Frequently asked questions

Short answers that route to the page carrying the full story. If your question is not here, the worked analyses and the User’s Guide are the two front doors.

Getting started

What data can I give bgm()?

Binary, ordinal, and continuous variables, in a matrix or a data frame; variable_type says which columns are which, and the default reads every column as ordinal. Binary items are ordinal items with two categories, so they need no special treatment. More: bgm() and Getting Started.

Why does the documentation run bgm() so much longer than the default?

The defaults (2,000 warmup and 2,000 sampling iterations) are tuned for exploration. What you read off a fit is evidence, and the run has to be long enough to pin each inclusion probability down to the precision the evidence claim needs, which is why the worked analyses run iter = 1e4, warmup = 5e3. More: run-length settings and Reading the output.

Why does bgm() take minutes when EBICglasso takes seconds?

Because it computes a posterior over network structures and weights rather than one penalized point estimate. The credible intervals, the evidence per edge, and the undecided category all come out of those draws, with no bootstrap loop to run afterwards; the comparison page prices the trade honestly. More: Coming from bootnet and qgraph.

Reading the results

What does an inclusion Bayes factor of 10 mean?

The data have multiplied the prior odds in favor of the edge by ten. It is a ratio, so it runs in both directions: 1/10 is tenfold evidence against the edge, and values near one mean the data moved nothing either way. Ten is a reporting convention, not a law of nature; state the threshold you use. More: Edge Selection.

Can I say an edge is absent?

Yes, and that is much of the point: an inclusion Bayes factor at or below 1/10 is evidence of conditional independence, a positive finding you can report. What you cannot do is read absence off a blank in an estimated network, which cannot distinguish “ruled out” from “not enough data”. More: From estimation to evidence.

The posterior inclusion probability is high. Is that the evidence?

Not by itself. The inclusion probability mixes what the data showed with where the prior started, so under any edge prior other than 0.5 the posterior odds understate or overstate the evidence. The inclusion Bayes factor divides the prior odds out, which is why the verdicts are read from it. More: Edge Selection.

I reran the analysis and the numbers changed slightly. Is something wrong?

No: the estimates are Monte Carlo averages, so two unseeded runs differ by sampling error, and summary(fit) reports each estimate’s mcse. Set seed for a repeatable run, and read the diagnostics page for what a seed does and does not guarantee. A difference large enough to change what you would report is a mixing problem, not a seeding one. More: Results differ between runs.

Does a thicker edge mean stronger evidence?

No. Width encodes the estimated weight, and effect size and evidence are different quantities: a moderate weight can be estimated with overwhelming evidence, and a large one with next to none. That is why the edge evidence plot weights only its presence panel and draws the other two by classification alone. More: the edge evidence plot.

verdicts() flagged some verdicts as fragile. Is the analysis broken?

No: fragile means the verdict sits within two Monte Carlo standard errors of an evidence threshold, so a rerun or a longer run could change it. It is a statement about run length, not about the data, and the remedy is more iterations. More: verdicts().

Priors

Do the default priors matter?

They are defensible defaults, not neutral ones: the slab is the alternative hypothesis the inclusion Bayes factor tests against, so its scale has real leverage on the verdicts. Understand what the defaults assume, and state them when you report. More: Prior Basics.

A reviewer asks whether my conclusions depend on the prior. What do I answer?

Run prior_sensitivity_check(fit) and report what it prints: how many verdicts hold across the examined range of prior scales, which edges genuinely move, and the scale the data prefer next to the one you chose. One page walks the whole report, including the two-sentence manuscript template. More: Check your priors.

Practicalities

What happens to my missing data?

By default, rows with any missing value are dropped (na_action = "listwise"), and a message reports how many. Setting na_action = "impute" instead imputes the missing entries inside the MCMC, drawing each from its full conditional distribution at every iteration. More: Missing Data.

How do I cite bgms?

citation("bgms") prints the reference for your installed version, and the homepage has a citation picker with APA, Chicago, BibTeX, and RIS formats. Cite the version you actually used.

Where do I ask a question or report a bug?

On the bgms issue tracker. A report that can be acted on carries the bgm() call as you ran it, fit$nuts_diag$summary, and the output of sessionInfo(). More: Still stuck.