Study Design
The present analysis is a non-randomized, matched cohort analysis of longitudinal data, up to 2 years from baseline, among patients enrolled in THAOS. Details regarding the basic design and methodology of THAOS (ClinicalTrials.gov: NCT 00628745) are described elsewhere [20]. THAOS is an observational registry for which any individual with a confirmed TTR mutation or wild-type ATTR amyloidosis is eligible, and includes some patients participating in ongoing clinical trials [20]. All study sites received ethical or institutional review board approval prior to subject enrollment, and each subject provided written informed consent. The study followed the International Conference on Harmonisation Good Clinical Practice guidelines and the principles of the Declaration of Helsinki.
Study Population
This analysis included subjects with confirmed TTR mutations (subjects with wild-type ATTR were excluded) who were symptomatic at enrollment or became symptomatic post enrollment, and who had baseline data with at least one follow-up visit. Data were extracted from the THAOS database on January 6, 2015. Subjects with TTR mutations known to have a predominantly cardiac phenotype (i.e., Val122Ile, Ile68Leu, Leu111Met, and Thr60Ala) or who received any treatment (other than tafamidis) intended to target the course of the disease (i.e., liver transplant or diflunisal) pre-enrollment were excluded. The analysis was limited to subjects in Coutinho stage 1 (as defined by the modified Polyneuropathy Disability [mPND]) score, with subjects either having sensory disturbances in their feet but able to walk without difficulty [mPND score I], or having some difficulties walking but able to walk without aid [mPND score II] [21].
For patients receiving tafamidis, dosage recommendations (including dosage strength and timing of the dosing) were based on the prescribing physician’s clinical judgment and the local approved product label, or the open-label study protocol for those patients receiving tafamidis as part of their participation in a clinical study (the indicated dose of tafamidis is 20 mg once daily) [22].
Disease duration was calculated from the earliest onset date of any symptoms commonly associated with TTR-FAP that the treating physician had flagged as definitively or possibly attributable to the disease. An examination of disease duration at enrollment in all THAOS subjects identified a subgroup with unusually long disease duration, raising the possibility that these unusually long disease duration values could distort the multiple imputation and treatment propensity score models. Because disease duration was an important factor in the matching process, subjects in the highest quartile (≥ 8.6 years) were excluded from the analysis due to these concerns and concerns that there would be an insufficient number of appropriate matches for subjects who had been symptomatic for many years prior to enrollment in THAOS.
Outcome Measures
Subject data from the participating THAOS registry sites included medical, laboratory, neurologic, and quality-of-life assessments collected as part of the local standard of care [20]. The focus of the present analysis was disease progression, based on neurologic and quality-of-life assessments, and mortality. Outcome assessments performed at regular clinic visits were examined based on approximately 6-month intervals (± 3 months). Using data from the clinical neurologic examination, the Neurologic Composite Score (NCS; range 0–294, with higher scores reflecting worsening neurologic functioning) and subscale scores (NCS-Reflex [range 0–10, with items rated as present or absent], NCS-Motor [range 0–160, with items rated from 0 = no contractions to 5 = full range of motion and maximum resistance], NCS-Sensory [range 0–124, with items rated as normal, decreased, or absent]) were calculated as the sum of quantitative ratings for subjects’ neurologic assessments of both upper and lower limbs. The mapped Neuropathy Impairment Score-Lower Limbs (NIS-LL, range 0 [normal] to 88 [total impairment]) was calculated from a subset of neurologic assessment items for lower limbs and assesses function of the extremities most affected early in disease progression [23]. Coutinho stage was derived using the mPND (mPND score I or II equating to Coutinho stage 1; mPND score IIIa or IIIb equating to Coutinho stage 2), and time to progression from Coutinho stage 1 to stage 2 (or higher) was determined [21]. Total Quality of Life (TQoL) score (range − 4 to 136) was derived from the patient-reported Norfolk Quality of Life-Diabetic Neuropathy questionnaire, with higher scores reflecting poorer quality of life [23, 24]. The Karnofsky Performance Status Scale score was used to quantify the subjects’ ability to perform normal daily life activities and their need for assistance (ranging from 0 [dead] to 100 [normal; no complaints]).
Statistics
Matching Methodology
The matching methodology was designed to ensure that the heterogeneity among TTR-FAP cases worldwide, including variability in mutation, age of onset, phenotype, and course of disease, was well balanced between the treated and untreated analysis populations. Matching was accomplished in a three-step process by matching on mutation group, country of origin, and by mean treatment propensity score to account for clinical status.
Subjects were matched within mutation groups, defined as Val30Met and non-Val30Met. Thus, treated Val30Met subjects were matched with untreated Val30Met subjects. Within the group of subjects with non-Val30Met mutations, it was not possible to match on the exact mutation owing to low numbers; therefore, treated subjects were matched to untreated subjects within the non-Val30Met mutation group. Although country of ancestry may be the closest approximation of genetic background, this information was not captured consistently; thus, data on country of birth, which were available for > 99% of subjects, were used as the indicator of genetic background.
Finally, propensity score estimation was used to balance the treated and untreated groups on disease characteristics (e.g., severity and duration). Propensity scores are frequently employed to reduce indication bias associated with treatment status in observational studies, thus facilitating causal inference [25]. Propensity score matching is widely used in observational studies and, when appropriately applied, has been shown to be an effective method [26, 27]. Indication bias occurs in real-world (i.e., non-randomized) scenarios, in which treatment decisions are made by clinicians based on disease severity and any number of other important demographic and clinical factors. The estimated treatment propensity scores reflect the probability of receiving treatment based on observed baseline covariates and allow imbalances observed in the study population to be addressed in the analysis [28].
Baseline for treated subjects was defined as the enrollment date (if tafamidis treatment started before enrollment) or the treatment start date (if tafamidis treatment started on or after enrollment). Baseline for untreated subjects was defined as any visit occurring after symptom onset and not more than 3 months prior to enrollment (the THAOS registry permits collection of retrospective data). Owing to the observational nature of the THAOS registry, clinical data are collected and reported as part of routine care. Given variability in practice patterns and the influence of disease status on care decisions, data are assumed not to be missing at random, that is, the probability of a data point being missing is related to other factors of interest at both the site and patient levels. Missing baseline data were imputed for the purpose of propensity score modeling to avoid possible selection bias associated with excluding subjects with incomplete data. Missing data were imputed via multiple imputation using a Markov chain approach. The propensity score models included eight selected baseline clinical variables: TQoL; disease duration; Karnofsky score; the three NCS sub-scores (Reflex, Sensory, and Motor); blood urea nitrogen (BUN); and modified body mass index (mBMI). Up to 101 imputations were performed, separately for Val30Met and non-Val30Met populations, resulting in a corresponding number of fully imputed datasets. Each imputed dataset was entered into a logistic regression model of treatment probability, as predicted by the same set of eight baseline clinical variables.
The mean propensity score for each subject at baseline was used for matching in a nearest-neighbor approach [29]. To enhance the quality of matched sets of patients, propensity scores for the treated and untreated subjects were required to be within 0.25 standard deviations (SD) on the logit scale (i.e., logit of the propensity score) [29]. Furthermore, multiple baseline dates were considered for the untreated population (i.e., any THAOS visit occurring after disease onset and otherwise meeting eligibility criteria). This allowed for the matching of treated and untreated at a point in time where clinical status was most similar rather than at an arbitrary point in time, such as enrollment. Finally, the quality of the matches was maintained through the use of matching “with replacement” of untreated subjects. Untreated subjects were allowed to be used in more than one matched set, but not more than once with a given treated subject.
Analyses
Descriptive statistics were calculated for demographic and baseline clinical characteristics. Treated subjects are presented at a given time point only if one or more of their matched controls also had data available. Untreated subjects are presented at a given time point only if their matched treated subject also had data available. Values for matched controls within each matched set are weighted. Treatment effects were evaluated by repeated measures analyses with appropriate covariates (age at baseline, gender, duration of symptoms, treatment propensity score, time and treatment-by-time interaction, and baseline values). Survival and time to progression from Coutinho stage 1 were analyzed using Cox proportional hazards regression models. Subjects who received alternative disease-modifying treatment (i.e., liver transplant or diflunisal) were censored at the earlier of liver transplant date or the start of diflunisal treatment. For time-to-event analyses, age at baseline and gender were used as covariates for the primary analysis.
Secondary and tertiary models were used to more fully characterize the survival and time-to-progression results. The secondary analysis included all subjects (unmatched) who were ever at Coutinho stage 1 during the THAOS registry. As the start time was unrelated to the treatment start date, the treatment effect was estimated through use of a time-varying covariate for tafamidis treatment. The tertiary analysis included the matched cohorts but was limited to subjects with mPND score I (having sensory disturbances in their feet but able to walk without difficulty) at baseline (i.e., excluding subjects with mPND score II; some difficulties walking but could walk without aid) to account for the truncated time to progression in subjects with a higher mPND score. Covariates for the secondary analysis also included age at baseline, genetic mutation, and birth region, while covariates for the tertiary analysis included age at the start of Coutinho stage 1 and gender.
Sensitivity Analysis
Assessment data for the period from the start of treatment to enrollment in the registry generally were not available for subjects treated with tafamidis prior to enrolling in THAOS. Furthermore, these subjects typically had a longer overall duration of treatment than those who began receiving treatment at or after enrollment. To evaluate the robustness of the model, given the uncertainty of the effect of these two factors, a sensitivity analysis was performed, analyzing matched cohorts that excluded subjects who received tafamidis treatment prior to THAOS enrollment.