In IJaljuli/metarep: Replicability-Analysis Tools for Meta-Analysis

The package metarep is an extension to the package meta, which allows incorporating replicability-analysis tools to quantify consistency and replicability of treatment effect estimates in a meta-analysis. The tool was proposed by Jaljuli et. al. (submitted) for the fixed-effect and for the random-effects meta-analyses, whit or without the common-effect assumption.

Regardless of the type of meta-analysis applied, metarep allows to perform replicability analysis with or with out the common-effect assumption of the fixed-effects model. We recommend applying the replicability analysis free of the common-effect assumption to guard from a possibly faulty assumption. At this case, the replicability analysis is performed based Fishers' combining function using truncated-Pearson's' test. If the user finds the common-effect assumption supported and wishes to incorporate it in the replicability analysis, they might as well do that. Whereas with this assumption, the replicability analysis is performed using the test statistic of the fixed-effects model, for the combining of every set of $n-u+1$ studies.

Package instalation:

Currently, both meta and metarep packages can be downloaded from GitHub, therefore make sure that the package devtools is installed. metarep also requires the latest version of meta (>= 4.11-0, available on GitHub)

Run the following commands in \code{console} to install the packages:

devtools::install_github( "guido-s/meta"    , force=T )
devtools::install_github( "IJaljuli/metarep", force=T )

Examples:

Here we demonstrate the approach by implementation with metarep, using examples from systematic reviews Cochrane library. These examples are detailed in the paper as well, along with a demonstration of a way to incorporate our suggestions in standard meta-analysis reporting system.

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = " "
)

We demonstrate the with an example based on a fixed-effects meta-analysis. This example was included in Jaljuli et. al. 2020, found in review number CD002943 in the Cochrane library. This analysis explores the effect of mammogram invitation on attendance during the following 12 months.

library(metarep)
library(meta)
data(CD002943_CMP001)

m2943 <- metabin( event.e = N_EVENTS1, n.e = N_TOTAL1, 
                     event.c = N_EVENTS2, n.c = N_TOTAL2,
                     studlab = STUDY, comb.fixed = T , comb.random = F,
                     method = 'Peto', sm = CD002943_CMP001$SM[1],
                     data = CD002943_CMP001)

m2943

summary(m2943)

In this meta-analysis, the effect of sending invitation letters was examined in five studies. The authors main result is that: "The odds ratio in relation to the outcome, attendance in response to the mammogram invitation during the 12 months after the invitation, was 1.66 (95% CI 1.43 to 1.92)".

We suggest reporting the replicability-analysis results alongside: the r-value and lower confidence bounds on the number of studies. Results of complete replicability analysis can be added to the contents of a meta object or using the function metarep(...), as well as to its summary using summary( metarep(...) ).

To perform assumption-free replicability-analysis requiring replicability in at least u = 2 (default) studies, we calculate $r(2)-value$ using truncated-Pearson's' test with truncation threshold t=0.05 (default):

(m2943.ra <- metarep(x = m2943 , u = 2 , common.effect = F ,t = 0.05 ,report.u.max = T))

The bottom two lines report the $r(2)-value$, lower bound on the number of studies with increased effect ($u^L_{max}$) and decreased effect ($u^R_{max}$) , respectively. The evidence towards an increased effect was replicable, with $r(2) − value = 0.0002$. Moreover, with $95\%$ confidence, we can conclude that at least two studies had an increased effect. For higher replicability requirement, compute $r(u')-value$ for $u'>2$ using metarep(u = u' , ... ).

The two-sided $r(u)-value$ of the model can be accessed via r.value:

m2943.ra$r.value

The replicability-analysis reported was performed with an assumption free test, based on truncated-Pearson's' test with truncation level set at the nominal hypothesis testing level (i.e., t=0.05, default). For ordinary Pearson's' test, use t=1.

Although the fixed-effect model assumes that all studies are estimates of the same common effect $\theta$, we recommend applying assumption-free replicability-analysis for protection against an (perhaps) unsupported assumption. Despite that, we extend our suggested method with the common-effect incorporation in section 7. This analysis can be performed via metarep( ... , common.effect = TRUE ).

metarep also allows adding replicability results to the conventional forest plots by meta. This can be done by simply applying forest() on a metarep object.

forest(m2943.ra, layout='revman5',digits.pval = 2 , test.overall = T )

The lower bounds $u^L_{max} \, \text{and} \, u^R_{max}$ are calculated with $1-\alpha = 95\%$ confidence level ( default), meaning that each of the null hypotheses $$H^{u^L_{max}/n}(L) \;\; \text{and}\;\; H^{u^L_{max}/n}(R)$$ is tested at level $\alpha /2 = 2.5\%$, resulting in bounds in overall type error rate $5\%$. Type I error rate can be controlled for any desired $\alpha$ using the argument confidence = 1 -$\alpha$.

The calculation of $u^L_{max} \, \text{and} \, u^R_{max}$ can also be calculated directly using the function find_umax() with the option to specify one-sided alternative, confidence level, truncation threshold and common-effect assumption. For example, let's compute $u^L_{max}$ with the same confidence level as produced by m2943.ra.bounds.

find_umax(x = m2943 , common.effect = F,alternative = 'less',t = 0.05,confidence = 0.975)

Note that this function produces 2 main types of results:

Worst-case scenario studies: A list of $n-u_{max}^L+1$ studies names yielding the maximum $$\max_{\forall {i_1,\dots , i_{n-u+1}} \subset {1,\dots , n} } {\,p^L_{i_1,\dots , i_{n-u+1}}\,}$$
Replicability-analysis results, including:
- $u^L_{max}$ or $u^R_{max}$. If alternative='two-sided', then $u_{max}=\max{u^L_{max}\, , \, u^R_{max}}$ is also reported.
- $r(u_{max}^L)-value$ or $r(u_{max}^R)-value$ if setting alternative='less' or alternative='greater', respectively. If alternative='two-sided', then $$rvalue = r(u_{max}) = 2\cdot\min{r^R(u_{max}^R) , r^L(u_{max}^L) }$$ is also reported.
For demonstration, see the following example.

find_umax(x = m2943 , common.effect = F,alternative = 'two-sided',t = 0.05,confidence = 0.95)

Replicability analysis using a test-statistic based on

The common-effect assumption

find_umax(x = m2943 , common.effect = T,alternative = 'two-sided', confidence = 0.95)

(m2943.raFE <- metarep(x = m2943 , u = 2 , common.effect = T ,report.u.max = T))

find_umax(x = m2943 , common.effect = T,alternative = 'two-sided', confidence = 0.95)

(m2943.raFE <- metarep(x = m2943 , u = 2 , common.effect = T ,report.u.max = T))

forest(m2943.raFE, layout='revman5',digits.pval = 2 , test.overall = T )