Investigating clinical heterogeneity in systematic reviews: a methodologic review of guidance in the literature

Systematic reviews sometimes apply statistical techniques to combine data from multiple studies resulting in a meta-analysis. Meta-analyses result in a point estimate, the summary treatment effect, together with a measure of the precision of results (e.g., a 95% confidence interval). These measures of precision represent the degree of variability or heterogeneity in the results among included studies. There are several possible sources of variability or heterogeneity among studies that are included in meta-analyses. Variability in the participants, the types or timing of outcome measurements, and intervention characteristics may be termed clinical heterogeneity; variability in the trial design and quality is typically termed methodological heterogeneity; variability in summary treatment effects between trials is termed statistical heterogeneity [1]. Methodological and clinical sources of heterogeneity contribute to the magnitude and presence of statistical heterogeneity [1].

Methodological heterogeneity hinges on aspects of implementation of the individual trials and how they differ from each other. For example, trials that do not adequately conceal allocation to treatment groups may result in overestimates in the meta-analytic treatment effects [2]. Significant statistical heterogeneity arising from methodological heterogeneity suggests that the studies are not all estimating the same effects due to different degrees of bias.

Clinical heterogeneity arises from differences in participant characteristics (e.g., sex, age, baseline disease severity, ethnicity, comorbidities), types or timing of outcome measurements, and intervention characteristics (e.g., dose and frequency of dose [1]). This heterogeneity can cause significant statistical heterogeneity, inaccurate summary effects and associated conclusions, misleading decision makers and others. As such, systematic reviewers need to consider how best to handle sources of heterogeneity [1]. For example, preplanned subgroup analyses, stratifying for similar characteristics of the intervention and participants, could tease-out important scientific and clinically relevant information [3].

Systematic reviews are frequently recognized as the best available evidence for decisions about health-care management and policy [3‐7]. Results of systematic reviews are often incorporated into clinical practice guidelines [5] and required in funding applications by granting agencies [6]. In spite of all this it appears health-care professionals and policy makers infrequently use systematic reviews to guide decision-making [8].

A limitation of many systematic reviews is that their content and format are frequently not useful to decision makers [8]. For example, while some guidance exists describing what to include in reports of systematic reviews (e.g., the PRISMA statement [9]), characteristics of the intervention that are necessary to apply their findings are infrequently provided [10‐13]. This has led to some preliminary work on how to extract clinically relevant information from systematic reviews [14]. Furthermore, systematic reviews commonly show substantial heterogeneity in estimated effects (statistical heterogeneity), possibly due to methodological, clinical or unknown features in the included trials [15]. While guidance exists on the assessment and investigation of methodological [1] and statistical heterogeneity [1, 16], little attention has been given to clinical heterogeneity.

We report a systematic review of suggested methods for investigating clinical heterogeneity in systematic reviews of controlled clinical trials. We also provide some guidance for systematic reviewers.

This project identified resources giving recommendations for investigating clinical heterogeneity in systematic reviews. We extracted their recommendations, assessed their risk for bias, and categorized and described the suggestions.

Our searches identified 2497 unique titles and abstracts; after screening, 101 papers were included in the review [17‐117]. These resources included statistical papers, methodological reviews, narrative reviews, expert opinions, learning guides, consensus-based guidelines and textbooks. Figure 1 describes details of the search and screening results. The very few disagreements on inclusion were easily resolved through discussion. Sixty-four (64.6%) of the resources (statistical, methodological, consensus guideline resources) were assessed for validation. Forty-one (64.1%) of these references were evaluated as being sufficiently validated.

Table 1 describes some basic characteristics of the included resources. The most common type of resource was statistical papers (42.4%), with narrative reviews/expert opinion papers being the next most common (29.3%). Most of the papers were published in the 2000s (70.1%), and statistical methods for investigating clinical heterogeneity were the most frequent types of suggestions across resources (73.7%). Table 2 reports a list of clinical variables suggested for investigating clinical heterogeneity and the number and types of resources suggesting each. General suggestions of clinically related variables, without identification of specific clinical covariates, were the most common across all included resources. Most suggestions were within distinct categories: participant level (e.g., age), intervention level (e.g., dose), or outcome level (e.g., event type, length of follow-up) covariates. A number of resources (N = 14) reported control event rate/baseline risk as being a covariate worth investigating.

Table 3 lists recommendations regarding the process of choosing clinical characteristics to investigate. Five or more resources suggested the following: a priori choice of clinical covariates (e.g., in the review protocol); look at forest plots for trials that may contribute to heterogeneity and then look for clinical characteristics therein; proceed with investigation regardless of results of formal testing for statistical heterogeneity; base clinical covariates on a clear scientific rationale (e.g., a pathophysiological argument); investigate a small number of covariates; base each covariate suggestion on an adequate number of trials (e.g., 10 trials was a common suggestion); use caution when interpreting the findings of investigations; consider the results of such investigations as exploratory, hypothesis generating and observational; and consider confounding between covariates.

Table 4 summarizes the types of statistical methods suggested for investigating clinical heterogeneity characteristics and the number of resources suggesting each. Many included resources made some mention of statistical methods of investigating aspects of clinical heterogeneity (N = 69/99, 69.7%). Also, many of these resources made general suggestions regarding the use of subgroup analyses (N = 18) and meta-regression (N = 16); however, the majority of these did not offer any specific recommendations. A wide variety of meta-regression techniques were suggested, many of which included simulated evidence or other forms of empirical testing. Several Bayesian approaches were suggested as well as several methods for individual patient data analysis [34, 48, 58, 63, 66, 69, 71, 75, 95]. Four textbooks appeared to be relatively comprehensive in their treatment of statistical recommendations [93‐95, 114].

Overall, we felt that there was some consensus across the resources regarding planning investigations, the use of clinical expertise, the rationale for choice of covariate, how to think through types of covariates, making a covariate hierarchy, post hoc covariate identification, statistical methods, data sources and interpretation of findings (See Table 5). We summarize the common recommendations that appeared in the literature to offer some preliminary guidance for systematic reviewers in Table 5 and we elaborate on several key areas in the discussion section below.

Sources appearing to be the most comprehensive in their discussion of recommendations for investigating clinical heterogeneity included the Cochrane Handbook[100] and the Centre for Reviews and Dissemination’s Guidance For Undertaking Reviews In Health Care[98] and the AHRQ Comparative effectiveness review methods: clinical heterogeneity[115].

A variety of decisions must be made when performing a systematic review. One such decision is how to deal with obvious differences among and within trials. Though a significant test for the presence of statistical heterogeneity (e.g., Q test) and a large degree of heterogeneity (e.g., I² > 75%) might obligate a reviewer to look for covariates to explain this variability, a nonsignificant test or a small I² (e.g., <25%) does not preclude the need to investigate covariate treatment effect interactions [35, 92, 97, 100]. That is, even with low statistical heterogeneity, there may still be factors that influence the size of the treatment effect, especially if there is a strong argument (i.e., pathophysiologic or otherwise) that some variable likely does have such an influence.

Observed or expected heterogeneity of treatment effects can be handled in several ways. The heterogeneity can be ignored and a meta-analysis conducted with a fixed-effects or random-effects model, or one can attempt to explain the heterogeneity through subgroup analyses, meta-regression or other techniques [25]. The latter moves the review away from overall statements of evidence to increasingly clinically applicable results and conclusions as well as new hypotheses for future research [75]. [Anello and Fleiss 28] make a clear distinction between meta-analyses with a goal of arriving at a common summary estimate of effect (“analytic meta-analyses”) and those focused on explaining why the effect sizes vary (“exploratory or causal meta-analyses”). The choice between these depends on the objective of the review, but it is clear that meta-analyses are more applicable to decision making (e.g., clinical, policy) when they are exploratory in nature [14, 28, 53, 75, 99]. The trials included in a systematic review may be so very similar that the summary effect estimate is the most reasonable and applicable metric [114]. But these cases are very rare, and therefore we would expect most questions asked and tested through meta-analytic methods should concern possible reasons for variation in effect [114].

Many resources were found that suggested methods for carrying out investigations of clinical heterogeneity in systematic reviews [17‐102]. There was great variety in the types of resources identified (statistical papers to commentaries) and in their potential for risk of bias. It was decided early to include any resource, no matter the design, methods, or publication type. For this reason many of the included resources might normally be considered at a high risk of bias (e.g., narrative reviews, expert opinions, learning guides and commentaries) and thus providing suggestions of questionable validity. But it was felt that these types of resources might provide the most valuable information on the subject of clinical heterogeneity. That is, investigating, and in particular choosing which clinical characteristics to investigate, requires clinical expertise, or at a minimum, knowledge of empirical evidence of some covariate of importance. The inclusion of these resources could be viewed as a drawback, but we saw it as a strength of this research. It was these resources that provided most of the suggestions regarding the methods for choosing or identifying clinical covariates to investigate (Table 3). The consensus-based guidelines provided most of the suggestions regarding the process of choosing or identifying clinical covariates, and the statistical papers, as might be expected, covered the majority of the specific statistical suggestions; but the textbooks also offered many suggestions in both areas. There was some consensus across resources, but only a small number of resources included a relatively comprehensive set of recommendations [15, 93, 94, 98]. Therefore, future research should be directed at developing a comprehensive and up to date set of guidelines to aid reviewers in investigating clinical heterogeneity. We summarize the common recommendations that appear in the literature to offer some preliminary guidance for systematic reviewers (Table 5).

We were surprised to see that the term clinical heterogeneity was relatively commonly used and consistently defined. We took our definition from several publications with which we were previously familiar [1, 3]. In some of the resources the term methodological heterogeneity was used synonymously with clinical heterogeneity, or clinical heterogeneity was considered to be one component of methodological heterogeneity. While this was infrequent in the literature, methodological aspects of heterogeneity include but go beyond clinical aspects or reasons for heterogeneity between trials. Thus, when describing reasons for heterogeneity that are related to the participants, intervention, outcomes or settings of the trial, these should be termed clinical aspects of heterogeneity. A consistency of terminology is mandatory for development of thought and investigation in this area. With terminology in place, the discussion can move to our recommendations.

When planning investigations of clinical heterogeneity in systematic reviews of controlled trials one should make such plans explicit, a priori, in the protocol for the review. We would suggest that protocols be published or registered in appropriate databases [118]. Next, it is reasonable and arguably beneficial, when organizing the review team, to include clinical experts or at a minimum, state a plan for consulting clinical experts during particular phases of the review (e.g., when choosing clinical covariates or during interpretation of findings). Furthermore, a set of clinical covariates should be chosen that have a clearly stated rationale for their importance (e.g., pathophysiological argument or reference to the results of a previous large trial). Review teams should think through the following categories to determine if related covariates might logically influence the treatment effect in their particular review: participant level, intervention level, outcome level, research setting, or others unique to the research question. Several resources offered conceptual mapping, idea webbing and causal modeling as possible methods for identifying important covariates and relationships between them [98, 112, 113]. Next, a hierarchy of clinical covariates should be formed and covariates investigated only if there is sufficient rationale and later a sufficient number of trials available. That is, covariates deemed more important than others on the basis of an explicitly stated rationale should be immediately included in such investigations, with other covariates being included when the number of trials is sufficient. A generally accepted rule of thumb is that 10 events per predictor variable (EPV) maintains bias and variability at acceptable levels. This rule derives from 2 simulation studies carried out for logistic and Cox modeling strategies [119‐121] and has been adapted to meta-regression [1, 114]. Therefore, it has been suggested that for each covariate there should be at least 10 trials to avoid potentially spurious findings [15]. Also, investigators should describe any plans to include additional covariates after looking at the data from included studies (e.g., forest plots). This might include an examination of summary tables or various types of plots [92, 93, 97, 98, 106], and it would be reasonable to include the clinical expert(s) at this stage to aid in the interpretation of the plotted data. Finally, how the results of any findings are going to be interpreted and used in the synthesis methods of the review needs to be explained. Most resources advise caution in interpreting these investigations, noting their exploratory nature, but when there is a clearly stated rationale, especially when derived from previous research, and sufficient trials are included, a priori planned investigations may improve applicability. Also, it was frequently suggested that the interpretation of the results of these investigations should consider confounds and important potential biases, the magnitude of the effect, confidence intervals and the directionality of the effect. Following these recommendations may lead to valid and reliable investigations of clinical heterogeneity and could improve their overall applicability and lead to future research that might test hypothesized subgroup effects.

A wide variety of statistical analyses are available for investigating clinical heterogeneity in systematic reviews of controlled clinical trials, and it is not within the scope of this paper to cover these in detail. Other resources cover this subject very well [15, 93, 95, 100, 114]. The sophistication of techniques is constantly growing, and an updated, precise summary of such methods is needed. Instead we will describe three available options frequently suggested by resources included in our review—subgroup analyses, meta-regression and the analogue to the analysis of variance (ANOVA)—and comment upon methods for exploring control group event rate.

Subgroup analyses involve separating trials into groups on the basis of some characteristic (e.g., intervention dose) and then performing separate meta-analyses for each group. This test provides an effect estimate within subgroups and a significance test for that estimate; it does not provide a test of variation in effect due to covariates. The greater the number of significant tests performed, the greater the likelihood of type 1 errors. There are some suggestions in the literature for how to control for this (e.g., Bonferroni adjustments [48]). To test for differences between subgroups a moderator analysis must be done. Moderator analyses include meta-regression and the analogue to the ANOVA, among other techniques (e.g., Z test [114]). Meta-regression is used to assess the impact of one or more independent variables (e.g., age or intervention dose) upon the dependent variable, the overall treatment effect [62]. Independent variables may be continuous or categorical, the latter expressed as a set of dummy variables with one omitted category. Several modeling strategies are available for performing meta-regression [100, 108, 122]. The results of meta-regression indicate which variables influence the summary treatment effect, how much the summary effect changes with each unit change in the variable and the p-value of this influence. It has been suggested that at least 10 trials per covariate are needed to limit spurious findings, due to the low statistical power of meta-regression, and a nonparametric test has been suggested when this tenet is not fulfilled [30] Also, one needs to consider the problems associated with ecological bias when performing meta-regressions on patient levels variables [40]. Finally, the analogue to the ANOVA examines the difference in the effect between categorical levels of some variable using identical statistical methods as a standard ANOVA [94].

The literature suggests many methods for examining the influence of the control event rate or baseline risk, which is considered an aggregate measure of known (e.g., age and disease severity) and unknown variables [15, 43, 93]. It has been argued that these examinations provide little import to clinical practice since the influence of any possible causative variables is aggregated and therefore the effect of individual covariates is unknown [15]. Also, the influence of the control event rate on the summary affect is affected by regression to the mean, and sophisticated statistical procedures are required to deal with this [15, 43, 93].

Bayesian approaches to meta-regression and hierarchical Bayes modeling, among other areas, appear to be well represented in the literature [66, 71, 95], as well as more general resources for Bayesian meta-analytic techniques [95, 123]. These methods are developing rapidly; therefore, frequent summaries of these important techniques are required as a resource to reviewers.

Finally, we would like to note suggestions in the literature concerning the utility of aggregate patient data (APD) versus individual patient data (IPD). Several resources give general recommendations regarding use of IPD when exploring characteristics that could be considered aspects of clinical heterogeneity [15, 74‐76, 95, 97]. Some empirical evidence supports these recommendations [40, 66, 124, 125]. When IPD is available, it should be used as a basis to investigate aspects of clinical heterogeneity at the patient level (e.g., demographic characteristics) so as to avoid ecological bias associated with summary APD. It is reasonable to use APD for trial-level covariates (e.g., intervention characteristics) that can be considered aspects of clinical heterogeneity. In addition, there may be opportunities to strategically use APD together with IPD to avoid the significant, and sometimes insurmountable, effort required to collect complete IPD [71].

Finally, in relation to the suggestions above for including clinical expertise in systematic reviews, we feel it is the responsibility of each therapeutic discipline to create a repository of variables to consider when exploring effect variation in systematic reviews. Such warehousing of clinically important covariates would serve as an important resource, allowing systematic reviewers and clinical trialists to explore nuances in treatment effect that might inform clinical decision making, and allowing for increased applicability of findings.

In summary, although many recommendations are available for investigating clinical heterogeneity in systematic reviews of controlled clinical trials, there is a need to develop a comprehensive set of recommendations for how to perform valid, applicable, and appropriate investigations of clinical covariates [7, 14]. This will improve the applicability and utilization of systematic reviews by policy makers, clinicians, and other decision makers and researchers who wish to build on these findings.