Hi all,
Using Stata9 and survey data I am having a few concerns about the
utilisation of nosvyadjust when running test models.
With my data I have the following design specs:
pweight: pps
VCE: linearized
Strata 1: sex
SU 1: <observations>
FPC 1: <zero>
I ran a tab:
svy: tab ec12 qmc22, count col pearson null wald noadjust
Number of strata = 1 Number of obs =
4364
Number of PSUs = 4364 Population size =
4365.2309
Design df =
4363
-------------------------------
| other p. pain-12m
ec12 | no yes Total
----------+--------------------
no | 3368 841.9 4210
| .9684 .9494 .9645
|
yes | 110 44.85 154.8
| .0316 .0506 .0355
|
Total | 3478 886.8 4365
| 1 1 1
-------------------------------
Key: weighted counts
column proportions
Pearson:
Uncorrected chi2(1) = 7.4145
D-B (null) F(1, 4363) = 3.7006 P = 0.0545
Design-based F(1, 4363) = 4.9500 P = 0.0261
Wald (Pearson):
Unadjusted chi2(1) = 3.6951
Unadjusted F(1, 4363) = 3.6951 P = 0.0546
Adjusted F(1, 4363) = 3.6951 P = 0.0546
Then I ran the svy: proportion of the model and a linear hypothesis test
(with and without nosvyadjust):
svy: proportion ec12, over(qmc22)
--------------------------------------------------------------
| Linearized Binomial Wald
Over | Proportion Std. Err. [95% Conf. Interval]
-------------+------------------------------------------------
no |
no | .9683775 .0034443 .9616249 .9751301
yes | .9494275 .0091864 .9314175 .9674375
-------------+------------------------------------------------
yes |
no | .0316225 .0034443 .0248699 .0383751
yes | .0505725 .0091864 .0325625 .0685825
--------------------------------------------------------------
test [no]no=[no]yes
Adjusted Wald test
( 1) [no]no - [no]yes = 0
F( 1, 4363) = 3.73
Prob > F = 0.0535
test [no]no=[no]yes, nosvyadjust
Unadjusted Wald test
( 1) [no]no - [no]yes = 0
F( 1, 4363) = 3.73
Prob > F = 0.0535
1. As this is survey data, when I do linear tests SHOULD I use the
nosvyadjust options?
2. Why is my D-B model F(1, 4363) = 4.9500; P = 0.0261 significant yet
the linear test is not: F(1, 4363) = 3.73; Prob > F = 0.0535
3. This difference becomes a little concerning when one of my variables
say year5 has more than 2 categories (e.g., 5 year groups 16-20, 21-25
etc). When I want to test for linear differences between groups I
question the results of the linear test (see below). In particular the
proportions test for group 2 (_prop_2) has overlapping CIs. Yet the
linear tests suggest that these are statistically different (this is the
same for using or not using the option nosvyadjust; see below). This
raises the question for me: Can I use the linear test output to
highlight which groups are contributing to any overall differences?
svy: proportion year5, over(qmc16)
--------------------------------------------------------------
| Linearized Binomial Wald
Over | Proportion Std. Err. [95% Conf. Interval]
-------------+------------------------------------------------
_prop_1 |
no | .1029562 .0143045 .0749072 .1310051
yes | .1531804 .0108347 .1319352 .1744257
-------------+------------------------------------------------
_prop_2 |
no | .0917431 .0122425 .0677374 .1157488
yes | .1311121 .0087642 .1139269 .1482972
-------------+------------------------------------------------
_prop_3 |
no | .1202854 .0117523 .097241 .1433298
yes | .1224578 .0076194 .1075173 .1373983
-------------+------------------------------------------------
_prop_4 |
no | .1569827 .0127744 .1319341 .1820312
yes | .1557767 .0082218 .139655 .1718985
-------------+------------------------------------------------
_prop_5 |
no | .1712538 .012996 .1457708 .1967368
yes | .1691908 .0085773 .1523722 .1860095
-------------+------------------------------------------------
_prop_6 |
no | .2038736 .0154416 .1735951 .2341521
yes | .1639983 .0090101 .1463308 .1816657
-------------+------------------------------------------------
_prop_7 |
no | .1529052 .0140525 .1253505 .1804599
yes | .1042839 .0072473 .0900731 .1184947
-------------+------------------------------------------------
test ([_prop_1]no=[_prop_1]yes) ([_prop_2]no=[_prop_2]yes)
([_prop_3]no=[_prop_3]yes) ([_prop_4]no=[_prop_4]yes)
([_prop_5]no=[_prop_5]yes) ([_prop_6]no=[_prop_6]yes)
([_prop_7]no=[_prop_7]yes), mtest(noadjust)
---------------------------------------
| F(df,2695) df p
-------+-------------------------------
(1) | 7.83 1 0.0052 #
(2) | 6.84 1 0.0090 #
(3) | 0.02 1 0.8768 #
(4) | 0.01 1 0.9367 #
(5) | 0.02 1 0.8946 #
(6) | 4.97 1 0.0258 #
(7) | 9.46 1 0.0021 #
-------+-------------------------------
all | 4.12 6 0.0004
---------------------------------------
Sorry about the output but I thought it would provide some insight.
Jase
Jason Ferris
Research Officer
Australian Research Centre in Sex, Health and Society
Level 1, 215 Franklin St
Melbourne
VIC 3000
P: 61 (0)3 9285 5282
F: 61 (0)3 9285 5220
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
