DRModels.jl

Appearance

Handling missing data ​

Status — Partial

Related drmTMB article: Handling missing data. drmTMB's missing-data routes are much broader than DRModels.jl's today. In DRModels.jl today: rows with a missing response are dropped with a warning; drm_listwise drops incomplete rows explicitly; and an experimental mi() route models one or two missing predictors for a Gaussian response. There is no general multiple imputation and no full-information route for missing responses in bivariate models.

Real data sets have gaps. What DRModels.jl does with them depends on whether the gap is in the response (the variable on the left of ~) or in a predictor (a variable on the right).

Missing responses ​

If some response values are missing (or NaN), drm fits the rows whose response is observed and warns how many rows it dropped. This is the same default as glmmTMB's na.action. Dropping these rows gives unbiased estimates when missingness depends only on the predictors in the model, not on the unobserved response value itself.

julia
using DRModels, Random
Random.seed!(3)

n = 60
x = randn(n)
y = 1 .+ 0.5 .* x .+ 0.4 .* randn(n)
y_obs = Union{Missing,Float64}[i % 10 == 0 ? missing : y[i] for i in 1:n]   # 6 gaps

fit = drm(bf(@formula(y ~ x), @formula(sigma ~ 1)), Gaussian();
          data = (; y = y_obs, x))
coef(fit, :mu)
2-element Vector{Float64}:
 1.0208992850603846
 0.4983631174351413

For a bivariate model, a row is dropped when either response is missing. That discards the observed half of each incomplete pair.

Missing predictors ​

A missing predictor value makes drm drop that row and warn how many rows it dropped and which columns caused it (R's na.omit, as drmTMB does; #1089). nobs(fit) counts the rows used, and prior weights are dropped with their rows. A NaN predictor is not dropped: it is an input error. A relmat, animal or spatial model whose levels the drop would reorder is refused by name, because its matrix is matched to the levels by position. You have two choices.

Complete-case analysis. Either let drm drop the rows, as above, or do it yourself first. drm_listwise removes every row with a missing or non-finite value in any column the model uses, warns how many rows it dropped and which columns caused it, and returns data ready for drm. Use it when you want the drop to be explicit, when predictors contain NaN or Inf, or when you need the complete-case data for bootstrap_ci. The example turns the warning off and reports the row counts instead:

julia
z = randn(n)
z_obs = Union{Missing,Float64}[i % 7 == 0 ? missing : z[i] for i in 1:n]
dat = (; y, x, z = z_obs)

form  = bf(@formula(y ~ x + z), @formula(sigma ~ 1))
clean = drm_listwise(form, dat; verbose = false)   # verbose = true (default) warns
(rows_before = n, rows_after = length(clean.y))
(rows_before = 60, rows_after = 52)
julia
fit_cc = drm(form, Gaussian(); data = clean)
coef(fit_cc, :mu)
3-element Vector{Float64}:
 1.030332944413597
 0.4780888766993422
 0.00010975427586892312

Complete-case analysis is unbiased when the rows are missing completely at random, and it wastes the observed values in the dropped rows.

Model the missing predictor (experimental). For a Gaussian response, writing mi(x) in the mean formula and giving a model for x integrates the missing x values out of the likelihood instead of dropping their rows. This supports one Gaussian or binary predictor, or two independent Gaussian predictors, with fixed effects only. See Modelled missing predictors for the syntax, a worked example, and the full list of limits.

What is not available yet ​

  • multiple imputation, or an na.action-style option to choose the behaviour;

  • full-information likelihood that keeps the observed half of an incomplete bivariate pair;

  • missing-predictor models for non-Gaussian responses, for count, ordinal or proportion predictors, or together with random or structured effects (drmTMB supports several of these);

  • missing-predictor models through the R bridge beyond the narrow route above.

If you need one of these, fit the model in drmTMB, or impute outside the package and analyse each imputed data set separately.

See also ​