Abstract
The identification of patient subgroups with differential treatment effects is the first step towards individualised treatments. A current draft guideline by the EMA discusses potentials and problems in subgroup analyses and formulated challenges to the development of appropriate statistical procedures for the data-driven identification of patient subgroups. We introduce model-based recursive partitioning as a procedure for the automated detection of patient subgroups that are identifiable by predictive factors. The method starts with a model for the overall treatment effect as defined for the primary analysis in the study protocol and uses measures for detecting parameter instabilities in this treatment effect. The procedure produces a segmented model with differential treatment parameters corresponding to each patient subgroup. The subgroups are linked to predictive factors by means of a decision tree. The method is applied to the search for subgroups of patients suffering from amyotrophic lateral sclerosis that differ with respect to their Riluzole treatment effect, the only currently approved drug for this disease.
1 Introduction
With the rise of personalised medicine, the search for individual treatments poses challenges to the development of appropriate statistical methods. Subgroup analyses following a traditional statistical assessment of an overall treatment effect of a new therapy aim at identifying three groups of patients: (1) those who benefit from the new therapy, (2) those who do not benefit, and (3) those whose clinical outcome under the new therapy is worse than under alternative therapies. Such post-hoc subgroup analyses potentially lead to better benefit-risk decisions and treatment recommendations but are subject to all kind of biases and can hardly be performed under full statistical error control. Therefore, the European Medicines Agency (EMA) recently published a draft of a guideline for the investigation of subgroups in confirmatory clinical trials [1] that discusses potential areas of application, necessity, pitfalls, and good practice in subgroup analyses. In the guideline draft, three scenarios in which exploratory investigation of subgroups is of special interest were identified:
Scenario 1: “The clinical data presented are overall statistically persuasive with therapeutic efficacy demonstrated globally. It is of interest to verify that the conclusions of therapeutic efficacy (and safety) apply consistently across subgroups of the clinical trial population.”
Scenario 2: “The clinical data presented are overall statistically persuasive but with therapeutic efficacy or benefit/risk which is borderline or unconvincing and it is of interest to identify post-hoc a subgroup, where efficacy and risk-benefit is convincing.”
Scenario 3: “The clinical data presented fail to establish statistically persuasive evidence but there is interest in identifying a subgroup, where a relevant treatment effect and compelling evidence of a favourable risk-benefit profile can be assessed.”
Especially in trials with highly heterogeneous study populations, subgroup analyses can help to reduce the variability of the estimated overall treatment effect by splitting the study population into more homogeneous subgroups.
Information about the individual treatment effect might be available from cross-over trials or from counterfactual analyses of parallel-group designs [2, 3]. These individual effects can then be linked to potentially predictive variables. In the absence of such information, most importantly in the case of parallel-group designs studied here, subgroup analyses can be seen as the search for or specification of treatment × covariate interactions and we proceed along this path. A covariate measures a patient characteristic that potentially explains the patient’s individual treatment effect. In the commonly applied models with linear predictors, such as the linear, generalised linear or linear transformation models, the specification of higher-order interaction terms and especially the subsequent inference are known to be burdensome. For non-categorical covariates, it is a priori unclear how one can derive a subgroup from a significant treatment × covariate interaction.
Automated interaction detection [4], today known as recursive partitioning methods or simply “trees”, was suggested as an interaction search procedure more than 50 years ago, and has had a very active development community ever since. Although the application of trees for subgroup identification seems to be straightforward, no generally applicable method is available [5]. The main technical problem is that classical trees were developed for identifying higher-order covariate interactions but additional work is required to restrict interactions to treatment × covariate interactions. Due to the non-parametric nature of most tree models, blending trees with the linear models typically used to describe the treatment effect is challenging.
While setting up such automated procedures for subgroup identification, one has to bear in mind that the impact of a covariate on the endpoint can be prognostic, predictive, or both. Prognostic factors have a direct impact on the endpoint, independent of the treatment applied. This corresponds to a main effect. A predictive factor explains a differential treatment effect, i.e. a treatment × covariate interaction term. Both the main and the treatment interaction terms are important for factors that are prognostic and predictive at the same time [6].
In our analysis, we aimed at detecting subgroups of patients suffering from amyotrophic lateral sclerosis (ALS) in which the subgroups differ in the effect of treatment with Riluzole, the only approved drug for ALS treatment today. The two endpoints of interest are a functional endpoint assessing the patient’s ability to handle daily life and the overall survival time. We estimated the overall treatment effect of Riluzole using four different base models; the choice of the model depended on the measurement scale of the endpoint. A normal generalised linear model (GLM) with log-link was used for the sum-score of the functional endpoint, and item-specific proportional odds models were used for the decomposed score. For the right-censored survival times, we used a parametric Weibull model and a semiparametric Cox model. Our aim was to partition these linear models with respect to the treatment effect parameter and to develop a segmented model that includes treatment × covariate interactions that describe the relevant subgroups.
We applied model-based recursive partitioning [7] to the functional and survival models describing the effect of Riluzole on ALS patients in order to obtain subgroups with a differential treatment effect. The main advantage of embedding our subgroup analysis into this general framework of model partitioning is that one can partition the base model used for analysing the overall treatment effect, regardless of the measurement scale of the endpoint. The method allows us to focus attention on predictive factors, while other terms, such as the effects of strata or nuisance parameters, can be held fixed.
Section 2 introduces the general framework for subgroup identification and compares the new procedure to methods published previously in the light of this general theoretical framework. In Section 3, we present results of our subgroup analysis of Riluzole treatment of ALS patients and discuss the patient subgroups and corresponding differential treatment effects found.
2 Model-based recursive partitioning for subgroup identification
Subgroup analyses require the definition of a parameter describing the treatment effect. In clinical trials, this parameter is typically already contained in the model that was defined in the study protocol for the analysis of the primary endpoint. The treatment effect was estimated in the primary analysis under the assumption that the corresponding parameter is universally applicable to all patients. In the presence of subgroups, this assumption does not hold and these patient subgroups differ in their treatment effect. If we assume that the different treatment effects can be understood as a function of patient characteristics, the patient subgroups can be identified by estimating this treatment effect function. Model-based recursive partitioning can be employed as a procedure for the estimation of such a treatment effect function and the identification of the corresponding patient subgroups. The name of the procedure comes from the nature of the algorithm that recursively partitions the initial model used for the analysis of the primary endpoint.
2.1 Model and algorithm
We started with a model
Estimating
where
In the presence of patient subgroups that differ in their treatment effect
Since we are searching for predictive and prognostic factors, we are only interested in subgroups that differ in the intercept or the treatment effect or both as explained in Section 2.2. With
where
Without any a priori knowledge about the partition
for a given model
If we can reject at least one of the
Since model-based recursive partitioning is a tree method, in the following we use topic-specific vocabulary, such as nodes. The root node contains all patients and is the basis for the initial model, inner nodes represent splits and leaf nodes contain the patients of the different subgroups and specify the partition-specific models. The paths from root to leaf nodes define the subgroups.
2.2 Content interpretation
A clearer picture of the interpretation of subgroup-dependent model parameters and distribution of the partial scores under unstable parameters is best given by means of a partitioned linear model discussed in the following.
Here
The segmented model we want to fit using model-based recursive partitioning reads
where
and thus to the same solution. Note that the partial score function with respect to the intercept is proportional to the least-square residuals and all further scores are proportional to the product of the residuals and the respective variable.
A partitioning variable can be predictive, prognostic, or both, and we have to consider the parameters in the model to understand the nature of a partitioning variable. Figure 1 shows examples for mean primary endpoints and the corresponding intercept
It is also interesting to take a closer look at the partial scores. Figure 2a shows the partial scores with respect to intercept and treatment parameter that result from a linear model
with
The partial scores with respect to the intercept
Here the procedure would split the partial score with respect to the intercept, although
If we focus on the prognostic variable
we see non-random patterns in both scores (see Figure 2c). Since the partial scores with respect to the treatment parameter are set to zero for treatment arm
These three examples show that splitting in the partial score with respect to the intercept does not give any information about whether the partitioning variable is predictive or prognostic. It also does not make sense to choose to split only in the score with respect to the treatment parameter because one might miss important cut-points. In order to be able to say whether a partitioning variable is predictive or prognostic, it is not enough to know which partial scores are responsible for the split. It is necessary to consider the model parameters in the segmented model. If the treatment parameter
2.3 Relation to established procedures
Traditional approaches for subgroup identification are also based on a model for the primary endpoint, but the segmentation is implemented by means of varying coefficients. More precisely, the model includes interactions between treatment and the patient characteristics
with
RECPAM [17, 18] goes a step further and fits such models by trees. In every node, a likelihood-ratio test is computed that compares the segmented model
to the constant model
for every possible segment
A method that is similar in spirit to model-based recursive partitioning, but is limited to normal linear models, is the Gs method [14] based on the GUIDE algorithm [19, 20]. Instead of using partial scores with respect to intercept and treatment effect, Gs uses only the least-square residuals (that is, only the partial score with respect to the intercept). In contrast to model-based recursive partitioning, Gs looks at the dichotomised (at zero) residuals separately in the two treatment arms. The independency between positive/negative residual signs and each partitioning variable is tested using a chi-squared test separately for each treatment. If the partitioning variable is at least ordinal, it is dichotomised by splitting at the mean. The optimal split variable chosen is the one that induces the highest sum of chi-squared statistics. Looking at the left panels of Figures 2a and 2b, one can imagine that in these situations the procedure may successfully find the subgroups. However, in a less clear situation and where the optimal cut-point is not near the mean of
Another recently proposed tree algorithm is qualitative interaction trees (QUINT [21]). QUINT searches for instabilities in the treatment parameter
SIDES (subgroup identification based on differential effect search) [23] and SIDEScreen [24] aim at identifying subgroups of patients with high benefit from a novum compared to the standard treatment. Although the subgroups are linked to hypercubes in the sample space of
More general approaches blending recursive partitioning with traditional models (known as hybrid, model, or functional trees in machine learning Gama [25]), include M5 [26], GUIDE [19], CRUISE [27], LOTUS [28] and maximum likelihood trees [29]. (Bayesian approaches can be found in Chipman, George, and McCulloch [30]) and (Bernardo, Bayarri, Berger, Dawid, Heckerman, Smith, and West [31]). Except GUIDE, none of these methods has been studied in the specific context of subgroup analyses so far.
3 Partitioning effects of Riluzole on ALS patients
ALS is a neurodegenerative disease that causes weakness, muscle waste and paralysis. Currently the only drug on the market for treating ALS is Riluzole (Rilutek). It slows down disease progression but only modestly prolongs life expectancy by about two months [32]. A more thorough investigation of the treatment effect of Riluzole in ALS patients is of great importance since a cure is not yet available and patients usually die within
Our analysis is based on patient information obtained from the PRO-ACT (Pooled Resource Open-Access ALS Clinical Trials) database [34], which contains data of ALS patients that were involved in one of several publicly- and privately-conducted clinical trials. The database provides information on patient survival, functional endpoint (the ALS functional rating scale), Riluzole use, demographics, family history, patient history, forced and slow vital capacity, laboratory data and vital signs. The data were fully de-identified and therefore the centres of data ascertainment are not given in the data set. The participants gave their informed consent, and study protocols were approved in the respective medical centres. The database was initiated by the non-profit organisation Prize4Life that aims at accelerating cure and drug development for ALS, for example through the DREAM-Phil Bowen ALS Prediction Prize4Life challenge [35].
The ALS Functional Rating Scale (ALSFRS [36]), is a widely used instrument for evaluating the functional status of patients with ALS even though the uni-dimensionality of the score seems questionable [37]. It is a sum-score of the following ten items: speech, salivation, swallowing, handwriting, cutting food and handling utensils, dressing and hygiene, turning in bed and adjusting bed clothes, walking, climbing stairs, and breathing. Each of these items can have values from zero to four, where four is normal and zero indicates the inability of performing the respective action. Hence, if the ALSFRS has a value
The survival time of patients was measured in days starting with the patient’s enrolment in one of the trials. For patients without survival information, we used the latest follow-up time given for the patient in the data as censoring time.
Model-based recursive partitioning was applied to models for the functional and survival endpoints. We allowed parameter instabilities in both the intercept and the Riluzole treatment effect. Bonferroni-adjusted permutation tests using test statistics of a quadratic form [13] were applied for assessing independence of the partial score functions and the partitioning variables and also for cut-point selection. The use of permutation tests for cut-point selection improves speed compared to the original suggestion of fitting and comparing models for all reasonable partitions [7]. We restricted the depth of the trees to two levels. Parameter estimates including confidence intervals are given for the final subgroups. Note that we are computing the confidence intervals after applying model selection through splitting into subgroups and thus the intervals should be interpreted with caution. For both endpoints, we used partitioning variables available at patient enrolment from the following groups of variables: demographics, family history, patient history, forced and slow vital capacity, laboratory data, and vital signs. We excluded patient records with missing values at the endpoints; the sample size was
3.1 ALSFRS
The ALSFRS six months after treatment start (
describes the expected relative change in the ALSFRS over the first six months under treatment. The treatment (Riluzole/no Riluzole) is indicated by
The time between disease onset and start of treatment, the forced vital capacity (FVC), and the phosphorus balance are the three partitioning variables selected for the tree given in Figure 3. The FVC value gives the volume of air in liters that can forcibly be blown out after full inspiration to the lung. A normal phosphorus balance is between
3.2 ALSFRS items
The model for the ALSFRS sum-score assumes that the effect of Riluzole is the same for the ten items that define the score. In a more fine-grained analysis, we decomposed the score into its ten items (each ranging between zero and four) and modelled each item by means of a proportional odds model. For one of the ten items assessed at six months, e.g.
where
As in the previous example, we needed to adjust for the baseline value
Therefore, we had a total of five different treatment parameters and
The implementation of the non-standard model in the theoretical and computational framework of model-based recursive partitioning was straightforward. For every node, we computed the five separate models for the respective baseline values for each of the ten items and extracted the partial scores. A stratified permutation test using the baseline values as independent blocks was used to assess parameter instability. The same procedure was applied for cut-off selection.
The resulting tree (on top of Table 1) contains splits in time between disease onset and treatment start and in the FVC value. The tree is in good agreement with the tree based on the ALSFRS (Figure 3). The third split variable is the lymphocyte percentage. Normal lymphocyte concentrations range from
Item | No. | Start | Node 3 | Node 4 | Node 6 | Node 7 |
Speech | 1 | 0 | ||||
1 | ||||||
2 | –0.27 (–1.15, 0.60) | |||||
3 | 0.33 (–0.33, 1.00) | 0.04 (–0.40, 0.47) | −0.27 (–1.12, 0.56) | −0.06 (–0.60, 0.48) | ||
4 | 0.84 (0.08, 1.59) | 0.11 (–0.28, 0.48) | ||||
Salivation | 2 | 0 | ||||
1 | ||||||
2 | 0.22 (–0.94, 1.40) | 0.43 (–0.48, 1.36) | 1.36 (–0.31, 3.11) | −0.20 (–1.28, 0.88) | ||
3 | 0.15 (–0.57, 0.87) | −0.26 (–0.76, 0.23) | −0.24 (–0.96, 0.47) | −0.05 (–0.58, 0.48) | ||
4 | 0.49 (–0.11, 1.07) | −0.03 (–0.39, 0.32) | ||||
Swallowing | 3 | 0 | ||||
1 | ||||||
2 | 0.35 (–0.62, 1.32) | −0.89 (–2.06, 0.21) | 1.51 (–0.33, 3.51) | −0.75 (–1.96, 0.43) | ||
3 | 0.57 (–0.06, 1.21) | −0.36 (–0.85, 0.12) | −0.40 (–1.17, 0.36) | 0.35 (–0.22, 0.93) | ||
4 | 0.62 (0.01, 1.23) | 0.15 (–0.49, 0.75) | 0.28 (–0.22, 0.75) | |||
Handwriting | 4 | 0 | −1.45 (–3.21, 0.37) | |||
1 | −1.15 (–2.54, 0.20) | −0.36 (–1.93, 1.25) | −0.08 (–1.26, 1.12) | |||
2 | −0.54 (–1.33, 0.25) | 0.04 (–0.71, 0.79) | ||||
3 | −0.13 (–0.78, 0.51) | 0.14 (–0.22, 0.49) | −0.08 (–0.70, 0.52) | −0.28 (–0.74, 0.17) | ||
4 | −0.10 (–0.71, 0.49) | 0.04 (–0.34, 0.42) | −0.14 (–0.65, 0.36) | |||
Cutting | 5 | 0 | ||||
1 | −0.01 (–1.19, 1.15) | −0.79 (–1.68, 0.07) | ||||
2 | 0.15 (–0.54, 0.85) | 0.48 (–0.14, 1.12) | ||||
3 | −0.03 (–0.76, 0.70) | −0.07 (–0.44, 0.30) | 0.10 (–0.60, 0.79) | 0.52 (0.03, 1.02) | ||
4 | 0.13 (–0.45, 0.72) | −0.09 (–0.49, 0.31) | −0.14 (–0.82, 0.53) | −0.21 (–0.74, 0.30) | ||
Hygiene | 6 | 0 | ||||
1 | ||||||
2 | −0.11 (–0.65, 0.43) | −0.37 (–0.89, 0.15) | ||||
3 | −0.22 (–0.88, 0.44) | −0.37 (–0.72, –0.03) | 0.14 (–0.50, 0.78) | 0.27 (–0.17, 0.71) | ||
4 | 0.26 (–0.40, 0.92) | 0.01 (–0.42, 0.44) | 0.14 (–0.71, 0.98) | 0.30 (–0.31, 0.90) | ||
Bed | 7 | 0 | ||||
1 | ||||||
2 | −0.03 (–0.96, 0.89) | 0.29 (–0.47, 1.04) | ||||
3 | 0.15 (–0.57, 0.87) | −0.32 (–0.71, 0.08) | −0.12 (–0.74, 0.49) | −0.05 (–0.45, 0.35) | ||
4 | −0.21 (–0.80, 0.36) | −0.10 (–0.45, 0.24) | −0.11 (–0.81, 0.57) | −0.35 (–0.90, 0.18) | ||
Walking | 8 | 0 | ||||
1 | ||||||
2 | 0.48 (–0.22, 1.16) | −0.04 (–0.56, 0.46) | ||||
3 | 0.11 (–0.33, 0.55) | 0.46 (–0.12, 1.04) | ||||
4 | 0.51 (–0.16, 1.18) | 0.13 (–0.27, 0.52) | ||||
Stairs | 9 | 0 | ||||
1 | −0.02 (–0.49, 0.45) | −0.01 (–0.72, 0.68) | −0.39 (–0.89, 0.10) | |||
2 | −0.80 (–2.10, 0.48) | 0.07 (–0.97, 1.12) | ||||
3 | 0.26 (–0.19, 0.72) | −0.65 (–1.46, 0.15) | −0.16 (–0.79, 0.46) | |||
4 | 1.01 (0.27, 1.77) | 0.06 (–0.35, 0.48) | 0.72 (–0.11, 1.55) | 0.29 (–0.32, 0.89) | ||
Respiratory | 10 | 0 | ||||
1 | ||||||
2 | ||||||
3 | ||||||
4 | 0.58 (–0.06, 1.24) | −0.08 (–0.44, 0.28) |
3.3 Survival time
We used both a Weibull model and a Cox model to identify subgroups with differing effects of Riluzole on the survival endpoint. The application of the model-based recursive partitioning framework in the Weibull model is straightforward and was introduced by Zeileis et al. [7]. Since the Cox model is a semiparametric model, where the intercept is a function of time, treated as a nuisance parameter omitted in the partial likelihood, there is no direct way of obtaining
3.3.1 Weibull model
The Weibull model is a transformation model of the form
where
Figure 4 shows that the patient’s age and again the time between onset and treatment start play a role in the partitioning. Older patients (
3.3.2 Cox model
The use of the Cox model in model-based recursive partitioning is a rather special case, since the baseline hazard in the Cox model is treated as an infinite-dimensional nuisance parameter and estimation is performed by minimisation of the negative partial log-likelihood. The Cox proportional hazards model is given by
where
Age and the time between disease onset and start of Riluzole treatment form the segments in this example. The tree in this example has almost the same splits as the tree in the previous example. Also estimates support the results of the Weibull example. Again, we did not see much difference between treated and untreated very young patients. For all other groups, Riluzole treatment led to a slight tendency for a lower risk of death (Figure 5).
4 Discussion
Model-based recursive partitioning allows the direct segmentation of the model describing the overall treatment effect as specified in the study protocol. This is the most important benefit of embedding subgroup analysis into this framework because it would be hard to explain why the overall treatment effect and the partitioned treatment effect have to be estimated by two different procedures. This renders the application of suboptimal models unnecessary, such as when a change score is analysed using linear models [21].
Although we are conceptually only interested in finding predictive factors, we think it is necessary to allow splits in the partial scores with respect to both intercept and treatment parameter. This procedure will also detect prognostic factors, but there is a higher chance of including all relevant predictive factors since one might miss prognostic factors when only the treatment scores are split. In our analysis, we decided on the nature of the partitioning variables (prognostic or predictive) only when we interpreted the results of the analysis.
In a model with more covariates than the treatment (e.g. strata), we would still split the partial scores with respect to intercept and treatment parameter for subgroup analyses. A theoretical assumption is then that the parameters that are not split stay constant. In practice, this assumption usually does not hold. It is generally also possible to split more than just the scores with respect to intercept and treatment parameter. Then the split variables are not restricted to being predictive or prognostic but may have an association with the effect of the other covariates.
In model-based recursive partitioning, the variable selection in each node is error controlled, i.e. the probability of selecting a partitioning variable for splitting, when actually all partitioning variables are independent of the scores, is at most as large as the nominal level. The only drawback of using multiple testing procedures is in cases where there are many possible partitioning variables that do not contain information, because with increasing number of noise variables the chance of detecting an actually existing subgroup goes down. The application of permutation tests has the advantage of taking the correlation structure among the partitioning variables into account. Furthermore, for small studies or small subgroups, the exact conditional
The statistical properties of the confidence intervals derived from the segmented model await further attention. Leeb and Pötscher ([40]) discuss the validity of inference after variable selection and claim that it is difficult if at all possible. Bai and Perron [41], who discuss the construction of confidence intervals after splitting up the data based on a break point in a single partitioning variable, argue that it is possible. In our approach we first search for the most appropriate partitioning variable (variable selection) and then search for the optimal split point (break point selection). To our knowledge there is no literature on inference after variable and break point selection and thus it is unclear if or how valid confidence intervals can be computed. In any case the results of such a subgroup analysis have to be confirmed in follow-up trials, which lowers the necessity of confidence intervals. To be conservative one can see the confidence intervals for parameters in the subgroup-specific models as shown in our examples as a range of possible values and hence as a measure of variability rather than significance [42].
It would be interesting to extend the framework of the PRO-ACT database of ALS studies to models for non-independent data, such as mixed models for longitudinal observations. This would allow ALS disease progression to be modelled over time, and also a potentially time-varying treatment effect to be assessed. In our way of modelling the functional endpoint, we include no information about patients that died within the first six months after treatment start. Joint modelling of the longitudinal functional endpoint and the survival endpoint is a means of combining all possible information [43].
Despite the deficits of model-based recursive partitioning for subgroup analysis discussed in this section, we think that the procedure as introduced and illustrated in this paper rather closely resembles the requirements for statistical procedures in this field as outlined in the EMA guideline [1]. In particular, it is the most generally applicable procedure with statistical error control and unbiased variable selection [13, 7]. With the available open-source implementation (see following section for details), the method can be applied straightforwardly elsewhere.
Computational details
An open-source implementation of all methods discussed in this paper and beyond is available in the partykit package hothorn_partykit:_2014. PRO-ACT data are available at https://nctu.partners.org/ProACT/ [44]. The source code for reading and cleaning the database is provided in the TH.data package [45]. The source code for the analyses is provided in the supplementary material. All computations were conducted using partykit (version 0.8-2) in the R system for statistical computing [46], version 3.1.2).
Listing 1: Code snippet for Weibull model in model-based recursive partitioning using the function ctree() from the partykit package.
## Function to compute Weibull model and return score matrix mywb <- function(data, weights, parm) { mod <- survreg(Surv(survival.time, cens) Riluzole, data = data, subset = weights > 0, dist = "weibull ")ef <- as.matrix(estfun(mod)[,parm]) ret <- matrix(0, nrow = nrow(data), ncol = ncol(ef)) ret[weights > 0,] <- ef ret } ## Compute tree tree <- ctree(fm, data = data, ytrafo = my.wb, control = ctree_control(maxdepth = 2, testtype = "Bonferroni"))
Acknowledgements
We are thankful to the organisers and participants of the “Workshop on Classification and Regression Trees” (March 2014), sponsored by the Institute for Mathematical Sciences of the National University of Singapore, for helpful feedback and stimulating discussions and to Karen A. Brune for improving the language. Financial support by the Swiss National Science Foundation (grant 205321_163456) is gratefully acknowledged.
References
1. European Medicines Agency. EMA guideline on the investigation of subgroups in confirmatory clinical trials (draft). Available at: http://www.ema.europa.eu/docs/en_GB/document_library/Scientific_guideline/2014/02/WC500160523.pdf, 2014.Search in Google Scholar
2. Gadbury GL, Iyer HK. UnitTreatment Interaction and Its Practical Consequences. Biometrics 2000;56:882–5.10.1111/j.0006-341X.2000.00882.xSearch in Google Scholar
3. Holland PW. Statistics and causal inference. J Am Stat Assoc 1986;81:945–60.10.1080/01621459.1986.10478354Search in Google Scholar
4. Morgan JN, Sonquist JA. Problems in the analysis of survey data, and a proposal. J Am Stat Assoc 1963;58:415–34.10.1080/01621459.1963.10500855Search in Google Scholar
5. Doove LL, Dusseldorp E, Van Deun K, Van Mechelen I. A comparison of five recursive partitioning methods to find person subgroups involved in meaningful treatment-subgroup interactions. Adv Data Anal Class 2014;8:403–25.10.1007/s11634-013-0159-xSearch in Google Scholar
6. Italiano A. Prognostic or predictive? It’s time to get back to definitions!. J Clin Oncol 2011;29:4718–4718.10.1200/JCO.2011.38.3729Search in Google Scholar PubMed
7. Zeileis A, Hothorn T, Hornik K. Model-based recursive partitioning. J Comput Graph Stat 2008;17:492–514.10.1198/106186008X319331Search in Google Scholar
8. Hastie T, Tibshirani R. Varying-Coefficient Models. J R Stat Soc Series B (Methodological) 1993;55:757–96. Available at: http://www.jstor.org/stable/2345993.10.1111/j.2517-6161.1993.tb01939.xSearch in Google Scholar
9. Zeileis A, Hornik K. Generalized M-fluctuation tests for parameter instability. Stat Neerl 2007;61:488–508.10.1111/j.1467-9574.2007.00371.xSearch in Google Scholar
10. Zeileis A, Hothorn T. A toolbox of permutation tests for structural change. Stat Papers 2013;54:931–54.10.1007/s00362-013-0503-4Search in Google Scholar
11. Hothorn T, Hornik K, Van de Wiel MA, Zeileis A. A Lego system for conditional inference. Am Stat 2006a;60:257–63.10.1198/000313006X118430Search in Google Scholar
12. Hothorn T, Hornik K, van de Wiel MA, Zeileis A. Implementing a class of permutation tests: The coin package. J Stat Software 2008;28:1–23.10.18637/jss.v028.i08Search in Google Scholar
13. Hothorn T, Hornik K, Zeileis A. Unbiased recursive partitioning: A conditional inference framework. J Comput Graph Stat 2006b;15:651–74.10.1198/106186006X133933Search in Google Scholar
14. Loh W-Y, He X, Man M. A regression tree approach to identifying subgroups with differential treatment effects. Stat Med 2015;34:1818–33.10.1002/sim.6454Search in Google Scholar
15. Foster JC, Taylor JM, Ruberg SJ. Subgroup identification from randomized clinical trial data. Stat Med 2011;30:2867–80.10.1002/sim.4322Search in Google Scholar
16. Kehl V, Ulm K. Responder identification in clinical trials with censored data. Comput Stat Data Anal 2006;50:1338–55.10.1016/j.csda.2004.11.015Search in Google Scholar
17. Ciampi A, Negassa A, Lou Z. Tree-structured prediction for censored survival-data and the Cox model. J Clin Epidemiol 1995;48:675–89.10.1016/0895-4356(94)00164-LSearch in Google Scholar
18. Negassa A, Ciampi A, Abrahamowicz M, Shapiro S, Boivin JF. Tree-structured subgroup analysis for censored survival data: Validation of computationally inexpensive model selection criteria. Stat Comput 2005;15:231–9.10.1007/s11222-005-1311-zSearch in Google Scholar
19. Loh W-Y. Regression trees with unbiased variable selection and interaction detection. Stat Sinica 2002;12:361–86.Search in Google Scholar
20. Loh W-Y. Improving the precision of classification trees. Ann Appl Stat 2009;3:1710–37.10.1214/09-AOAS260Search in Google Scholar
21. Dusseldorp E, Van Mechelen I. Qualitative interaction trees: A tool to identify qualitative treatment-subgroup interactions. Stat Med 2013;33:219–37.10.1002/sim.5933Search in Google Scholar PubMed
22. Dusseldorp E, Doove L, Van Mechelen I. quint: Qualitative Interaction Trees. Available at: http://CRAN.R-project.org/package=quint, R package version 1.0., 2013Search in Google Scholar
23. Lipkovich I, Dmitrienko A, Denne J, Enas G. Subgroup identification based on differential effect search – A recursive partitioning method for establishing response to treatment in patient subpopulations. Stat Med 2011;30:2601–21.10.1002/sim.4289Search in Google Scholar PubMed
24. Lipkovich I, tri enko AD. Strategies for identifying predictive biomarkers and subgroups with enhanced treatment effect in clinical trials using SIDES. J Biopharm Stat 2014;24:130–53.10.1080/10543406.2013.856024Search in Google Scholar PubMed
25. Gama J. Functional trees. Mach Learn 2004;55:219–50.10.1007/3-540-45650-3_9Search in Google Scholar
26. Quinlan JR. C4.5: Programs for Machine Learning. San Francisco, CA, USA: Morgan Kaufmann Publishers Inc, 1993.Search in Google Scholar
27. Kim H, Loh W-Y. Classification trees with unbiased multiway splits. J Am Stat Assoc 2001;96:589–604.10.1198/016214501753168271Search in Google Scholar
28. Chan K-Y, Loh W-Y. Lotus: An algorithm for building accurate and comprehensible logistic regression trees. J Comput Graph Stat 2004;13:826–52.10.1198/106186004X13064Search in Google Scholar
29. Su X, Wang M, Fan J. Maximum likelihood regression trees. J Comput Graph Stat 2004;13:586–98.10.1198/106186004X2165Search in Google Scholar
30. Chipman H, George E, McCulloch R. Bayesian treed models. Mach Learn 2002;48:299–320.10.1023/A:1013916107446Search in Google Scholar
31. Bernardo JM, Bayarri MJ, Berger JO, Dawid AP, Heckerman D, Smith AFM, et al. Bayesian Treed Generalized Linear Models. Bayesian Statistics 7: Proceedings of the Seventh Valencia International Meeting, 85, 2003.Search in Google Scholar
32. European Medicines Agency. Riluzole Zentiva: EPAR summary for the public. Available at:http://www.ema.europa.eu/docs/en_GB/document_library/EPAR_-_Summary_for_the_public/human/002622/WC500127609.pdf, 2012.Search in Google Scholar
33. Chiò A, Logroscino G, Hardiman O, Swingler R, Mitchell D, Beghi E, et al., and On Behalf of the Eurals Consortium. Prognostic factors in ALS: A critical review. Amyotroph Lateral Scler 2009;10:310–23.10.3109/17482960802566824Search in Google Scholar PubMed PubMed Central
34. Atassi N, Berry J, Shui A, Zach N, Sherman A, Sinani E, et al. The PROACT database: Design, initial analyses, and predictive features. Neurology 2014;83:1719–25.10.1212/WNL.0000000000000951Search in Google Scholar PubMed PubMed Central
35. Küffner R, Zach N, Norel R, Hawe J, Schoenfeld D, Wang L, et al. Crowd sourced analysis of clinical trial data to predict amyotrophic lateral sclerosis progression. Nat Biotechnol 2014;33:51–7.10.1038/nbt.3051Search in Google Scholar PubMed
36. Brooks BR, Sanjak M, Ringel S, England J, Brinkmann J, Pestronk A, et al. The amyotrophic lateral sclerosis functional rating scale – assessment of activities of daily living in patients with amyotrophic lateral sclerosis. Arch Neurol 1996;53:141–7.10.1001/archneur.1996.00550020045014Search in Google Scholar
37. Franchignoni F, Mora G, Giordano A, Volanti P, Chio A. Evidence of multidimensionality in the ALSFRS-R scale: A critical appraisal on its measurement properties using Rasch analysis. J Neurol Neurosurg Psychiatry 2013;84:1340–5.10.1136/jnnp-2012-304701Search in Google Scholar PubMed
38. Hothorn T, Jung HH. RandomForest4life: A Random Forest for predicting ALS disease progression. Amyotroph Lateral Scler Frontotemporal Degener 2014;15:444–52.10.3109/21678421.2014.893361Search in Google Scholar PubMed
39. Strobl C, Malley J, Tutz G. An Introduction to Recursive Partitioning: Rationale, Application and Characteristics of Classification and Regression Trees, Bagging and Random Forests. Psychol Methods 2009;14:323–48.10.1037/a0016973Search in Google Scholar PubMed PubMed Central
40. Leeb H, Pötscher BM. Model selection and inference: Facts and fiction. Economet Theory 2005;21:21–59.10.1017/S0266466605050036Search in Google Scholar
41. Bai J, Perron P. Computation and analysis of multiple structural change models. J Appl Economet 2003;18:1–22.10.1002/jae.659Search in Google Scholar
42. Lagakos SW. The challenge of subgroup analyses–reporting without distorting. N Eng J Med 2006;354:1667–9.10.1056/NEJMp068070Search in Google Scholar PubMed
43. Henderson R, Diggle P, Dobson A. Joint modelling of longitudinal measurements and event time data. Biostatistics 2000;1:465–80.10.1093/biostatistics/1.4.465Search in Google Scholar PubMed
44. Massachusetts General Hospital, N. C. R. I. Pooled resource open-access ALS clinical trials database. Available at: https://nctu.partners.org/ProACT/, 2013.Search in Google Scholar
45. Hothorn T. TH.data: TH’s Data Archive. Available at: http://CRAN. R-project.org/package=TH.data, R package version 1.0-4, 2014.Search in Google Scholar
46. R Core Team. R: A Language and Environment for Statistical Computing, R Foundation for Statistical Computing, Vienna, Austria. Available at: http://www.R-project.org/ 2014.Search in Google Scholar
47. Hothorn T, Zeileis A. partykit: A modular toolkit for recursive partitioning in R, Technical report. Available at: http://eeecon.uibk.ac.at/wopec2/repec/inn/wpaper/2014-10.pdf, 2014.Search in Google Scholar
©2016 by Seibold et al., Walter de Gruyter GmbH, Berlin/Boston
This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 3.0 License.