Title
umeta - U-statistics-based random-effects meta-analyses
Description
The umeta command performs U-statistics-based random-effects meta-analysis on a dataset of univariate, bivariate or trivariate point estimates, sampling variances, and for bivariate or trivariate data, within-study correlations or covariances. The methodology is described in Ma and Mazumdar (2011).
For each outcome, umeta calculates the overall effect and a confidence interval for the effect. The command also displays the between-study variance (or alternatively between-study standard deviation), between-study correlation(s) for bivariate or trivariate data and inconsistency (I-squared) statistics.
umeta Syntax
umeta yvar* svar* [wsvar*] [if] [in] [, covvar(string) level(#) predint tscale(logit|log|asin) noestimates bssd zci i2]
where the data are arranged with one line per study: the point estimates are held in variables yvar*, the sampling variances are held in svar*, and within-study correlations (or covariances) for 2 or 3 outcomes are held in variable wsvar*.
For univariate data, yvar* is yvar and svar* is svar
For bivariate data, yvar* is yvar1 yvar2, svar* is svar1 svar2 and wsvar* is wsvar12
For trivariate data, yvar* represents yvar1 yvar2 yvar3, svar* is svar1 svar2 svar3 and wsvar* is wsvar12, wsvar13 wsvar23
For any unreported outcomes, umeta sets the outcome and its variance at 0 and 1E12, respectively.
Options for umeta
covvar(string) For bivariate or trivariate data analysis, you must specify covvar(rho) or covvar(cov) depending on whether you are using within-study correlation(s) or covariance(s).
level(#) specifies the significance level for probability intervals.
predint displays outcome-specific mean estimates with the probability interval of the approximate predictive distribution of a future trial, based on the extent of heterogeneity. No method has been developed as yet for multivariate predictive distribution.
tscale(logit|log|asin) transformation of estimates to original scale, if data was transformed prior to analysis.
bssd reports the between-study standard deviations with confidence intervals (calculated as a function of inconsistency statistic and typical within-study variance as by White(2009)) instead of the default between-study variances.
noestimates prevents display of mean estimates, between-study variances (or standard deviations) and correlation(s)
zci uses z-statistics instead of default t-statistics for confidence interval calculation. This is overriden if option predint specified.
i2 reports I-squared statistic for each outcome, together with confidence intervals as is described in White(2009).
umeta, typed without specifying varlist, redisplays the latest estimation results. All the output options listed above may be used
by...: or statsby...: may be used with umeta to perform subgroup analyses; see help by or statsby.
Remarks
Multivariate meta-analysis is used to synthesize multiple outcomes simultaneously taking into account the correlation between the outcomes (Riley(2009)). Likelihood based approaches, in particular, Restricted Maximum Likelihood (REML) method is commonly utilized in this context. REML assumes a multivariate normal distribution for the random-effects model. This assumption is difficult to verify, especially for meta-analysis with small number of component studies. Use of REML also requires iterative estimation between parameters, needing moderately high computation time, especially when the dimension of outcomes is large (White(2009)). Jackson, White and Thompson(2010) have developed a multivariate method of moments (MMM) which has been shown to perform equally well to REML.
Ma and Mazumdar recently proposed a new method for multivariate meta-analysis based on the theory of U-statistic. The motivation for using U-statistic stems from the fact that it provides a a robust, nonparametric and noniterative approach. Additionally, the asymptotic behavior of the related statistics and their estimates are easy to derive being based on theorems already available for U-statistics.
Since the between-study variance matrix for the random-effects meta-analysis model involves second order moments, U-statistic formulation is especially beneficial. It is easily applied to estimate the variance matrix components and to develop their joint asymptotic distribution for related inference. Because the U-statistic-based method does not depend on parametric distributional assumptions for both random effects and sampling errors, it provides robust inference irrespective of the data distribution
For a detailed description of the u-statistic methodology, see Ma and Mazumdar (2011).
By convention, the within-study variances are assumed known and replaced by their sample estimates. Thus imprecision in within-study variance estimates may affect the estimation of pooled effect size especially when the size of within-study variation is relatively large.
This program does not assume that variables need log, logit or arcsin or other transformation(s). However, if study-level outcome data are available as odds ratios, risk ratios or proportions, the user may choose to log-, logit-or arcsin-transform them first. Then tscale option may be used to change back to the original scale for reporting if so desired.
The probability interval of the approximate predictive distribution of a future trial, is based on the extent of heterogeneity. This incorporates uncertainty in the location and spread of the random effects distribution using the formula t(df) x sqrt(se2 + tau2) where t is the t-distribution with n-2 degrees of freedom, se2 is the squared standard error and tau2 the heterogeneity statistic and n is the number of observations(studies). This is applied to each outcome separately. For further information see Higgins, Thompson and Spiegelhalter(2009)
I-squared formulated by Higgins and Thompson (2002), describes the percentage of total variation across studies that is attributable to heterogeneity rather than chance and measures impact of heterogeneity. . Negative values of I-squared are made equal to 0 so that I-squared lies between 0% and 100%. A value of 0% indicates no observed heterogeneity, and values greater than 50% may be considered substantial heterogeneity. The main advantage of I-squared is that it does not inherently depend on the number of the studies in the meta-analysis
Examples
Example 1: Univariate Data
. use umeta_example1, clear
. list yvar svar, clean noobs . umeta yvar svar
Example 2: Bivariate logit-transformed Data, No within-study correlation
. use umeta_example2, clear
. list yvar* svar* rho*, clean noobs
. umeta yvar* svar* rho*, p
. umeta yvar* svar* rho*, z bssd p tscale(logit)
Example 3: Bivariate Outcomes with missing Data
. use umeta_example3, clear
. list yvar* svar* rho*, clean noobs
. umeta yvar* svar* rho*
. umeta yvar* svar* rho*, pred
. umeta, noest i2 z q
Example 4: Trivariate Outcomes with Zero within-study covariance matrix
. use umeta_example4, clear
. list yvar* svar* rho*, clean noobs
. umeta yvar* svar* rho*
. umeta, noest i2 z q
Example 5: Trivariate Outcomes with within-study correlations
. use umeta_example5, clear
. list yvar* svar* rho*, clean noobs
. umeta yvar* svar* rho*, pred
umeta saves the following in e():
Scalars e(N) number of observations e(dims) number of outcomes for meta-analysis e(df_r) degrees of freedom for meta-analysis estimation e(Qdf) degrees of freedom for homogeneity testing
Macros e(cmd) umeta e(cmdline) command as typed e(properties) b V e(yvars) names of study-specific outcome variables (point estimates) e(svars) names of study-specific sampling variances e(predict) program used to implement predict
Matrices e(b) coefficient vector e(V) variance-covariance matrix of the estimators e(Isqmat) matrix of outcome-specific I^2 values e(Qmat) matrix of outcome-specific heterogeneity statistic e(Vtyp) typical within-study variance e(Sigma) between-study variance-covariance matrix e(svars) matrix of study-specific sampling variances e(rho) matrix of between-study correlation e(yvars) matrix of study-specific point estimates
Functions e(sample) marks estimation sample
Authors Ben A. Dwamena, Department of Radiology, Division of Nuclear Medicine, University of Michigan Medical School, Ann Arbor, Michigan
Yan Ma, Hospital for Special Surgery, Weill Medical College of Cornell University, New York, New York
programming problems: bdwamena@umich.edu.
u-statistic-based questions: yam2007@med.cornell.edu.
-------------------------------------------------------------------------------
Title
umeta postestimation -- Postestimation tools for umeta
Description
umeta is programmed as an Stata estimation command and so supports many of the commands listed under help estcom and postest. The following standard postestimation commands may be particularly useful:
Command Description ------------------------------------------------------------------------- estat VCE and estimation sample summary. See help estat estimates Cataloging estimation results. See help estimates lincom Point estimates, standard errors, testing, and inference for linear combinations of coefficients. See lincom nlcom Point estimates, standard errors, testing, and inference for nonlinear combinations of coefficients. See nlcom predict predictions, residuals, influence statistics, and other diagnostic measures test Wald tests of linear hypotheses. See help test testnl Wald tests of non-linear hypotheses. See help testnl -------------------------------------------------------------------------
predict Syntax
The syntax of predict following umeta is
syntax 1:
predict [type] newvarname [if exp] [in range] [, statistic ]
syntax 2:
predict newvarname [if exp] [in range] [, statistic show(string)]
------------------------------------------------------------------------- statistic Description ------------------------------------------------------------------------- fixed prediction of fixed-effects; the default stfixed standard error of the fixed-effects prediction fitted prediction including random effects stfit standard error of fitted stdf standard error of the forecast reffects predicted random effects reses standard error of predicted random effects rstandard standardized predicted random effects lev leverage (diagonal elements of projection matrix) cooksd Cook's influence measure -------------------------------------------------------------------------
These statistics are available both in and out of sample; type "predict ... if e(sample) ..." if wanted only for the estimation sample.
------------------------------------------------------------------------- show Description ------------------------------------------------------------------------- clean force table format with no divider or separator lines table force table format abbreviate(#) abbreviate variable names to # characters; default is ab(8) noobs do not list observation numbers divider draw divider lines between columns separator(#) draw a separator line every # lines; default is separator(5) -------------------------------------------------------------------------
Options for predict
fixed calculates the linear prediction for the fixed portion of the model.
stfixed calculates the outcome-specific standard error of the fixed-portion linear prediction
stfitted calculates the outcome-specific standard error of the prediction including random effects.
fitted calculates the outcome-specific prediction including random effects, Xb[i] + u[i], also known as the empirical Bayes estimates of the effects in each study.
stdf calculates the outcome-specific standard error of the forecast. This gives the standard deviation of the predicted distribution of the true value of depvar in a future study stdf^2 = stdp^2 + tau2.
reffects calculates the outcome-specific best linear unbiased predictions (BLUPs) of the random effects, also known as the posterior mean or empirical Bayes estimates of the random effects, or as shrunken residuals.
reses calculates the outcome-specific standard error of predicted random effects.
rstandard calculates the outcome-specific standardized predicted random effects, i.e. the predicted random effects u[i] divided by their (unconditional) standard errors. These may be useful for diagnostics and model checking.
lev calculates the study-specific leverages
cooksd calculates the study-specific Cook's influence statistic.
Remarks
Similar to other types of data, it is not uncommon to observe extreme effect size values when conducting a meta-analysis. As the main objective of a meta-analysis is to provide a reasonable summary of the effect sizes of a body of empirical studies, the presence of such outliers may distort the conclusions of a meta-analysis. Moreover, if the conclusions of a meta-analysis hinge on the data of only one or two influential studies, then the robustness of the conclusions are called into question. Researchers, therefore, generally agree that the effect sizes should be examined for potential outliers and influential cases when conducting a meta-analysis.
The most thorough treatment of outlier diagnostics in the context of meta-analysis to date can be found in the classic book by Hedges and Olkin, who devoted a whole chapter to diagnostic procedures for effect size data. However, the methods developed by Hedges and Olkin(1985) are only applicable to fixed-effects models. Given that random- and mixed-effects models are gaining popularity in the meta-analytic context, corresponding methods for outlier and influential case diagnostics need to be developed.
Viechtbauer and Cheung(2010) have introduced several outlier and influence diagnostic procedures for the random- and mixed-effects model in meta-analysis. These procedures are logical extensions of the standard outlier and case-deletion influence diagnostics for regular regression models as in Demidenko and Stukel(2005) and take both sampling variability and between-study heterogeneity into account. The proposed measures provide a simple framework for evaluating the potential impact of outliers or influential cases in meta-analysis.
Examples
. use umeta_example5, clear
. umeta yvar* svar* rho*
. predict lev, lev show(clean)
. predict cook, cooksd show(clean)
. predict fit, fit
. predict fix
. predict reff, reff show(clean noobs)
. predict res, res
. predict rst, rst
. predict stpred, stfit
. predict double stdf, stdf
Author Ben A. Dwamena, Department of Radiology, Division of Nuclear Medicine, University of Michigan Medical School, Ann Arbor, Michigan. bdwamena@umich.edu.
References
Demidenko, E., T. A. Stukel. 2005 Influence analysis for linear mixed-effects models Statistics in Medicine 24: 893–909
DerSimonian, R., and N. Laird. 1986. Meta-analysis in clinical trials. Controlled Clinical Trials 7: 177-188.
Hedges LV, I. Olkin. 1985. Statistical Methods for Meta-Analysis Academic Press: New York.
Higgins, J. P. T., and S. G. Thompson. 2002. Quantifying heterogeneity in a meta-analysis. Statistics in Medicine 21: 1539-1558.
Higgins, J. P. T., S. G. Thompson, and D. J. Spiegelhalter. 2009. A re-evaluation of random-effects meta-analysis. Journal of the Royal Statistical Society, Series A 172: 137-159.
Jackson, D., I. R. White, and S. G. Thompson. 2010. Extending DerSimonian and Laird's methodology to perform multivariate random effects meta-analyses. Statistics in Medicine 29: 1282-1297.
Ma, Y., and M. Mazumdar. 2011. Multivariate meta-analysis: a robust approach based on the theory of U-Statistic. Statistics in Medicine 30: 2911-2929.
Riley, R. D. 2009. Multivariate meta-analysis: The effect of ignoring within-study correlation. Journal of the Royal Statistical Society, Series A 172: 789-811.
Viechtbauer, W., M. W.-L. Cheung. 2010. Outlier and influence diagnostics for meta-analysis. Research Synthesis Methods 1: 112-125.
White, I. R. 2009. Multivariate random-effects meta-analysis. Stata Journal 9: 40-56.
Also see
Help: mvmeta (if installed)