--- title: "Fitting Network Scale-up Models" output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{networkscaleup-fitting-vignette} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r, include = FALSE} knitr::opts_chunk$set( collapse = TRUE, comment = "#>" ) ``` ## Overview This packaged fits several different network scale-up models (NSUM) to Aggregated Relational Data (ARD). ARD represents survey responses about how many people each respondents knows in different subpopulations through "How many X's do you know?" questions. Specifically, if $N_i$ respondents are asked how many people they know in $N_k$ subpopulations, then ARD is an $N_i$ by $N_k$ matrix, where the $(i,j)$ element represents how many people respondent $i$ reports knowing in subpopulation $j$. NSUM leverages these responses to estimate the unknown size of hard-to-reach populations. See Laga, et al. (2021) for a more details. In this package, we provide functions to estimate the size and accompanying parameters (e.g. degrees) from 4 papers: * Killworth, P. D., Johnsen, E. C., McCarty, C., Shelley, G. A., and Bernard, H. R. (1998) plug-in MLE * Killworth, P. D., McCarty, C., Bernard, H. R., Shelley, G. A., and Johnsen, E. C. (1998) MLE * Zheng, T., Salganik, M. J., and Gelman, A. (2006) overdispersed model * Laga, I., Bao, L., and Niu, X (2021) uncorrelated, correlated, and covariate models This vignette introduces each model and shows how to fit all of these models on one data set. ## Instructions First, load the library: ```{r setup} library(networkscaleup) ``` We will simulate a data set from the Binomial formulation in Killworth, P. D., Johnsen, E. C., McCarty, C., Shelley, G. A., and Bernard, H. R. (1998). ```{r simulation} set.seed(1998) N_i = 50 N_k = 5 N = 1e5 sizes = rbinom(N_k, N, prob = runif(N_k, min = 0.01, max = 0.15)) degrees = round(exp(rnorm(N_i, mean = 5, sd = 1))) ard = matrix(NA, nrow = N_i, ncol = N_k) for(k in 1:N_k){ ard[,k] = rbinom(N_i, degrees, sizes[k] / N) } ## Create some artificial covariates for use later x = matrix(sample(1:5, size = N_i * N_k, replace = T), nrow = N_i, ncol = N_k) z = cbind(rbinom(N_i, 1, 0.3), rnorm(N_i), rnorm(N_i), rnorm(N_i)) ``` We should also prepare the data for modeling by scaling all covariates to have standard deviation 1 and mean 0. Additionally, we need to define which columns of $Z$ belong to $z_{subpop}$ and $z_{global}$. ```{r prep} x = scale(x) z = scale(z) z_subpop = z[,1:2] z_global = z[,3:4] ``` ### PIMLE The plug-in MLE estimator from Killworth, P. D., Johnsen, E. C., McCarty, C., Shelley, G. A., and Bernard, H. R. (1998) is a two-stage estimator that first estimates the degrees for each respondent $d_i$ by maximizing the following likelihood for each respondent: \begin{equation} L(d_i;y, \{N_k\}) = \prod_{k=1}^L \binom{d_i}{y_{ik}} \left(\frac{N_k}{N} \right)^{y_{ik}} \left(1 - \frac{N_k}{N} \right)^{d_i - y_{ik}}, \end{equation} where $L$ is the number of subpopulations with known $N_k$. For the second stage, the model plugs in the estimated $d_i$ into the equation \begin{equation} \frac{y_{ik}}{d_i} = \frac{N_k}{N} \end{equation} and solves for the unknown $N_k$ for each respondent. These values are then averaged to obtain a single estimate of $N_k$. To summarize, stage 1 estimates $\hat{d}_i$ by \begin{equation} \hat{d}_i = N \cdot \frac{\sum_{k=1}^L y_{ik}}{\sum_{k=1}^L N_k} \end{equation} and then these estimates are used in stage 2 to estimate the unknown $\hat{N}_k$ by \begin{equation} \hat{N}_k^{PIMLE} = \frac{N}{n} \sum_{i=1}^n \frac{y_{ik}}{\hat{d}_i} \end{equation} These estimates are obtained using the following call to the `killworth` function. ```{r pimle} pimle.est = killworth(ard, known_sizes = sizes[c(1, 2, 4)], known_ind = c(1, 2, 4), N = N, model = "PIMLE") plot(degrees ~ pimle.est$degrees, xlab = "Estimated PIMLE degrees", ylab = "True Degrees") abline(0, 1, col = "red") round(data.frame(true = sizes[c(3, 5)], pimle = pimle.est$sizes)) ``` Note that the function provides a warning saying that at least $\hat{d}_i$ was 0. This occurs when a respondent does not resport knowing anyone in the known subpopulations. This is an issue for the PIMLE since a 0 value is in the denominator for $\hat{N}_u^{PIMLE}$. Thus, we ignore the responses from respondents that correspond to $\hat{d}_i = 0$. ### MLE Next, we analyze the data from the Killworth, P. D., McCarty, C., Bernard, H. R., Shelley, G. A., and Johnsen, E. C. (1998) MLE estimator. This is also a two-stage model, which an identical first stage, i.e. \begin{equation} \hat{d}_i = N \cdot \frac{\sum_{k=1}^L y_{ik}}{\sum_{k=1}^L N_k}. \end{equation} However, the second stage estimates $\hat{N}_k$ by maximizing the Binomial likelihood with respect to $\hat{N}_k$, fixing $d_i$ at the estimated $\hat{d}_i$. Thus, the estimate for the unknown subpopulation size is given by \begin{equation} \hat{N}_k^{MLE} = N \cdot \frac{\sum_{i=1}^n y_{ik}}{\sum_{i=1}^n \hat{d}_i}. \end{equation} These estimates are also obtained using a single call to the `Killworth` function. ```{r mle} mle.est = killworth(ard, known_sizes = sizes[c(1, 2, 4)], known_ind = c(1, 2, 4), N = N, model = "MLE") plot(degrees ~ mle.est$degrees, xlab = "Estimated MLE degrees", ylab = "True Degrees") abline(0, 1, col = "red") round(data.frame(true = sizes[c(3, 5)], pimle = mle.est$sizes)) ``` Note that there is no warning here since the denominator depends on the summation of $\hat{d}_i$. ## Bayesian Models Now we introduce the two Bayesian estimators implemented in this package. ### Overdispersed Model The overdispersed model proposed in Zheng et al. (2006) assumes the following likelihood: \begin{equation} y_{ik} \sim \text{negative-binomial}(\text{mean} = e^{\alpha_i + \beta_k}, \text{overdispersion} = \omega_k) \end{equation} Please see the original manuscript for more details on the model structure and priors. This package fits this overdispersed model either via the Gibbs-Metropolis algorithm provided in the original manuscript (`overdispersed`) or via Stan (`overdispersedStan`). We suggest using the Stan version since convergence and effective sample sizes are more satisfactory in the Stan implementation, and does not require tuning jumping scales for Metropolis updates. In order to identity the $\alpha_i$ and $\beta_k$ as log-degrees and log-prevalences, respectively, the overdispersed model requires scaling the parameters. In order to scale the parameters, the user must supply at least one subpopulation with known size and the column index corresponding to that known size. Additionally, a two secondary groups may be supplied which can adjust for differences in gender or other binary group classifications. More details of the scaling procedure can be found in the original manuscript. Below we fit both the overdispersed and overdispersedStan implementations to the ARD and compare estimates. Note that in practice, both warmup and iter should be set to higher values. ```{r overdisp} overdisp_gibbs_metrop_est = overdispersed( ard, known_sizes = sizes[c(1, 2, 4)], known_ind = c(1, 2, 4), G1_ind = 1, G2_ind = 2, B2_ind = 4, N = N, warmup = 500, iter = 1000, verbose = TRUE, init = "MLE" ) overdisp_stan = overdispersedStan( ard, known_sizes = sizes[c(1, 2, 4)], known_ind = c(1, 2, 4), G1_ind = 1, G2_ind = 2, B2_ind = 4, N = N, chains = 2, cores = 2, warmup = 250, iter = 500, ) round(data.frame(true = sizes, gibbs_est = colMeans(overdisp_gibbs_metrop_est$sizes), stan_est = colMeans(overdisp_stan$sizes))) plot(degrees ~ colMeans(overdisp_stan$degrees), xlab = "Overdispersed Degree Estimates", ylab = "True Degrees") abline(0, 1, col = "red") ``` ### Correlated Models The correlated model proposed in Laga et al. (2022+) assumes the following likelihood \begin{equation} y_{ik} \sim Poisson\left(exp\left\{\delta_i + \rho_k + {\beta}_{global} {z}_{i,global} + {\beta}_{k,subpop} {z}_{i,subpop} + \alpha_k x_{ik} + b_{ik} \right\} \right), \end{equation} where critically, \begin{equation} \mathbf{b}_i \sim \mathcal{N}_k\left({\mu}, \Sigma\right), \end{equation} i.e. the responses for each respondent are correlated across subpopulations. Again, $\delta_i$ and $\rho_k$ need to be scaled, and they can either be scaled using the same procedure as for the overdispersed model (providing indices corresponding to different groups), by using all known subpopulation sizes, or by weighting groups according to their correlation with other groups. More details about these scaling procedures are provided in Laga et al. (2022+). In this package, model parameters are estimated via Stan. Note that while the full model likelihood depends on ${X}$, ${Z}_{global}$, and ${Z}_{subpop}$, any combination of these covariates can be provided. Additionally, we can assume that $\Sigma$ is a diagonal matrix (i.e. no correlation) by setting the argument `model = uncorrelated` in the `correlatedStan` function. ```{r correlated} correlated_cov_stan = correlatedStan( ard, known_sizes = sizes[c(1, 2, 4)], known_ind = c(1, 2, 4), model = "correlated", scaling = "weighted", x = x, z_subpop = z_subpop, z_global = z_global, N = N, chains = 2, cores = 2, warmup = 250, iter = 500, ) correlated_nocov_stan = correlatedStan( ard, known_sizes = sizes[c(1, 2, 4)], known_ind = c(1, 2, 4), model = "correlated", scaling = "all", N = N, chains = 2, cores = 2, warmup = 250, iter = 500, ) uncorrelated_cov_stan = correlatedStan( ard, known_sizes = sizes[c(1, 2, 4)], known_ind = c(1, 2, 4), model = "uncorrelated", scaling = "all", x = x, z_subpop = z_subpop, z_global = z_global, N = N, chains = 2, cores = 2, warmup = 250, iter = 500, ) uncorrelated_x_stan = correlatedStan( ard, known_sizes = sizes[c(1, 2, 4)], known_ind = c(1, 2, 4), model = "uncorrelated", scaling = "all", x = x, N = N, chains = 2, cores = 2, warmup = 250, iter = 500, ) round(data.frame(true = sizes, corr_cov_est = colMeans(correlated_cov_stan$sizes), corr_nocov_est = colMeans(correlated_nocov_stan$sizes), uncorr_cov_est = colMeans(uncorrelated_cov_stan$sizes), uncorr_x_est = colMeans(uncorrelated_x_stan$sizes))) plot(degrees ~ colMeans(correlated_cov_stan$degrees), xlab = "Correlated Covariate Degree Estimates", ylab = "True Degrees") abline(0, 1, col = "red") ## Examine parameter estimates colMeans(correlated_cov_stan$alpha) colMeans(correlated_cov_stan$beta_global) colMeans(correlated_cov_stan$beta_subpop) ```