 Review
 Open Access
 Published:
Patientreported outcomes in metaanalyses Part 2: methods for improving interpretability for decisionmakers
Health and Quality of Life Outcomes volume 11, Article number: 211 (2013)
Abstract
Systematic reviews and metaanalyses of randomized trials that include patientreported outcomes (PROs) often provide crucial information for patients, clinicians and policymakers facing challenging health care decisions. Based on emerging methods, guidance on improving the interpretability of metaanalysis of patientreported outcomes, typically continuous in nature, is likely to enhance decisionmaking. The objective of this paper is to summarize approaches to enhancing the interpretability of pooled estimates of PROs in metaanalyses. When differences in PROs between groups are statistically significant, decisionmakers must be able to interpret the magnitude of effect. This is challenging when, as is often the case, clinical trial investigators use different measurement instruments for the same construct within and between individual randomized trials. For such cases, in addition to pooling results as a standardized mean difference, we recommend that systematic review authors use other methods to present results such as relative (relative risk, odds ratio) or absolute (risk difference) dichotomized treatment effects, complimented by presentation in either: natural units (e.g. overall depression reduced by 2.4 points when measured on a 50point Hamilton Rating Scale for Depression); minimal important difference units (e.g. where 1.0 unit represents the smallest difference in depression that patients, on average, perceive as important the depression score was 0.38 (95% CI 0.30 to 0.47) units less than the control group); or a ratio of means (e.g. where the mean in the treatment group is divided by the mean in the control group, the ratio of means is 1.27, representing a 27% relative reduction in the mean depression score).
Introduction
Clinical trials evaluating medical treatments and health interventions increasingly incorporate selfreported measures from patients, often referred to as patientreported outcomes (PROs). A PRO is defined as “any report of the status of a patient’s health condition that comes directly from the patient without interpretation of the patient’s response by a clinician or anyone else” [1]. Systematic reviews and metaanalyses of clinical trials often include PROs. In Part 1 of this series, we addressed the importance of PROs for health care decisionmaking, illustrated the key risk of bias issues that systematic reviews of PROs should consider and provided guidance on combining PROs in metaanalyses [2]. Part 1 used examples of PROs employed in assessing and summarizing postoperative pain and chronic obstructive pulmonary disease outcomes, while in this article, in addition to using examples of chronic obstructive pulmonary disease–we primarily use an example of summarizing PROs from clinical trials in depression. The structure of this article borrows from a recent article we published on preparing Summary of Findings tables for systematic reviews of continuous outcomes prepared by the Grading of Recommendations Assessment, Development and Evaluation (GRADE) working group [3]. Summary of Findings tables are an approach the GRADE working group developed for the presentation of findings particular to each outcome of interest in systematic reviews and metaanalysis. The tables have been developed for the presentation of continuous and dichotomous outcomes [3, 4]. The purpose of this article is to summarize five presentation approaches to enhancing the interpretability of pooled estimates of PROs.
Methods for improving the interpretability of pooled data
Metaanalyses of clinical trials routinely provide enough information for decisionmakers to evaluate the extent to which chance can explain apparent differences between interventions. The interpretation of the magnitude of treatment effects is typically more challenging. First, if trials have used the same instrument, decisionmakers may have difficulty interpreting the size of the effect. For instance, if told that the weighted mean difference between rehabilitation and standard care in a series of randomized trials using the Chronic Respiratory Questionnaire (CRQ) was 1.0 (95% CI 0.61.5), many readers would have no idea if this represents a trivial, small but important, moderate, or large effect [5–7].
The situation becomes even more challenging when trials use different instruments to measure the same or similar constructs. For instance, there are at least five instruments available to measure healthrelated quality of life (HRQoL) in patients with chronic obstructive respiratory disease (Chronic Respiratory Questionnaire, Clinical COPD Questionnaire, Pulmonary Functional Status and Dyspnea Questionnaire, Seattle Obstructive Lung Disease Questionnaire, St George’s Respiratory Questionnaire) [8]. We will deal with these two situations–all trials having used the same instrument, and trials having used different instruments–in turn.
Summarizing a single PRO: beyond a mean difference and statistical significance
On occasion, individual studies using continuous variables will provide data that facilitate creating meaningful dichotomies. For example, studies of the impact of thrombolytic therapy after stroke typically use the Rankin instrument that classifies patients into one of six categories of disability from no symptoms to severe handicap requiring constant attention. Authors of a systematic review evaluating the impact of thrombolytic therapy in patients with stroke dichotomized the Rankin instrument, creating a “bad outcome” category of those dead, or moderately or severely disabled (which they labelled as “dependent”) and a “good outcome” category of those with no symptoms, no significant disability, slight disability or moderate disability [9]. The reviewers were therefore able to present results showing that thrombolytic therapy significantly reduced the proportion of patients who were dead or dependent at the end of 3 to 6 months of followup (OR 0.81, 95% CI 0.73 to 0.90). This presentation facilitates interpretation by the review’s readers. A priori, reviewers should choose and justify their threshold when dichotomizing PROs, and consider conducting sensitivity analyses providing results for reasonable and extreme thresholds to support a better understanding of the generalizability of the results.
When authors do not provide information that would facilitate meaningful dichotomies, the systematic reviewer can aid interpretation by reporting the range of possible results, and the range of means in treatment and control groups in the studies. Particularly useful–if it is available–is an estimate of the smallest difference that patients are likely to consider important (the minimally important difference or MID). There are a variety of methods for generating estimates of the MID [10, 11], the application of which can lead to statements such as the following in a systematic review of the impact of respiratory rehabilitation in patients with chronic lung disease on HRQoL: “for each of the Chronic Respiratory Questionnaire domains (dyspnea, fatigue, emotional function and mastery), the common effect size exceeded the MID (0.5 point on the 7point scale).” Authors also reported that for each of the Chronic Respiratory Questionnaire domains, the lower limit of the confidence interval around the common treatment effect exceeded the MID (e.g. dyspnea domain: 1.0; 95% CI 0.81.2) [12].
Although this is very helpful, it potentially tempts clinicians to make inappropriate inferences. If the MID is 0.5 and the mean difference between treatments is 0.4, clinicians may infer that nobody benefits from the intervention; if the mean difference is 0.6, they may conclude that everyone benefits. Both inferences are a misinterpretation as they ignore the distribution of benefit between individuals. We suggest the following guide for interpretation given a 0.5 MID: if the pooled estimate is greater than 0.5, and one accepts that the estimate of effect is accurate, many patients may gain important benefits from treatment. If the estimate of effect lies between 0.25 and 0.5, the treatment may benefit an appreciable number of patients. As the pooled estimate falls below 0.25 (i.e. 50% of the MID), it becomes progressively less likely that an appreciable numbers of patients will achieve important benefits from treatment.
More than one PRO: beyond a standardized mean difference & statistical significance
As the prior discussion pointed out, when pooling across different PROs that measure a common construct the weighted mean difference is much more challenging to generate and we therefore replace it with a unitless measure of effect called the standardized mean difference (SMD) or “effect size”. This involves dividing the difference between the intervention and control means in each trial (i.e., the mean difference) by the estimated betweenperson standard deviation (SD) for that trial [13]. The SMD expresses the intervention effect in SD units rather than the original units of measurement, with the value of a SMD depending on both the size of the effect (the difference between means) and the SD of the outcomes (the inherent variability among participants). This approach has a number of limitations. First, decisionmakers will not have an intuitive sense of the importance of the effect on the basis of the SD unit report. Second, it has statistical limitations (the same effect will appear different if population heterogeneity across eligible trials differs) [14].
Unfortunately, there is no fully satisfactory way of providing a sense of the magnitude of effect for a PRO when one has had to resort to effect sizes to generate a summary estimate. One can offer readers standard rules of thumb in interpretation of effect sizes (for instance 0.2 represents a small effect, 0.5 a moderate effect, and 0.8 a large effect [15] or some variation (for instance, <0.40 = small, 0.40 to 0.70 = moderate, >0.70 = large). However, effect size interpretations are often diseasespecific and contextspecific, further warranting an explanation for the reader. Another, perhaps even less satisfactory approach suggests that a standardized mean difference of 0.5 approximates, in many cases, the MID [16, 17]. It is, however, very unlikely that a single SD ratio (explained below) will apply to all instruments.
When at least one instrument has an established anchorbased MID, the MID to SD ratio (SD ratio) may provide an estimate of MID values for instruments without an established MID. For a given PRO instrument, the SD ratio is the anchorbased MID divided by the baseline SD (or, if not reported, the endoftreatment SD for the control group). When several SDs are available from a number of trials, a median SD ratio can be computed, and can be used to estimate the MID for a PRO instrument for which an anchorbased MID is not established. This is done by multiplying the SD by the median SD ratio [18]. This method assumes that the SD ratio is relatively constant across a range of PRO instruments measuring the same or similar constructs (e.g. diseasespecific quality of life) in similar populations. For instance, the SD ratios based on the four instruments with known anchorbased MIDs were 0.26 (St George’s Respiratory Questionnaire), 0.51 (Chronic Respiratory Questionnaire), 0.34 (Montgomery Asberg Depression Rating Scale (MADRS)) and 0.86 (17item Hamilton Rating Scale for Depression (HRSD)). These findings suggest that a single SD ratio based on average ratios between MIDs and baseline SDs is very unlikely to apply to all instruments and advocate the need for sensitivity analyses to explore the extent to which pooled estimates are robust to a variety of MID estimates [18].
Many authors have proposed alternatives to the SMD that produce summary estimates that clinicians can interpret more easily, some of which rely on standard deviations being similar across trials, and some of which do not [14, 18–22]. Thus far, alternatives to the SMD have seen limited use and few studies have compared the SMD approach to the available alternatives [14, 19, 23–26].
Despite their limited use, the alternative approaches to summarizing results to enhance interpretability can be very useful. Consider, for instance, a systematic review assessing paroxetine vs placebo for the treatment of major depression in adults, which included 34 randomized trials employing the HRSD (n = 30) and the MADRS (n = 4) [27]. The MADRS ratings can be added to form an overall score ranging from 0 to 60; whereas for the HRSD, a number of versions exist, the most common being the 17item HRSD, with overall scores ranging from 0 to 50 [28]. The majority of the included trials employing the HRSD used the 17item version. Investigators have established 3 as the anchorbased MID for the 60point Montgomery Asberg Depression Rating Scale [29] and 7 as the anchorbased MID for the 50point Hamilton Rating Scale for Depression [30]. Providing pooled estimates of effect and making results interpretable for decisionmakers mandates use of one of five available presentation approaches that we will summarize here, the merits of which–and our associated recommendations–are presented in Table 1. The five presentation approaches discussed are: standard deviation units (i.e. the standardized mean difference); conversion to the natural units of the most common instrument; conversion to dichotomized relative and absolute effects; ratio of means; and minimally important difference units.
Presentation approaches
Standard deviation units  standardized mean difference
One way of generating a pooled estimate when trials have measured the same construct with different instruments is to divide the difference between the intervention and control means (i.e., the difference in means) in each trial by the estimated betweenperson standard deviation (SD) (see row A in Table 2 [13]. This measure is often referred to as the standardized mean difference (SMD) or Cohen’s effect size.
Presenting results in SD units (as an SMD) is by far the longest standing and most widely used approach and is recommended in the Cochrane Handbook [13]. Calculating and presenting results in SD units has, however, major limitations. First, clinicians and their patients are unlikely to be able to relate to this way of presenting results [26]. Second, if the variability or heterogeneity in the severity of patients’ condition (and thus the variability in scores on the chosen outcome) varies between trials, the SDs will also vary. As a result, trials that enrol heterogeneous groups of patients will yield smaller SMDs than trials enrolling less heterogeneous patients, even if the actual (not standardized) mean difference estimates–and thus the absolute estimate of the magnitude of treatment effect–is similar across all trials. Finally, if very homogenous populations are enrolled, SD units can give a misleading, inflated impression of the magnitude of treatment effect.
In Table 2, the presentation in SD units suggest a small treatment effect. The structure of the Summary of Findings table, however, is not well suited to this presentation. If authors use the SMD, it is not sensible to present absolute values in the intervention and comparison groups because studies have used different measurement instruments with different units. One approach to this dilemma, presented in Table 2, is to present the SMD in place of the two columns usually devoted to absolute rates. An alternative is to present the median value from the studies that used the most familiar measure of the concept in the control group column, and the SMD in the intervention group column. To aid interpretability of a metric unfamiliar to clinicians or patients, a comment provides a ruleofthumb guide to the significance of various effect sizes [15] (see row A, Table 2).
Conversion into units of the most commonly used instrument
A second approach (see row B in Table 2) converts the effect size back into the natural units of the outcome measure most familiar to the target audience(s). There are two statistical approaches to making the conversion. One calculates the absolute difference in means by multiplying the SMD by an estimate of the SD associated with the most familiar instrument. For example, one might assume that the HRSD, a 0 to 50 point measure with evidence of reliability and validity, is the most familiar depression instrument among decisionmakers [31]. In this case the magnitude of effect for the chosen instrument is 2.47. This result would be of limited use without knowledge of the MID, and thus the comment includes the estimated MID (7 points) [30], suggesting a small, and perhaps very small, effect (row B, Table 2).
The other statistical approach makes a simple conversion–before pooling and without calculating the SMD–of other instruments to the units of the most familiar instrument [25]. In this case, we chose the Hamilton Rating Scale for Depression, and rescaled the mean and SD of the other instruments to HRSD units. Given the MID of the HRSD (7 units), the mean difference in change of 2.50 suggests a small treatment effect of paroxetine [30].
This second approach, presenting in units of the most familiar instrument, may be the most desirable when the target audience(s) have extensive experience with that instrument, particularly if the MID is well established [10]. Nevertheless, the natural unit presentation may, in relation to the MID, still be misleading. In this case, the absolute difference is less than half the MID. This may lead clinicians to conclude the effect of treatment is trivial. While it is correct that the effect is small, as indicated above, it may still be important. For instance, a mean difference of 2.50 units in the HRSD (in which the MID is 7.0) is translated into a difference of the proportion of patients benefiting in experimental and control groups of 9.2%, and thus a number needed to treat (NNT) of approximately 11.
Conversion to dichotomized relative and absolute effects
A third approach (see row C in Table 2) converts the continuous measure into a dichotomy and thus allows calculation of relative and absolute effects on a binary scale. One method to generate a dichotomy from continuous data relies on the SMD and assumes that results of both treatment and control group are normally distributed and have equal variances [21, 32]. Metaanalysts usually make these assumptions when they calculate SMDs [21]. We have used this approach in Table 2, row C, and it suggests a small relative effect and a small but still potentially important absolute effect. This approach has the advantage that you can apply it easily by consulting Tables 3 and 4, which provides the relation between the SMD, control group response rate, and the resulting risk difference. Table 3 presents the conversion when the outcome is undesirable (e.g. depression) and Table 4 when the outcome is desirable (e.g. response to treatment).
This approach suffers from three important limitations. First, the dichotomous outcome that the intervention is decreasing is often not selfevident from the continuous outcome from which it is derived. We obtain a difference in the proportion of patients in intervention and control groups above some threshold, but the choice of that threshold is often arbitrary. In this example (Table 2, row C), fortunately, we can specify the threshold as an important improvement in depression (i.e. a change of 1 MID or more is representative of a 7 point difference on the HRSD). Second, the method requires investigators to specify the proportion of control patients with an improvement of at least one MID. Choosing this proportion may also be difficult. For instance, if one knows that control group depression scores varied from 23 to 44, with standard deviations around 12, how is one to decide the proportion of patients who failed to experience an important improvement with placebo? One possible approach would be, as a first step, to convert the mean value of the PRO in the control group into proportion of patients experiencing an improvement of at least one MID, for each of the studies included in the metaanalysis [32]. Reviewers could then use the median proportion across all studies for the conversion of the overall SMD [21]. The latter problem is ameliorated to some extent because only at the extremes of control proportions do the proportions benefiting change substantially. A third limitation is that the approach, by relying on the SMD, is vulnerable to whether study populations had very similar scores on the outcome of interest, or whether scores were widely variable.
Other statistical approaches also rely on the SMD to generate dichotomous presentations for continuous outcomes [22, 33]. They share similar limitations, with the exception that they do not require specification of the control group response rate, and one approach becomes unstable when the underlying control group response rate is less than 20% or greater than 80% [22].
Another strategy for creating dichotomies and generating estimates of relative and absolute effect relies on knowledge of the MID. In applying the approach, we assume normal distributions of data and then calculate the proportions of participants in the intervention and control groups in each study that demonstrated an improvement greater than the MID [25]. The results are then pooled across studies. Applying this approach in Table 2, findings suggest small to moderate relative and absolute benefit in depression as a result of paroxetine therapy (Odds Ratio 1.64; 95% CI 1.47 to 1.84; Risk Difference 0.11; 95% CI 0.07 to 0.16, in favor of paroxetine).
If one only has posttest data (rather than magnitude of change), one can apply this approach if evidence exists regarding meaningful thresholds. For instance, if one knows that people with scores of less than 8 on the HRSD are considered to be not depressed, one could examine the proportion of individuals below that threshold.
If such meaningful thresholds do not exist, one can still use posttest data if one assumes that the minimally important change within an individual corresponds, on average, to the minimally important difference between individuals. Making this assumption, one can calculate the difference in the proportion who benefit in intervention and control. To do this, one takes the mean value in the control group plus one MID unit, and calculates the proportion of patients in each group above that threshold.
Ratio of means
A fourth approach (see row D in Table 2) may appeal to clinicians: calculate a ratio of means (RoM) between the intervention and control groups [20]. Advantages of RoM include the ability to pool studies with outcomes expressed in different units, avoiding the vulnerability of heterogeneous populations that limits approaches that rely on SD units, and ease of clinical interpretation. However, a limitation of this RoM method is that it is designed for posttest scores only.
It is possible to calculate a ratio of change score if both intervention and control groups change in the same direction in each relevant study, and this ratio may sometimes be informative. Limitations include: i) the unlikelihood of intervention and control group changes in the same direction in all studies and ii) the possibility of misleading results if the control group change is very small–in which case, even a modest change in the intervention group will yield a large and therefore misleading ratio of mean changes.
In the paroxetine for depression example (Table 2), the ratio of means approach suggests a 27% increase in the mean depression score–meaning that those receiving paroxetine have a 27% decrease in depression manifestations relative to the placebo control group, an effect that strikes us as moderate and important.
Minimally important difference units
A final strategy pools across studies in the same way as the SMD, but instead of dividing the mean difference of each study by its SD, it divides by the MID associated with that outcome [14]. The final output, instead of being in SD units, is in MID units. This approach avoids the problem of varying SDs across studies that may distort estimates of effect in approaches that rely on the SMD. It may, in addition, be more easily interpretable though it risks the possibility that a difference less than the MID may be interpreted as trivial when a substantial proportion of patients have achieved an important benefit. This is almost certainly an inaccurate interpretation, as conversion into an absolute risk difference and NNT would demonstrate (in this case a risk difference of 0.11 equates to an NNT of 9). In addition, to the extent that the MID estimate is not based on secure evidence, the approach becomes more questionable [18]. As stated in the comment in Table 2 (row E), the result for paroxetine for depression is an effect less than half of one MID, suggesting a small treatment effect. We suggest the following guide for interpretation: if the pooled estimate is greater than 1 MID unit, many patients are likely to gain important benefits from treatment. If the estimate of effect lies between 0.5 and 1 MID unit, the treatment may benefit an appreciable number of patients. As the pooled estimate falls below 0.5 MID units it becomes progressively less likely that an appreciable numbers of patients will achieve important benefits from treatment.
Natural frequencies and numbers needed to treat
A systematic review of the literature suggests that natural frequencies (× of 100 people not taking any osteoporotic drug will suffer a hip fracture over a three year period) optimizes understanding for most patients and health professionals [34]. Another approach for readers who are familiar with the measure is to present the NNT (the inverse of the proportion benefiting) [23]. Any approach that yields a proportion can be converted to NNTs. Furukawa offers an approach based on the binomial and equal variance assumptions, which metaanalysts usually presuppose when they resort to standardized mean differences [21, 35]. Tables 3 and 4 shows the results of this method, which provides the relation between the SMD, control group response rate, and the resulting risk difference. Table 3 presents the conversion when the outcome is undesirable (e.g. depression) and Table 4 when the outcome is desirable (e.g. response to treatment). The NNT can be derived from the inverse of the risk difference.
Summary and recommendations for enhancing interpretation
We have provided an overview of available methods, including the strengths and limitations of the approaches, for improving the interpretability of pooled estimates of PROs when trials measure effects using the same instrument as well as a diversity of instruments. A complete summary of strengths and limitations of each of the methods for pooling diverse instruments, including details of the underlying statistical assumptions and methods, is available in an earlier review [25].
When trials all use the same PRO it is important to report results beyond a mean difference and statistical significance as suggested above. When primary studies have employed more than one instrument it will almost certainly be informative to report one or more alternatives to the SMD. Calculation and reporting of several approaches will, if the estimate of effect is of apparently similar magnitude, be reassuring (and if they are not, will present a challenge that reviewers should address). Of the two approaches for converting to natural units of the most familiar instrument, we recommend rescaling the observed means and SDs in the intervention and control groups over multiplying the SMD by an estimate of the SD associated with the most familiar instrument because the former approach does not depend on similarity of patient heterogeneity between studies.
Because of its familiarity to clinicians, in most instances reviewers might choose to present one of the measures that generates relative and/or absolute measures of effect. Of these approaches, if all instruments have an established MID, we recommend presenting results as a risk difference with corresponding risks, presented as a natural frequency, in the experimental group and control group as illustrated in Table 2. Consideration of the relative advantages and disadvantages of each presentation method when pooling PRO data will help ensure that data is interpretable to patients, clinicians and other key decisionmakers in the health care domain.
Abbreviations
 CRQ:

Chronic respiratory questionnaire
 HRSD:

Hamilton rating scale for depression
 HRQoL:

Healthrelated quality of life
 MID:

Minimal important difference
 MADRS:

Montgomery Asberg depression rating scale
 SMD:

Standardized mean difference.
References
US Food and Drug Administration (FDA) Guidance for Industry: PatientReported Outcome Measures: Use in Medical Product Development to Support Labeling Claims. Rockville, MD: Department of Health and Human Services (US), Food and Drug Administration, Center for Drug Evaluation and Research (CDER); 2009. Accessed 5 Sept, 2013 http://www.fda.gov/downloads/Drugs/GuidanceComplianceRegulatoryInformation/Guidances/UCM193282.pdf]
Johnston BC, Patrick DL, Busse JW, Schünemann HJ, Agarwal A, Guyatt GH: Patientreported outcomes in metaanalyses–part 1: assessing risk of bias and combining outcomes. BMC Health Qual Life Outcomes 2013,11(1):109. 10.1186/1477752511109
Guyatt GH, Thorlund K, Oxman AD, Walter SD, Patrick D, Furukawa TA, Johnston BC, Karanikolas P, Vist G, Kunz R, Brozek J, Meerpohl J, Akl EA, Christensen R, Schünemann HJ: Preparing summary of findings tables: continuous outcomes. J Clin Epidemiol 2013,66(2):173–183. 10.1016/j.jclinepi.2012.08.001
Guyatt GH, Oxman AD, Santesso N, Helfand M, Vist G, Kunz R, Brozek J, Norris S, Meerpohl J, Djulbegovic B, AlonsoCoello P, Post PN, Busse JW, Glasziou P, Christensen R, Schünemann HJ: GRADE guidelines: 12: preparing summary of findings tablesbinary outcomes. J Clin Epidemiol 2013,66(2):158–172. 10.1016/j.jclinepi.2012.01.012
Brozek JL, Guyatt GH, Schünemann HJ: How a wellgrounded minimal important difference can enhance transparency of labelling claims and improve interpretation of a patientreported outcome measure. BMC Health Qual Life Outcomes 2006, 27: 4–69.
Schünemann HJ, Akl EA, Guyatt GH: Interpreting the results of patientreported outcome measures in clinical trials: the clinician’s perspective. BMC Health Qual Life Outcomes 2006, 4: 62. 10.1186/14777525462
Guyatt GH, Juniper EF, Walter SD, Griffith LE, Goldstein RS: Interpreting treatment effects in randomised trials. BMJ 1998,316(7132):690–693. 10.1136/bmj.316.7132.690
Patient Reported Outcomes and Quality of Life Instruments Database. . Accessed 3 Sept, 2013 http://www.proqolid.org/proqolid/]
Wardlaw JM, Murray V, Berge E, Del Zoppo GJ: Thrombolysis for acute ischaemic stroke. Cochrane Database Syst Rev 2009, 4: CD000213.
Guyatt GH, Osoba D, Wu AW, Wyrwich KW, Norman GR, Clinical Significance Consensus Meeting Group: Methods to explain the clinical significance of health status measures. Mayo Clin Proc 2002, 77: 371–383. 10.4065/77.4.371
Schünemann HJ, Goldstein R, Mador MJ, McKim D, Stahl E, Puhan MA, Griffith LE, Grant B, Austin P, Collins R, Guyatt GH: A randomised trial to evaluate the selfadministered standardised chronic respiratory questionnaire. Europ Respirat J 2005, 25: 31–40. 10.1183/09031936.04.00029704
Lacasse Y, Goldstein R, Lasserson TJ, Martin S: Pulmonary rehabilitation for chronic obstructive pulmonary disease. Cochrane Database Syst Rev 2006,18(4):CD003793.
Deeks J, Higgins J, Altman D: Chapter 9–Analysing Data and Undertaking Metaanalyses: Cochrane Handbook for Systematic Reviews of Interventions Version 5.1.0 [updated March 2011]. The Cochrane Collaboration; 2011. Available from: . [Accessed 3 Sept, 2013] http://www.cochranehandbook.org
Johnston BC, Thorlund K, Schunemann HJ, Xie F, Murad MH, Montori VM, Guyatt GH: Improving the interpretation of quality of life evidence in metaanalysis: the application of minimally important difference units. BMC Health Qual Life Outcomes 2010, 8: 116. 10.1186/147775258116
Cohen J: Statistical Power Analysis in the Behavioral Sciences. 2nd edition. Hillsdale (NJ): Lawrence Erlbaum Associates, Inc; 1988.
Norman GR, Sloan JA, Wyrwich KW: Interpretation of changes in healthrelated quality of life: the remarkable universality of half a standard deviation. Med Care 2003, 41: 582–592.
Farivar SS, Liu H, Hays RD: Half standard deviation estimate of the minimally important difference in HRQOL scores. Expert Rev Pharmacoenconomics Outcomes Res. 2004,4(5):515–523. 10.1586/14737167.4.5.515
Johnston BC, Thorlund K, Da Costa BR, Furukawa TA, Guyatt GH: New methods can extend the use of minimal important difference units in metaanalyses of continuous outcome measures. J Clin Epidemiol 2012,65(8):817–826. 10.1016/j.jclinepi.2012.02.008
AnzuresCabrera J, Sarpatwari A, Higgins JP: Expressing findings from metaanalyses of continuous outcomes in terms of risks. Stat Med 2011,30(25):2967–2985. 10.1002/sim.4298
Friedrich JO, Adhikari NKJ, Beyene J: The ratio of means method as an alternative to mean differences for analyzing continuous outcome variables in metaanalysis: a simulation study. BMC Med Res Methodol 2008,8(32):1–15.
Furukawa T: From effect size into number needed to treat. Lancet 1999, 353: 1680.
Hasselblad V, Hedges LV: Metaanalysis of screening and diagnostic tests. Psychol Bull 1995, 117: 167–178.
Da Costa BR, Rutjes AW, Johnston BC, Reichenbach S, Nüesch E, Tonia T, Gemperli A, Guyatt GH, Jüni P: Methods to convert continuous outcomes into odds ratios of treatment response and numbers needed to treat: metaepidemiological study. Int J Epidemiol 2012,41(5):1445–1459. 10.1093/ije/dys124
Friedrich JO, Adhikari NKJ, Beyene J: Ratio of means for analyzing continuous outcomes in metaanalysis performed as well as mean difference methods. J Clin Epidemiol 2011, 64: 556–564. 10.1016/j.jclinepi.2010.09.016
Thorlund K, Walter SD, Johnston BC, Furukawa TA, Guyatt GH: Pooling continuous outcomes in metaanalysis: a comprehensive review of methods for enhancing interpretability. Res Synth Meth 2011, 2: 188–203. 10.1002/jrsm.46
Johnston BC, Bandayrel K, Friedrich JO, Akl EA, Da Costa BR, Neumann I, Adhikari NKJ, AlonsoCoello P, Crawford MW, Mustafa RA, Svendrovski A, Thabane L, Tikkinen KAO, Vandvik PO, Guyatt GH: Presentation of continuous outcomes in metaanalysis: a survey of clinicians’ understanding and preferences: 21st Cochrane Colloquium. Suppl 1–212 edition. Quebec City, Canada: Cochrane Database Syst Rev; 2013. 4.097
Barbui C, Furukawa TA, Cipriani A: Effectiveness of paroxetine in the acute phase treatment of adults with major depression: a systematic reexamination of published and unpublished randomised data. CMAJ 2008,178(3):296–305.
Furukawa TA: Assessment of mood: guides for clinicians. J Psychosom Res 2010,68(6):581–589. 10.1016/j.jpsychores.2009.05.003
Bandelow B, Baldwin DS, Dolberg OT, Andersen HF, Stein DJ: What is the threshold for symptomatic response and remission for major depressive disorder, panic disorder, social anxiety disorder, and generalized anxiety disorder? J Clin Psychiatry 2006,67(1):428–434.
Furukawa TA, Akechi T, Azuma H, Okuyama T, Higuchi T: Evidencebased guidelines for interpretation of the Hamilton rating scale for depression. J Clin Psychopharmacol 2007,27(5):531–534. 10.1097/JCP.0b013e31814f30b1
Hamilton M: Development of a rating scale for primary depressive illness. Br J Soc Clin Psychol 1967, 6: 278–296. 10.1111/j.20448260.1967.tb00530.x
Suissa S: Binary methods for continuous outcomes: a paremetric alternative. J Clin Epidemiol 1991, 44: 241–248. 10.1016/08954356(91)900358
Cox D, Snell E (Eds): Analysis of binary data. London: Chapman and Hall; 1989.
Akl EA, Oxman AD, Herrin J, Vist GE, Terrenato I, Sperati F, Costiniuk C, Blank D, Schünemann H: Using alternative statistical formats for presenting risks and risk reductions. Cochrane Database Syst Rev 2011, 3: CD006776.
Furukawa TA, Leucht S: How to obtain NNT from Cohen’s d: comparison of two methods. PLoS One 2011,6(4):e19070. 10.1371/journal.pone.0019070
Funding
No funds were received for the preparation of this manuscript. Dr. Busse is funded by a New Investigator Award from the Canadian Institutes of Health Research and Canadian Chiropractic Research Foundation.
Author information
Authors and Affiliations
Corresponding author
Additional information
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
BCJ: concept, data analysis, interpretation of data, manuscript drafting, preparation and approval, administrative support. DLP: concept, interpretation of data, manuscript preparation and approval. KT: concept, data analysis, manuscript preparation and approval. JWB: interpretation of data, manuscript preparation and approval. BRDC: interpretation of data, manuscript preparation and approval. HJS: interpretation of data, manuscript preparation and approval. GHG: concept, data analysis, interpretation of data, manuscript preparation and approval. All authors read and approved the final manuscript.
Rights and permissions
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited.
About this article
Cite this article
Johnston, B.C., Patrick, D.L., Thorlund, K. et al. Patientreported outcomes in metaanalyses Part 2: methods for improving interpretability for decisionmakers. Health Qual Life Outcomes 11, 211 (2013). https://doi.org/10.1186/1477752511211
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/1477752511211
Keywords
 Standardize Mean Difference
 Standard Deviation Unit
 Montgomery Asberg Depression Rate Scale
 Standard Deviation Ratio
 Chronic Respiratory Questionnaire