| Title: | Exact Matching and Matching-Adjusted Indirect Comparison (MAIC) |
|---|---|
| Description: | The second version (0.2.0) contains implementation for exact matching which is an alternative to propensity score matching (see Glimm & Yau (2025)). The initial version (0.1.2) contains a collection of easy-to-implement tools for checking whether a MAIC can be conducted, as well as an alternative way of calculating weights (see Glimm & Yau (2021) <doi:10.1002/pst.2210>.) |
| Authors: | Lillian Yau [aut, cre], Ekkehard Glimm [aut], Xinlei Deng [aut]
|
| Maintainer: | Lillian Yau <[email protected]> |
| License: | GPL (>= 3) |
| Version: | 0.2.0 |
| Built: | 2025-03-03 17:25:06 UTC |
| Source: | https://github.com/cran/maicChecks |
Three artificial scenarios serves as the ad cases. This is used in Glimm & Yau (2021)
data(eAD)data(eAD)
corresponds to scenarios A, B, and C in the reference manuscript (Glimm & Yau (2021)). Scenario A is very close to IPD center (see data(ipd)) and is within the IPD convex hull; scenario B is further away from IPD center but otherwise still inside the IPD convex hull; scenario C is outside IPD convex hull.
a numeric vector
a numeric vector
Glimm & Yau (2021)
data(eAD)data(eAD)
an articial data set serves as the IPD set. this is used in Glimm & Yau (2021)
data(eIPD)data(eIPD)
y1a numeric vector
y2a numeric vector
Glimm & Yau (2021)
data(eIPD)data(eIPD)
Checks whether two IPD datasets can be matched with lpSolve::lp
exmLP.2ipd( ipd1, ipd2, vars_to_match = NULL, cat_vars_to_01 = NULL, mean.constrained = FALSE )exmLP.2ipd( ipd1, ipd2, vars_to_match = NULL, cat_vars_to_01 = NULL, mean.constrained = FALSE )
ipd1 |
a dataframe with n1 row and p column, where n1 is number of subjects of the first IPD, and p is the number of variables used in standardization. |
ipd2 |
a dataframe with n2 row and p column, where n2 is number of subjects of the second IPD, and p is the number of variables used in standardization. |
vars_to_match |
variables used for matching. if NULL, use all variables. |
cat_vars_to_01 |
variable names for the categorical variables that need to be converted to indicator variables. |
mean.constrained |
whether to restrict the weighted means to be within the ranges of observed means. Default is FALSE. When it is TRUE, there is a higher chance of not having a solution. |
If dummy variables are already created for the categorical variables in the data set, and are present in ipd1 and ipd2, then cat_vars_to_01 should be left as NULL.
lp.check |
0 = OS can be conducted; 2 = OS cannot be conducted |
Lillian Yau
## Not run: ipd1 <- sim110[sim110$study == 'IPD A',] ipd2 <- sim110[sim110$study == 'IPD B',] x <- exmLP.2ipd(ipd1, ipd2, vars_to_match = paste0('X', 1:5), cat_vars_to_01 = paste0('X', 1:3), mean.constrained = FALSE) ## End(Not run)## Not run: ipd1 <- sim110[sim110$study == 'IPD A',] ipd2 <- sim110[sim110$study == 'IPD B',] x <- exmLP.2ipd(ipd1, ipd2, vars_to_match = paste0('X', 1:5), cat_vars_to_01 = paste0('X', 1:3), mean.constrained = FALSE) ## End(Not run)
Exact matching for two IPD's
exmWt.2ipd( ipd1, ipd2, vars_to_match = NULL, cat_vars_to_01 = NULL, mean.constrained = FALSE )exmWt.2ipd( ipd1, ipd2, vars_to_match = NULL, cat_vars_to_01 = NULL, mean.constrained = FALSE )
ipd1 |
a dataframe with n row and p column, where n is number of subjects and p is the number of variables used in matching. |
ipd2 |
the other IPD with the same number of columns |
vars_to_match |
variables used for matching. if NULL, use all variables. |
cat_vars_to_01 |
a list of variable names for the categorical variables that need to be converted to indicator variables. |
mean.constrained |
whether to restrict the weighted means to be within the ranges of observed means. Default is FALSE. When it is TRUE, there is a higher chance of not having a solution. |
If dummy variables are already created for the categorical variables in the data set, and are present in ipd1 and ipd2, then cat_vars_to_01 should be left as NULL.
ipd1 |
re-scaled weights of the exact matching by maximizing ESS for IPD 1, and the input IPD 1 data with categorical variables converted to 0-1 indicators |
ipd2 |
re-scaled weights of the exact matching by maximizing ESS for IPD 2, and the input IPD 2 data with categorical variables converted to 0-1 indicators |
wtd.summ |
ESS for IPD 1, ESS for IPD 2, and weighted means of the matching variables |
Lillian Yau
## Not run: ipd1 <- sim110[sim110$study == 'IPD A',] ipd2 <- sim110[sim110$study == 'IPD B',] x <- exmWt.2ipd(ipd1, ipd2, vars_to_match = paste0('X', 1:5), cat_vars_to_01 = paste0('X', 1:3), mean.constrained = FALSE) ## End(Not run)## Not run: ipd1 <- sim110[sim110$study == 'IPD A',] ipd2 <- sim110[sim110$study == 'IPD B',] x <- exmWt.2ipd(ipd1, ipd2, vars_to_match = paste0('X', 1:5), cat_vars_to_01 = paste0('X', 1:3), mean.constrained = FALSE) ## End(Not run)
Checks if AD is within the convex hull of IPD using lp-solve
maicLP(ipd, ad)maicLP(ipd, ad)
ipd |
a dataframe with n row and p column, where n is number of subjects and p is the number of variables used in matching. |
ad |
a dataframe with 1 row and p column. The matching variables should be in the same order as that in |
lp.check |
0 = AD is inside IPD, and MAIC can be conducted; 2 = otherwise |
Glimm & Yau (2021). "Geometric approaches to assessing the numerical feasibility for conducting matching-adjusted indirect comparisons", Pharmaceutical Statistics, 21(5):974-987. doi:10.1002/pst.2210.
## eAD[1,] is the scenario A in the reference paper, ## i.e. when AD is within IPD convex hull maicLP(eIPD, eAD[1,2:3]) ## eAD[3,] is the scenario C in the reference paper, ## i.e. when AD is outside IPD convex hull maicLP(eIPD, eAD[3,2:3])## eAD[1,] is the scenario A in the reference paper, ## i.e. when AD is within IPD convex hull maicLP(eIPD, eAD[1,2:3]) ## eAD[3,] is the scenario C in the reference paper, ## i.e. when AD is outside IPD convex hull maicLP(eIPD, eAD[3,2:3])
Should only be used when all matching variables are normally distributed
maicMD(ipd, ad, n.ad = Inf)maicMD(ipd, ad, n.ad = Inf)
ipd |
a dataframe with n row and p column, where n is number of subjects and p is the number of variables used in matching. |
ad |
a dataframe with 1 row and p column. The matching variables should be in the same order as that in |
n.ad |
default is Inf assuming |
When AD does not have the largest Mahalanobis distance, in the original scale AD can still be outside of the IPD convex hull. On the other hand, when AD does have the largest Mahalanobis distance, in the original scale, AD is for sure outside the IPD convex hull.
Prints a message whether AD is furthest away from 0, i.e. IPD center in terms of Mahalanobis distance. Also returns ggplot object for plotting.
md.dplot |
dot-plot of AD and IPD in Mahalanobis distance |
md.check |
0 = AD has the largest Mahalanobis distance to the IPD center; 2 = otherwise |
Glimm & Yau (2021). "Geometric approaches to assessing the numerical feasibility for conducting matching-adjusted indirect comparisons", Pharmaceutical Statistics, 21(5):974-987. doi:10.1002/pst.2210.
## Not run: ## eAD[1,] is the scenario A in the reference paper, ## i.e. when AD is perfectly within IPD md <- maicMD(eIPD, eAD[1,2:3]) md ## a dot-plot of IPD Mahalanobis distances along with AD in the same metric. ## End(Not run)## Not run: ## eAD[1,] is the scenario A in the reference paper, ## i.e. when AD is perfectly within IPD md <- maicMD(eIPD, eAD[1,2:3]) md ## a dot-plot of IPD Mahalanobis distances along with AD in the same metric. ## End(Not run)
Checks whether AD is outside IPD in principal component (PC) coordinates
maicPCA(ipd, ad)maicPCA(ipd, ad)
ipd |
a dataframe with n row and p column, where n is number of subjects in IPD set and p is the number of variables used in matching. |
ad |
a dataframe with 1 row and p column. The matching variables should be in the same order as that in |
When AD is within the IPD PC ranges, AD can still be outside the IPD convex hull in the original scale. On the other hand, if AD is outside the IPD PC ranges, in the original scale AD is for sure outside the IPD convex hull.
Prints a message whether AD is inside or outside IPD PC coordinates. Also returns a ggplot object to be plotted.
pc.dplot |
dot-plot of AD and IPD both in IPD's PC coordinates |
pca.check |
0 = AD within the ranges of IPD's PC coordinates; 2 = otherwise |
Glimm & Yau (2021). "Geometric approaches to assessing the numerical feasibility for conducting matching-adjusted indirect comparisons", Pharmaceutical Statistics, 21(5):974-987. doi:10.1002/pst.2210.
## Not run: ## eAD[1,] is the scenario A in the reference paper, ## i.e. when AD is perfectly within IPD a1 <- maicPCA(eIPD, eAD[1,2:3]) a1 ## the dot plots of PC's for IPD and AD ## eAD[3,] is the scenario C in the reference paper, ## i.e. when AD is outside IPD a3 <- maicPCA(eIPD, eAD[3,2:3]) a3 ## the dot plots of PC's for IPD and AD ## End(Not run)## Not run: ## eAD[1,] is the scenario A in the reference paper, ## i.e. when AD is perfectly within IPD a1 <- maicPCA(eIPD, eAD[1,2:3]) a1 ## the dot plots of PC's for IPD and AD ## eAD[3,] is the scenario C in the reference paper, ## i.e. when AD is outside IPD a3 <- maicPCA(eIPD, eAD[3,2:3]) a3 ## the dot plots of PC's for IPD and AD ## End(Not run)
Hotelling's T-square test to check whether maic is needed
maicT2Test(ipd, ad, n.ad = Inf)maicT2Test(ipd, ad, n.ad = Inf)
ipd |
a dataframe with n row and p column, where n is number of subjects and p is the number of variables used in matching. |
ad |
a dataframe with 1 row and p column. The matching variables should be in the same order as that in |
n.ad |
default is Inf assuming |
When n.ad is not Inf, the covariance matrix is adjusted by the factor n.ad/(n.ipd + n.ad)), where n.ipd is nrow(ipd), the sample size of ipd.
T.sq.f |
the value of the T^2 test statistic |
p.val |
the p-value corresponding to the test statistic. When the p-value is small, matching is necessary. |
Glimm & Yau (2021). "Geometric approaches to assessing the numerical feasibility for conducting matching-adjusted indirect comparisons", Pharmaceutical Statistics, 21(5):974-987. doi:10.1002/pst.2210.
## eAD[1,] is the scenario A in the reference paper, ## i.e. when AD is perfectly within IPD maicT2Test(eIPD, eAD[1,2:3])## eAD[1,] is the scenario A in the reference paper, ## i.e. when AD is perfectly within IPD maicT2Test(eIPD, eAD[1,2:3])
Estimates the MAIC weights for each individual in the IPD. Should only be used after it is ascertained that AD is indeed within the convex hull of IPD.
maicWt(ipd, ad, max.it = 25)maicWt(ipd, ad, max.it = 25)
ipd |
a dataframe with n row and p column, where n is number of subjects and p is the number of variables used in matching. |
ad |
a dataframe with 1 row and p coln. The matching variables should be in the same order as that in |
max.it |
maximum iteration passed to optim(). if |
The main code are taken from Philippo (2016). It returns the following:
optim.out |
results of optim() |
maic.wt |
MAIC un-scaled weights for each subject in the IPD set |
maic.wt.rs |
re-scaled weights which add up to the original total sample size, i.e. nrow(ipd) |
ipd.ess |
effective sample size |
ipd.wtsumm |
weighted summary statistics of the matching variables after matching. they should be identical to the input AD when AD is within the IPD convex hull. |
Phillippo DM, Ades AE, Dias S, et al. (2016). Methods for population-adjusted indirect comparisons in submissions to NICE. NICE Decision Support Unit Technical Support Document 18.
## eAD[1,] is scenario A in the reference manuscript m1 <- maicWt(eIPD, eAD[1,2:3])## eAD[1,] is scenario A in the reference manuscript m1 <- maicWt(eIPD, eAD[1,2:3])
Estimates an alternative set of weights which maximizes effective sample size (ESS) for a given set of variates used in the matching. Should only be used after it is ascertained that AD is indeed within the convex hull of IPD.
maxessWt(ipd, ad)maxessWt(ipd, ad)
ipd |
a dataframe with n row and p column, where n is number of subjects and p is the number of variables used in matching. |
ad |
a dataframe with 1 row and p column. The matching variables should be in the same order as that in |
The weights maximize the ESS subject to the set of baseline covariates used in the matching.
maxess.wt |
maximum ESS weights. Scaled to sum up to the total IPD sample size, i.e. nrow(ipd) |
ipd.ess |
effective sample size. It is no smaller than the ESS given by the MAIC weights. |
ipd.wtsumm |
weighted summary statistics of the matching variables after matching. they should be identical to the input AD when AD is within the IPD convex hull. |
Glimm & Yau (2021). "Geometric approaches to assessing the numerical feasibility for conducting matching-adjusted indirect comparisons", Pharmaceutical Statistics, 21(5):974-987. doi:10.1002/pst.2210.
## eAD[1,] is scenario A in the reference manuscript m0 <- maxessWt(eIPD, eAD[1,2:3])## eAD[1,] is scenario A in the reference manuscript m0 <- maxessWt(eIPD, eAD[1,2:3])
sim110 is one of the simulated data presented in the simulation study in Glimm & Yau (2025).The covariates used in matching are X1 to X15. A response variable Y is simulated to depend on 6 of the 15 covariates.
data(sim110)data(sim110)
Covariates used in matching
Response variable
IPD A and IPD B
Glimm & Yau (2025)
data(sim110)data(sim110)