the "other" part has a weight of 0 and is excluded), you only need to create a weight for the subset you want to include.

1. Set up the adjustment variables(s) for the weighting scheme of one of the sub-samples using Targets are > Percentages.
2. Change Make weights sum to > target values (this will make the weighted base for each sub-sample 100 rather than the total sample of the data file, which is essential for later steps).
3. Factor in the sub-sample grouping:
   1. If you have the necessary sample for each target category across all sub-samples, you can select your grouping variable that identifies the sub-samples in the Recompute weights for field. It will compute weights for all sub-samples (note this is for all respondents, including those marked as missing data), but in a later step, we will pull off only the weight values from this variable for the appropriate sub-samples. In this example, the variable is called Interview Date
      
   2. If you don't have respondents in each target for all sub-samples, for example, a wave doesn't have respondents in every category in the weighting scheme, or you are weighting on a region variable within different countries, click Additional adjustment variable set, select the variable that identifies your sub-samples, and set the particular sub-sample group associated with the weighting scheme to 100%. For example, below I'm only calculating this weighting scheme for July 2019 respondents, so I've added Interview Date as a separate adjustment variable and set the July 2019 date to 100:
      
   3. If you're not using or can't use the Recompute weights for feature, repeat the previous step until you have one weighting variable for each sub-sample. 

Step 2. Combine the weights

Now we will pull off the values needed for each sub-sample from the appropriate weight variables created above, and combine them into one single weighting variable. The way you do this depends on whether you'd like 1) your sub-samples to have the same sample size or 2) to match the sample sizes of the original sub-sample groups. 

Option 1: If you'd like your sub-samples' populations to be evenly weighted with the same sample size

1. In the Data Sources tree, hover over a variable and click + > Custom Code > JavaScript > Numeric.
2. Paste in the code below in the JavaScript Code editor and edit as described in the comments (denoted by //):

//pull in sub-sample grouping variable (wave)
//replace wave with your own sub-sample grouping variable
var group = wave;

//////
//edit the if-else below to test for the appropriate code values and assign them
//the appropriate weight variable values
//if weighting part of a data set you will have only one group else NaN
/////

//if the respondent is in the 1st group use the weight for the 1st group
if (group== 1)
    weight1; //replace with the sub-sample weight variable for the 1st group
//else if the respondent is in the 2nd group use the weight for the 2nd group
else if (group == 2)
    weight2; //replace with the sub-sample weight variable for the 2nd group
//else if the respondent is in the 3rd group use the weight for the 3rd group
else if (group == 3)
    weight3; //replace with the sub-sample weight variable for the 2nd group
//if any respondents don't fall into a group above, they will be given a blank weight
//which will filter them out when the weight is applied
else
    NaN;

3. Give the new variable a label and check the Usable as a weight box in the object inspector to make it available in the Weight dropdowns.

Option 2: If you'd like each of your sub-samples' populations to be equal to the unweighted sample size within the unweighted data, you can combine the individual weights and gross up/down the respondent-level weights to balance out the sample size using a new R variable.

1. In the Data Sources tree, hover over a variable and click + > Custom Code > R > Numeric.
2. Paste in the code below in the R Code editor and edit as described in the comments (denoted by #). Be sure to use the Label of each category in the "group" variable:

#replace YourGroupVariable with the variable used for the sub-sample
#with your cursor below you can click on it in the Data Sources tree 
#to insert its name into the code
group = YourGroupVariable

#create an empty combined weight variable which we will fill in with values below
combweight = rep(NA,length(group))

######
#for each group/sub-sample, reference the correct individual weight and gross up/down
#based on the sample size of that group in the unweighted data
######

#hover over your group variable highlighted at the top of the code to confirm
#the labels R "sees" for the conditions below
#fill in the combined weight for respondents in the 1st group with the weight value of the
#weight for the 1st group (weight1) and gross up based on original sample size
#replace 1st with the category label of your first group and weight1 with the first's weight variable
combweight[group == "Group 1"] = weight1[group == "Group 1"] * sum(group == "Group 1",na.rm=T)/100
#replace 2nd with the label of your second group and weight2 with the second's weight variable
combweight[group == "Group 2"] = weight2[group == "Group 2"] * sum(group == "Group 2",na.rm=T)/100
#replace 3rd with the label of your third group and weight3 with the third's weight variable
combweight[group == "Group 3"] = weight3[group == "Group 3"] * sum(group == "Group 3",na.rm=T)/100
#any respondent not in the groups above will retain the initial missing value (NA)
#and be filtered out when the weight is applied, use combweight[group == "X"] = 1 

#return the final result
combweight

3. Give the new variable a label and check the Usable as a weight box in the object inspector to make it available in the Weight dropdowns.

Step 3. Check the combined weight

1. Drag one of the adjustment variables you used in the weights onto the page to create a table.
2. Cross it by the variable used to group the sub-samples.
3. Select your combined weight variable from Step 2. Combine the Weights in the Data > Filters & Weight > Weight dropdown in the object inspector .
4. Confirm the target percentages match for each sub-sample.
5. If necessary, compare the Statistics > Below > Column Sample Size and Weighted Column Sample Size to confirm they are what you expect.

For example, in the table below, the combined weight uses a 50-50 target scheme for all months except July 2012 where a different 55-45 weighting scheme was used:

Worked Example

Weight each quarter of data based on age.

Targets within each quarter are as follows:

We want the final weighted population for each quarter to be equal to the # of respondents in the unweighted data.

To do so:

1. Create the weight for 1/1/2019

2. Create a weight for 4/1/2019. Duplicate the above weight and edit the Age and date targets.

3. Create the weight for 7/1/2019. Duplicate the above weight and edit the Age and date targets.

4. Insert the combined weight variable below the weight3 variable. To do so, click the + button to the right of weight3 and > Custom Code > R > Numeric.

5. Paste in R Code similar to below (line 2 has been edited to use the aggregated date values)

#identify group variable and use aggregated date names for each quarter using attr()
group=attr(date, "QDate")

#create empty combined weight variable
combweight = rep(NA,length(group))

#for each group, reference the correct individual weight and gross up based on subgroup sample size
combweight[group == "Jan-Mar 19"] = weight1[group == "Jan-Mar 19"] * sum(group == "Jan-Mar 19",na.rm=T)/100
combweight[group == "Apr-Jun 19"] = weight2[group == "Apr-Jun 19"] * sum(group == "Apr-Jun 19",na.rm=T)/100
combweight[group == "Jul-Sep 19"] = weight3[group == "Jul-Sep 19"] * sum(group == "Jul-Sep 19",na.rm=T)/100

#return final result
combweight

6. Press the Calculate button to run the code.

7. In the Object Inspector , in the Data > Properties, check Usable as a weight.
8. In the Object Inspector , select Data > Filters & Weight and apply the weight to the Age by Interview Date table.

The results confirm that the weights have been combined successfully.

Next

How to Configure a Weight from Variable(s)

Common Errors in Weights