VALIDITA INTERNA. La misura in cui uno studio riesce a cogliere la relazione «vera» fra due variabili ERRORE CASUALE ERRORE SISTEMATICO (BIAS)

Transcript

1

2

3 VALIDITA INTERNA La misura in cui uno studio riesce a cogliere la relazione «vera» fra due variabili ERRORE CASUALE ERRORE SISTEMATICO (BIAS)

4 ERRORE CASUALE Errore che si verifica per effetto del caso Replicazioni multiple della stessa misurazione producono differenti risultati in tutte le direzioni per variazioni casuali ma la media dà il risultato corretto ERRORE SISTEMATICO Errore che si verifica per la presenza di un fattore che distorce sistematicamente le osservazioni nella stessa direzione Replicazioni multiple della stessa misurazione producono risultati sempre nella stessa direzione e sbagliati

5 Errore sistematico e validità interna di uno studio I risultati di uno studio sono tanto più validi (probabilmente veri) quanto meno esso è affetto da errori sistematici Gli errori sistematici vanno previsti ed evitati o ridotti in fase di disegno dello studio

6 Bias Systematic distortion of the estimated intervention effect away from the truth, caused by inadequacies in the design, conduct, or analysis of a trial, or in the publication of its results. In other words, in a biased trial, the results observed reflect other factors in addition to (or, in extreme cases, instead of) the effect of the tested therapeutic procedure alone. Altman DG, Schulz KF, Moher D, et al. The revised CONSORT statement for reporting randomized trials: explanation and elaboration. Ann Intern Med 2001;134:

7

8

9

10 Randomizzazione

11 End of a clinical trial Why randomise? We find a difference in outcomes between intervention and control groups Possible explanations: the intervention exhibits a real effect the outcome difference is due to chance there is a systematic difference (or bias) between the groups due to factors other than the intervention Randomisation prevents the third possibility Randomisation ensures similar levels of all risk factors (known and unknown)

12 RANDOMIZATION BIAS Randomisation (coin-toss, computer) Alternate, days of week, record number Allocation schedule? Pre-vedibili Allocation Allocation Intervention Control Intervention Control

13 RANDOMIZATION BIAS Recruiting selected Randomisation (cointoss, computer) individuals due to knowledge of the next allocation Allocation schedule Allocation Manipulating allocations of people based on personal believing Intervention Control Exclusion of certain patients based on their prognosis

14 RANDOMIZATION COMPONENTS Item Descriptor Sequence generation Allocation concealment Method used to generate the random allocation sequence, including details of any restriction (eg, blocking, stratification) Method used to implement the random allocation sequence (eg, numbered containers or central telephone), clarifying whether the sequence was concealed until interventions were assigned Implementation Who generated the allocation sequence, who enrolled participants, and who assigned participants to their groups

15 Selection bias 1. generazione della sequenza di randomizzazione Adequate methods :random number table; computer random number generator; coin tossing; shuffling cards or envelopes; throwing dice. (Low risk of bias) Inadequate methods: odd or even date of birth; date (or day) of admission; hospital or clinic record number; alternation; judgement of the clinician; results of a laboratory test or a series of tests; availability of the intervention (High risk of bias). «quasi randomised studies «

16 Baron Ja et al. A Trial of Calcium and Vitamin D for the Prevention of Colorectal Adenomas. N Engl J Med Oct 15;373(16). Randomization randomization by the coordinating center was performed with the use of computergenerated random numbers with permuted blocks and stratification according to clinical center, sex, anticipated colonoscopic examination at 3 years or 5 years, and full factorial randomization.

17 Selection bias 2) Mascheramento della assegnazione Chi recluta i pazienti e verifica se rispondono ai criteri di inclusione non sa a che gruppo verranno assegnati Chi assegna i pazienti ai gruppi non sa chi sono i pazienti

18 Selection bias 2. Mascheramento della assegnazione Adequate methods: Investigators enrolling participants could not foresee assignment : central allocation (including telephone, web-based, and pharmacycontrolled, randomisation); sequentially numbered drug containers of identical appearance; sequentially numbered, opaque, sealed envelopes. Low risk of bias Inadequate methods: open random allocation schedule (e.g. a list of random numbers); assignment envelopes without appropriate safeguards (e.g. if envelopes were unsealed or nonopaque or not sequentially numbered); alternation or rotation; date of birth; case record number; any other explicitly unconcealed procedure. High risk of bias

19

20 Concelament: randomizzazione telefonica

21 Concealment: drug containers

22 Ratios of odds ratios comparing estimates of intervention effects 532 trials with inadequate or unclear allocation concealment versus 272 trials with adequate concealment Wood, L. et al. BMJ 2008;336:

23 CECITA

24 6 ragioni per introdurre la cecità Se dite al paziente che è stato randomizzato al placebo, non è contento Se dite alle persone che l efficacia del trattamento è dovuto all effetto placebo, si arrabbiano Se dite al clinico che il paziente prende il trattamento, il clinico vedrà un miglioramento (anche in assenza di cambiamento) Se dite al paziente che non si dovrebbe grattare, si gratta uguale, ma vi dice che si gratta di meno (Effetto Rosenthal) Illusione di specifici effetti come le tradizioni millenarie sono molto radicate (agopuntura nei meridiani vs a caso) Avete inventato la panacea che, ogni volta che la somministrate, fallisce miseramente cercate cercate fino a analizzare il beneficio su 100 variabili (così funziona la statistica)

25 COSA POTREBBE FARE Usually reduces differential assessment May improve compliance and retention May reduce biased supplemental care or treatment (co-intervention) [and testing]

26 Blinding or Masking Different terms to describe the same procedures Masking may be more appropriate semantics: Participants with impaired vision Less confusing with blindness an outcome Blinding conveys strong bias prevention Lasagna used the term double blindfold in 1955

27 Blinding or Masking We prefer blinding because: it rests on a long history maintains worldwide recognition If you use masking someone using PubMed in Asia or Africa may not know what you did creates strong visual imagery permeates the ICH guidelines

28 Confused Terminology of Single, Double, and Triple Blinding Permeates the Literature Physicians, textbooks, and journal articles vary greatly in interpretations and definitions [Devereaux et al. JAMA 2001; 285: ] Define double-blind inconsistently Authors frequently fail to report their definitions clearly When I use double-blind, participants, investigators, and assessors are blinded In reporting RCTs, authors should explicitly state what steps were taken to keep whom blinded

29

30

31 Performance bias (cointervention) The interpretation of a randomized controlled trial relies on the assumption that any differences in outcome are the result of either chance (whose effects can be quantified) or of inherent differences between treatments. This assumption is invalid if the treatment groups are not handled equally with regard to all of the study procedures, a part the experimental treatment

32 Performance bias Blinding of participants and providers Rischio di bias dipende dal tipo di outcome!! Low risk of bias : Blinding of participants and providers and unlikely that the blinding could have been broken No blinding or incomplete blinding, but the outcome is not likely to be influenced by lack of blinding (e.g. mortality, cancer incidence) High risk of bias: No blinding or incomplete blinding, and the outcome is likely to be influenced by lack of blinding; Blinding of key study participants and personnel attempted, but likely that the blinding could have been broken, and the outcome is likely to be influenced by lack of blinding

33

34 Detection bias When knowledge of the treatment assignment (by participants already recruited into a trial, investigators, or persons who analyze and report trial results) leads to systematic differences on the way the outcomes are assessed 34

35 Detection bias Blinding of outcome assessor Rischio di bias dipende dal tipo di outcome!! Low risk of bias : Blinding of outcome assessment ensured, and unlikely that the blinding could have been broken No blinding of outcome assessment, but the outcome measurement is not likely to be influenced by lack of blinding High risk of bias: No blinding of outcome assessment, and the outcome measurement is likely to be influenced by lack of blinding; Blinding of outcome assessment, but likely that the blinding could have been broken, and the outcome measurement is likely to be influenced by lack of blinding

36 Detection bias Blinding? Double blinding? Triple blinding? Who needs to be blinded? Is the outcome sensitive to blinding? Blinding: clearly very difficult in many intervention trials (i.e. surgical) Solution: Blinded assessors should be used routinely for measuring outcome

37 Outcome assessor Participants ( subiective outcomes) Investigator who collects outcome data Data manager Statistician Quando l intervento non può essere fatto in cieco ma l outcome è soggettivo è fondamentale cercare di garantire la cecità di chi rileva i dati Non tutela dal detection bias del paziente Non tutela dal performance bias del medico

38 Open studies (unblinded) Quando la cecità non è praticamente realizzabile (chirurgia, interventi educativi, psicosociali, riabilitazione, prevenzione primaria) Quando la cecità non è rilevante per il tipo di outcome ( mortalità, incidenza di tumore, recidiva) Risk of bias: patients might under- or overreport treatment effects and side-effects, based on their confidence on the intervention (detection bias) Providers may give advice or prescribe additional therapy to the control group if they feel that these patients are disadvantaged in comparison to the active group, (performance bias)

39 Single-blinded studies the patient should be unaware of which treatment they are taking the investigators are aware Risk of bias: Providers may give advice or prescribe additional therapy to the control group if they feel that these patients are disadvantaged in comparison to the active group( performance bias)

40 Double-blinded studies neither the patient nor the provider knows the identity of the assigned intervention the validity of the study depends on the providers and participants remaining really blinded throughout the study. A study of a drug is easily unblinded if the medications are not identical in appearance

41 Double blind - double dummy Double dummy is a technique for retaining the blind when administering supplies in a clinical trial, when the two treatments cannot be made identical. Supplies are prepared for Treatment A (active and indistinguishable placebo) and for Treatment B (active and indistinguishable placebo). Subjects then take two sets of treatment; either A (active) and B (placebo), or A (placebo) and B (active). 41

42 Triple-blinded studies Providers blinded Participants blinded All the sponsor s project team (eg, the project clinician, outcome assessor, statistician, and data manager) blinded Triple blinding is appropriate for studies in which the risk of adverse events due to the new or standard treatment is low, and should not be used for treatments where safety is a critical issue 42

43 Assessing trial blindness The degree to which the blinding was maintained in a study can be estimated by asking the patients to guess which group they were assigned to. If the mean result of the guesses is close to being 50% correct, the study was well blinded. A similar enquiry could be done with providers also. 43

44 Ratios of odds ratios comparing intervention effect estimates in 314 non-blinded trials versus 432 blinded trials. Wood, L. et al. BMJ 2008;336:

45 Allocation concealment It prevents selection bias in intervention assignment by protecting the allocation sequence before and until assignment It can always be successfully implemented regardless of the study topic Blinding It seeks to prevent performance and detection bias by protecting the sequence after assignment Not always feasible for example, in trials comparing surgical with medical interventions

46 Attrition bias Quando non tutti i soggetti randomizzati completano lo studio i soggetti non escono a caso dallo studio: è possibile che quelli che escono siano sistematicamente diversi da quelli che non escono: i gruppi non sono più randomizzati Validità esterna : es: escono tutti i più giovani, o i meno gravi, o i maschi: posso trarre conclusioni solo su quelli che rimangono Validità interna (Bias): se la probabilità di uscire dallo studio è legata all intervento o all outcome, cioè se quelli che escono hanno sistematicamente probabilità più alte o più basse di avere l outcome di quelli che restano

47 Attrition bias Persi al follow up: il soggetto sparisce non si hanno più info Uscito dallo studio il soggetto interrompe il trattamento ma è reperibile ( eventi avversi? Non efficace? ) Bassa compliance: il soggetto riceve il trattamento ma in dosi e modalità diverse da quelle prescritte (eventi avversi? Trattamento poco accettabile?) Missing data: misurazioni ripetute: il soggetto riceve il trattamento ma non è presente a tutte le misurazioni dell outcome (TD non consegnano le urine quando sono positive) 47

48 Low risk of bias Attrition bias Numero di persi (piccolo) ma quanto? (<5-10%) Bilanciati fra i gruppi Riportate le ragioni (non differenti fra gruppi e non attribuibili agli interventi) Intention to treat Imputation of missing data 48

49 Attrition bias Intention to treat analysis: all subjects analysed in the treatment group they were originally randomized, regardless if they actually received the assigned treatment or not Imputation of missing data : es: considerare gli usciti come fallimenti terapeutici (TD); last observation carried forward Per protocol analysis: only patients who received the treatment as described in the prtocol were analysed 49

50 Intention to treat: effectiveness ( efficacia in pratica, efficacia del trattamento prescritto) Tiene conto anche della scarsa compliance, della difficoltà a somministrare il trattamento Tutela da attrition bias (mantiene la similitudine dei gruppi ottenuta con la randomizzazione Per protocol: efficacy (efficacia in condizioni ottimali, efficacia della trattamento ricevuto nelle modalità previste) Può dare stime distorte se la non compliance e l uscita dallo studio è legata al trattamento o all outcome 50

51 Attrition bias Low risk of bias No missing outcome data; the proportion of missing outcomes compared with observed event risk not enough to have a relevant impact on the intervention effect; Missing outcome data balanced in numbers across intervention groups, with similar reasons across groups; Missing data imputed using appropriate methods All patients analysed in the group they were allocated to by randomisation irrespective of non-compliance and cointerventions (intention to treat) High risk of bias: the proportion of missing outcomes compared with observed event risk enough to induce relevant bias in intervention effect estimate Reason for missing outcome data likely to be related to true outcome, with either imbalance in numbers or reasons for missing data across intervention groups; 51

52 What is publication bias (1)? Definition Publication bias refers to the greater likelihood that studies with positive results will be published JAMA 2002;287:

53 What is publication bias (2)? An alternative definition: Publication bias is the selective or multiple publication or suppression of trial results so that the scientific record is distorted Extension: applied to trial parts - outcomes, subgroups, adverse events REPORTING BIAS The likelihood of finding studies is related to the results of those studies (positive vs negative/detrimental)

54 Why does it matter? Distorts the scientific record Hides the truth Influences doctors decision making Misleads policy makers Causes harm to patients Costly for the health service A form of scientific and research misconduct TO U: It will matter if the studies you don t find differ systematically from the ones you have found You might arrive at different answers, or even THE WRONG ANSWER

55 Publication of All Trials

56 Publication Bias Asymmetrical appearance of the funnel plot with a gap in a bottom corner of the graph

57 Funnel plots A funnel plot is a scatter plot of treatment effect against a measure of study size / precision. Precision in the estimation of the true treatment effect increases as the sample size increases. Small studies scatter more widely at the bottom of the graph In the absence of bias the plot should resemble a symmetrical inverted funnel

58 Publication Bias In this situation the effect calculated in a meta-analysis will overestimate the treatment effect The more pronounced the asymmetry, the more likely it is that the amount of bias will be substantial.

59 Outcome reporting bias

60 Reporting bias is selection bias Reporting bias is perhaps the greatest source of selection bias Originally defined as the publication or nonpublication of studies depending on the direction and statistical significance of the results Is a complex phenomenon

61 Reported outcomes Full Partial Qualitative n and effect size, plus precision / p- value for continuous data Effect size or precision (± n or p-value) p-value Incompletely reported outcomes Unreported Hierarchy of the levels of outcome reporting (Chan, 2004)

62 Eterogeneità

63 E efficace?

64 Forest plot (meta-graph) analitico

65 META-ANALYSIS General Number of studies Number of participants (62607) OR (MH) - Fixed effect model Meta-analysis outcome 95% CI lower limit 95% CI upper limit 1,0063 0,9482 1,068 z p-value (two-tailed) 0,2073 0,8358 Heterogeneity Q H p-value (two-tailed) 95% CI lower limit 95% CI upper limit 47,1363 < 0,0001 1,7727 1,3675 2,2979 I^2 95% CI lower limit 95% CI upper limit 68,18% 46,53% 81,06%

66 Could we just add the data from all the trials together? One approach to combining trials would be to add all the treatment groups together, add all the control groups together, and compare the totals This is wrong for several reasons, and it can give the wrong answer

67 If we add up the columns we get 34.3% vs 32.5%, a RR of 1.06, a higher chance of death in the steroids group From a meta-analysis, we get RR=0.96, a lower chance of death in the steroids group

68

69 Va a scua il mar Mettere insieme studi diversi che testano quesiti diversi considerando popolazione diverse usando interventi lievemente diversi ma partendo da protocolli profondamente diversi e dando risultati Eterogeneità

70 What is heterogeneity? Heterogeneity is variation between the studies results

71 What is heterogeneity? Differences between studies with respect to: Clinical heterogeneity (clinical diversity) Participants e.g. conditions under investigation, eligibility criteria for trials, geographical variation Interventions e.g. intensity / dose / duration, sub-type of drug, mode of administration, experience of practitioners, nature of the control (placebo/none/standard care) Outcomes e.g. definition of an event, follow-up duration, ways of measuring outcomes, cut-off points on scales

72 What is heterogeneity? Differences between studies with respect to: Methodological heterogeneity (methodological diversity) Design e.g. randomised vs non-randomised, crossover vs parallel group vs cluster randomised, pre-test and long follow up Conduct e.g. allocation concealment, blinding etc, approach to analysis, imputation methods for missing data

73 What is heterogeneity? What do we do if there is statistical heterogeneity? Variation in the true effects underlying the studies...which may manifest itself in more observed variation than expected by chance alone May be due to clinical diversity (different treatment effects) or methodological diversity (different biases)

74 Come si misura questa eterogeneità?

75 Confidence interval overlapping Eyeball test Cochran s Q: to assess whether observed differences in results are compatible with change alone 2 distribution; low power (small number of studies, small sample size) Q P=<0.10 p=<0.10 (heterogeneity) I 2 quantifing heterogeneity (describes the percentage of variation across studies that is due to heterogeneity rather than chance) I 2 = 20% - 50% 0-40% might not be important 30-60% may represent moderate heterogeneity 50-90% may represent substantial heterogeneity % considerable heterogeneity Tau.

76 How to deal with heterogeneity 1. Do not pool at all 2. Ignore heterogeneity: use fixed effect model 3. Allow for heterogeneity: use random effects model 4. Explore heterogeneity: subgroups analysis or metaregression (tricky)

77 Example: PC Games for intelligence Study Omar Einstein Margie Bart Ombroso Estimates and 95% confidence intervals I 2 = 0 % Standardized mean difference Favours Control Favours PC Games

78 Example: PC Games for intelligence Study Omar Einstein Margie Bart Ombroso Estimates and 95% confidence intervals I 2 = 80 % Standardized mean difference Favours Control Favours PC Games

79 Fixed and random effects

80 The Fixed Effects assumption

81 The Fixed Effects assumption Observed in studies True

82 Fixed effects model In un modello a effetti fissi si assume che tutti gli studi provengano dalla stessa popolazione di studi Si assume che ci sia un parametro (es.media) unico, fisso Il peso degli studi è funzione della variabilità intra-studio Gli intervalli di confidenza del parametro sono ridotti Popolazione di riferimento unica, omogenea

83 The Random Effects assumption

84 The Random Effects assumption Observed in studies True in studies True

85 Popolazioni di riferimento molteplici, eterogenee Random effects model In un modello a effetti random gli studi potrebbero provenire da popolazioni di studi diverse I pesi sono ridistribuiti in modo più omogeneo tra studi grandi e piccoli (il peso non è dovuto solo alla variabilità intrastudio) Gli intervalli di confidenza del parametro sono aumentati

86 Quale modello? Fixed effect Random effect

87 Quale modello? Fixed effect Random effect Potente (IC ristretti) Assume un solo parametro, non facile in ambito biomedico Più facile per sottogruppi Semplicistico Dà luogo a un aggiustamento dei pesi grezzo (ridistribuzione senza tener conto di nessuna co-variata) IC realistici I 2 = 20% - 50% I 2 = 50% - 70% I 2 = > 70%

88

89 Strutturazione del Quesito Clinico sec. modello P.I.C.O. P Nei Pazienti con Specifiche caratteristiche di malattia (stadio, classe di rischio, ecc.) I C l Intervento Non necessariamente coincidenti con gli outcome di efficacia delle evidenze disponibili considerare anche gli outcome di tollerabilità! (è suscettibile di impiego) in Confronto con Intervento terapeutico oggetto del quesito clinico Trattamento altrimenti considerabile in alternativa all intervento in esame O riguardo agli Outcome di beneficio/danno Parametri clinico-laboratoristici ritenuti essenziali per la decisio-ne terapeutica

90 Adjuvant RT PCa

91 Surrogate endpoints Issue: Quicker, less expensive, less clinically relevant endpoint or More expensive, clinically definitive endpoint? Mark Conaway, June 2006

92 Surrogate endpoints

93 Surrogate endpoints Issue: Quicker, less expensive, less clinically relevant endpoint or More expensive, clinically definitive endpoint? Hesitate to use the term "surrogate" Has a specific technical definition Mark Conaway, June 2006

94

95

96 Quando si hanno dati di molti RCT si deriva un modello di regressione: - che possa predire la magnitudine - dell effetto del trattamento sull endpoint vero - in base all effetto del trattamento sull endpoint (candidato) surrogato Il surrogato è tale se la predizione è sufficientemente precisa

97 TRIAL LEVEL CORRELATION BETWEEN EFFECTS threshold for surrogacy Burzykowski and Buyse, Pharmaceutical Statist 2006;5:173

98

99

100

101

102 Removal of Patients from Therapy for Disease Progression COU-AA-301 AFFIRM

103

104 Pappagallo Cinquini - Moschetti Tutti

105

106 The best? No head-to-head comparison

107

108

109 Population: previously untreated any age and race histologically proven NSCLC harbouring activating EGFR-mutation Intervention: EGFR-TKIs (Erlotinib, Gefitinib, Afatinib) Comparison: Platinum-based chemotherapy

110 Outcomes: PFS (whenever possible independently reviewed data) PFS in exon 19 deletion PFS in L858R mutation OS ORR (complete and/or partial and/or stable) Treatment related toxic events

111 Search strategy PubMed, Cancer-Lit, Embase-databases and Cochrane-Library were searched for RCTs up to June 2014 with no language or publication status restrictions. Search terms were TKI [Substance Name] and Carcinoma, NSCLC [Substance Name]. The proceedings of the conferences of the American Society of Clinical Oncology(ASCO), European Society of Medical Oncology (ESMO)and International Association for the Study of Lung Cancer (IASLC), World Conference of Lung Cancer were also searched for relevant abstracts. Any unpublished RCTs were considered for inclusion.

112

113

114 Data synthesis: HR for PFS and OS RR for the Others

115 PFS

116

117 OS

118 Skin reactions Diarrhea Hypertransaminasemia

119 So, who s the best?

120 Head to Head vs. Indirect Comparisons Head to Head comparison comes from a trial where A was directly compared to B. Indirect Comparison comes from multiple studies where A and B may have been compared to the same comparator (i.e., C) but have never been compared to each other in the same study, What is indirect comparison? Fujian Song BMed MMed PhD Reader in Research Synthesis, Faculty of Health, University of East Anglia

121 Indirect Comparisons Indirect comparison refers to a comparison of different healthcare interventions using data from separate studies, in contrast to a direct comparison within randomized controlled trials. Indirect comparison is often used because of a lack of, or insufficient, evidence from head-to-head comparative trials. Naive indirect comparison is a comparison of the results of individual arms from different trials as if they were from the same randomized trials. This method provides evidence equivalent to that of observational studies and should be avoided in the analysis of data from randomized trials. Adjusted indirect comparison (including mixed treatment comparison) is an indirect comparison of different treatments adjusted according to the results of their direct comparison with a common control, so that the strength of the randomized trials is preserved. Empirical evidence indicates that results of adjusted indirect comparison are usually, but not always, consistent with the results of direct comparison. What is indirect comparison? Fujian Song BMed MMed PhD Reader in Research Synthesis, Faculty of Health, University of East Anglia

122 Indirect Comparisons Basic assumptions underlying indirect comparisons include: homogeneity assumption for standard meta-analysis, similarity assumption for adjusted indirect comparison and consistency assumption for the combination of direct and indirect evidence. It is essential to fully understand and appreciate these basic assumptions in order to use adjusted indirect and mixed treatment comparisons appropriately. What is indirect comparison? Fujian Song BMed MMed PhD Reader in Research Synthesis, Faculty of Health, University of East Anglia

123 HOMOGENEITY ASSUMPTION When multiple trials are available for a given comparison, the results from multiple trials can be pooled in metaanalyses before an adjusted indirect comparison is conducted. For a meta-analysis to be valid, it is commonly established that results from different trials should be sufficiently homogeneous from a clinical and statistical perspective. This is usually demonstrated by a 2-tailed p value for homogeneity at Pearson chi-squared test or Cochran Q test > 0.10 and a I 2 (inconsistency) < 50%. When homogeneity is unlikely (e.g. I 2 >50%) than heterogeneity and inconsistency are likely. Song, What is? 2009; Higgins et al, BMJ 2003

124 PFS

125 CONSISTENCY ASSUMPTION When both direct and indirect evidence is available, an assumption of evidence consistency is required to quantitatively combine the direct and indirect estimates. It is important to investigate possible causes of discrepancy between the direct and indirect evidence, such as the play of chance, invalid indirect comparison, bias in head-to-head comparative trials, and clinically meaningful heterogeneity When the direct comparison differs from the adjusted indirect comparison, we should usually give more credibility to evidence from head-to-head comparative trials. However, evidence from direct comparative trials may not always be valid. Song, What is? 2009; Song et al, J Clin Epidemiol 2008

126 No head-to head comparisons

127 SIMILARITY ASSUMPTION For an adjusted indirect comparison (A vs B) to be valid, a similarity assumption is required in terms of moderators of relative treatment effect. That is, patients included should be sufficiently similar in the two sets of control arms (C 1 from the trial comparing A vs C 1, and C 2, from the trial comparing B vs C 2 ). This is crucial as only a large theoretical overlap between patients enrolled in C 1 and C 2 enables the relative effect estimated by trials of A versus C 1 to be generalizable to patients in trials of B versus C 1, and the relative effect estimated by trials of B versus C 2 to be generalizable to patients in trials of A versus C 2. Song, What is? 2009

128 Study FIRST-SIGNAL Cisplatin 75 mg/m2 day 1&8 Gemcitabine 1,250 mg/m2 day 1 IPASS Carboplatin (AUC 5.0/6.0) mg/millimeter per minutes Paclitaxel 200mg/m2 day 1 NEJG002 WJTOG3405 EURTAC OPTIMAL Carboplatin(AUC 6.0)mgmm Paclitaxel 200mg/m2 day 1 Cisplatin 80 mg/m2 Docetaxel 60mg/m2 Cisplatin 75 mg/m2 or Carbo Docetaxel 75mg/m2 day 1 or Gemcitabine 1250 day 1&8 Carboplatin(AUC 5.0)mgmm Gemcitabine 1000 mg/m2 day 1&8 TORCH Cisplatin 80 mg/m2 day 1 Gemcitabine 1,200 mg/m2 LUX-LUNG III LUX-LUNG VI Cisplatin 75 mg/m2 Pemetrexed 500mg/m2 Cisplatin 75 mg/m2 Gemcitabine 1000 mg/m2 day 1&8 i.v. every 3 weeks Max 9 cycles i.v. every 3 weeks up to 6 weeks i.v. 3 cycles i.v. every 3 weeks up to 6 weeks i.v. 4 cycles i.v. every 3 weeks up to 6 weeks i.v. 6 cycles i.v. Up to 6 cycles

129 COMPUTATIONS The log relative risk of the adjusted indirect comparison of A and B (lnrr A vs B ) can be estimated by: and its standard error is: ln RR A vs B = ln RR A vs C1 ln RR B vs C2 SE ( ln RR A vs B ) = [ SE ( ln RR A vs C1 ) 2 + SE ( ln RR B vs C2 ) 2 ] Similar computations can be envisioned for odds ratio, absolute risk reductions, weighted mean differences, and standardized mean differences. Higgins et al, BMJ 2003; Song, What is? 2009;

130

131 Interpretation - Quality Rubbish studies = unbelievable results If all the trials in a meta-analysis were of very low quality, then you should be less certain of your conclusions. Instead of Treatment X cures Y disease, try There is some evidence that Treatment X cures Y disease, but the data should be interpreted with caution.

132

133 TAKE HOME MESSAGES Adjusted indirect comparison meta-analysis represents a simple yet robust tool to make statistical and clinical inference despite the lack of conclusive evidence from headto-head randomized clinical trials. Despite being not at the uppermost level of the hierarchy of evidence based medicine, it can often provide results equivalent to those of subsequent direct comparisons.

134

135 What? So What? Now What? Riflettete da soli per 5 minuti e annotate sul modulo le vostre risposte (indicate almeno 2 argomenti) Confrontatevi con un collega del vostro tavolo e trovate almeno due argomenti condivisi (5 minuti) Confrontatevi con i colleghi del vostro tavolo per 15 minuti, e rispondete in modo condiviso ai tre punti W3 e delegate un portavoce Riportate sulla lavagna il vostro W3 condiviso su almeno due aspetti ritenuti rilevanti ed «impattanti» sulla vostra professione (5min) Presentate ai colleghi dei vostri tavoli il vostro W3 condiviso