Do pilocarpine drops help dry mouth in palliative care patients: a protocol for an aggregated series of n-of-1 trials

Symptomatic improvement will be measured by NRS score for average dry mouth (answer to the question – how dry was your mouth on average over the last 24 hours?). A clinically significant response to pilocarpine will be defined as a ≥2 point improvement in xerostomia NRS score compared to placebo. This is in view of previous work, where a 20% (2 cm) improvement or more against the baseline score was considered to be a positive improvement [21, 22]. NRS data will be collected on days 2 and 3 of each cycle pair to allow wash-out during day 1.

The trial will be reported according to the Consort statement [25], and analysis will be on an intention-to-treat basis.

Power calculations were based on the primary outcome variable, the NRS score, and a pooled mean (standard deviation) baseline score across groups of 3.250 (2.975) from Davies et al., 1998 [15] (with correction, 1999). Thomson and Williams’ 2000 [31] article on testing the xerostomia inventory was used to estimate patients’ ICC (intra class correlation). In their table IV, Thomson and Williams give a correlation between standard-question responses that range from 0.20 to 0.76, so we assumed a value of 0.3 (approximately the median). From Bland 2000 [32] [page 204], recognising that r ≈ ICC = s _b ²/ (s _b ² + s _w ²) where s _b ² is the between subject variance and s _w ² is the within subject variance, we can thus determine that s _w ² ≈ 4.752. Of the 70 palliative care patients randomized by Davies et al. [15], 36 (51%) completed phase I (14 days in length) and 26 (37%) completed phase II (21 days in length after phase I completion). Based on this attrition rate, but recognizing the shorter duration of our study, we expect that approximately 60% of our patients will complete cycle I, 50% will complete cycle II, and 45% will complete cycle III.

If we assume no period effect or treatment × time interaction, then simulations of size N = 10,000 in Stata (Statacorp, College Station, TX) programmed to model the repeated measures design, attrition rates, and s _b ² and s _w ² variances above, then n = 70 patients need to be recruited to detect a 2 cm change in NRS scores between treatments with 80% power at the 5% level of significance. With n = 70, we expect 31 patients will complete all 3 cycles, 4 will complete cycles I and II, 7 will complete cycle I, and 28 will fail to complete any cycles. In contrast, a conventional RCT with similar attrition would require 120 patients.

Data preparation and descriptive reporting will follow that recommended by the CONSORT statement [25].

For each cycle, data from day 1 will be discarded to allow for a wash-out period, and data from days 2 and 3 data will be analysed. All patients with at least one completed treatment cycle will be included in analyses. An effect size will then be calculated between active medication cycles and placebo, thus providing a population measure of effect commensurate with an RCT.

Both individual and population treatment differences will be estimated using hierarchical Bayesian methods and employing noninformative priors using the methods described in Zucker et al. [33], and Schluter and Ware [34]. The likelihood distributions for each model will be assessed for violations and data transformations undertaken, where necessary. Conventional burn-in periods, model convergence and stability diagnostics, and residual checks will be employed [35]. WinBUGS [35] will be used for the Bayesian analysis.

To describe participants’ overall response, three types of Bayesian results will be presented: (i) the mean of the posterior distribution of the mean difference between placebo and stimulant scores, which gives the best estimate of the overall effect size difference between treatments; (ii) the associated 95% credible region, which give intervals of uncertainty (in this case the 2.5 and 97.5 percentile) of the posterior distributions used in (i); and (iii) the posterior probability of the mean difference that stimulant scores were better than placebo scores, which describes the likelihood that the patients will favour the active treatment in future cycles [34]. A patient will be defined to be a ‘responder’ when these estimated values exceed predefined threshold values [34].

Important confounding variables, such as anti-cholinergic load and cause of xerostomia (eg prior radiotherapy), will in included in adjusted analyses and report treatment effects (or success differences) over the various variable stratifications and combinations, following the method advocated by Zucker and colleagues [33].