New bug: `get_weights.svyglm()` is broken on Github

[vincentarelbundock]

I think that the latest version of insight breaks marginaleffects tests because it no longer retrieves weights from survey objects.

This appears to be a bug introduced recently. Versions of insight as recent as 6 commits ago behave as expected.

library(insight)
library(survey)
packageVersion("survey")

[1] '4.4.8'

packageVersion("insight")

[1] '1.4.2.8'

data("fpc", package = "survey")
svyd <- survey::svydesign(
    weights = ~weight,
    ids = ~psuid,
    strata = ~stratid,
    fpc = ~Nh,
    variables = ~ x + nh,
    data = fpc,
    nest = TRUE
)
mod <- survey::svyglm(x ~ nh, design = svyd)
find_weights(mod)

[1] "weight"

get_weights(mod)

NULL

[vincentarelbundock]

Also, I think we were getting this name for weights previously:

"(weights)"

Changing that name would also be a breaking change.

[bwiernik]

"(weights)" is the name that survey gives to all weight columns in its model frames. This makes the returned model frames incompatible with the original model call.

It's a breaking change from previous behavior, but previous behavior seems not thoughtful and at odds with typical insight principles.

Perhaps add the weights column name as an attribute or include both the new behavior and the "(weights)" column?

[strengejacke]

Let me check. The recent changes were initiated by this issue: #1139

[strengejacke]

Maybe we can also ping @ngreifer for this discussion.

[strengejacke]

Maybe we can also use the source argument here, which either extracts data directly from the model.frame (the new behaviour), or from the "restored" model.frame (the former behaviour)?

If we go this route, we have to decide about the default, and which option gets which name.

library(insight)
library(survey)
data("fpc", package = "survey")
svyd <- survey::svydesign(
    weights = ~weight,
    ids = ~psuid,
    strata = ~stratid,
    fpc = ~Nh,
    variables = ~ x + nh,
    data = fpc,
    nest = TRUE
)
mod <- survey::svyglm(x ~ nh, design = svyd)

# works when recovered from "environment" - which falls back
# to classical "model frame" behaviour in this case
find_weights(mod, source = "environment")
#> [1] "(weights)"

get_weights(mod, source = "environment")
#> [1] 0.8888889 0.8888889 0.8888889 0.8888889 0.8888889 1.1851852 1.1851852
#> [8] 1.1851852

get_data(mod, source = "environment")
#>     x nh (weights)
#> 1 2.8  5 0.8888889
#> 2 4.1  5 0.8888889
#> 3 6.8  5 0.8888889
#> 4 6.8  5 0.8888889
#> 5 9.2  5 0.8888889
#> 6 3.7  3 1.1851852
#> 7 6.6  3 1.1851852
#> 8 4.2  3 1.1851852

# weights variable not included in "model.frame()" method - not sure why?
find_weights(mod)
#> [1] "weight"

get_weights(mod)
#> NULL

get_data(mod)
#>     x nh
#> 1 2.8  5
#> 2 4.1  5
#> 3 6.8  5
#> 4 6.8  5
#> 5 9.2  5
#> 6 3.7  3
#> 7 6.6  3
#> 8 4.2  3



# this example works more or less the other way round
set.seed(123)
n <- 5
x <- runif(n)
y <- runif(n)
z <- runif(n)
w <- rnorm(n, 1, 0.05)

dat <- data.frame(w, x, y, z)
des <- svydesign(~1, weights = ~w, data = dat)
svy_fit <- svyglm(y ~ poly(x, 2), design = des)


find_weights(svy_fit, source = "environment")
#> [1] "(weights)"

get_weights(svy_fit, source = "environment")
#> [1] 1.0203994 0.8761746 1.0400410 0.9831547 1.0802303

get_data(svy_fit, source = "environment")
#>           y (weights)          x         x.1         w
#> 1 0.0455565 1.0203994 -0.6354621  0.47997516 1.0640277
#> 2 0.5281055 0.8761746  0.2151147 -0.57707056 0.9136365
#> 3 0.8924190 1.0400410 -0.4292432 -0.42647447 1.0845092
#> 4 0.5514350 0.9831547  0.3760007  0.01922542 1.0251906
#> 5 0.4566147 1.0802303  0.4735898  0.50434445 1.1264168


find_weights(svy_fit)
#> [1] "w"

get_weights(svy_fit)
#> [1] 1.0640277 0.9136365 1.0845092 1.0251906 1.1264168

get_data(svy_fit)
#>           w         x         y
#> 1 1.0640277 0.2875775 0.0455565
#> 2 0.9136365 0.7883051 0.5281055
#> 3 1.0845092 0.4089769 0.8924190
#> 4 1.0251906 0.8830174 0.5514350
#> 5 1.1264168 0.9404673 0.4566147

^{Created on 2025-10-14 with reprex v2.1.1}

[vincentarelbundock]

What is the logic for preferring this?

get_weights(mod)
NULL

Are we protecting the user from some specific error?

Otherwise, if there are weights, I think we should return them by default without requiring a specialized argument.

[strengejacke]

What is the logic for preferring this?

That was due to #1139, where the behaviour was changed. In #1145 we have the possibility to choose between options, but there's no preference right now.

I think we should probably keep the old behaviour as default, and thereby not introducing breaking changes.

Otherwise, if there are weights, I think we should return them by default without requiring a specialized argument.

See 2nd example, where we have "two weights". I think, @ngreifer would like to extract the "original" names and values (as implemented in #1139), so for the following example, we indeed would need an argument to control which weights are returned (and which name for the weight-variable).

find_weights(svy_fit, source = "environment")
#> [1] "(weights)"

get_weights(svy_fit, source = "environment")
#> [1] 1.0203994 0.8761746 1.0400410 0.9831547 1.0802303

find_weights(svy_fit)
#> [1] "w"

get_weights(svy_fit)
#> [1] 1.0640277 0.9136365 1.0845092 1.0251906 1.1264168

[strengejacke]

To summarize: insight now supports both extraction methods. The default is like the old default (source = "environment"), the options implemented in #1139 are available with source = "mf".

[strengejacke]

@vincentarelbundock can you confirm that the latest version of insight no longer breaks marginaleffects?

[vincentarelbundock]

Yep, looks great. Thanks!!