Register for free services | Subscribe | Sign In
BMJ 1998; 316:7126 (17 January)

Education and debate

Meta-analysis

Meta-Analysis Software

This article complements a series of six articles that have appeared in the BMJ between 22 November 1996 and 17 January 1998 examining the procedures in conducting reliable meta- analysis in medical research

Matthias Egger, Jonathan AC Sterne, George Davey Smith

Over the past few years, computer software entirely devoted to meta-analysis has increasingly become available, and meta-analytic procedures have been introduced in general statistical software packages. We have identified five commercial packages which are sold with handbooks and three programs which are in the public domain. In this article we briefly review each program and, in table 1 and table 2, provide a comparative summary of features.

Commercial software

Fast*Pro
Fast*Pro is entirely devoted to the appraisal of evidence. Results from a wide range of experimental and observational study designs can be analysed using dichotomous, categorical or continuous effect measures, including odds ratios, relative risks and risk differences. The pros and cons of these measures were discussed in a previous article in this series. (1) The program accommodates studies conducted in single groups of patients (e.g. natural history studies) and studies in which groups received different doses of the same treatment. The software is based on the confidence profile method (2) which uses Bayesian statistics (3) to calculate posterior probability distributions for parameters of interest. The probability distributions, referred to as profiles, are graphically displayed to provide a visual picture of the uncertainty attached to a parameter. Different models can be used to combine the profiles constructed from individual studies. This includes Bayesian models that accommodate the fixed and random effects assumption, conventional fixed effects models (e.g. variance-weighted, Mantel-Haenszel,(4) Yusuf-Peto(5)) and the DerSimonian-Laird(6) random effects model. Unfortunately, the conventional graph showing effect measures and confidence intervals of individual studies along with the combined results which we discussed previously1 cannot be produced. Also, cumulative meta-analysis (7,8) is not available. Finally, the address given in the manual for technical support is invalid. Technical support is available from one of the authors (see table 1).

Table 1 Summary of features of five commercial meta-analysis software packages

FAST*PRO STATA True Epistat DSTAT DESCARTES
Version tested 1.0 5.0 5.0 1.10 available in spring 1998
Operating system MS-DOS Windows MS-DOS MS-DOS Windows
Distributor Academic Press
24-28 Oval Road
London NW1 7DX
UK 1
Timberlake Consultants
47 Hartfield Crescent West
Wickham
Kent BR4 9DW
UK 2
www.stata.com
Epistat Services
2011 Cap Rock
Circle
Richardson
Texas 75080-3417
USA
Lawrence Erlbaum Ass.
365 Broadway
Hillsdale
New Jersey 07642
USA
Update Software Ltd
Summertown Pavilion
Middle Way
Summertown
Oxford OX2 7LG
UK
www.update-software.com
Price £ 250 £ 345 US$ 533 US$ 99 not yet available
Input data 2 x k tables, odds ratios, relative risks, risk difference, differences of means, slopes and others Any (package has sophisticated data management and modelling facilities). Data are analysed using standard methods and the results used as input for meta analysis commands. 2 x 2 tables, differences of means, p values, z-, t-, F-, 02- statistics, correlation coefficients 2 x 2 tables, means, p values, z-, t-, F-, 02- statistics, correlation coefficients 2x2 tables, means, covariates
Statistical models Bayesian and classical fixed effects and random effects Fixed effects and random effects. Empirical Bayes. Regression models. Fixed effects and random effects Fixed effects only Fixed and random effects. Regression models.
Output effect measures Odds ratio, relative risk, risk difference, difference of means and others Odds ratio, relative risk, risk difference, difference of means and others Odds ratio, relative risk, risk difference, standardised difference and others Standardised difference Odds ratio, relative risk, risk difference, NNT, weighted mean difference, standardised mean difference
Test for Homogeneiy Yes Yes Yes Yes Yes
Manual Yes Yes Yes Yes Yes
Formulae given No Yes3 Yes Yes Yes
Technical support is available from Dr Vic Hasselblad, Durham, NC, USA (Fax +1 919 286 5601)
The meta-analysis macro for use within STATA is free and can be downloaded from http://www.stata.com
Sharp SJ, Sterne JAC. Sbe16: Meta-analysis. Stata Technical Bulletin 1997;38: 9-14

STATA
Stata is a comprehensive statistics, data management and graphics package for which a meta- analysis command has recently been written. (9) Individual-level or study-level data are analysed using standard methods to provide an effect estimate (e.g. odds ratio, risk difference or difference between means) and corresponding standard error for each study. The meta-analysis command then calculates fixed-effects (variance-weighted) and random-effects (DerSimonian- Laird (6)) estimates, together with the standard 02 test for heterogeneity between studies and estimate of between-study variance. The classical meta-analysis graph is displayed with either the fixed-effects or random-effects combined estimate. Empirical Bayes estimates of the true effect in each study given the random-effects model can be calculated, displayed and graphed. Results and graphical displays can be shown either on the original scale or on the ratio scale (when the original effect estimates are on a log scale). Funnel plots can be displayed using the standard graphics facilities of the package. A meta-regression (10,11) command which can be used to explore sources of heterogeneity between studies has been written and will be available soon. Cumulative meta-analysis will also be available soon.

True Epistat
This is a statistics package which also offers a number of meta-analysis capabilities. Studies comparing two groups and using odds ratios, relative risks, risk differences (dichotomous outcomes), or standardised differences (continuous outcomes) can be analysed in a variance- weighted fixed effects model or a DerSimonian-Laird random effects model. Data are entered in a two-by-two table or as group variances along with the difference between two means. Correlation coefficients, test statistics from widely used distributions and p-values (and mixtures of the former) can also be combined. Cumulative meta-analysis is not available. The results are given in tabular form or in the typical graphical display showing effect measures and confidence intervals for each study and for the overall result. Funnel plots can be drawn. The graphics can be edited on the screen and printed.

DSTAT
DSTAT was developed for meta-analysis in the psychological sciences. It combines studies comparing two groups. The data are entered as correlation coefficients, test statistics, p-values or mixtures. These statistics are then converted into a standardised (scale-free) effect measure, the effect size (Hedges' g) which is defined as the difference between the two groups expressed in (pooled) standard deviation units (12). If the user wishes to do so, a bias-adjusted effect size can be calculated (Hedges' d). Adjusted or unadjusted effect sizes of individual studies are then combined to produce an overall value. Clinically more relevant quantities such as the difference in risk, the relative risk or the odds ratio cannot be calculated with DSTAT. Also, results cannot be graphically displayed. These drawbacks limit the usefulness of DSTAT for meta-analysis in medical research.

DESCARTES
DESCARTES is a set of software tools for writing systematic reviews and performing meta- analysis which is being developed by Update Software, the company which is responsible for the Cochrane Library and the Cochrane Collaboration's RevMan package (see below). The package will be released in spring 1998. It is an interactive guide through all the steps involved, from protocol through data collection and analysis to publication in paper and electronic formats. The emphasis is on producing a finished document (protocol, systematic review or individual meta-analysis) so output is geared towards publication quality graphics, text and tables. Data for meta-analyses can be imported from other packages (for example RevMan), entered directly into a spreadsheet or entered interactively via guided data-entry screens which check for impossible or unlikely values. Standard fixed effects and random effects models and meta-regression models are available for dichotomous, continuous and individual patient data. Graphical output allows a wide range of plots (for example funnel, L'Abbé, (13) and Galbraith (14) plots). Cumulative meta-analysis and sensitivity analyses are also available. Suggested interpretations of calculated statistics can be generated automatically in textual form and included in the output. All DESCARTES output can be pasted directly into other Windows packages.

Public domain software

The Cochrane Collaboration's Review Manager (RevMan)
RevMan is a software package designed to enter review protocols or completed reviews in Cochrane format, as described in detail in the Cochrane Collaboration handbook. (15) This includes a structured text of the review and tables of included as well as excluded studies. The program is based on the Windows operating system and is easy to use. Dichotomous or continuous data can be entered and analysed in fixed and random effects models for odds ratio, relative risk, risk difference and weighted mean difference. Different comparisons and outcomes can be accommodated in the same data sheet. The classical meta-analysis graph is displayed with or without raw data, weights and year of individual studies. The graphics can be edited on the screen and printed. Funnel plots and cumulative meta-analysis are not available. The package is available from an internet site (see table 2).

Table 2 Summary of features of three public domain meta-analysis software packages.

RevMan Easy MA Meta-Analyst
Version tested 3.0 97b 0.988
Operating system Windows MS-DOS MS-DOS
Distributor The Cochrane Collaboration Dr M Cucherat
Dept. of Clinical Pharmacology
162 Av. Lacassagne
F-69003 Lyon
France
Dr J Lau
New England Medical Center,
Box 63
750 Washington St
Boston,
MA 02111
USA
Internet address hiru.mcmaster.ca/cochrane www.spc.univ-lyon1.fr/citccf/easyma www.spc.univ-lyon1.fr/~mcu/easyma available from Dr Lau via electronic mail (joseph.lau{at}es.nemc.org)
Input data 2 x 2 tables, means 2 x 2 tables 2 x 2 tables
Statistical models Fixed effects and random effects Fixed effects and random effects Fixed effects and random effects
Output effect measures Odds ratio, relative risk, risk difference Odds ratio, relative risk, risk difference, NNT Odds ratio, relative risk, risk difference
Test for Homogenei y Yes Yes Yes
Manual Yes Yes No
Formulae given in preparation Yes No

Easy MA
EasyMA was developed by Michel Cucherat from the University of Lyon. It is described in more detail elsewhere. (16) The package can be down loaded from an internet site (table 2). All menu headings are written in English but contextual help is at present available only in French. An English version is in preparation. EasyMA was developed for meta-analysis of clinical trials with two arms and one or several dichotomous outcomes. It is menu driven and offers fixed effects (e.g. Yusuf-Peto, (5) Mantel-Haenszel (4)) and random effects models (e.g. DerSimonian and Laird (6)) for calculation of overall odds ratios, relative risks and risk differences. In the latter case the number of patients needed to treat to prevent one event (NNT) (1,17) is also given. Other useful features include a table ranking studies according to control group event rates, and weighted and unweighted regression analysis of control group against treatment group rates. EasyMA produces the classical meta-analysis graphs both for standard and cumulative meta- analysis as well as radial and funnel plots. Rosenthal's number of unpublished negative trials needed to render the combined results non-significant (18) and Begg and Mazumdar's test for publication bias (19) are also available. Graphs are of high quality but editing possibilities are somewhat limited.

Meta-Analyst
Meta-analyst was programmed by Joseph Lau from the New England Medical Center. Interested readers should write to the author if they would like to obtain a copy (see table for details). Like EasyMA, this software was developed for conventional and cumulative meta- analysis of clinical trials with two arms and dichotomous outcomes. Only one outcome can be entered at a time. The programme is easy to use and offers the widely used fixed effects and random effects models for combining odds ratios, relative risks and risk differences. The graphs produced are of excellent quality, but, as in EasyMA, they cannot easily be exported for editing in another programme. Additional features include regression analyses of treatment effect against control event rates and sample size estimation of a hypothetical future randomised trial.

Table 3 Data used in meta-analysis of 17 trials of the effect of beta-blockade on mortality after myocardial infarction
Trial Beta-blocker Control

N No of deaths N No of deaths
Reynolds 38 3 39 3
Wilhemsson 114 7 116 14
Ahlmark 69 5 93 11
MIS 1533 102 1520 127
Baber 355 28 365 27
Ahnve 59 4 52 6
Norwegian MC 945 98 939 152
Taylor 632 60 471 48
Hansteen 278 25 282 37
BHAT 1916 138 1921 188
Julian 873 64 583 52
Austr & Swed 263 45 266 47
Manger Cats 291 9 293 16
EIS 858 57 883 45
Olsson 154 25 147 31
LIT 1195 65 1200 62
Boissel 298 17 309 34

Comparing the output from different software packages

In our series we repeatedly discussed a meta-analysis of 17 clinical trials of beta-blockers in secondary prevention after myocardial infarction. (1,8,20) We used the same set of trials to compare the combined effect estimates and 95% confidence intervals produced by the software packages reviewed in this article (except DESCARTES which is not yet available). The data used in these analyses are given in table 3. Calculations were performed for odds ratio, relative risk and risk difference using conventional (non-Bayesian) fixed effects and random effects models. The fixed effect models used Mantel-Haenszel or Yusuf-Peto modified Mantel- Haenszel methods (4,5) whereas random-effect models used DerSimonian and Laird's method. (6) The results obtained from the different software packages are identical or very similar. However, the confidence intervals surrounding these estimates are erroneously wide in FAST*PRO 1.0. This has been corrected in an improved version (Dr Vic Hasselblad, personal communication). At present, Academic Press continues to distribute version 1.0.

Department of Social Medicine
University of Bristol
Bristol BS8 2PR,
U.K.

Matthias Egger,
Reader in Social Medicine and Epidemiology

George Davey Smith,
Professor of Clinical Epidemiology

Department of Public Health Medicine
United Medical and Dental Schools
St Thomas's Hospital
London

Jonathan AC Sterne,
Senior Lecturer in Medical Statistics

Address for correspondence:
Dr Matthias Egger
Department of Social Medicine
University of Bristol
Canynge Hall,
Whiteladies Road
Bristol BS8 2PR,
UK

Tel 0117 928 73 87
Fax 0117 928 73 25
Email m.egger@bristol.ac.uk

References

1. Egger M, Davey Smith G, Phillips AN. Meta-analysis: principles and procedures. BMJ 1997; 315:1533-1537. [Full text]

2. Eddy DM, Hasselblad V, Shachter R. Meta-analysis by the confidence profile method. The statistical synthesis of evidence. Boston: Academic Press, 1992.

3. Lilford RJ, Braunholtz D. The statistical basis of public policy: a paradigm shift is overdue. BMJ 1996; 313:603-607.

4. Mantel N, Haenszel W. Statistical aspects of the analysis of data from retrospective studies of disease. J Natl Cancer Inst 1959; 22:719-748.

5. Yusuf S, Peto R, Lewis J, Collins R, Sleight P. Beta blockade during and after myocardial infarction: an overview of the randomized trials. Progr Cardiovasc Dis 1985; 17:335-371.

6. DerSimonian R, Laird N. Meta-analysis in clinical trials. Controlled Clinical Trials 1986; 7:177-188.

7. Lau J, Antman EM, Jimenez-Silva J, Kupelnick B, Mosteller F, Chalmers TC. Cumulative meta-analysis of therapeutic trials for myocardial infarction. New Engl J Med 1992; 327:248- 254.

8. Egger M, Davey Smith G. Meta-analysis: potentials and promise. BMJ 1997; 315:1371- 1374.[Full text]

9. Sharp SJ, Sterne JAC. sbe16: Meta-analysis. Stata Technical Bulletin 1997; 38:9-14.

10. Thompson S.G. and Sharp S.J. 1997. Explaining heterogeneity in meta-analysis: a comparison of methods. Stat Med (submitted).

11. Egger M, Davey Smith G, Schneider M, Minder CE. Bias in meta-analysis detected by a simple, graphical test. BMJ 1997; 315:629-634. [Full text]

12. Hedges LV, Olkin I. Statistical methods for meta-analysis. Boston: Academic Press, 1985.

13. L'Abbe KA, Detsky AS, O'Rourke K. Meta-analysis in clinical research. Ann Intern Med 1987; 107:224-233.

14. Galbraith R. A note on graphical presentation of estimated odds ratios from several clinical trials. Stat Med 1988; 7:889-894.

15. Oxman A. VI. Preparing and maintaining systematic reviews. In: Sackett DL, Oxman A, editors. Cochrane Collaboration Handbook. Oxford: The Cochrane Collaboration, 1995:

16. Cucherat M, Boissel J-P, Leizorovicz A, Haugh MC. EasyMA: a program for the meta- analysis of clinical trials. Computer Methods and Programs in Biomedicine 1997; 53:187- 190.

17. Laupacis A, Sackett DL, Roberts RS. An assessment of clinically useful measures of the consequences of treatment. New Engl J Med 1988; 318:1728-1733.

18. Rosenthal R. The 'file drawer problem' and tolerance for null results. Psychol Bull 1979; 86:638-641.

19. Begg CB, Mazumdar M. Operating characteristics of a rank correlation test for publication bias. Biometrics 1994; 50:1088-1099.

20. Egger M, Davey Smith G. Meta-analysis: bias in location and selection of studies. BMJ 1998; 316:61-66. [Full text]