CODEWITHSUNDEEP
rsurvey
s.stats
s.stats

Reputation: 136

Negative lower confidence interval limit for a proportion using svyby() for survey weighted data

How do you ensure the lower confidence interval limit is non-negative for a proportion when dealing with survey weighted data?

This is the code:

svy_design <- svydesign(ids = ~block, strata = ~strata_study, weights = ~wt, data = data, nest = TRUE, fpc = NULL)

svyby(~binary_outcome,by=~binary_exposure, design=svy_design, FUN=svymean, data=data,keep.var = TRUE, vartype = "ci")

Result as follows:

  binary_exposure binary_outcome0 binary_outcome1 ci_l.binary_outcome0 ci_l.binary_outcome1 ci_u.binary_outcome0 ci_u.binary_outcome1
1               1        0.9858986       0.01410138             0.9782720           0.006474771             0.9935252            0.02172798
2               2        0.9264685       0.07353151             0.8489539          -0.003983028             1.0039830            0.15104605

The lower confidence interval for the proportion is reported as -0.003983028. How to ensure that the estimates are bounded 0 and above?

Upvotes: 2

Views: 349

Answers (2)

Thomas Lumley
Thomas Lumley

Reputation: 2765

You can use svyciprop: eg, compare

> svyby(~I(emer==0), ~stype, svyciprop, design=dclus1, method="mean",vartype="ci")
  stype I(emer == 0)        ci_l      ci_u
E     E    0.2083333  0.08601926 0.3306474
H     H    0.2142857 -0.04681564 0.4753871
M     M    0.0800000 -0.04023591 0.2002359

> svyby(~I(emer==0), ~stype, svyciprop, design=dclus1, method="logit",vartype="ci")
  stype I(emer == 0)       ci_l      ci_u
E     E    0.2083333 0.11139102 0.3558571
H     H    0.2142857 0.05467787 0.5625459
M     M    0.0800000 0.01669209 0.3081661

Upvotes: 3

MendelYev
MendelYev

Reputation: 174

svyby uses confint function with method='Wald' argument to process confidence interval. Wald method can lead to negative values. You should use method='likelihood' instead. I haven't find a way to change the method directly from svyby in the documentation, but you can call confint() directly on a svyglm object with method='likelihood'.

Upvotes: 1

Related Questions