
Coefficient inference options in mars
mars_inference_options.RdSummary-time choices select inference for mean coefficients without changing
estimates, variance components or the original fit. Use matching options in
confint() to obtain the same ordinary or network coefficient intervals.
Details
Ordinary mars() fits support normal model inference and cluster-robust
inference with explicit covariance and df choices. Standard and modified HKSJ
are restricted to independent univariate ML/REML fits with estimated constant
heterogeneity and inverse-variance GLS weights. Model test='t' uses n-p df
without rescaling covariance. HKSJ rescales covariance by \(Q_{\mathrm{res}}/(n-p)\).
hksj_adjust='modified' uses \(\max(1,Q_{\mathrm{res}}/(n-p))\).
Model-based test='Satterthwaite' uses expected covariance-parameter
information, not the inverse joint observed Hessian. For ML this information
is \(I_{ab}=\mathrm{tr}(WD_a WD_b)/2\). For REML, replace \(W\) by
\(P=W-WX(X'WX)^{-1}X'W\). Here \(D_a\) denotes a derivative of the total
observation covariance. For contrast variance v and its covariance-parameter
gradient \(g\), \(\mathrm{df}=2v^2/(g'I^{-1}g)\). Covariance derivatives for ordinary
fits are evaluated numerically in the fitted covariance coordinates;
network and repeated-effects models use their analytic variance kernels.
Singular information and variance boundaries produce errors, not repaired
or substituted approximations. Scale models are excluded from these overrides.
Cluster-robust Satterthwaite is a different calculation using cluster
influence contributions and the selected sandwich adjustment. CR2 and
CR3 must be estimable; summary overrides do not silently fall back.
Use cluster only for independent units that contain all dependent rows.
A stored robust fit supplies its existing cluster and covariance type.
network_meta_coherent() and mars_repeated() summaries support model
normal/Satterthwaite and robust inference, but not HKSJ or residual-df model t.
Network overrides update only basic coefficients and total treatment
contrasts (including the fitted p-value adjustment). Direct/indirect tests,
diagnostic tests, rankings and predictions retain their fitted calculations.
masmr() stage summaries and hsmd() underlying models inherit ordinary
options; stored MASMR substantive joint tests and benchmark uncertainty
are not recalculated. Selection models, penalized models, forests and
downstream path/factor models do not support these coefficient overrides.
Summary options do not launch bootstrap, jackknife or permutation fitting.
Use mars_alt_estimation() or explicit replicate functions and summarize
their returned results. Jackknife omission ranges remain sensitivity
descriptions rather than sampling confidence intervals.
Model compatibility
The table concerns explicit summary-time overrides, not every inference method available when fitting a model. Yes means supported subject to the requirements below; it does not guarantee estimability for every dataset.
| Model/workflow | Model z | Model t | HKSJ | Model Satterthwaite | Robust tests |
| Independent univariate | Yes | Yes | Yes | Yes | Yes |
| Dependent univariate | Yes | No | No | If covariance is reconstructible | Yes |
| Multivariate/correlation synthesis | Yes | No | No | Yes | Yes |
| Multilevel | Yes | No | No | Yes | Yes |
| Location-scale | Yes | No | No | No | No summary override |
| Coherent network, including repeated networks | Yes | No | No | Yes | Yes |
| Repeated-effects (mars_repeated) | Yes | No | No | Common heterogeneity only | Yes |
| MASMR / HSMD | Inherit | Inherit | Inherit | Inherit | Inherit |
| Selection, LASSO, forest, downstream path/factor | No override | No | No | No | No override |
Model t requires one observation per study, a full-rank mean design and positive n-p residual df. HKSJ additionally requires ML/REML-estimated constant heterogeneity and matched inverse-variance GLS covariance; it excludes common-effect and user-fixed tau2 fits. A zero estimated variance does not itself exclude HKSJ.
Model-based Satterthwaite requires estimated, interior, identifiable heterogeneity and conditional GLS covariance. Common-effect/none and user-fixed heterogeneity, active variance boundaries, nonconvergence and singular covariance information are excluded. Ordinary fits must retain their sampling covariance, outcome indices, design and covariance parameters. Correlation synthesis means inference on fitted pooled coefficients, not automatic Satterthwaite tests for a subsequent path or factor model.
Robust overrides require retained aligned design, residuals and covariance, independent clusters, an estimable sandwich adjustment and positive df. These conditions also apply to common-effect or user-fixed-heterogeneity fits: retaining a model class alone is not enough. An explicit cluster cannot split observations that are dependent in the working covariance. Location-scale fits retain their usual summaries, including stored robust results where available; selecting a new robust summary override is unsupported. MASMR and HSMD check each underlying fitted model separately.
Argument combinations
- inference
Use model for model covariance or robust for sandwich inference. Ordinary summary defaults retain the existing tables; explicit selections return one selected coefficient table.
- test
Model z is normal Wald inference. Model t is restricted to eligible independent univariate fits. Model Satterthwaite uses expected covariance-parameter information; robust Satterthwaite uses cluster influences. Robust z, naive-t (clusters minus one df) and naive-tp (clusters minus coefficients df) are available; robust t aliases naive-t. HKSJ/knha is model-only and follows the compatibility table.
- vcov_method
Ordinary summary/confint model-covariance selection: gls or observed. Observed requires a valid interior ML fit. HKSJ and model Satterthwaite require gls. Do not combine vcov_method with robust inference. Network and repeated-effects summaries do not accept it.
- robust_type and cluster
For robust inference only. Covariance types are CR0, CR1, CR1p, CR1S, CR2 and CR3. An ordinary fit uses its stored robust type/cluster when available; otherwise supply cluster, and the covariance defaults to CR2. Networks and repeated-effects models default to study clustering. A network fit already using CR1 retains CR1 and naive-t unless explicitly overridden.
- hksj_adjust
Only with model HKSJ for eligible independent univariate fits: standard or modified. Unsupported on network and repeated-effects models; it is not a robust covariance adjustment.
- level
Pointwise coefficient confidence level. Network summaries also update total treatment-contrast confidence intervals. This does not set prediction coverage, simultaneous coverage or resampling coverage.
References
Hedges, L. V., Tipton, E., and Johnson, M. C. (2010). Robust variance estimation in meta-regression with dependent effect size estimates. Research Synthesis Methods, 1(1), 39–65. doi:10.1002/jrsm.5 .
Tipton, E. (2015). Small sample adjustments for robust variance estimation with meta-regression. Psychological Methods, 20(3), 375–393. doi:10.1037/met0000011 .
Pustejovsky, J. E., and Tipton, E. (2018). Small-sample methods for cluster-robust variance estimation and hypothesis testing in fixed effects models. Journal of Business & Economic Statistics, 36(4), 672–683. doi:10.1080/07350015.2016.1247004 .
Sidik, K., and Jonkman, J. N. (2002). A simple confidence interval for meta-analysis. Statistics in Medicine, 21(21), 3153–3159. doi:10.1002/sim.1262 .
Knapp, G., and Hartung, J. (2003). Improved tests for a random effects meta-regression with a single covariate. Statistics in Medicine, 22(17), 2693–2710. doi:10.1002/sim.1482 .
Röver, C., Knapp, G., and Friede, T. (2015). Hartung-Knapp-Sidik-Jonkman approach and its modification for random-effects meta-analysis with few studies. BMC Medical Research Methodology, 15, 99. doi:10.1186/s12874-015-0091-1 .
Examples
fit <- mars(school, formula=effect~1, studyID="study", variance="var",
varcov_type="univariate", structure="univariate")
summary(fit, inference="model", test="hksj", hksj_adjust="modified")
#> Results generated with MARS:v 0.6.0.8
#> Friday, November 20, 2026
#>
#> Model Type:
#> univariate
#>
#> Estimation Method:
#> Restricted Maximum Likelihood
#>
#> Coefficient covariance: gls
#> Summary inference: model; covariance: gls; test: hksj; level: 0.95
#> Hartung-Knapp adjustment: modified; variance multiplier: 1
#> Coefficient/contrast inference only; heterogeneity, scale, prediction and model-fit tests are unchanged.
#> Optimizer convergence code: 0
#>
#>
#> Model Formula:
#> effect ~ 1
#> <environment: 0x5c66ab5ccc20>
#>
#> Data Summary:
#> Number of Effect Sizes: 56
#> Number of Fixed Effects: 1
#> Number of Random Effects: 1
#>
#> Random Components:
#> term var SD
#> intercept 0.08844 0.2974
#>
#> Fixed Effects Estimates:
#> attribute estimate SE t_test df p_value lower upper
#> (Intercept) 0.1279 0.04387 2.916 55 0.00512 0.04001 0.2158
#>
#> Model Fit Statistics:
#> logLik Dev AIC BIC AICc
#> -16.85 33.69 37.69 41.71 38.89
#>
#> Q Error: 578.864 (55), p < 0.0001
#>
#> I2 (General): 94.70491
#>
#> Residual Diagnostics:
#> Diagnostic Value
#> Observations 56
#> Finite raw residuals 56
#> Mean raw residual -0.007932
#> SD of raw residuals 0.3258
#> Root mean squared residual 0.323
#> Mean absolute residual 0.2463
#> Sum of squared Pearson residuals 53.68
#> Mean |studentized residual| 0.7463
#> Maximum |studentized residual| 3.419
#> Proportion |studentized residual| > 2 0.03571
#> Proportion |studentized residual| > 3 0.01786
#>
#> Normality screen (Shapiro-Wilk, whitened residuals):
#> N W p-value
#> 56 0.9094 0.0004762
summary(fit, inference="model", test="Satterthwaite")
#> Results generated with MARS:v 0.6.0.8
#> Friday, November 20, 2026
#>
#> Model Type:
#> univariate
#>
#> Estimation Method:
#> Restricted Maximum Likelihood
#>
#> Coefficient covariance: gls
#> Summary inference: model; covariance: gls; test: Satterthwaite; level: 0.95
#> Expected-information delta-method Satterthwaite
#> Coefficient/contrast inference only; heterogeneity, scale, prediction and model-fit tests are unchanged.
#> Optimizer convergence code: 0
#>
#>
#> Model Formula:
#> effect ~ 1
#> <environment: 0x5c66ab5ccc20>
#>
#> Data Summary:
#> Number of Effect Sizes: 56
#> Number of Fixed Effects: 1
#> Number of Random Effects: 1
#>
#> Random Components:
#> term var SD
#> intercept 0.08844 0.2974
#>
#> Fixed Effects Estimates:
#> attribute estimate SE t_test df p_value lower upper
#> (Intercept) 0.1279 0.04387 2.916 52.81 0.005191 0.03993 0.2159
#>
#> Model Fit Statistics:
#> logLik Dev AIC BIC AICc
#> -16.85 33.69 37.69 41.71 38.89
#>
#> Q Error: 578.864 (55), p < 0.0001
#>
#> I2 (General): 94.70491
#>
#> Residual Diagnostics:
#> Diagnostic Value
#> Observations 56
#> Finite raw residuals 56
#> Mean raw residual -0.007932
#> SD of raw residuals 0.3258
#> Root mean squared residual 0.323
#> Mean absolute residual 0.2463
#> Sum of squared Pearson residuals 53.68
#> Mean |studentized residual| 0.7463
#> Maximum |studentized residual| 3.419
#> Proportion |studentized residual| > 2 0.03571
#> Proportion |studentized residual| > 3 0.01786
#>
#> Normality screen (Shapiro-Wilk, whitened residuals):
#> N W p-value
#> 56 0.9094 0.0004762
summary(fit, inference="robust", cluster="district", robust_type="CR2",
test="Satterthwaite")
#> Results generated with MARS:v 0.6.0.8
#> Friday, November 20, 2026
#>
#> Model Type:
#> univariate
#>
#> Estimation Method:
#> Restricted Maximum Likelihood
#>
#> Coefficient covariance: gls
#> Summary inference: robust; covariance: CR2; test: Satterthwaite; level: 0.95
#> Coefficient/contrast inference only; heterogeneity, scale, prediction and model-fit tests are unchanged.
#> Optimizer convergence code: 0
#>
#>
#> Model Formula:
#> effect ~ 1
#> <environment: 0x5c66ab5ccc20>
#>
#> Data Summary:
#> Number of Effect Sizes: 56
#> Number of Fixed Effects: 1
#> Number of Random Effects: 1
#>
#> Random Components:
#> term var SD
#> intercept 0.08844 0.2974
#>
#> Fixed Effects Estimates:
#> attribute estimate SE t_test df p_value lower upper
#> (Intercept) 0.1279 0.0782 1.636 7.373 0.1437 -0.05509 0.311
#>
#> Model Fit Statistics:
#> logLik Dev AIC BIC AICc
#> -16.85 33.69 37.69 41.71 38.89
#>
#> Q Error: 578.864 (55), p < 0.0001
#>
#> I2 (General): 94.70491
#>
#> Residual Diagnostics:
#> Diagnostic Value
#> Observations 56
#> Finite raw residuals 56
#> Mean raw residual -0.007932
#> SD of raw residuals 0.3258
#> Root mean squared residual 0.323
#> Mean absolute residual 0.2463
#> Sum of squared Pearson residuals 53.68
#> Mean |studentized residual| 0.7463
#> Maximum |studentized residual| 3.419
#> Proportion |studentized residual| > 2 0.03571
#> Proportion |studentized residual| > 3 0.01786
#>
#> Normality screen (Shapiro-Wilk, whitened residuals):
#> N W p-value
#> 56 0.9094 0.0004762
confint(fit, inference="model", test="hksj", hksj_adjust="modified")
#> 2.5% 97.5%
#> (Intercept) 0.04001423 0.2158499