Results from 17 trials comparing post-operative radiation therapy with and without adjuvant chemotherapy in patients with malignant gliomas.

dat.fine1993

## Format

The data frame contains the following columns:

 study numeric study number nei numeric sample size in the experimental group receiving radiotherapy plus adjuvant chemotherapy nci numeric sample size in the control group receiving radiotherapy alone e1i numeric number of survivors at 6 months in the experimental group c1i numeric number of survivors at 6 months in the control group e2i numeric number of survivors at 12 months in the experimental group c2i numeric number of survivors at 12 months in the control group e3i numeric number of survivors at 18 months in the experimental group c3i numeric number of survivors at 18 months in the control group e4i numeric number of survivors at 24 months in the experimental group c4i numeric number of survivors at 24 months in the control group

## Details

The 17 trials report the post-operative survival of patients with malignant gliomas receiving either radiation therapy with adjuvant chemotherapy or radiation therapy alone. Survival was assessed at 6, 12, 18, and 24 months in all but one study (which assessed survival only at 12 and at 24 months).

The data were reconstructed by Trikalinos and Olkin (2012) based on Table 2 in Fine et al. (1993) and Table 3 in Dear (1994). The data can be used to illustrate how a meta-analysis can be conducted of effect sizes reported at multiple time points using a multivariate model.

## Source

Dear, K. B. G. (1994). Iterative generalized least squares for meta-analysis of survival data at multiple times. Biometrics, 50(4), 989–1002. https://doi.org/10.2307/2533438

Trikalinos, T. A., & Olkin, I. (2012). Meta-analysis of effect sizes reported at multiple time points: A multivariate approach. Clinical Trials, 9(5), 610–620. https://doi.org/10.1177/1740774512453218

## References

Fine, H. A., Dear, K. B., Loeffler, J. S., Black, P. M., & Canellos, G. P. (1993). Meta-analysis of radiation therapy with and without adjuvant chemotherapy for malignant gliomas in adults. Cancer, 71(8), 2585–2597. https://doi.org/10.1002/1097-0142(19930415)71:8<2585::aid-cncr2820710825>3.0.co;2-s

## Author

Wolfgang Viechtbauer, wvb@metafor-project.org, https://www.metafor-project.org

## Concepts

medicine, oncology, odds ratios, longitudinal models

## Examples

### copy data into 'dat' and examine data
dat <- dat.fine1993
dat
#>    study nei nci e1i c1i e2i c2i e3i c3i e4i c4i
#> 1      1  19  22  16  20  11  12   4   8   4   3
#> 2      2  34  35  22  22  18  12  15   8  15   6
#> 3      3  72  68  44  40  21  15  10   3   3   0
#> 4      4  22  20  19  12  14   5   5   4   2   3
#> 5      5  70  32  62  27  42  13  26   6  15   5
#> 6      6 183  94 130  65  80  33  47  14  30  11
#> 7      7  26  50  24  30  13  18   5  10   3   9
#> 8      8  61  55  51  44  37  30  19  19  11  15
#> 9      9  36  25  30  17  23  12  13   4  10   4
#> 10    10  45  35  43  35  19  14   8   4   6   0
#> 11    11 246 208 169 139 106  76  67  42  51  35
#> 12    12 386 141 279  97 170  46  97  21  73   8
#> 13    13  59  32  56  30  34  17  21   9  20   7
#> 14    14  45  15  42  10  18   3   9   1   9   1
#> 15    15  14  18  14  18  13  14  12  13   9  12
#> 16    16  26  19  21  15  12  10   6   4   5   1
#> 17    17  74  75  NA  NA  42  40  NA  NA  23  30

# \dontrun{

### load metafor package
library(metafor)

### calculate log(ORs) and sampling variances for each time point
dat <- escalc(measure="OR", ai=e1i, n1i=nei, ci=c1i, n2i=nci, data=dat, var.names=c("y1i","v1i"))
dat <- escalc(measure="OR", ai=e2i, n1i=nei, ci=c2i, n2i=nci, data=dat, var.names=c("y2i","v2i"))
dat <- escalc(measure="OR", ai=e3i, n1i=nei, ci=c3i, n2i=nci, data=dat, var.names=c("y3i","v3i"))
dat <- escalc(measure="OR", ai=e4i, n1i=nei, ci=c4i, n2i=nci, data=dat, var.names=c("y4i","v4i"))

### calculate the covariances (equations in Appendix of Trikalinos & Olkin, 2012)
dat$v12i <- with(dat, nei / (e1i * (nei - e2i)) + nci / (c1i * (nci - c2i))) dat$v13i <- with(dat, nei / (e1i * (nei - e3i)) + nci / (c1i * (nci - c3i)))
dat$v14i <- with(dat, nei / (e1i * (nei - e4i)) + nci / (c1i * (nci - c4i))) dat$v23i <- with(dat, nei / (e2i * (nei - e3i)) + nci / (c2i * (nci - c3i)))
dat$v24i <- with(dat, nei / (e2i * (nei - e4i)) + nci / (c2i * (nci - c4i))) dat$v34i <- with(dat, nei / (e3i * (nei - e4i)) + nci / (c3i * (nci - c4i)))

### create dataset in long format
dat.long <- data.frame(study=rep(1:nrow(dat), each=4), time=1:4,
yi=c(t(dat[c("y1i","y2i","y3i","y4i")])),
vi=c(t(dat[c("v1i","v2i","v3i","v4i")])))

### var-cov matrices of the sudies
V <- lapply(split(dat, dat$study), function(x) matrix(c( x$v1i, x$v12i, x$v13i, x$v14i, x$v12i,  x$v2i, x$v23i, x$v24i, x$v13i, x$v23i, x$v3i, x$v34i, x$v14i, x$v24i, x$v34i,  x$v4i), nrow=4, ncol=4, byrow=TRUE)) ### remove rows for the missing time points in study 17 dat.long <- na.omit(dat.long) ### remove corresponding rows/columns from var-cov matrix V[[17]] <- V[[17]][c(2,4),c(2,4)] ### make a copy of V Vc <- V ### replace any (near) singular var-cov matrices with ridge corrected versions repl.Vi <- function(Vi) { res <- eigen(Vi) if (any(res$values <= .08)) {
round(res$vectors %*% diag(res$values + .08) %*% t(res\$vectors), 12)
} else {
Vi
}
}
Vc <- lapply(Vc, repl.Vi)

### do not correct var-cov matrix of study 17
Vc[[17]] <- V[[17]]

### construct block diagonal matrix
Vc <- bldiag(Vc)

### multivariate fixed-effects model
res <- rma.mv(yi, Vc, mods = ~ factor(time) - 1, method="FE", data=dat.long)
print(res, digits=3)
#>
#> Multivariate Meta-Analysis Model (k = 66; method: FE)
#>
#> Variance Components: none
#>
#> Test for Residual Heterogeneity:
#> QE(df = 62) = 62.447, p-val = 0.460
#>
#> Test of Moderators (coefficients 1:4):
#> QM(df = 4) = 15.631, p-val = 0.004
#>
#> Model Results:
#>
#>                estimate     se   zval   pval   ci.lb  ci.ub
#> factor(time)1     0.251  0.143  1.755  0.079  -0.029  0.532    .
#> factor(time)2     0.428  0.114  3.742  <.001   0.204  0.652  ***
#> factor(time)3     0.343  0.134  2.560  0.010   0.080  0.606    *
#> factor(time)4     0.281  0.138  2.035  0.042   0.010  0.552    *
#>
#> ---
#> Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
#>

### multivariate random-effects model with heteroscedastic AR(1) structure for the true effects
res <- rma.mv(yi, Vc, mods = ~ factor(time) - 1, random = ~ time | study,
struct="HAR", data=dat.long, control=list(optimizer="hjk"))
print(res, digits=3)
#>
#> Multivariate Meta-Analysis Model (k = 66; method: REML)
#>
#> Variance Components:
#>
#> outer factor: study (nlvls = 17)
#> inner factor: time  (nlvls = 4)
#>
#>            estim   sqrt  k.lvl  fixed  level
#> tau^2.1    0.000  0.000     16     no      1
#> tau^2.2    0.000  0.015     17     no      2
#> tau^2.3    0.029  0.172     16     no      3
#> tau^2.4    0.197  0.443     17     no      4
#> rho        1.000                   no
#>
#> Test for Residual Heterogeneity:
#> QE(df = 62) = 62.447, p-val = 0.460
#>
#> Test of Moderators (coefficients 1:4):
#> QM(df = 4) = 14.555, p-val = 0.006
#>
#> Model Results:
#>
#>                estimate     se   zval   pval   ci.lb  ci.ub
#> factor(time)1     0.254  0.143  1.772  0.076  -0.027  0.534    .
#> factor(time)2     0.408  0.115  3.536  <.001   0.182  0.634  ***
#> factor(time)3     0.365  0.145  2.521  0.012   0.081  0.650    *
#> factor(time)4     0.351  0.187  1.878  0.060  -0.015  0.718    .
#>
#> ---
#> Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
#>

### profile the variance components
par(mfrow=c(2,2))
profile(res, tau2=1, xlim=c( 0,.2))
profile(res, tau2=2, xlim=c( 0,.2))
profile(res, tau2=3, xlim=c( 0,.2))
profile(res, tau2=4, xlim=c(.1,.3))

### profile the autocorrelation coefficient
par(mfrow=c(1,1))
profile(res, rho=1)

# }