|
<< Click to Display Table of Contents >> Navigation: User Guide > Weighting > Respondent weighting |
For more information online about how to use the Weight data window, see https://netmr-download.scot/user-documentation/weighting.
This refers to the application of a weight to each individual record during analysis so that the figures produced are based on weighted data. For example when a table (cross tab) is accumulated, the weight for an individual is added to the relevant cells instead of the value 1.
Respondent weighting is also known as:
•Target weighting: weighting to a full matrix of figures
•Rim weighting: weighting to various targets without the full matrix
•Grossing up or down: simple weighting that gives many respondents the same weight
Respondent weights are applied for various reasons including:
•So that the tables are based on known universe figures
•To rebalance the data to reflect the target population
•To adjust for minor errors in sampling
•To make the bases the same as an earlier survey
The weights are calculated by setting targets for particular subsets of the data. The targets can be in the form of a full matrix or just the rims (rim weighting).
The targets are set using the Weight targets window and calculated using the Weight data window. The weights calculated are stored in a float question with the other raw data for each record.
You can choose the precision needed when calculating weights. The greater the precision required, the more computations (iterations) that will be needed.
IMPORTANT: when data is weighted there is usually some "loss" of information, see Effective sample size (ESS).
IMPORTANT: whenever data is weighted the weighting report, including the ESS (efficiency), values should be kept and delivered to the user of the data.
A number of different weight schemes can be allocated to respondents but in general only one of these can be used for any one report. Usually only one respondent weighting scheme will be used for each project.
Constraints can be set on the weights applied to individual respondents, for example they can be constrained to a value between 0.5 and 2.0.
If a relevant entry in the project is named "weight" it will automatically be set to be used for respondent weighting in tables.
The entry to use for respondent weighting can be any float question or variable. Integer entries can also be used. The entry for respondent weighting is set in Tables overall.
Targets can be set for:
•Overall weighted
•Single-coded entry responses
•Multi-coded entry responses
•Filters, true and false
•Entry bases, true or false for filtered entries
•An accumulated quantity (integer or float), see important note below
TIP: for full matrix weighting a variable will need to be created with a response for each cell in the matrix, see Variable responses matrix.
When the targets are set there is no distinction between the various types of respondent weighting and you can mix any types together in one scheme.
Where multi-coded and single-coded entries are used you do not need to allocate a target for every response. For example if the targets are a brand list you could set targets only for the five most important brands.
IMPORTANT: When weighting using a value (integer or weight) this does not apply variable weights depending on the individual values. This facility only grosses up or down treating all relevant records in the same way.
Sometimes you need different weighting schemes for different subsets of the data. This is achieved by filtering the target entries so that those targets are only used for the subset described by the filter. This may mean duplicating the target variables.
You can mix separated weighting and global weighting within the same weighting scheme.
One common example of separated weighting is a repeat survey where the same or similar weighting is applied to each repeat. In this case you would set "Keep existing" for previous waves that have already been weighted.
Sometimes the weighting is applied in two stages. The first stage weighting is applied and then without resetting (preparing) the weights the second stage is applied.
There are two ways to do this:
•The second or subsequent stages are independent of the first, these later stages will cause the earlier stage weighting to be "distorted" so that the targets for these earlier stages are no longer met
•The second or subsequent stages include the first earlier stage targets in addition to the new targets, in this case the earlier stages are not "distorted"
If the calculations do not succeed in achieving your targets the usual reasons for failure are:
That there is more than one overall target value and the targets are not the same. In particular you may have a target for the total sample and separate targets for a set of all-inclusive responses that do not add to same figure.
Similarly, you have allocated two different targets to the same subset of the data.
You have only allocated percentage targets without any overall control over the required total. This can cause overall total to spiral indefinitely.
It may be that you have given it an impossible task. The chances of this happening are higher for projects with less records or with more rims (target entries). If this happens you may have to combine some of the responses for weighting.