--- title: "Sample size calculation for negative binomial outcomes" output: rmarkdown::html_vignette bibliography: gsDesignNB.bib vignette: > %\VignetteIndexEntry{Sample size calculation for negative binomial outcomes} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r, include=FALSE} knitr::opts_chunk$set( collapse = TRUE, comment = "#>" ) ``` ```{r, message=FALSE, warning=FALSE} library(gsDesignNB) ``` This vignette describes the methodology used in the `sample_size_nbinom()` function for calculating sample sizes when comparing two treatment groups with negative binomial outcomes. We present the notation, compare the underlying methods from the literature, derive the average exposure calculations, discuss statistical information, and provide simulation verification. ## Notation We use the following notation throughout this vignette and the package documentation. | Symbol | Description | |:-------|:------------| | $\lambda_g$ | Event rate (events per unit time) for group $g$ ($g = 1$ for control, $g = 2$ for treatment) | | $k$ (or $k_g$) | Dispersion parameter such that $\mathrm{Var}(Y) = \mu + k \mu^2$; equivalent to $1/\texttt{size}$ in R's `rnbinom` | | $t_i$ | Exposure (follow-up) time for subject $i$ | | $\bar{t}_g$ | Average exposure time for group $g$: $\bar{t}_g = \mathrm{E}[t_i \mid \text{group } g]$ | | $\mu_{g} = \lambda_g \bar{t}_g$ | Expected event count per subject in group $g$ | | $\theta = \log(\lambda_2 / \lambda_1)$ | Log rate ratio (treatment effect) | | $\theta_0 = \log(\text{rr}_0)$ | Log rate ratio under the null hypothesis | | $p_g = n_g / n_{\text{total}}$ | Allocation proportion for group $g$ | | $r = n_2 / n_1$ | Allocation ratio | | $Q_g = \mathrm{E}[t^2_g] / (\mathrm{E}[t_g])^2$ | Variance inflation factor for variable follow-up in group $g$ | | $\mathcal{I}$ | Statistical information: $\mathcal{I} = 1 / \mathrm{Var}(\hat\theta)$ | | $z_\alpha, z_\beta$ | Standard normal quantiles at levels $\alpha$ and $\beta$ | ## Methodology We wish to test for differences in rates of recurrent events between two treatment groups using a negative binomial model to account for overdispersion. The underlying test evaluates the log rate ratio $\theta = \log(\lambda_2 / \lambda_1)$ between the two groups. The usual null hypothesis is $H_0\!: \theta \ge \theta_0$ with $\theta_0 = 0$ (i.e., equal rates) and a one-sided alternative $H_1\!: \theta < 0$ (treatment reduces the event rate). Non-inferiority or super-superiority tests can also be performed by specifying $\theta_0 \ne 0$. ### Negative binomial distribution We assume the outcome $Y$ for a subject with exposure time $t$ follows a negative binomial distribution with mean $\mu = \lambda t$ and dispersion parameter $k$, such that the variance is: $$ \mathrm{Var}(Y) = \mu + k \mu^2 $$ Note that in R's `rnbinom` parameterization, $k = 1/\texttt{size}$. When $k = 0$, the negative binomial reduces to the Poisson distribution. ### Gamma-Poisson mixture motivation The negative binomial distribution arises naturally as a Gamma-Poisson mixture. For each subject $i$, suppose the subject-specific event rate $\Lambda_i$ follows a Gamma distribution: $$ \Lambda_i \sim \text{Gamma}\!\left(\text{shape} = \frac{1}{k},\; \text{rate} = \frac{1}{k\lambda}\right) $$ so that $\mathrm{E}[\Lambda_i] = \lambda$ and $\mathrm{Var}(\Lambda_i) = k\lambda^2$. Given $\Lambda_i$, the event count over exposure time $t_i$ is Poisson: $$ Y_i \mid \Lambda_i \sim \text{Poisson}(\Lambda_i\, t_i) $$ Marginalizing over $\Lambda_i$, the unconditional distribution of $Y_i$ is negative binomial with: $$ \mathrm{Var}(Y_i) = \mathrm{E}[\Lambda_i t_i] + \mathrm{Var}(\Lambda_i t_i) = \lambda t_i + k\lambda^2 t_i^2 = \mu_i + k\mu_i^2 $$ This connects the event rate $\lambda$ (hypothesis testing framework) with the expected count $\mu$ (negative binomial parameterization). ### Graphical illustration The following panel illustrates the distribution of event counts with mean $\mu = 5$ for dispersion parameters $k \in \{0, 0.5, 1\}$. As $k$ increases, the variance increases and the distribution becomes more spread out. ```{r fig.width=8, fig.height=4} par(mfrow = c(1, 3)) k_values <- c(0, 0.5, 1) for (k in k_values) { mu <- 5 x <- 0:15 if (k == 0) { probs <- dpois(x, lambda = mu) } else { size <- 1 / k probs <- dnbinom(x, size = size, mu = mu) } barplot(probs, names.arg = x, horiz = TRUE, main = paste("k =", k), xlab = "Probability", ylab = "Event Count", las = 1, xlim = c(0, 0.2) ) } ``` ### Sample size formula The default sample size formula is based on the asymptotic normality of the maximum likelihood estimator of the log rate ratio $\hat\theta$, using a Wald test. The required sample size for group 1 is: $$ n_1 = \frac{(z_{\alpha/s} + z_{\beta})^2 \cdot \tilde{V}}{(\theta - \theta_0)^2} $$ where $n_2 = r \cdot n_1$, $n_{\text{total}} = n_1 + n_2$, and $\tilde{V}$ is the per-subject variance factor: $$ \tilde{V} = \left(\frac{1}{\mu_1} + k_1\right) + \frac{1}{r}\left(\frac{1}{\mu_2} + k_2\right) $$ Here $\mu_g = \lambda_g \bar{t}_g$ is the expected event count per subject in group $g$, and $k_g$ is the (possibly inflated) dispersion parameter for group $g$. In the simplest case where both groups share a common dispersion $k$ and a common average exposure $\bar{t}$, with equal allocation ($r = 1$, $p_1 = p_2 = 1/2$), this simplifies to: $$ n_{\text{total}} = \frac{(z_{\alpha/s} + z_{\beta})^2 \left[\left(\frac{1}{\mu_1} + k\right) + \left(\frac{1}{\mu_2} + k\right)\right]}{(\theta - \theta_0)^2} $$ When `test_type = "score"`, `sample_size_nbinom()` instead uses a two-variance score-test formula: $$ n_1 = \frac{(z_{\alpha/s}\sqrt{V_0} + z_\beta\sqrt{V_1})^2} {(\theta - \theta_0)^2}, $$ where $V_1$ is the alternative variance factor above and $V_0$ is the corresponding null variance factor under the constrained null model. This aligns the design calculation with a score statistic evaluated under $H_0$. It is most useful as a diagnostic and alternative sizing rule when finite-sample Type I error control is the dominant objective, for example in lower-information designs, high-dispersion settings, event-gap settings, non-inferiority or super-superiority hypotheses, or adaptive/group sequential designs that will use a score statistic at analysis. The simulation grid in `vignette("score-vs-wald-simulation")` shows that the final test statistic can matter more than the small difference between Wald and score sample sizes: in the displayed superiority scenarios, the traditional Zhu--Lakkis Wald sample size paired with the score test preserved Type I error and gave slightly higher score-test power than score sizing. Thus, if the planned analysis remains the Wald log-rate-ratio test, the Zhu--Lakkis Wald formula is the directly matched calculation. If the planned analysis is the score test, the practical default is to compare Wald and score sizing, often retaining Wald sizing as a modest sample-size margin, and then verify both Type I error and power by simulation. ### Relationship between Zhu-Lakkis, Friede-Schmidli, and Mutze et al. The default Wald sample size formula implemented in `sample_size_nbinom()` corresponds to **Method 3** of @zhu2014sample, which is the Wald test for the log rate ratio under a negative binomial model with a log link and an exposure offset. This same test statistic and variance formula are used by @friede2010blinded in the context of blinded sample size reestimation and by @mutze2019group for group sequential designs. **What the three references share:** All three references use the same underlying asymptotic result for the variance of the MLE of the log rate ratio. For a single subject $i$ in group $g$ with exposure $t_i$ and expected count $\mu_i = \lambda_g t_i$, the Fisher information contribution is: $$ \mathcal{I}_i = \frac{\mu_i}{1 + k\,\mu_i} $$ The total information for the log rate ratio is: $$ \mathcal{I} = \frac{1}{1/W_1 + 1/W_2}, \quad W_g = \sum_{i \in \text{group } g} \frac{\mu_i}{1 + k\,\mu_i} $$ When all subjects in a group have the same exposure $\bar{t}$, this simplifies to $W_g = n_g \cdot \mu_g / (1 + k\,\mu_g)$, giving the familiar formula above. **How the references differ:** | Aspect | @zhu2014sample | @friede2010blinded | @mutze2019group | |:-------|:--------------|:-------------------|:----------------| | **Focus** | Fixed sample size calculation | Blinded sample size reestimation | Group sequential designs | | **Methods compared** | Five different test statistics (Methods 1--5) | One method (equivalent to Zhu Method 3) | One method (same Wald test) | | **Dispersion** | Common $k$ across groups | Common $k$; estimated from blinded data | Common $k$ | | **Exposure** | Fixed $\bar{t}$ for all subjects | Fixed $\bar{t}$ | Fixed $\bar{t}$ | | **Key contribution** | Systematic comparison of NB sample size methods | Blinded estimation of nuisance parameters ($k$, $\lambda$) during a trial | Extension to interim analyses with spending functions | The `sample_size_nbinom()` function extends these methods by allowing: - **Group-specific dispersion** ($k_1 \ne k_2$), - **Variable exposure** across subjects (piecewise accrual, dropout, max follow-up), - **Variance inflation** ($Q$ factor) to account for non-constant exposure, following the approach in @zhu2014sample, and - **Event gaps** that reduce effective exposure, and - **Score-test sizing** with separate null and alternative variance factors when `test_type = "score"`. ### Non-inferiority and super-superiority The parameter `rr0` ($\theta_0$) allows for testing non-inferiority or super-superiority hypotheses. The null hypothesis is $H_0\!: \lambda_2 / \lambda_1 \ge \text{rr}_0$ (assuming lower rates are better). * **Superiority (default):** $\text{rr}_0 = 1$ ($\theta_0 = 0$). We test if the treatment rate is lower than the control rate. * **Non-inferiority:** $\text{rr}_0 > 1$ (e.g., 1.1). We test if the treatment rate is not worse than the control rate by more than a specified margin. * **Super-superiority:** $\text{rr}_0 < 1$ (e.g., 0.8). We test if the treatment rate is better than the control rate by at least a specified margin. ## Average exposure In practice, subjects have variable follow-up times due to staggered enrollment, administrative censoring at the end of the study, dropout, and maximum follow-up caps. The sample size formula requires the *expected* exposure $\bar{t}_g = \mathrm{E}[t_i \mid \text{group } g]$ for each group. This section derives the average exposure computation implemented in `sample_size_nbinom()`. ### Setup Let $T$ denote the total trial duration. Suppose recruitment occurs in $J$ segments, where segment $j$ has accrual rate $R_j$ and duration $D_j$. The start time of segment $j$ is $S_j = \sum_{i=1}^{j-1} D_i$ (with $S_0 = 0$). The expected number of patients in segment $j$ is $N_j = R_j D_j$. A patient enrolled at calendar time $s$ has a *potential* follow-up time of $u = T - s$ before administrative censoring. For patients in segment $j$, the potential follow-up ranges from $u_{\min} = T - (S_j + D_j)$ to $u_{\max} = T - S_j$. ### Calendar exposure without dropout When there is no dropout ($\delta = 0$), a patient enrolled at time $s$ is followed for exactly $u = T - s$ time units (or until `max_followup`, whichever comes first). Under uniform enrollment within each segment, the average follow-up for segment $j$ is: $$ E_j = \frac{u_{\min} + u_{\max}}{2} = T - S_j - \frac{D_j}{2} $$ This is simply the midpoint of the follow-up range for that segment. ### Calendar exposure with dropout Dropout is modeled as a piecewise exponential process with hazard rates $\delta_1, \delta_2, \ldots$ over successive intervals $[0, c_1), [c_1, c_2), \ldots$ (the last interval extends to $\infty$). The survival function is $S(t) = \exp\!\bigl(-\sum_j \delta_j \ell_j\bigr)$ where $\ell_j$ is the time spent in interval $j$. The expected exposure for a patient with potential follow-up $u$ is $$ m(u) = \int_0^u S(t)\, dt $$ which decomposes into a sum of closed-form exponential integrals over each piece. For a single constant rate $\delta > 0$ (the most common special case) this simplifies to $$ m(u) = \frac{1 - e^{-\delta u}}{\delta} $$ The average exposure for segment $j$ is obtained by averaging $m(u)$ over the uniform distribution of enrollment times within the segment. For a single constant rate this yields: $$ \bar{t}_j = \frac{1}{D_j}\int_{u_{\min}}^{u_{\max}} m(u)\, du = \frac{1}{\delta} - \frac{e^{-\delta u_{\min}} - e^{-\delta u_{\max}}}{\delta^2 D_j} $$ Note that as $\delta \to 0$, this reduces to $(u_{\min} + u_{\max})/2$, consistent with the no-dropout case. The `dropout_rate` argument accepts a scalar (common constant rate), a length-2 vector (per-group constant rates), or a data frame with columns `rate` and `duration` (and optionally `treatment`) for the full piecewise specification. ### Maximum follow-up If `max_followup` ($F$) is specified, the follow-up time is capped at $F$ for each patient. The expected exposure for a patient with potential follow-up $u$ becomes: $$ m^*(u) = \mathrm{E}[\min(u, F, \text{dropout time})] $$ For segment $j$, three cases arise: 1. **No truncation** ($u_{\max} \le F$): All patients have potential follow-up $\le F$. The calculation is as above. 2. **Full truncation** ($u_{\min} \ge F$): All patients have potential follow-up $\ge F$, so each patient's exposure is $\min(F, \text{dropout time})$, giving $m^*(F) = (1 - e^{-\delta F})/\delta$ (or simply $F$ when $\delta = 0$). 3. **Partial truncation** ($u_{\min} < F < u_{\max}$): Patients enrolled earlier (with $u \ge F$) are capped, while later enrollees are not. The average is a weighted combination: $$ \bar{t}_j = \frac{(u_{\max} - F)\, m^*(F) + \int_{u_{\min}}^{F} m(u)\, du}{D_j} $$ ### Weighted average across segments The overall average exposure for group $g$ is: $$ \bar{t}_g = \frac{\sum_{j=1}^{J} N_j\, \bar{t}_{j,g}}{\sum_{j=1}^{J} N_j} $$ When dropout rates are the same for both groups, $\bar{t}_1 = \bar{t}_2$. When they differ, each group has its own average exposure. ### Group-specific parameters The function supports specifying `dropout_rate` as a scalar, a length-2 vector, or a data frame with piecewise rates (see above), and `max_followup` as a scalar or vector of length 2, corresponding to the control and treatment groups respectively. When these differ between groups, the average exposure $\bar{t}_g$, the second moment $\mathrm{E}[t_g^2]$, and the variance inflation factor $Q_g$ are all computed separately for each group. ### Variance inflation for variable follow-up ($Q$ factor) When follow-up times vary across subjects, using $\bar{t}$ alone in the variance formula underestimates the true variance. This is because the negative binomial variance depends *non-linearly* on exposure through the $k\mu^2 = k\lambda^2 t^2$ term. @zhu2014sample derive a correction factor: $$ Q_g = \frac{\mathrm{E}[t_g^2]}{(\mathrm{E}[t_g])^2} \ge 1 $$ by Jensen's inequality. The effective dispersion used in the sample size formula is $k_{g,\text{adj}} = k_g \cdot Q_g$. When exposure is constant ($t_i = \bar{t}$ for all $i$), $Q = 1$ and no adjustment is needed. Variable follow-up inflates $Q$ above 1, increasing the required sample size. **Computing $\mathrm{E}[t^2]$:** The second moment is computed analogously to $\mathrm{E}[t]$ by integrating $m_2(u) = \mathrm{E}[\min(u, \text{dropout time})^2]$ over the enrollment distribution. Without dropout, $m_2(u) = u^2$. With a single constant dropout rate $\delta$: $$ m_2(u) = \frac{2}{\delta^2}\left(1 - e^{-\delta u}(1 + \delta u)\right) $$ For piecewise dropout, $m_2(u)$ is computed by integrating $2 t\, S(t)$ over the piecewise survival function, analogously to $m(u)$. The average of $m_2(u)$ across enrollment segments (with appropriate handling of `max_followup` truncation) gives $\mathrm{E}[t^2_g]$. ### Event gaps In some recurrent event trials, there is a mandatory "dead time" or gap of duration $g$ after each event, during which no new events can occur (e.g., a recovery period). This reduces the effective time at risk. **Naive approximation.** For a *single subject* with fixed event rate $\lambda$ and gap $g$, the long-run effective rate from renewal theory is: $$ \lambda_{\text{naive}} = \frac{\lambda}{1 + \lambda g} $$ **Jensen's inequality correction.** In the negative binomial model, subjects do not share the same rate. Each subject's rate $\Lambda_i$ is drawn from a Gamma distribution with mean $\lambda$ and variance $k\lambda^2$ (the Gamma-Poisson mixture). The population-level effective rate is $\mathrm{E}[\Lambda_i / (1 + \Lambda_i g)]$, which by Jensen's inequality is strictly less than $\lambda / (1 + \lambda g)$ because $f(x) = x/(1+xg)$ is concave. The naive formula overestimates the effective rate (and thus the expected event count), leading to underestimated variance and slightly underpowered designs. The bias scales with both the dispersion $k$ and the gap $g$. To correct for this, `sample_size_nbinom()` applies a second-order Taylor expansion of $f(\Lambda)$ around $\mathrm{E}[\Lambda] = \lambda$: $$ \mathrm{E}\!\left[\frac{\Lambda}{1+\Lambda g}\right] \approx \frac{\lambda}{1+\lambda g} + \frac{f''(\lambda)}{2}\,\mathrm{Var}(\Lambda) = \frac{\lambda}{1+\lambda g} \left(1 - \frac{k\lambda g}{(1+\lambda g)^2}\right) $$ where $f''(\lambda) = -2g/(1+\lambda g)^3$ and $\mathrm{Var}(\Lambda) = k\lambda^2$. This corrected effective rate is used in the sample size formula: $$ \mu_g = \lambda_{g,\text{eff}} \cdot \bar{t}_g $$ The function also reports the *at-risk exposure*, which is the calendar exposure reduced by the expected gap time: $$ \bar{t}_{g,\text{risk}} = \frac{\bar{t}_g}{1 + \lambda_g \cdot g} $$ Since the gap correction depends on $\lambda_g$, the at-risk exposure differs between groups even when the calendar exposure is the same. **Magnitude of the correction.** The following table illustrates the correction factor $1 - k\lambda g / (1+\lambda g)^2$ for selected parameter combinations. The correction is negligible when $k$ or $g$ is small, but can reach 5--10% for high-dispersion, high-rate, large-gap scenarios. ```{r gap-correction-table} # Show correction magnitude for various parameter combos params <- expand.grid( lambda = c(0.3, 0.5, 1.0, 2.0), k = c(0.1, 0.3, 0.5, 1.0), gap = c(0.25, 0.5, 1.0) ) params$naive_eff <- params$lambda / (1 + params$lambda * params$gap) params$correction <- 1 - params$k * params$lambda * params$gap / (1 + params$lambda * params$gap)^2 params$corrected_eff <- params$naive_eff * params$correction params$pct_reduction <- round(100 * (1 - params$correction), 2) # Show a subset subset_df <- params[params$pct_reduction > 0.5, ] subset_df <- subset_df[order(-subset_df$pct_reduction), ] knitr::kable( head(subset_df[, c("lambda", "k", "gap", "naive_eff", "corrected_eff", "pct_reduction")], 12), digits = 4, row.names = FALSE, col.names = c("lambda", "k", "gap", "Naive eff. rate", "Corrected eff. rate", "Reduction (%)") ) ``` ### Simulation verification of average exposure We verify the average exposure computation by simulating trials with known parameters and comparing the simulated average exposure to the theoretical prediction. ```{r exposure-verification, fig.width=7, fig.height=5} set.seed(42) # Design parameters lambda1 <- 0.5 lambda2 <- 0.3 dispersion <- 0.3 dropout_rate_val <- 0.05 max_followup_val <- 8 trial_duration <- 12 # Piecewise accrual: 5/month for 4 months, then 15/month for 4 months accrual_rate_vec <- c(5, 15) accrual_duration_vec <- c(4, 4) # Theoretical calculation design <- sample_size_nbinom( lambda1 = lambda1, lambda2 = lambda2, dispersion = dispersion, power = 0.8, accrual_rate = accrual_rate_vec, accrual_duration = accrual_duration_vec, trial_duration = trial_duration, dropout_rate = dropout_rate_val, max_followup = max_followup_val ) theo_exposure <- design$exposure[1] # Same for both groups (same dropout) # Simulate many trials and compute average exposure enroll_rate <- data.frame(rate = design$accrual_rate, duration = accrual_duration_vec) fail_rate <- data.frame( treatment = c("Control", "Experimental"), rate = c(lambda1, lambda2), dispersion = dispersion ) dropout <- data.frame( treatment = c("Control", "Experimental"), rate = rep(dropout_rate_val, 2), duration = rep(100, 2) ) n_sims <- 100 sim_exposures <- numeric(n_sims) for (i in seq_len(n_sims)) { sim_data <- nb_sim( enroll_rate = enroll_rate, fail_rate = fail_rate, dropout_rate = dropout, max_followup = max_followup_val, n = design$n_total ) cut <- cut_data_by_date(sim_data, cut_date = trial_duration) sim_exposures[i] <- mean(cut$tte_total) } cat("Theoretical average exposure:", round(theo_exposure, 4), "\n") cat("Simulated average exposure (mean over", n_sims, "trials):", round(mean(sim_exposures), 4), "\n") cat("Simulation SE:", round(sd(sim_exposures) / sqrt(n_sims), 4), "\n") cat("Relative difference:", round(100 * (mean(sim_exposures) - theo_exposure) / theo_exposure, 2), "%\n") # Histogram of simulated average exposures hist(sim_exposures, breaks = 30, col = "steelblue", border = "white", main = "Distribution of simulated average exposure", xlab = "Average exposure per subject" ) abline(v = theo_exposure, col = "red", lwd = 2, lty = 2) legend("topright", legend = c("Theoretical"), col = "red", lty = 2, lwd = 2) ``` The theoretical prediction from `sample_size_nbinom()` closely matches the simulation average, confirming the correctness of the average exposure computation under piecewise accrual, dropout, and maximum follow-up truncation. ## Statistical information The *statistical information* $\mathcal{I}$ for the log rate ratio $\theta$ is the inverse of the variance of the MLE $\hat\theta$. It plays a central role in sample size calculation, group sequential design, and information monitoring during a trial. ### Per-subject Fisher information For a single subject $i$ in group $g$ with exposure $t_i$ and expected count $\mu_i = \lambda_g t_i$, the contribution to the Fisher information is (from the negative binomial log-likelihood with log link): $$ \mathcal{I}_i = \frac{\mu_i}{1 + k\,\mu_i} $$ This formula has intuitive limits: - When $k = 0$ (Poisson): $\mathcal{I}_i = \mu_i = \lambda t_i$ (information is proportional to expected events). - When $k\mu_i \gg 1$ (large overdispersion): $\mathcal{I}_i \approx 1/k$ (information saturates regardless of exposure). ### Total information The total statistical information for the log rate ratio is: $$ \mathcal{I} = \frac{1}{\mathrm{Var}(\hat\theta)} = \frac{1}{1/W_1 + 1/W_2} $$ where $W_g = \sum_{i \in \text{group } g} \mathcal{I}_i$ is the total information contributed by group $g$. When all subjects in a group have the same average exposure $\bar{t}_g$: $$ W_g = n_g \cdot \frac{\mu_g}{1 + k_g \mu_g} = \frac{n_g}{1/\mu_g + k_g} $$ and thus: $$ \mathrm{Var}(\hat\theta) = \frac{1/\mu_1 + k_1}{n_1} + \frac{1/\mu_2 + k_2}{n_2} $$ This is exactly the alternative variance used in the Wald sample size formula. For score-test sizing, the null variance is also computed under the constrained null and used in the Type I error component of the sample size calculation. ### Information in practice: blinded and unblinded estimation During a clinical trial, the statistical information can be estimated at interim analyses for information monitoring. The package provides four variants: | Method | Blinding | Estimation | Function | |:-------|:---------|:-----------|:---------| | Blinded ML | Blinded | Maximum likelihood via `MASS::glm.nb()` | `calculate_blinded_info()` | | Unblinded ML | Unblinded | Maximum likelihood via `MASS::glm.nb()` | `mutze_test()` (SE gives info) | | Blinded MoM | Blinded | Method of moments | `estimate_nb_mom()` | | Unblinded MoM | Unblinded | Method of moments | `estimate_nb_mom(group = "treatment")` | **Blinded estimation** [@friede2010blinded] fits a single model to pooled data (ignoring treatment assignment) and then "splits" the estimated overall rate $\hat\lambda$ into group-specific rates using the *planned* rate ratio: $\hat\lambda_1 = \hat\lambda / (p_1 + p_2 \cdot \text{RR}_{\text{plan}})$ and $\hat\lambda_2 = \hat\lambda_1 \cdot \text{RR}_{\text{plan}}$. This preserves the blind while allowing nuisance parameter ($k$) re-estimation. **Unblinded estimation** fits a model with a treatment covariate, directly estimating the log rate ratio and its standard error. The information is $\mathcal{I} = 1/\text{SE}^2$. ### Connection to sample size The sample size formula can be expressed in terms of information. For the Wald formula, the target information for a fixed design is: $$ \mathcal{I}_{\text{target}} = \frac{(z_{\alpha/s} + z_\beta)^2}{(\theta - \theta_0)^2} $$ The sample size is then the smallest $n_{\text{total}}$ such that the expected information at the end of the trial exceeds $\mathcal{I}_{\text{target}}$. For score-test sizing, the equivalent calculation separates the information scale under the null from the information scale under the planned alternative through $V_0$ and $V_1$. ## Examples ### Basic calculation [@zhu2014sample] Calculate sample size for: * Control rate $\lambda_1 = 0.5$ * Treatment rate $\lambda_2 = 0.3$ * Dispersion $k = 0.1$ * Power = 80% * Alpha = 0.025 (one-sided) * Accrual over 12 months * Trial duration 12 months (implying average exposure of 6 months) ```{r} sample_size_nbinom( lambda1 = 0.5, lambda2 = 0.3, dispersion = 0.1, power = 0.8, alpha = 0.025, sided = 1, accrual_rate = 10, # arbitrary, just for average exposure accrual_duration = 12, trial_duration = 12 ) ``` ### Piecewise constant accrual Consider a trial where recruitment ramps up: * 5 patients/month for the first 3 months * 10 patients/month for the next 3 months * Total trial duration is 12 months The function automatically calculates the average exposure based on this accrual pattern. ```{r} sample_size_nbinom( lambda1 = 0.5, lambda2 = 0.3, dispersion = 0.1, power = 0.8, accrual_rate = c(5, 10), accrual_duration = c(3, 3), trial_duration = 12 ) ``` ### Accrual with dropout and max follow-up Same design as above, but with a 5% dropout rate per unit time and a maximum follow-up of 6 months. ```{r} sample_size_nbinom( lambda1 = 0.5, lambda2 = 0.3, dispersion = 0.1, power = 0.8, accrual_rate = c(5, 10), accrual_duration = c(3, 3), trial_duration = 12, dropout_rate = 0.05, max_followup = 6 ) ``` ### Group-specific dropout rates Suppose the control group has a higher dropout rate (10%) than the treatment group (5%). ```{r} sample_size_nbinom( lambda1 = 0.5, lambda2 = 0.3, dispersion = 0.1, power = 0.8, accrual_rate = c(5, 10), accrual_duration = c(3, 3), trial_duration = 12, dropout_rate = c(0.10, 0.05), # Control, Treatment max_followup = 6 ) ``` ### Calculating power for fixed design Using the accrual rates and design from the previous example, suppose we want to calculate the power if the treatment effect is smaller ($\lambda_2 = 0.4$ instead of $0.3$). We use the `accrual_rate` computed in the previous step. ```{r} # Store the result from the previous calculation design_result <- sample_size_nbinom( lambda1 = 0.5, lambda2 = 0.3, dispersion = 0.1, power = 0.8, accrual_rate = c(5, 10), accrual_duration = c(3, 3), trial_duration = 12, dropout_rate = 0.05, max_followup = 6 ) # Use the computed accrual rates to calculate power for a smaller effect size sample_size_nbinom( lambda1 = 0.5, lambda2 = 0.4, # Smaller effect size dispersion = 0.1, power = NULL, # Request power calculation accrual_rate = design_result$accrual_rate, # Use computed rates accrual_duration = c(3, 3), trial_duration = 12, dropout_rate = 0.05, max_followup = 6 ) ``` ### Unequal allocation Sample size with a 2:1 allocation ratio ($n_2 = 2 n_1$). ```{r} sample_size_nbinom( lambda1 = 0.5, lambda2 = 0.3, dispersion = 0.1, ratio = 2, accrual_rate = 10, accrual_duration = 12, trial_duration = 12 ) ``` ### Example with event gap Calculate sample size assuming a 20-day gap after each event (approx 0.055 years). Note how the at-risk exposure differs between groups because the group with the higher event rate ($\lambda_1 = 2.0$) spends more time in gap periods than the group with the lower rate ($\lambda_2 = 1.0$). ```{r} sample_size_nbinom( lambda1 = 2.0, lambda2 = 1.0, dispersion = 0.1, power = 0.8, accrual_rate = 10, accrual_duration = 12, trial_duration = 12, event_gap = 20 / 365.25 ) ``` In this example, even though the calendar time is the same for both groups, the effective at-risk exposure is lower for Group 1 because they have more events and thus more gap time. ## References