The merlin package allows the fitting of multi-outcome
models and with any number of nested random effects, outcomes can be
modelled jointly, or with shared random effects.
For n record to be included in the model there must be at least one observation per level in the model. For example in a joint survival and longitudinal biomarker model, if the survival time for a patient is available, but all longitudinal biomarker observations are missing, the individual will be excluded from the analysis.
For each individual outcome in the model, a complete case analysis is used. The response, covariates and level indicators are required to fit the model, which may vary by outcome. For example if a model simultaneously models two separate biomarkers, which are sampled at different time points, data in each record will only be included in the model for the appropriate outcome, potentially resulting in different sample sizes being used for each outcome.
There is a separate variance-covariance matrix for each level of
random effects. The structure of the variance-covariance matrix can be
varied to allow for correlation between parameters at each level. The
possible structures are identity, where all diagonal
elements are estimated and constrained to be the same,
diagonal where all diagonal elements are estimated uniquely
(the default), unstructured where all elements of the
variance-covariance matrix are estimated, and exchangeable
which assumes a common variance and a common covariance.
Predictions using the fixed-effects only can be found using the
predict() function with the argument
type = fixedonly. Marginal predictions, where the random
effects are integrated out, can be obtained using
type = marginal.
In order to illustrate the potential uses of merlin a
number of increasingly advanced models have been fitted to the commonly
used (in the area of joint modelling of longitudinal and survival data)
primary biliary cirrhosis dataset.
This data set needs some re-formatting in order to fit joint models
library(merlin)
data("pbc")
pbc[1:11,c("id","years","status","status2","drug","serBilir","prothrombin","year")]
#> id years status status2 drug serBilir prothrombin year
#> 1 1 1.09517 dead 1 D-penicil 14.5 12.2 0.0000000
#> 2 1 1.09517 dead 1 D-penicil 21.3 11.2 0.5256817
#> 3 2 14.15234 alive 0 D-penicil 1.1 10.6 0.0000000
#> 4 2 14.15234 alive 0 D-penicil 0.8 11.0 0.4983025
#> 5 2 14.15234 alive 0 D-penicil 1.0 11.6 0.9993429
#> 6 2 14.15234 alive 0 D-penicil 1.9 10.6 2.1027270
#> 7 2 14.15234 alive 0 D-penicil 2.6 11.3 4.9008871
#> 8 2 14.15234 alive 0 D-penicil 3.6 11.5 5.8892783
#> 9 2 14.15234 alive 0 D-penicil 4.2 11.5 6.8858833
#> 10 2 14.15234 alive 0 D-penicil 3.6 11.5 7.8907020
#> 11 2 14.15234 alive 0 D-penicil 4.6 11.5 8.8325485The event times are given in the years variable and the
event indicator in the status variable which can take the
values alive, dead or transplanted or the status2 variable
which is 0 for alive and 1 for dead. Each survival observation should
only appear once for each individual, otherwise each observation will be
treated as a separate event, this allows for recurrent events to be
modelled. The survival observations can be reformatted as follows:
pbc$stime <- pbc$years
pbc$stime[duplicated(pbc$id)] <- NA
pbc$died <- pbc$status2
pbc$died[duplicated(pbc$id)] <- NAWe are also going to work with the log of biomarkers prothrombin index and serum bilirubin throughout, the time of this measurements will be recorded in the variable time. We will also change the treatment variable to be numeric rather than a factor variable.
pbc$logb <- log(pbc$serBilir)
pbc$logp <- log(pbc$prothrombin)
pbc$time <- pbc$year
pbc$trt <- as.numeric(pbc$drug) - 1The data now looks like this. Failing to set the data up in this way will lead to errors in the parameter estimates.
pbc[1:11,c("id","stime","died","logb","logp","time")]
#> id stime died logb logp time
#> 1 1 1.09517 1 2.67414865 2.501436 0.0000000
#> 2 1 NA NA 3.05870707 2.415914 0.5256817
#> 3 2 14.15234 0 0.09531018 2.360854 0.0000000
#> 4 2 NA NA -0.22314355 2.397895 0.4983025
#> 5 2 NA NA 0.00000000 2.451005 0.9993429
#> 6 2 NA NA 0.64185389 2.360854 2.1027270
#> 7 2 NA NA 0.95551145 2.424803 4.9008871
#> 8 2 NA NA 1.28093385 2.442347 5.8892783
#> 9 2 NA NA 1.43508453 2.442347 6.8858833
#> 10 2 NA NA 1.28093385 2.442347 7.8907020
#> 11 2 NA NA 1.52605630 2.442347 8.8325485We can fit a simple linear model
lin.model <- merlin(logb ~ time, family = "gaussian", data = pbc)
lin.model
#> Merlin: mixed-effects model
#> Data: pbc
#>
#> Coefficients:
#> time _cons log_sd(resid.)
#> 0.01392 0.55956 0.10354
summary(lin.model)
#> Mixed effects regression model
#> Log likelihood = -2961.414
#>
#> Estimate Std. Error z Pr(>|z|) [95% Conf. Interval]
#> time 0.013916 0.008128 1.712 0.08687 -0.002014 0.029846
#> _cons 0.559560 0.035806 15.628 0.00000 0.489382 0.629739
#> log_sd(resid.) 0.103540 0.016032 6.458 0.00000 0.072119 0.134962We can add flexibility to the model using restricted cubic splines:
rcs.model <- merlin(logb ~ rcs(time, df = 3), family = "gaussian", timevar = "time", data = pbc)
summary(rcs.model)
#> Mixed effects regression model
#> Log likelihood = -2960.733
#>
#> Estimate Std. Error z Pr(>|z|) [95% Conf. Interval]
#> rcs():1 0.043260 0.025147 1.720 0.0854 -0.006028 0.092548
#> rcs():2 0.028845 0.025148 1.147 0.2514 -0.020443 0.078133
#> rcs():3 -0.009251 0.025147 -0.368 0.7130 -0.058539 0.040037
#> _cons 0.603291 0.025147 23.990 0.0000 0.554003 0.652579
#> log_sd(resid.) 0.103511 0.016037 6.455 0.0000 0.072079 0.134943By default, the restricted cubic splines are orthogonalised
(orthog = TRUE). The serum bilirubin observations are
clustered within individuals, so we can add a normally-distributed
random intercept term named M1. The coefficient of the
random-effect term will normally constrained to 1 using the
*1 notation, unless the random effect is being shared with
another outcome model:
r.int.model <- merlin(logb ~ rcs(time, df = 3) + M1[id]*1,
family = "gaussian",
levels = "id",
timevar = "time",
data = pbc)
summary(r.int.model)
#> Mixed effects regression model
#> Log likelihood = -1978.61
#>
#> Estimate Std. Error z Pr(>|z|) [95% Conf. Interval]
#> rcs():1 0.266362 0.013699 19.444 0.00000 0.239513 0.293211
#> rcs():2 0.058592 0.013014 4.502 0.00001 0.033086 0.084098
#> rcs():3 0.016696 0.012695 1.315 0.18845 -0.008186 0.041578
#> _cons 0.614076 0.023175 26.498 0.00000 0.568655 0.659498
#> log_sd(resid.) -0.611477 0.016776 -36.450 0.00000 -0.644357 -0.578597
#> log_sd(M1) -0.127962 0.019616 -6.523 0.00000 -0.166409 -0.089515
#>
#> Integration method: Non-adaptive Gauss-Hermite quadrature
#> Integration points: 7We can also add an additional random-slope term (M2) to
the model by forming an interaction between the time variable and
random-effect using :. We can increase the number of
quadrature nodes to improve estimation of the likelihood using the
ip option within the control argument.
r.slope.model <- merlin(logb ~ rcs(time, df = 3) + M1[id]*1 + time:M2[id]*1,
family = "gaussian",
timevar = "time",
levels = "id",
data = pbc,
control = list(ip = 15))
summary(r.slope.model)
#> Mixed effects regression model
#> Log likelihood = -1718.679
#>
#> Estimate Std. Error z Pr(>|z|) [95% Conf. Interval]
#> rcs():1 -0.024415 0.016517 -1.478 0.13935 -0.056787 0.007957
#> rcs():2 -0.020279 0.010947 -1.852 0.06396 -0.041736 0.001177
#> rcs():3 -0.025302 0.009456 -2.676 0.00745 -0.043835 -0.006769
#> _cons 0.234964 0.019792 11.872 0.00000 0.196172 0.273756
#> log_sd(resid.) -1.017033 0.017687 -57.502 0.00000 -1.051698 -0.982367
#> log_sd(M1) -0.276734 0.021426 -12.916 0.00000 -0.318728 -0.234740
#> log_sd(M2) -1.432000 0.023614 -60.642 0.00000 -1.478283 -1.385717
#>
#> Integration method: Non-adaptive Gauss-Hermite quadrature
#> Integration points: 15If a model has random effects, merlin will fit fixed
effects models to obtain starting values and then assume an identity
matrix for the variance of the random effects. Initial values can be
manually set using the from option.