Sample weights column in dataset - what is it for

[Mosi Fagbemide]

Please help. I have a variable column in my dataset named weight, is it to be included in my regression?

[Rich Goldstein]

depends on what it is and how that relates to the purpose of your regression; if it is the weight of a physical object (e.g., weight of a car), then only include if that's substantively relevant; if it is some other kind of weight, we need to know what kind of weight it is and what the purpose of your regression is

[Marcos Almeida]

You may type in the command window:

. help weights

There, you get an intersting introduction of sample weights "in a nutshell".

[Mosi Fagbemide]

Thank you Rich and Marcos. I have looked at the "help weights", im still unclear. the definitions on the different types of weights are clear but its relevance to my regression commands is still unclear.

By definition/description of my dataset which is sourced from a survey, the weights are sampling weights.

[Marcos Almeida]

Hello Mosi,

The Stata manual (http://www.stata.com/manuals13/svy.pdf) provides the information on the "relevance" of using weights in a survey data analysis.

These are excerps worth reading:

In page 4, we read:

Sampling weights. In sample surveys, observations are selected through a random process, but different observations may have different probabilities of selection. Weights are equal to (or proportional to) the inverse of the probability of being sampled. Various postsampling adjustments to the weights are sometimes made, as well. A weight of wj for the jth observation means, roughly speaking, that the jth observation represents wj elements in the population from which the sample was drawn. Omitting weights from the analysis results in estimates that may be biased, sometimes seriously so. Sampling weights also play a role in estimating standard errors. [...] To put it succinctly: using sampling weights is important to get the point estimates right. We must consider the weighting, clustering, and stratification of the survey design to get the standard errors right. If our analysis ignores the clustering in our design, we would probably produce standard errors that are smaller than they should be. Stratification can be used to get smaller standard errors for a given overall sample size.

In pages 67-68, there is an example of what happens if you use - or not - weights on survey data analysis.


Best,

Marcos

[Steve Samuels]

When you have survey data, you must svyset the data. The data will often include not only the sampling weights, but a variable to designate "sampling strata" and one or more variables that designate the units sampled at each survey stage. The highest level units are "primary sampling units". See the remarks and examples in the Manual entry for survey.

[Mosi Fagbemide]

Thank you very much Marcus and Steve. warm regards.