Re: Formula to weight survey data

Posted by Mark Schneider on April 13, 2000 at 02:06:47:

In Reply to: Formula to weight survey data posted by Kathi on April 13, 2000 at 02:06:18:

Kathi, here's a weighting plan based on "fractional" weighting (it's also possible to do "integer" weighting using whole-number weights, but I don't think it's as common a practice any more). You'll need to weight based on one or more known demographic variables that were also asked/coded in the survey. Since you asked about weighting to households, not demographic variables, I'll use geography in my example as your weighting variable -- say, states or regions if a national study, counties or zip code groups if a state-wide or metro area study. A spreadsheet works very well in designing a weighting model.

So if we assume a metro study using counties as your weighting variable.... for each county, you first figure the percent that the county's households are of the total households in the MSA being surveyed, using Census or good household estimates (County A households divided by Total MSA households). Then you do the same percentaging for the county data in the survey results, first making sure that every respondent has a county code (missing data can complicate everything).

To figure your weights, use this formula for each county....

Census %
---------- = Weight
Survey %

Some of the weights will be >1.00 and some <1.00. Depending on how well-representative your sample is of the area you've interviewed in, hopefully none will be much higher than a weight of 2.00 or 3.00. (I think the issue of how much is too much weighting has been discussed on this forum, so check the archive if you're interested.)

Apply the appropriate weight to each respondent in the survey data depending on which county he/she lives in. Then run your tabs by the weighted data variable that you set up in your tab package. I usually use SPSS and it is fairly straight forward using its WEIGHT BY (variable name) command.

With one weighting variable it's pretty simple. But for most random respondent polling and consumer research surveys you'd probably need to weight by geography, gender, age groups, often race, and maybe even the # of adults in the household and perhaps other variables.

A piece of advice: string out all the percentages and the weights to 4-6 decimal points. What you want is the weighted data to add up to the same total as your total actual sample size. If your sample size is 1,000, without enough detail in your weighting model, you could end up with a weighted sample of 1,001 or 998 -- messy. This is just the tip of the weighting iceberg, but good luck.

Follow Ups:

Subject: