svy: regress
Regression that respects the survey design.
What we have
- weighted least squares with Taylor-linearised standard errors
- the adjusted Wald F test that every slope is zero
- checked against R's
surveypackage
Coming later
svy:for the other models —svy: logitand the rest are refused
How you write it
svy: regress depvar [indepvars] [if]
Square brackets mean optional. A vertical bar means choose one.
What it does
svy: regress mpce age i.sector fits the same model as regress, but the coefficients are weighted and the standard errors are computed the way survey statisticians do (Taylor linearisation, also called the sandwich estimator). Confidence intervals use the design degrees of freedom.
The F test at the top is the adjusted Wald test that every slope is zero. Rows with a missing value leave the estimate but stay in the design, so the cluster structure is not damaged by a blank.
Watch out. Fitting an ordinary
regress with weights is not the same thing. It gets the coefficients right and the standard errors wrong.Try it
Every example below was run after sysuse vira.sample02, clear.
. svyset village [pw=multiplier], strata(stratum)
pweight: multiplier
VCE: linearized
Single unit: missing
Strata 1: stratum
Sampling unit 1: village
FPC 1: <zero>
The same thing in pandas and in R
pandas
design = dict(psu='village', strata='stratum', weight='multiplier')
R
library(survey)
design <- svydesign(ids = ~village, strata = ~stratum,
weights = ~multiplier, data = df, nest = TRUE)
# This R was run in a real R (4.6.1) and gave the same numbers.
. svy: regress mpce age i.sector
Survey: Linear regression
Number of strata = 12 Number of obs = 6,000
Number of PSUs = 400 Population size = 10,612,910
Design df = 388
F( 2, 387) = 272.76
Prob > F = 0.0000
R-squared = 0.0887
------------------------------------------------------------------------------
| Linearized
mpce | Coefficient Std. err. t P>|t| [95% conf. interval]
-------------+----------------------------------------------------------------
sector=Urban | 946.0186 41.52655 22.78 0.000 864.3734 1027.664
age | 5.704484 1.030673 5.53 0.000 3.678081 7.730887
_cons | 3125.311 50.91935 61.38 0.000 3025.199 3225.424
------------------------------------------------------------------------------
The same thing in pandas and in R
pandas
# survey regression: weighted least squares, Taylor-linearised covariance # design: [' VCE: linearized', ' Single unit: missing', ' Strata 1: stratum', ' Sampling unit 1: village', ' FPC 1: <zero>']
R
library(survey) model <- svyglm(mpce ~ age + factor(sector), design = design) summary(model) # This R was run in a real R (4.6.1) and gave the same numbers.
The full built-in help (help svy: regress)
svy: regress depvar [indepvars] [if]
Linear regression that respects the survey design from svyset: the weights,
the strata and the primary sampling units. Point estimates are weighted
least squares; the standard errors are Taylor-linearised (the "sandwich"),
and confidence intervals use the design degrees of freedom, PSUs minus strata.
. svyset village [pw=multiplier], strata(stratum)
. svy: regress mpce age i.sector
Survey: Linear regression
Number of strata = 12 Number of obs = 6,000
Number of PSUs = 400 Population size = 10,612,910
Design df = 388
F( 2, 387) = 272.76
...
* an `if` is a subpopulation: the rest of the sample stays in the design.
Dropping those rows first would lose PSUs and shrink the standard errors
* rows with a missing value leave the estimate but not the design
* the weights come from svyset, so `[pw=...]` on this command is refused
* with no regressors it estimates a constant, which is `svy: mean`
* the F test is Stata's adjusted Wald test that every slope is zero
Not compared against a real Stata run yet: see docs/semantics.md.Nearby
svyset svy: mean svy: total svy: proportion svy: ratio svy: tabulate