MLSurvey fits machine learning models – elastic net,
XGBoost, and random forest – to complex survey data, incorporating
sampling weights, strata, and clusters via a replicate-weights approach.
This vignette walks through the core elastic net workflow:
wElnet()alpha values in parallel with
par_wElnet()ggplot2To keep this vignette fast and self-contained, we use a small
synthetic survey-style data set rather than a real NHANES extract. The
same workflow applies directly to real survey data with a
cluster, strata, and weights
column (or an equivalent survey::svydesign() object).
A complex survey data set typically includes:
set.seed(1)
n <- 240
survey_data <- data.frame(
psu = rep(rep(1:3, each = n / 12), times = 4),
strata = rep(1:4, each = n / 4),
weight = runif(n, 500, 5000),
x1 = rnorm(n),
x2 = rnorm(n),
x3 = rnorm(n),
x4 = rnorm(n),
y = rep(c(0, 1), length.out = n) # binary outcome
)
head(survey_data)
#> psu strata weight x1 x2 x3 x4 y
#> 1 1 1 1694.789 -0.5059575 -2.5923277 -0.3410670 2.0057186 0
#> 2 1 1 2174.558 1.3430388 1.3140022 1.5024245 -2.0705715 1
#> 3 1 1 3077.840 -0.2145794 -0.6355430 0.5283077 3.0557424 0
#> 4 1 1 4586.935 -0.1795565 -0.4299788 0.5421914 -0.2613506 1
#> 5 1 1 1407.569 -0.1001907 -0.1693183 -0.1366734 -0.4543933 0
#> 6 1 1 4542.754 0.7126663 0.6122182 -1.1367339 0.1575606 1wElnet()wElnet() fits a weighted elastic net model for a single
value of alpha (the mixing parameter between ridge and
LASSO penalties), selecting the optimal lambda by
replicate-weight cross-validation.
fit <- wElnet(
data = survey_data,
col.y = "y",
col.x = c("x1", "x2", "x3", "x4"),
family = "binomial",
alpha = 0.5,
cluster = "psu",
strata = "strata",
weights = "weight",
method = "dCV",
k = 4, R = 1
)
fit
#> Weighted Elastic Net (wElnet)
#>
#> alpha: 0.5
#> lambda (min): 0.05908
#> Weighted error at min: 0.6964
#> Nonzero coefs: 0
#> Lambda grid size: 29Printing the fitted object (via print.w.elnet())
summarizes the key results: the alpha used, the selected
lambda.min, its weighted error, and the number of nonzero
coefficients at that lambda.
par_wElnet()In practice, you’ll usually want to compare several
alpha values (e.g. a grid spanning ridge to LASSO) rather
than fixing one in advance. par_wElnet() fits
wElnet() once per alpha, in parallel across CPU cores, and
reports the best (alpha, lambda) combination.
alpha_grid <- c(0, 0.25, 0.5, 0.75, 1)
par_fit <- par_wElnet(
alpha = alpha_grid,
n.cores = 2,
data = survey_data,
col.y = "y",
col.x = c("x1", "x2", "x3", "x4"),
family = "binomial",
cluster = "psu",
strata = "strata",
weights = "weight",
method = "dCV",
k = 4, R = 1
)
par_fit#> Parallel Weighted Elastic Net over 5 alpha value(s)
#>
#> alpha lambda.min min.error
#> 1 0.00 0.0847 0.251
#> 2 0.25 0.0391 0.248
#> 3 0.50 0.0312 0.246
#> 4 0.75 0.0288 0.247
#> 5 1.00 0.0265 0.249
#>
#> Best alpha: 0.5
(This chunk is not evaluated when the vignette is built, since
spawning a parallel cluster is best avoided in automated/non-interactive
build environments such as R CMD check. It runs normally in
an interactive R session or script.)
plot() and autoplot() methods are also
available for par.w.elnet objects, showing both the error
profile for every alpha and a summary of the best error achieved by
each:
plot(par_fit) # both panels: profile + summary
plot(par_fit, type = "profile") # just the error-vs-lambda curves, one per alpha
plot(par_fit, type = "summary") # just the best-error-vs-alpha curve
autoplot(par_fit) # combined (requires patchwork)
autoplot(par_fit, type = "profile", facet = TRUE) # one panel per alphasurvey::svydesign() object insteadIf you already have a survey.design object, you can pass
it directly via the design argument instead of specifying
cluster/strata/weights
separately:
method also accepts "JKn" (jackknife),
"BRR", "bootstrap",
"subbootstrap", "split", and
"extrapolation", each implemented via
[replicate_weights()]. For example, a jackknife fit:
fit_jkn <- wElnet(
data = survey_data, col.y = "y", col.x = c("x1", "x2", "x3", "x4"),
family = "binomial", alpha = 0.5,
cluster = "psu", strata = "strata", weights = "weight",
method = "JKn"
)
fit_jkn
#> Weighted Elastic Net (wElnet)
#>
#> alpha: 0.5
#> lambda (min): 0.05908
#> Weighted error at min: 0.6958
#> Nonzero coefs: 0
#> Lambda grid size: 29See ?replicate_weights for details on each method and
how its training/test replicate-weight columns are constructed.
Note: Global options to avoid only one PSU in a stratum in a particular domain or subpopulation can be set at the beginning when getting errors with other methods:
#> options(survey.adjust.domain.lonely=TRUE)
#> options(survey.lonely.psu="adjust")
?wXGBoost and ?wRandomforest extend the
same replicate-weight framework to gradient boosting and random
forests.?optim_wxgb_para automates hyperparameter search for
wXGBoost().