Skip to contents

drmSEM

📖 Documentation & articles: https://itchyshin.github.io/drmSEM

The drmSEM package is a distributional piecewise SEM framework built on drmTMB, where causal paths can target not only the expected response but also scale, shape, zero-inflation, hurdle probability, random-effect scale, and residual correlation.

Status: early / experimental (version 0.5.0). See Status.

Component-labelled DAG for the canonical size, abundance, and survival example. Black solid arrows target the mean, green dashed arrows target sigma, and orange dotted arrows target zero inflation.
Component-labelled DAG for the canonical size, abundance, and survival example. Black solid arrows target the mean, green dashed arrows target sigma, and orange dotted arrows target zero inflation.

Installation

drmSEM is built on the drmTMB fitting engine, which compiles TMB (C++) from source, so you need a working C++ toolchain (Rtools on Windows, Xcode command-line tools on macOS, a standard build toolchain on Linux). Install the engine first, then drmSEM:

# install.packages("pak")
pak::pak("itchyshin/drmTMB")
pak::pak("itchyshin/drmSEM")

Or with remotes:

# install.packages("remotes")
remotes::install_github("itchyshin/drmTMB")
remotes::install_github("itchyshin/drmSEM")

Why drmSEM

Engine / layer split. drmTMB is the fitting engine; drmSEM is the SEM layer on top of it. drmSEM never fits its own likelihoods. Each endogenous node is one drmTMB fit, the system is piecewise, and the graph must be a DAG.

Component-labelled paths. A causal path does not have to point at the mean. It can target any modelled distributional component of a node: mu, sigma, nu, zi, hu, sd(group), or rho12 (the residual correlation between the two responses of a bivariate node, i.e. eps_y1 <-> eps_y2 — not a directed y1 -> y2 path). For example, temperature can act on abundance through several distinct channels:

  • temp -> mu(abundance) — temperature shifts the expected abundance.
  • temp -> sigma(abundance) — temperature changes the dispersion / scale, not the mean.
  • temp -> zi(abundance) — temperature changes the probability of structural zeros, not the conditional mean.

These are different scientific claims, and drmSEM keeps them distinct everywhere: a path to sigma or zi is never reported as a mean effect.

Honest effects. Indirect and total effects are computed by Monte-Carlo g-computation over the fitted DAG (mean-mediated vs distribution-mediated), never by multiplying coefficients — coefficient products are invalid across non-Gaussian links and across distributional components. Effects are reported on the conditional (typical-group, random-effects-at-zero) response scale. The default direct is a controlled direct effect; the mean/distribution split is an interventional decomposition (the distribution-mediated row is a Jensen-gap term — Pearl; Imai et al.; VanderWeele), not a cross-world natural decomposition unless the outcome is linear with no exposure–mediator interaction (use effect = "natural" for that).

Compared to existing tools. lavaan does latent-variable Gaussian SEM; piecewiseSEM does piecewise SEM but works on the mean only; glmmTMB fits rich distributional GLMMs but is not an SEM; dsem does dynamic SEM. drmSEM is the piece that lets a piecewise SEM address scale, shape, zero-inflation, hurdle, random-effect scale, and residual correlation as first-class causal targets.

Quick start

The question. Suppose body size, local abundance, and survival form a chain, and you want to know: does temperature reach survival only by shifting the average size and abundance, or also by changing how variable size is and how often abundance collapses to zero? A mean-only SEM cannot even ask the second half of that question. drmSEM can, because a path can target a non-mean component.

The canonical example below encodes that chain: size -> abundance -> survival, with temp acting on the sigma (scale) of size and habitat on the zi (zero-inflation) of abundance. This block is illustrative and not executed here; the drmSEM intro vignette runs the same chain end to end with simulated data.

library(drmSEM)

sem <- drm_sem(
  size      = drm_node(drmTMB::bf(size ~ temp + habitat, sigma ~ temp),
                       family = stats::gaussian()),
  abundance = drm_node(drmTMB::bf(abundance ~ size + temp, zi ~ habitat),
                       family = drmTMB::nbinom2()),
  survival  = drm_node(drmTMB::bf(cbind(alive, dead) ~ abundance + size),
                       family = drmTMB::beta_binomial()),
  data = dat
)

paths(sem)        # component-labelled path table
basis_set(sem)    # independence claims implied by the DAG
dsep(sem)         # any-component LRT for each claim
fisher_c(sem)     # Fisher's C goodness-of-fit for the SEM

# Effects of temperature on survival, propagated through the fitted DAG
direct_effects(sem,   from = "temp", to = "survival")
indirect_effects(sem, from = "temp", to = "survival")
total_effects(sem,    from = "temp", to = "survival", method = "simulate")

plot(sem)         # DAG with component-labelled edges

Reading the output. indirect_effects() returns one row per quantity, each an effect on the response (here, probability) scale:

  • total_path — the full effect of temp on survival through the chain;
  • direct — the controlled direct effect (mediators held fixed);
  • indirect — total_path - direct;
  • mean_mediated — the part carried by mediator means;
  • distribution_mediated — the extra part that appears only when mediators pass realized draws, i.e. signal flowing through their sigma, zi, or nu (indirect ≈ mean_mediated + distribution_mediated).

The distribution_mediated row is the answer to the second half of our question: if it is clearly non-zero, temperature reaches survival partly by changing the spread of size and the zeros of abundance, not only their means. A mean-only SEM would report that channel as zero or fold it silently into the mean; drmSEM keeps it visible and correctly labelled. (Effects are computed by Monte-Carlo g-computation, so honest numbers require the fitted engine — see the vignette for a worked run.)

More

Status

Early and experimental. The kernel logic — d-separation bookkeeping, the any-component LRT and Fisher’s C, and the simulation-based effect calculus — is validated by recovery tests that run without the engine. The full drmTMB-integration path (fitting nodes end to end) is validated in the cloud / CI environment where drmTMB is compiled and installed. APIs may change before a stable release.

Methodological background

drmSEM builds on a number of distinct literatures and is rigorous about crediting them at the point of use (see the @references blocks on each help page and the bibliography: of each vignette). The canonical entry points are:

  • piecewise SEM and local-likelihood d-separation — Shipley (2000, 2009,
    1. and Lefcheck (2016, piecewiseSEM);
  • d-separation as a graphical criterion — Pearl (2009);
  • counterfactual mediation — Pearl (2001), Robins & Greenland (1992), Imai et al. (2010), VanderWeele (2014, 2015), Vansteelandt & Daniel (2017);
  • distributional regression — Rigby & Stasinopoulos (2005, GAMLSS), Brooks et al. (2017, glmmTMB);
  • phylogenetic comparative methods — Felsenstein (1985), Martins & Hansen (1997), Pagel (1999), van der Bijl (2018, phylopath), Thorson & van der Bijl (2023, phylosem);
  • information criteria on Fisher’s C — Shipley (2013, CICc), Schwarz (1978, BIC), Burnham & Anderson (2002).

Three constructions inside the package are explicitly drmSEM contributions built on this foundation — the any-component d-separation test, the distribution-mediated effect row in indirect_effects(), and the BIC-style CBIC information criterion that is now the default in compare(). Each is flagged as such in its help page and the corresponding vignette, with the underlying components cited.

The bibliography lives in inst/REFERENCES.bib (and the JOSS-facing paper.bib); ?drmSEM lists the principal sources.

License

GPL (>= 3).