1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
|
% Generated by roxygen2: do not edit by hand
% Please edit documentation in R/rescale_weights.R
\name{rescale_weights}
\alias{rescale_weights}
\title{Rescale design weights for multilevel analysis}
\usage{
rescale_weights(
data,
probability_weights = NULL,
by = NULL,
nest = FALSE,
method = "carle"
)
}
\arguments{
\item{data}{A data frame.}
\item{probability_weights}{Variable indicating the probability (design or
sampling) weights of the survey data (level-1-weight), provided as character
string or formula.}
\item{by}{Variable names (as character vector, or as formula), indicating
the grouping structure (strata) of the survey data (level-2-cluster
variable). It is also possible to create weights for multiple group
variables; in such cases, each created weighting variable will be suffixed
by the name of the group variable. This argument is required for
\code{method = "carle"}, but optional for \code{method = "kish"}.}
\item{nest}{Logical, if \code{TRUE} and \code{by} indicates at least two group
variables, then groups are "nested", i.e. groups are now a combination from
each group level of the variables in \code{by}. This argument is not used when
\code{method = "kish"}.}
\item{method}{String, indicating which rescale-method is used for rescaling
weights. Can be either \code{"carle"} (default) or \code{"kish"}. See 'Details'. If
\code{method = "carle"}, the \code{by} argument is required.}
}
\value{
\code{data}, including the new weighting variable(s). For \code{method = "carle"}, new
columns \code{rescaled_weights_a} and \code{rescaled_weights_b} are returned, and for
\code{method = "kish"}, the returned data contains a column \code{rescaled_weights}.
These represent the rescaled design weights to use in multilevel models (use
these variables for the \code{weights} argument).
}
\description{
Most functions to fit multilevel and mixed effects models only
allow the user to specify frequency weights, but not design (i.e., sampling
or probability) weights, which should be used when analyzing complex samples
(e.g., probability samples). \code{rescale_weights()} implements two algorithms,
one proposed by \cite{Asparouhov (2006)} and \cite{Carle (2009)}, to rescale
design weights in survey data to account for the grouping structure of
multilevel models, and one based on the design effect proposed by
\cite{Kish (1965)}, to rescale weights by the design effect to account for
additional sampling error introduced by weighting.
}
\details{
\itemize{
\item \code{method = "carle"}
Rescaling is based on two methods: For \code{rescaled_weights_a}, the sample
weights \code{probability_weights} are adjusted by a factor that represents the
proportion of group size divided by the sum of sampling weights within each
group. The adjustment factor for \code{rescaled_weights_b} is the sum of sample
weights within each group divided by the sum of squared sample weights
within each group (see Carle (2009), Appendix B). In other words,
\code{rescaled_weights_a} "scales the weights so that the new weights sum to the
cluster sample size" while \code{rescaled_weights_b} "scales the weights so that
the new weights sum to the effective cluster size".
Regarding the choice between scaling methods A and B, Carle suggests that
"analysts who wish to discuss point estimates should report results based
on weighting method A. For analysts more interested in residual
between-group variance, method B may generally provide the least biased
estimates". In general, it is recommended to fit a non-weighted model and
weighted models with both scaling methods and when comparing the models,
see whether the "inferential decisions converge", to gain confidence in the
results.
Though the bias of scaled weights decreases with increasing group size,
method A is preferred when insufficient or low group size is a concern.
The group ID and probably PSU may be used as random effects (e.g. nested
design, or group and PSU as varying intercepts), depending on the survey
design that should be mimicked.
\item \code{method = "kish"}
Rescaling is based on scaling the sample weights so the mean value is 1,
which means the sum of all weights equals the sample size. Next, the design
effect (\emph{Kish 1965}) is calculated, which is the mean of the squared
weights divided by the squared mean of the weights. The scaled sample
weights are then divided by the design effect. This method is most
appropriate when weights are based on additional variables beyond the
grouping variables in the model (e.g., other demographic characteristics),
but may also be useful in other contexts.
Some tests on real-world survey-data suggest that, in comparison to the
Carle-method, the Kish-method comes closer to estimates from a regular
survey-design using the \strong{survey} package. Note that these tests are not
representative and it is recommended to check your results against a
standard survey-design.
}
}
\examples{
\dontshow{if (all(insight::check_if_installed(c("lme4", "parameters"), quietly = TRUE))) (if (getRversion() >= "3.4") withAutoprint else force)(\{ # examplesIf}
data(nhanes_sample)
head(rescale_weights(nhanes_sample, "WTINT2YR", "SDMVSTRA"))
# also works with multiple group-variables
head(rescale_weights(nhanes_sample, "WTINT2YR", c("SDMVSTRA", "SDMVPSU")))
# or nested structures.
x <- rescale_weights(
data = nhanes_sample,
probability_weights = "WTINT2YR",
by = c("SDMVSTRA", "SDMVPSU"),
nest = TRUE
)
head(x)
\donttest{
# compare different methods, using multilevel-Poisson regression
d <- rescale_weights(nhanes_sample, "WTINT2YR", "SDMVSTRA")
result1 <- lme4::glmer(
total ~ factor(RIAGENDR) + log(age) + factor(RIDRETH1) + (1 | SDMVPSU),
family = poisson(),
data = d,
weights = rescaled_weights_a
)
result2 <- lme4::glmer(
total ~ factor(RIAGENDR) + log(age) + factor(RIDRETH1) + (1 | SDMVPSU),
family = poisson(),
data = d,
weights = rescaled_weights_b
)
d <- rescale_weights(
nhanes_sample,
"WTINT2YR",
method = "kish"
)
result3 <- lme4::glmer(
total ~ factor(RIAGENDR) + log(age) + factor(RIDRETH1) + (1 | SDMVPSU),
family = poisson(),
data = d,
weights = rescaled_weights
)
d <- rescale_weights(
nhanes_sample,
"WTINT2YR",
"SDMVSTRA",
method = "kish"
)
result4 <- lme4::glmer(
total ~ factor(RIAGENDR) + log(age) + factor(RIDRETH1) + (1 | SDMVPSU),
family = poisson(),
data = d,
weights = rescaled_weights
)
parameters::compare_parameters(
list(result1, result2, result3, result4),
exponentiate = TRUE,
column_names = c("Carle (A)", "Carle (B)", "Kish", "Kish (grouped)")
)
}
\dontshow{\}) # examplesIf}
}
\references{
\itemize{
\item Asparouhov T. (2006). General Multi-Level Modeling with Sampling
Weights. Communications in Statistics - Theory and Methods 35: 439-460
\item Carle A.C. (2009). Fitting multilevel models in complex survey data
with design weights: Recommendations. BMC Medical Research Methodology
9(49): 1-13
\item Kish, L. (1965) Survey Sampling. London: Wiley.
}
}
|
