svy- different se and confident intervals when using subpop or over

[Jennifer Perez]

Hi,

Here is the svyset for my survey data:
svyset psu [pw=wt_final], strata(strata1) fpc(fpc1) || school_id, strata(strata2) fpc(fpc2) || id, strata(strata3) fpc(fpc3) singleunit(centered)

I use svy mean and the subpop statement to get the mean for grade 4 and 6 and region A. I used again svy mean but now using over(grade region) to compare the results. I get the same estimate but different standard error (slightly) and confidence intervals. Could someone explain to me why?. I really appreciate your input.

Thanks!
Jennifer

. svy, subpop(if grade==4 & region=="REGION A"): mean adj_orf
(running mean on estimation sample)

Survey: Mean estimation

Number of strata = 1 Number of obs = 1574
Number of PSUs = 10 Population size = 182028
Subpop. no. obs = 503
Subpop. size = 63187.1
Design df = 9

--------------------------------------------------------------
| Linearized
| Mean Std. Err. [95% Conf. Interval]
-------------+------------------------------------------------
adj_orf | 4.637633 1.727553 .7296358 8.54563
--------------------------------------------------------------
Note: 2 strata omitted because they contain no subpopulation
members.
Note: Strata with single sampling unit centered at overall
mean.

. svy, subpop(if grade==6 & region==1): mean adj_orf
(running mean on estimation sample)

Survey: Mean estimation

Number of strata = 1 Number of obs = 1577
Number of PSUs = 10 Population size = 182253
Subpop. no. obs = 517
Subpop. size = 47116.9
Design df = 9

--------------------------------------------------------------
| Linearized
| Mean Std. Err. [95% Conf. Interval]
-------------+------------------------------------------------
adj_orf | 24.52088 4.222985 14.96783 34.07394
--------------------------------------------------------------
Note: 2 strata omitted because they contain no subpopulation
members.
Note: Strata with single sampling unit centered at overall
mean.


. svy: mean adj_orf, over(region grade)
(running mean on estimation sample)

Survey: Mean estimation

Number of strata = 3 Number of obs = 3428
Number of PSUs = 28 Population size = 461175
Design df = 25

Over: region grade
_subpop_1: REGION A 4
_subpop_2: REGION A 6
_subpop_3: REGION B 4
_subpop_4: REGION B 6
_subpop_5: REGION C 4
_subpop_6: REGION C 6

--------------------------------------------------------------
| Linearized
Over | Mean Std. Err. [95% Conf. Interval]
-------------+------------------------------------------------
adj_orf |
_subpop_1 | 4.637633 1.727522 1.079735 8.195532
_subpop_2 | 24.52088 4.226485 15.81627 33.22549
_subpop_3 | 10.21037 1.837862 6.425223 13.99552
_subpop_4 | 35.90704 5.374154 24.83876 46.97532
_subpop_5 | 10.59444 2.29463 5.868564 15.32032
_subpop_6 | 33.77606 4.203415 25.11897 42.43316
--------------------------------------------------------------
Note: Strata with single sampling unit centered at overall
mean.

.

[Jeff Pitblado (StataCorp)]

In the analysis using subpop() there was a footnote following the estimation table.
The footnote was

Note: 2 strata omitted because they contain no subpopulation
members.

This means that observations were dropped from the estimation sample, and were not
counted when computing the standard errors.

In the analysis using over(), the above mentioned strata were not omitted because
they overlapped with on of the other subpopulations identified by the over() variables.

[John Havens]

Hi Jeff,
Does the note "2 strata omitted because they contain no subpopulation members." raise any concern to you? Do you mean that the -over()- option is better in this case?
Thanks

[Jeff Pitblado (StataCorp)]

Not really.

For the subpop case, this just means that 2 strata were not relevant (i.e. did not provide any information)
for the specified subpop.

Using over benefits those of us who want to compare parameter estimates between multiple subpopulations.