ASA Connect

Hello All,

When using Stata, is it ever possible to obtain standard errors, confidence intervals, etc for parameter estimates for variables that were declared as strata or PSU in when setting the survey design for a data set?

A collaborator wishes to use as an explanatory variable in analysis, a variable that was used in the survey design as strata. Bivariate analyses using variables used to define the design will not yield p-values. Is there a way to get around this, separate and apart from changing the manner in which the design is declared?

I look forward to your feedback.

Regards

Novie

Strata are fixed by design, so they have no sampling variability. It
might be that the investigator is trying to do something that the
design cannot support.

Mike Elliott

Michael Elliott

Professor of Biostatistics
Dept. of Biostatistics
M4124 SPH II
1415 Washington Heights
Ann Arbor, MI 48109
(734) 647-5160

Research Professor of Survey Methodology
Survey Methodology Program
Rm. 4068, 426 Thompson Street
Ann Arbor, MI 48109
(734) 647-5563

------Original Message------

Hello All,

When using Stata, is it ever possible to obtain standard errors, confidence intervals, etc for parameter estimates for variables that were declared as strata or PSU in when setting the survey design for a data set?

A collaborator wishes to use as an explanatory variable in analysis, a variable that was used in the survey design as strata. Bivariate analyses using variables used to define the design will not yield p-values. Is there a way to get around this, separate and apart from changing the manner in which the design is declared?

I look forward to your feedback.

Regards

Novie

I do not follow the question. Let me explain: When we use stratified sampling the variables which we use to define the strata are known in the population. For example, if we stratify by gender, the gender distribution in the population is known (otherwise we wouldn't be able to stratify by gender). Stratification is the exact matching of a sample characteristic to the population. Thus there is no point in estimating something which is already known. On the other hand, the distribution of estimates which were derived from stratified samples differs from the distribution of estimates derived from simple random samples. So what is the real question?

------------------------------
Yiannis Bassiakos
Associate Professor
University Of Athens, Economics Department
------------------------------