Multiple respondent weights in SMRT

Question

Hi,

Since my survey sample was not representative in three dimensions (gender, region and age) vs the national approved standard, I would like to apply 3 sets of weights in SMRT to do the simulations. I have imported the variables into SMRT, which works fine. However, when I set the weights for gender and then move to age the weights for gender are lost. Are multiple sets of weights not supported? Is there are work-around for this?

To illustrate my example:
Weights  1.0559 for male and 0.9505 for female, while at the same time  1.1782 for age group 1, 0.9480 for age group 2, etc.

Thanks for a quick reply.

Comments

Is it valid if I create a new variable in the .csv file import that assigns a number between 1 and 40 to all combinations of the three dimensions (40 because 5 regions * 2 genders * 4 age groups = 40 combinations) and use this to put the weight. Then I calculate the weights by multiplying the individual weights for each combination.

Is it somehow possible to batch assign the weights or copy/paste the weights for all my 160 scenarios? Otherwise I have enter 6400 numbers manually. Thanks for your help.

Answer 1

I'm assuming you are using the Respondent Weights window and have the 'Assign weights to segment categories' option selected.

SMRT only uses a single set of respondent weights in simulations.   In that window, when you select a demographic variable from the left hand set of variables, you are selecting that variable as the weighting variable.  The weights you enter when a variable is selected are lost when another variable is selected.  I'll admit this is not intuitive or ideal.

Our new Choice Simulator (which replaces the SMRT simulator) works much better in this regard, as your weights would be columns in the demographic data, or configured as logic according to the segment.  So they would not be lost when another variable is selected.  However, still only a single weighting variable can be used in simulations.

The combinatorial idea you mention would be about the only way I can think to achieve what you're looking for.  There is no batch method for assigning the weights, but they could be easily processed with Excel and then imported.

Comments

Hi Walter,
Thanks for your answer. Yes I am using the respondent weights window and "assign weights to segment categories" via either the simulation scenario or via Analysis > Analysis Settings > Respondent Weights in the toolbar.
How do I import the weights from Excel? I can't find an option to import and the copy function of the keyboard (CTRL+v) doesn't work.

---

What you want to do is export the Excel data to CSV and then import using File | Merged Variables.  It will walk you through bringing in the data.

Once you have the new variables imported, you can select 'Use merged variable value as weight' from the Respondent Weights window.

---

Thanks Walter. I think I get it now: I create a a column with the weight per respondent in Excel (instead of creating a column with the categorization per respondent and assigning a weight per category in SMRT). I tested it and it takes a while longer now to calculate the simulations, but my results are rather similar. I guess my survey not being interely representative didn't affect the results too much in the end.

---

Hi Walter,
Just to be sure: using respondent weighting in SMRT doesn't affect the correctness of the  reported standard errors in SMRT? I read on the internet (https://www.decisionanalyst.com/blog/dataweighting/) that respondent weighting might increase variance, which is fine for my purpose, because my sample size is large. However, of course it is important for me to be able to rely on the correctness of the reported standard errors. Thanks for your confirmation that SMRT deals with this correctly.

---

The standard errors do take the weighting into account, where N in the common equation is replaced by the sum of the weights.

In the new Choice Simulator, we have made a change in that we now use Kish's Effective Sample Size rather than the sum of weights.  It is equal to the sum of weights squared divided by the sum of squared weights.

The ESS in essence is giving you the equivalent power compared to unweighted respondents.  If you interviewed 100 respondents, and your ESS is 96.4, your weighting is giving you the 'power'/'accuracy' if you had 96.4 unweighted respondents.

By using the ESS we believe the standard errors of the Choice Simulator are more proper.