collecting accurate confidence intervals for binomial proportion estimate from cross-sectional cluster survey

[Brett Collins]

Hi there,

I have been working mainly with multiple cross-sectional surveys that are multi-stage cluster sampling. I have been trying to figure out the best way to calculate the means along with the confidence intervals. The method I have been using has been: 'proportion exbf, vce(cluster clusternumber)' where exbf = variable of analysis and clusternumber = primary sampling unit variable. Would this provide me the correct results? It's giving me the 95% Logit CI and not sure if this is what I'm needing.

Thanks in advance,
-Brett

[Clyde Schechter]

No, that's not quite right. First, -svyset- your data set. Then use the -svy: proportion- command.

-help svyset-
-help svy_estimation- and follow the link for -proportion-.

[Brett Collins]

Hi Clyde,

Okay that's great. Thanks very much for your insight. I will try this out.

[Brett Collins]

Hi there,

So I managed to work out the svyset and svy for proportions. This is what I came up with:


svyset

pweight: <none>
VCE: linearized
Single unit: missing
Strata 1: <one>
SU 1: clusternumber
FPC 1: <zero>


Where 'clusternumber' is cluster variable.

An example around bottle feeding use:

. svy:proportion bottle
(running proportion on estimation sample)

Survey: Proportion estimation

Number of strata = 1 Number of obs = 1,598
Number of PSUs = 43 Population size = 1,598
Design df = 42

--------------------------------------------------------------
| Linearized Logit
| Proportion Std. Err. [95% Conf. Interval]
-------------+------------------------------------------------
bottle |
0 | .7916145 .0206998 .7467682 .8303233
1 | .2083855 .0206998 .1696767 .2532318
--------------------------------------------------------------

So this should have the clusters as the primary sampling unit and calculate the CI taking into consideration the clustered sampling method. Is this more accurate/correct? Or am I still missing something?

Cheers,
-Brett

[Clyde Schechter]

Assuming that your -svyset- parameters accurately reflect the sampling design of the survey, yes, this should give you appropriate estimates and confidence intervals.

[Brett Collins]

Great thanks for your help again Clyde.

Cheers,
-Brett