Mapping R workflows to PyMetaAnalysis

This guide helps users translate conventional R meta and metafor workflows into PyMetaAnalysis calls. It is a terminology and configuration map, not a claim that similarly named functions always use identical formulas or defaults.

Always compare the resolved configuration and numerical output when porting an analysis. See validation for the R versions, calls, and fixtures used by this project.

Entry points

Analysis PyMetaAnalysis R metafor R meta
Generic effects and variances meta_analysis() rma.uni(yi, vi, ...) metagen(TE, seTE, ...)
Binary 2x2 tables, inverse variance meta_binary(..., method="IV") escalc() then rma.uni() metabin(..., method="Inverse")
Binary 2x2 tables, Mantel-Haenszel meta_binary(..., method="MH") rma.mh() metabin(..., method="MH")
Binary 2x2 tables, Peto OR meta_binary(..., measure="OR", method="Peto") rma.peto() metabin(..., sm="OR", method="Peto")
Continuous group summaries meta_continuous() escalc() then rma.uni() metacont()
Correlations and sample sizes meta_correlation() escalc(measure="ZCOR") then rma.uni() metacor(sm="ZCOR")
Subgroups subgroup= on a high-level call separate fits or a moderator model subgroup=
Leave-one-out result.leave_one_out() leave1out() for supported fits metainf()
Meta-regression influence regression.influence() influence(), rstudent(), cooks.distance(), dfbetas() —
Meta-regression collinearity regression.collinearity() vif() plus weighted design diagnostics —
Meta-regression linear contrasts regression.contrast(...) anova(..., X=..., rhs=...) —
Cumulative analysis result.cumulative() cumul() metacum()
Contour-enhanced funnel plot result.funnel(contour_levels=(...)) funnel(level=c(...), refline=...) funnel(contour.levels=c(...), ref=...)
Trim-and-fill result.trim_and_fill() trimfill() trimfill()
Begg-Mazumdar rank test result.begg_test() ranktest() metabias(..., method.bias="Begg") (definition may differ)
Classical Egger test result.egger_test() regtest(..., model="lm", predictor="sei") metabias(..., method.bias="Egger")
Harbord binary-OR test result.harbord_test() manual documented score regression metabias(..., method.bias="Harbord")
Peters binary-OR test result.peters_test() manual documented WLS metabias(..., method.bias="Peters")

begg_test(exact=True) matches metafor::ranktest()'s exact default for tie-free data. PyMetaAnalysis's exact=None automatic policy switches to the asymptotic calculation at 50 studies to avoid an unexpectedly expensive exact request; pass exact=True explicitly when that R-compatible behavior is required. Its optional continuity correction is independently checked against stats::cor.test(..., method="kendall", exact=FALSE, continuity=TRUE) on the same standardized response.

trim_and_fill() follows metafor::trimfill() for L0/R0 rank estimators, automatic side selection, source-model refitting, and mirrored variances. Its result deliberately labels imputed rows as synthetic, exposes the selection slope and iteration trace, and requires at least three included studies.

PyMetaAnalysis intentionally has no metabin, metacont, or rma aliases. One documented Python entry point per input shape keeps result types and provenance behavior consistent.

Input names

Meaning PyMetaAnalysis R metafor R meta
Study effect effect yi TE
Sampling variance variance vi square of seTE
Study label study or DataFrame index slab studlab
Treatment events event_treat ai event.e
Treatment total n_treat n1i n.e
Control events event_control ci event.c
Control total n_control n2i n.c
Treatment mean/SD mean_treat, sd_treat m1i, sd1i mean.e, sd.e
Control mean/SD mean_control, sd_control m2i, sd2i mean.c, sd.c
Correlation/sample size correlation, n ri, ni cor, n

PyMetaAnalysis accepts DataFrame column names or aligned one-dimensional array-like values. When study= is omitted for a DataFrame, its index supplies the display labels.

Measures and scales

PyMetaAnalysis measure Meaning metafor measure meta sm Model scale
"OR" Odds ratio "OR" "OR" log ratio
"RR" Risk ratio "RR" "RR" log ratio
"RD" Risk difference "RD" "RD" identity
"MD" Mean difference "MD" "MD" identity
"SMD" Exact-corrected Hedges' g "SMD" with the documented correction "SMD" with exact Hedges correction identity
"ZCOR" Fisher's z-transformed correlation "ZCOR" "ZCOR" Fisher's z

For OR and RR, estimate and ci remain on the log model scale; display_estimate and display_ci provide exponentiated ratios. For ZCOR, the model attributes remain Fisher's z values and the display attributes apply tanh to return correlations. This is similar to choosing transformed or untransformed printing in R, but both scales remain explicit attributes in Python.

Models, pooling, and heterogeneity

PyMetaAnalysis metafor analogue meta analogue Notes
model="common" rma.uni(..., method="EE") common=TRUE, random=FALSE Inverse-variance common-effect fit
model="random" random-effects rma.uni() random=TRUE Requires a tau-squared policy
method="IV" inverse-variance weighting method="Inverse" Binary API only; generic, continuous, and correlation fits are IV
method="MH" rma.mh() method="MH" Common-effect OR/RR/RD
method="Peto" rma.peto() method="Peto" Common-effect OR; raw pooling tables and O-minus-E heterogeneity
tau2_method=None (resolved as "REML") method="REML" method.tau="REML" PyMetaAnalysis random-effects default
tau2_method="PM" method="PM" method.tau="PM" Paule-Mandel
tau2_method="DL" method="DL" method.tau="DL" DerSimonian-Laird

The same label does not guarantee identical optimizer tolerances, boundary handling, or heterogeneity definitions. In particular, PyMetaAnalysis records i2_method; random-effects inverse-variance results use the documented tau-squared/typical-variance definition, while common-effect and MH results use the Q-based definition. Peto also reports Q-based inconsistency, but its Q is formed from the Peto observed-minus-expected contributions rather than ordinary study-log-OR inverse-variance residuals.

result.tau2_confidence_interval() corresponds to confint(fit, type="QP") for an rma.uni random-effects fit. Both return Q-profile bounds for tau-squared and monotonic tau, I-squared, and H-squared transformations. PyMetaAnalysis additionally exposes is_empty so a formal empty confidence set is distinguishable from its constrained [0, 0] display.

Confidence and prediction intervals

PyMetaAnalysis ci_method metafor R meta Behavior
"normal" default test="z" method.random.ci="classic" Normal mean interval
"hartung_knapp" test="knha" method.random.ci="HK" Unmodified HK variance and t quantile
"hartung_knapp_adhoc" test="adhoc" HK plus an explicitly selected ad hoc correction HK variance cannot fall below the classic variance

Eligible random-effects fits use the documented HTS prediction interval under normal inference. Either Hartung-Knapp choice instead uses its selected pooled-mean variance with k-2 degrees of freedom (HK-PR), matching predict(fit, predtype="Riley") in metafor. R packages offer additional prediction-interval choices, so method settings must still be compared explicitly.

Meta-regression uses a separate prediction rule. Its default corresponds to predict(fit) in metafor: normal inference uses a normal critical value and Hartung-Knapp inference uses t_(k-p). PyMetaAnalysis prediction_interval_method="riley" corresponds to predict(fit, predtype="Riley"), using t_(k-p-1) for the true-effect prediction interval while leaving the mean-effect confidence interval unchanged.

Sparse binary studies

The closest names are:

PyMetaAnalysis R meta Meaning
continuity_correction incr Increment used for corrected study effects
correction_scope="only_zero_studies" method.incr="only0" Correct only studies containing a zero cell
correction_scope="if_any_zero" method.incr="if0all" Correct every study when any study contains a zero cell
correction_scope="all_studies" method.incr="all" Correct every study
correction_scope="none" no increment Disable study-effect correction
mh_continuity_correction=None MH.exact=TRUE in intent Avoid correction for exact MH pooling where defined

These are conceptual mappings, not interchangeable switches. R meta also supports dataset-wide correction scopes and methods that PyMetaAnalysis does not implement. For OR/RR, double-zero and double-all studies are excluded from all model and heterogeneity calculations by default while remaining visible in the result table. RD uses its separate rd_zero_variance policy.

For measure="RD", PyMetaAnalysis and metafor::rma.mh(measure="RD") use the Sato-Greenland-Robins sampling variance. A separate explicit mh_continuity_correction changes the MH tables; the study-effect continuity_correction still affects only displayed study variances and heterogeneity.

For Peto, PyMetaAnalysis's default study-level correction corresponds to escalc(measure="PETO", add=0.5, to="only0", drop00=TRUE) for the displayed study rows. Its pooled fit corresponds to rma.peto() with raw pooling tables; the study correction never changes the pooled result. Compare Peto's rare-outcome, within-study arm-balance, and modest-effect assumptions before porting it solely because a table contains zeros.

Read zero-event studies before translating sparse analyses.

Worked generic translation

Python:

import pandas as pd
import meta_analyze as ma

studies = pd.DataFrame(
    {"yi": [0.12, 0.35, -0.08, 0.21], "vi": [0.04, 0.06, 0.03, 0.05]}
)
studies["sei"] = studies["vi"] ** 0.5
result = ma.meta_analysis(
    data=studies,
    effect="yi",
    variance="vi",
    model="random",
    tau2_method="REML",
    ci_method="hartung_knapp_adhoc",
)

The corresponding metafor configuration is conceptually:

rma.uni(
  yi = yi,
  vi = vi,
  method = "REML",
  test = "adhoc",
  data = studies
)

The corresponding R meta configuration starts from standard errors rather than variances. PyMetaAnalysis can accept that uncertainty column directly as standard_error=; no manual squaring step is required:

result = ma.meta_analysis(
    studies,
    effect="yi",
    standard_error="sei",
    model="random",
    tau2_method="REML",
    ci_method="hartung_knapp_adhoc",
)

metagen(
  TE = yi,
  seTE = sqrt(vi),
  common = FALSE,
  random = TRUE,
  method.tau = "REML",
  method.random.ci = "HK",
  data = studies
)

Check the R package's explicit ad hoc HK option before treating the last call as numerically equivalent.

Classical Egger regression

egger = result.egger_test()

corresponds to the classical metafor configuration:

regtest(fit, model = "lm", predictor = "sei")

PyMetaAnalysis does not currently expose metafor's default model="rma" version. The returned intercept is the tested asymmetry coefficient, while limit_estimate is the extrapolated effect as the standard error tends to zero. Review the dedicated small-study-effects guide before treating similarly named R functions as numerically interchangeable.

Harbord regression for binary odds ratios

binary_or_result = ma.meta_binary(
    event_treat=[1, 2, 4, 3, 5],
    n_treat=[80, 90, 100, 110, 120],
    event_control=[3, 5, 2, 4, 7],
    n_control=[85, 100, 110, 100, 125],
    measure="OR",
)
harbord = binary_or_result.harbord_test()

corresponds to:

metabias(binary_or_fit, method.bias = "Harbord")

PyMetaAnalysis calculates the null efficient score Z and score variance V directly from retained treatment/control counts. It fits the standardized form Z/sqrt(V) on sqrt(V), which is algebraically equivalent to the V-weighted Z/V on 1/sqrt(V) regression used by R meta. The intercept is the tested asymmetry coefficient, with multiplicative dispersion and t_(k-2) inference. Study-level continuity corrections and the source pooling method do not enter the score calculation.

Peters regression for binary odds ratios

peters = binary_or_result.peters_test()

corresponds to:

metabias(binary_or_fit, method.bias = "Peters")

PyMetaAnalysis requires an OR result created by meta_binary() because Peters regression needs the original treatment/control counts. It reconstructs conventional continuity-corrected study log odds ratios, uses inverse total sample size as predictor and S*F/N weights, and applies multiplicative dispersion with t_(k-2) inference. slope is the tested coefficient; limit_estimate is the extrapolated log OR at infinite total sample size. The source common/random and MH/IV/Peto pooling choice does not enter this separate regression.

Primary R references