The collective dynamics of neural ensembles create complex spike patterns with many spatial and temporal scales. Understanding the statistical structure of these patterns can help resolve fundamental questions about neural computation and neural dynamics. Spatiotemporal conditional inference (STCI) is introduced here as a semiparametric statistical framework for investigating the nature of precise spiking patterns from collections of neurons that is robust to arbitrarily complex and nonstationary coarse spiking dynamics. The main idea is to focus statistical modeling and inference not on the full distribution of the data, but rather on families of conditional distributions of precise spiking given different types of coarse spiking. The framework is then used to develop families of hypothesis tests for probing the spatiotemporal precision of spiking patterns. Relationships among different conditional distributions are used to improve multiple hypothesis-testing adjustments and design novel Monte Carlo spike resampling algorithms. Of special note are algorithms that can locally jitter spike times while still preserving the instantaneous peristimulus time histogram or the instantaneous total spike count from a group of recorded neurons. The framework can also be used to test whether first-order maximum entropy models with possibly random and time-varying parameters can account for observed patterns of spiking. STCI provides a detailed example of the generic principle of conditional inference, which may be applicable to other areas of neurostatistical analysis.

Statistical nonparametric modeling tools that enable the discovery and approximation of functional forms (e.g., tuning functions) relating neural spiking activity to relevant covariates are desirable tools in neuroscience. In this article, we show how stochastic gradient boosting regression can be successfully extended to the modeling of spiking activity data while preserving their point process nature, thus providing a robust nonparametric modeling tool. We formulate stochastic gradient boosting in terms of approximating the conditional intensity function of a point process in discrete time and use the standard likelihood of the process to derive the loss function for the approximation problem. To illustrate the approach, we apply the algorithm to the modeling of primary motor and parietal spiking activity as a function of spiking history and kinematics during a two-dimensional reaching task. Model selection, goodness of fit via the time rescaling theorem, model interpretation via partial dependence plots, ranking of covariates according to their relative importance, and prediction of peri-event time histograms are illustrated and discussed. Additionally, we use the tenfold cross-validated log likelihood of the modeled neural processes (67 cells) to compare the performance of gradient boosting regression to two alternative approaches: standard generalized linear models (GLMs) and Bayesian P-splines with Markov chain Monte Carlo (MCMC) sampling. In our data set, gradient boosting outperformed both Bayesian P-splines (in approximately 90% of the cells) and GLMs (100%). Because of its good performance and computational efficiency, we propose stochastic gradient boosting regression as an off-the-shelf nonparametric tool for initial analyses of large neural data sets (e.g., more than 50 cells; more than 10 5 samples per cell) with corresponding multidimensional covariate spaces (e.g., more than four covariates). In the cases where a functional form might be amenable to a more compact representation, gradient boosting might also lead to the discovery of simpler, parametric models.