Datasets:

SouthernCrossAI
/

Arxiv

Dataset card Data Studio Files Files and versions Community

Split (1)

train · 132k rows

Filename stringlengths 22 64	Paragraph stringlengths 8 5.57k
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	Publications of the Astronomical Society of Australia (PASA) c(cid:13) Astronomical Society of Australia 2021; published by Cambridge University Press. doi: 10.1017/pas.2021.xxx.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	Abstract The latest generation of radio astronomy interferometers will conduct all sky surveys with data products consisting of petabytes of spectral line data. Traditional approaches to identifying and parameterising the astrophysical sources within this data will not scale to datasets of this magnitude, since the performance of workstations will not keep up with the real-time generation of data. For this reason, it is necessary to employ high performance computing systems consisting of a large number of processors connected by a high- bandwidth network. In order to make use of such supercomputers substantial modiﬁcations must be made to serial source ﬁnding code. To ease the transition, this work presents the Scalable Source Finder Framework, a framework providing storage access, networking communication and data composition functionality, which can support a wide range of source ﬁnding algorithms provided they can be applied to subsets of the entire image. Additionally, the Parallel Gaussian Source Finder was implemented using SSoFF, utilising Gaussian ﬁlters, thresholding, and local statistics. PGSF was able to search on a 256GB simulated dataset in under 24 minutes, signiﬁcantly less than the 8 to 12 hour observation that would generate such a dataset.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	A critical stage of radio astronomy spectral-line image analysis is source ﬁnding, which identiﬁes the galax- ies present in the image and determines their position and other parameters. As surveys increase in size, with larger ﬁelds of view and greater resolution, they produce greater amounts of data. For example, the HIPASS sur- vey (Meyer et al. 2004) produced a total of 22GB of im- age data. By comparison the Wideﬁeld ASKAP L-band Legacy All-sky Blind surveY (WALLABY) survey us- ing the Australian Square Kilometre Array Pathﬁnder (ASKAP) telescope is expected to produce ﬁles of at least 256GB every 8 to 12 hours, with the entire all sky survey likely to total several petabytes.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	addition to meeting real-time performance, the rate of source ﬁnding would ideally be signiﬁcantly faster than the rate of production. This would allow reprocessing of the entire dataset should the source ﬁnder be improved during the survey.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	Memory issues can also slow a source ﬁnder. If the machine running the source ﬁnder has insuﬃcient phys- ical RAM to store the data needed by the source ﬁnder, either the excess data will be stored on the hard disk, making access much slower, or the system will fail to allocate suﬃcient memory, halting the program. It is possible to write a source ﬁnder that only examines a portion of the image at a time, reducing the memory required, but this involves processing part or all of the image more than once. Because supercomputers have large amounts of memory available, it is more eﬃcient to process the whole image at once.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	In order to overcome these limitations, it is desir- able to use multiple machines working together on the problem. Modern supercomputers consist of a cluster of computing nodes, where each node consists of one or more multi-core CPUs. A fast network is employed to connect the nodes to each other, and to a parallel ﬁle storage system. The scalability of the program across these nodes is important because future surveys will produce even greater amounts of data. It is desirable for the program to be able to expand and make eﬀec- tive use of a greater number of processors in order to search greater amounts of data.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	However, in order for a source ﬁnding program to make use of such systems they must be written such that the data and processing are partitioned across the nodes, with communication via the network, and using parallel ﬁle operations. Additionally, specialised code libraries and application programming interfaces must be used such as MPI (The MPI Forum 1993), MPI- IO and OpenMP (OpenMP Architecture Review Board 2008). Converting a serial program to run in parallel can thus take a signiﬁcant amount of eﬀort.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	This work describes the Scalable Source Finding Framework (SSoFF), a framework with functionality to ease this transition. SSoFF handles the distribution of processing by dividing the image into portions and as- signing them to a three-dimensional grid of processes. Each process performs the work required to search its portion of the image. SSoFF provides routines that al- low the processes to read and write their portions of the image from the storage system, and to exchange intermediary values with their neighbouring processes. With the functionality described above in place, ex- isting source ﬁnding analysis routines can be adapted to process a portion of image data, and added to SSoFF. To demonstrate this, the Parallel Gaussian Source Finder (PGSF) was built using the framework. The analysis step of PGSF applies a series of three- dimensional, Gaussian ﬁlters to the data. For each ﬁlter, a threshold is applied based on the local data around each voxel, and voxels are selected if they are above the threshold for a set number of diﬀerent ﬁlters. Addition- ally, voxel weightings can optionally be used if available. Section 2 provides a background to source ﬁnding in radio astronomy. Section 3 presents each component of the framework in detail. Section 4 then details the im- plementation of PGSF. Section 5 provides benchmark- ing and correctness testing result, which are then dis- cussed in Section 6. Finally, concluding remarks are in- cluded in Section 7.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	A source ﬁnder can form part of a pipeline for re- ducing and analysing data from telescopes. While the details of the conﬁguration of the pipeline are highly speciﬁc to the instrument, survey and science goals, a general overview of the main stages for a interfero- metric, spectral-line HI survey are as follows. The ﬁrst step is correlation, where the data from the diﬀerent receivers are combined into visibilities. The imaging takes the calibrated visibilities and converts them into an image through a Fourier transformation. The imag- ing step also includes continuum subtraction, where un- wanted continuum emission is removed from the data, and deconvolution, where sidelobes are removed from the image. Usually the source ﬁnder sits at the end of this pipeline, taking the images and searching them for sources, but in some cases they can be used to search the visibilities. The objects that are found are then mea- sured to determine their properties, a process called pa- rameterisation. The parameterised sources found by the source ﬁnder are then analysed to achieve the desired science goals for the survey.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	The main measure of merit of a source ﬁnder is its accuracy, which has two components, completeness and reliability. Completeness is the fraction of the sources in the image that have been found by the source ﬁnder. Re- liability is the fraction of sources reported by the source ﬁnder that are real sources in the image. Independent of the source ﬁnding accuracy is the accuracy of the parameterisation step.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	The source ﬁnding framework presented in this work is intended for spectroscopic images of neutral hydrogen (HI) emission. A radio astronomy image, also known as a data cube, breaks up the area of the sky being ob- served along three dimensions. The ﬁrst two are spatial dimensions that denote the direction in the sky relating to a particular part of the image. The third dimension is frequency, which gives the frequency range of the ob- served radiation for a particular element of the image. For nearby sources this value may also be speciﬁed in terms of velocity, as frequency and velocity are related through the Doppler eﬀect from the radial velocity of an object.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	The general process of searching an image is shown in Figure 1. The ﬁrst step of the program is to read the image from storage into memory, in the input step. This also involves any conversion of data to a format that the source ﬁnder uses.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	Figure 1. Source Finding Stages. The radio image is read by the source ﬁnder and analysed to create a measure that states the likelihood that each voxel is part of a real source. The results of this analysis are used in the source formation step to select the set of voxels that are likely to be true sources, which then merges the chosen voxels together to form objects. The positions of these objects and the original image data are used to determine the parameters of the objects, and a conﬁrmation step is applied to remove objects that appear to be false detections. The remaining objects and their parameters are produced as the output to the program.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	based on their characteristics above the noise, whereas analysis techniques use statistical techniques to calcu- late the likelihood that a particular voxel is part of a real source. This step performs the bulk of the work and has the greatest diversity among the current se- rial source ﬁnding programs. MultiFind (Meyer et al. 2004) searches for voxels that are above a threshold af- ter Hanning smoothing the data and Tophat (Meyer et al. 2004) searches for voxels that are above a thresh- old after convolving the data with top hat ﬁlters of dif- ferent channel widths. Duchamp (Whiting 2012) uses a choice of smoothing or the `a trous wavelet transform to reduce noise. The 2D-1D Wavelet source ﬁnder (Fl¨oer & Winkel 2012) uses 2D-1D wavelet transform to re- duce noise, with the 2D transform operating in the spatial dimensions and the 1D transform operating in the spectral dimension. The Smooth Plus Cut (S+C) source ﬁnder (Serra et al. 2012) applies a series of dif- ferent series of box ﬁlters, and takes the union of vox- els that are above the threshold for each ﬁlter. The Characterised Noise HI (CNHI) source ﬁnder (Jurek 2012) uses the Kuiper test to compare a test region of voxels to the noise of the image to locate regions with diﬀerent ﬂux properties, and the Gamma Finder (Winkel 2008) uses the Gamma test to search for dis- continuities in otherwise continuous, noisy data.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	Voxels that are considered likely to be part of a true detection are selected from the results of the ﬁltering or analysis, and then merged together to form objects in the source formation step. For many ﬁltering tech- niques, the selection of voxels involves calculating a threshold in ﬂux or signal-to-noise ratio (SNR), and se- lecting voxels that have a value greater than this thresh- old. Some analysis techniques may eﬀectively do this as part of their analysis algorithm.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	source fragmentation. Source confusion occurs when two or more real objects are considered by the source ﬁnder to be the same object. Although HI sources are separated in three dimensions it is still possible for con- fusion to occur, depending on the proximity of the ob- jects and the resolution of the image. Source fragmenta- tion occurs when a source ﬁnder splits up a real object into two or more objects.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	The objects that have been created in the source formation step are measured, using the original image data, to determine their parameters in the parameteri- sation step. These parameters include the position, size and brightness of the objects, and will be used to study the galaxies, and other objects, found in the image. Pa- rameterisation can be considered a separate, though re- lated, problem to source ﬁnding, where source ﬁnding involves locating the sources of emission in an image, and parameterisation measures the properties of the sources. Another reason that parameterisation may be considered as separate to source ﬁnding is that once the source object positions are known, several diﬀerent parameterisation techniques may be employed.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	With the parameters of the objects known, the ob- jects are passed through a conﬁrmation step. This com- ponent of the program examines the properties of each object and removes those whose properties suggest that they are likely to be false detections. The objects that survive the conﬁrmation step are ﬁnally written out in the output step of the program.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	There is limited work currently published on apply- ing source ﬁnding using HPC techniques. Whiting & Humphreys (2012) describes Selavy, a parallelisation of the Duchamp source ﬁnder. The framework presented in the following section is intended to make such parallel source ﬁnders easier to implement.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	Figure 2. Data Distribution between Nodes. Each node has its own section of the data cube, along with a portion from each adjacent node, so that it can correctly process the data assigned to it. The colour of the data in- dicates which node that data was assigned to. The white sections are padding that is placed around the edge of the image, so that the edge nodes can be processed in the same manner as the interior nodes. The division and allocation of image data to the processes can be changed, which may aﬀect the computational performance of a source ﬁnder, but not its accuracy. The data can also be partitioned along the spectral axis, which is not shown here for clarity.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	tures for each process to store its assigned image data and halo data as a three-dimensional array. This array data structure is used for the initial image data, inter- mediate values and the processed image. The manner in which the image data is divided between processes and stored is shown in Figure 2.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	to address a ﬁle may be unable to properly read in and access the entire ﬁle. This function also provides a con- venient way to bypass a limitation in MPI-IO, where it can only read in 2GB of data per call. The data input function used by SSoFF avoids this problem by using multiple MPI-IO function calls to read the data. The data structures used by SSoFF to store the image data are addressed using three 32-bit integers, so the entire image can be accessed using integers as long as each side length is less than 231 − 1 elements, the maximum value of a signed integer. Alternatively, functions that use the data structures provided by SSoFF may choose to access them using a single 64-bit integer index.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	The halo data for the edges of the cube is set to zero. When using weighted calculations, this will auto- matically assign those voxels a weight of zero. For un- weighted calculations, these voxels are treated the same as the voxels from the image. The eﬀect of the padded values is left to be considered in future work.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	There are two options of ensuring that each node has the halo data it needs. Either the nodes read both their assigned data and the halo data directly from the stor- age system, or they read only their assigned data from storage, then use the network to transfer the halo data. Because the same network is used to transfer data from the storage system and inter-node communication, both of these methods would result in the same performance if the network is the bottleneck. However, if the stor- age system is the bottleneck in the transfer, then the second method will be faster because it reads less data from storage. Therefore, the framework uses the net- work to transfer the halo data, as covered in the next section.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	Image analysis algorithms may require the image values around a voxel, in order to evaluate that voxel. In order for a process to analyse voxels near its border, it will require information that was assigned to its neighbour- ing processes. SSoFF provides a function to copy data from one process to the appropriate position in the halo data of its surrounding processes, using the array data structures mentioned above. The transfer is performed in three steps, as shown in Figure 3. First the halo data in the x axis is transferred, to the processes’ left and right neighbours. Once this is complete, data is trans- ferred in the y axis, to the top and bottom neighbours, including sending data that was received in the x axis transfer. Finally, the data is transferred in the z axis, be- tween the front and back neighbours. Transferring data that was received from other nodes, in addition to data from a process’s own node, ensures that processes still get the data they need even when they are not adjacent in the process grid.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	Figure 3. Halo Communication. SSoFF transfers halo data in three steps, one for each axis. The bold lines show the data transferred in each step. The processes send the data they hold, as well as data that they received in previous steps. Only the x axis and y axis transfers are shown here, the framework also does a third transfer along the z axis.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	Source ﬁnders often require statistical functions to pro- vide a measure of the probability that a voxel is part of a valid source. SSoFF provides several statistics func- tions that operate across a distributed dataset. These include calculating the mean and standard deviation of the dataset, for both weighted and unweighted data. There is an option for global statistics, that calculate the values based on the entire contents of the array, and a local version that calculates the mean and stan- dard deviation for each voxel individually based on the data within a user-speciﬁed range that is Lx, Ly and Lz voxels in size along the x, y, and z axes. Because the local statistics calculations use data from the surround- ing nodes, these functions impose a minimum halo size equal to the range that the local statistics are being calculated across. The global statistics are calculated by each node determining the sum of its own portion of the data then using MPI to perform a global sum reduction across the diﬀerent processes.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	The calculation of the local statistics requires that the node possesses the values surrounding a voxel up to the speciﬁed range, so the halo size must be at least as large as the local statistics size, and the program must perform a halo transfer so that each process has the information it needs. Once this transfer is com- plete, each node can calculate the individual sum for each voxel. The mean, µw, and the standard deviation, σw for weighted data are calculated as shown in Equa- tions 12 and 10, from the image data d and the weights data w. In the case of unweighted data, the mean, µu, and standard deviation, σu are calculated as in Equa- tions 17 and 15. With the mean and standard deviation values known, they are used to calculate the z score for each voxel in the data, as shown in Equations 9 and 14.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	The totals are calculated using moving sums, track- ing three values for Sw, W and µw. In order to minimise memory use the statistics are performed in-place, over- writing the original image values with the sigma value for each voxel. Small additional buﬀers are used to store intermediate results. Each of these temporary buﬀers are (dx + Lx − 1)(dy + Ly − 1) Lz 2 elements in size. This is the smallest possible size of the buﬀers because the algorithm overwrites the image data in the x direction as it progresses, so previous values along the x axis must be read from a buﬀer while following values can be read from the image array. Both the weighted and unweighted local statistics calculations use one buﬀer of ﬂoats to store the original image information, and two buﬀers of doubles to store the sums for the mean and standard deviation. The weighted calculation requires an additional buﬀer of doubles to store the summed weights.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	sults in a total of six ﬂoating point operations per voxel. Because the buﬀers are half the size of Lz along the z axis each voxel is initialised twice, so this initialisation requires twelve ﬂoating point operations per voxel.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	Once the moving sums have been initialised, it is ex- ecuted across the buﬀer data, adding in new values and subtracting old values. This part of the algorithm re- quires twice as many calculations as the initialisation due to subtracting previous values but is only performed once per voxel. Therefore, performing the moving sum across the z axis requires a total of 24 ﬂoating point operations per voxel. The z axis sums are performed on all of a node’s assigned voxels plus a halo of Lx voxels in the x axis and Ly voxels in the y direction, because these values will be used in later sums. The number of z-axis operations required to calculate the local statis- tics could be reduced by using larger buﬀers, but this would come at the cost of increased memory use.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	Once the z axis sums are complete, the sums are per- formed across the y axis, and then the x axis. These sums only require additions and subtractions, as all multiplications have been performed in calculating the sums across the z axis. Initialising these sums requires three operations per voxel, one addition for each of the three values. The y-axis initialisation only needs to be performed once for each x-z line in the image and is calculated over Ly voxels per line. Likewise, the x-axis moving sum initialisation is performed once for each y-z line in the image, across Lx voxels per line. Once ini- tialised, performing the moving sums requires six ﬂoat- ing point operations per voxel, as the previous values need to be subtracted from the sum. The y-axis sums are calculated across a node’s assigned voxels, and an additional halo in the x axis. The x axis sums require no halo. With the sums calculated across the three axes, they are used to calculate the ﬁnal z value. This re- quires ﬁve ﬂoating point operations, three divides and one square root per voxel assigned to a node.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	A source ﬁnder must decide which voxels in an image are considered part of a legitimate source of emission and to collect these voxels into data objects that represent these sources. These are two separate but related tasks, called selection and merging. SSoFF provides function- ality to perform these tasks in a parallel environment. Selecting voxels is often done by applying a threshold to a dataset. This framework allows for thresholding across both the image array data structure, and across a sparse image dataset in the form of a hash map. These functions use a ﬂood ﬁll algorithm to pick the voxels that are above the threshold, and merge adjacent voxels into source objects.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	The ﬂood ﬁll merges source objects within a single process but sources may be split across multiple pro- cesses. SSoFF merges these objects using a multi-step procedure. First each process sends the positions of vox- els along its borders to its neighbouring processes. This is used to ﬁnd where source objects are split between processes. Each part of a split object is given a destina- tion process. The destination for a split source object is the process with a part of that object, that has the low- est index in the process grid. The destination process index is propagated across the diﬀerent parts of a split source, from process to process, to ensure that each part of the source is sent to the same process.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	eter ﬁle to decide which functions to use at run time. The options can be used to specify values used inside the functions, such as the size of ﬁlters, or the value of the threshold to use when selecting voxels.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	The framework makes use of several libraries to pro- vide this functionality. MPI is used for the basis of the parallelisation, and to communicate data between pro- cesses. SimCList is used for linked lists, which are used to store collections of voxels and source objects, which vary in size depending on the contents of the dataset. uthash is used for hash tables, which are used to store sparse voxel information, and to store parameter ﬁle in- formation. These libraries do not prevent source ﬁnders from using other libraries.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	The functions described in this section provide a toolkit for writing a parallel source ﬁnder and can per- form common source ﬁnding tasks in a parallel environ- ment. Through this functionality, SSoFF reduces the diﬃculty of implementing additional functionality to a parallel source ﬁnder. The use of this framework is demonstrated in the next section, where it is used to implement a source ﬁnder.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	This section describes the Parallel Gaussian Source Finder (PGSF), a parallel source ﬁnder for HI spectral line images implemented using SSoFF. The analysis is based on the use of three-dimensional Gaussian ﬁlters, and voxels are selected if they are above the threshold for a set number of diﬀerent ﬁlters. Sources constructed from these voxels are then subject to a conﬁrmation step where only the sources whose spectral extent is is greater than the user-speciﬁed cut-oﬀ are written to the catalogue. PGSF can make use of an arbitrary num- ber of processes, up to the number of voxels in the im- age being searched. This source ﬁnder can also process large ﬁles, limited by the memory of the nodes used to search the image and to a maximum size of 231 − 1 vox- els in each dimension. It processes an image that con- sists of single-precision ﬂoating point numbers, but it can be easily extended to other data types. The details of PGSF are described below.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	sult, ﬁlter output is only calculated once for values in the since the output for values in the halos are either calculated by the adjacent node that is responsible for that region, or not at all for the values outside the im- age. Additionally, the ﬁltering process requires that the halo be at least the size of the largest ﬁlter used. When a process is ﬁltering the edges of its assigned image data it uses the halo image values, and weights values if they are available, that were loaded in the input step of the program.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	For each ﬁlter the selection criteria used is a threshold equal to the mean plus a user-speciﬁed constant multi- plied by the standard deviation, where the local mean and standard deviation are calculated from the ﬁltered image, as in Whiting & Humphreys (2012), using the local statistics functions of the framework. A count is kept for each voxel each time that voxel’s ﬁltered value above the threshold for a ﬁlter. After all of the ﬁlters have been applied, voxels are selected if their count is above a second user-deﬁned threshold.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	If a ﬁlter has a size of Fx, Fy and Fz elements, then the weighted version requires 4DxDyDzFxFyFz total ﬂoating point operations for that ﬁlter. This includes one multiply to combine the ﬁlter value and the weight for a voxel, one multiply to combine the ﬁlter weight value to the image value, one addition to update the sum of the convolution, and an addition to update the sum of the ﬁlter weight for that voxel. The unweighted version requires a total of 2DxDyDzFxFyFz operations, one multiply to combine the ﬁlter value and the image data value and one addition to update the sum of the ﬁlter values. This ﬁltering does not require any data transfer between nodes beyond what is done reading in the image.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	The parameterisation of the sources is performed once they have been selected and merged by SSoFF. Because the data ﬁle format used by this program only stores the ﬂux of the image, that is, the format stores no metadata, only a subset of the parameters can be determined. Performing a complete parameterisation of sources is considered to be outside the scope of this work. The parameters given by this program are the position of an object, as a ﬂux-weighted mean in units of the array indices, the peak ﬂux of the object and the sum of its ﬂux across its voxels, and the width of the object along the spectral axis. PGSF also handles re- reading the information from the image cube when it is needed, to parameterise sources that contain voxels that were received from other nodes during the source formation step.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	The conﬁrmation of sources makes use of the pa- rameterisation information to conﬁrm or reject poten- tial sources. PGSF rejects sources that are below a user- speciﬁed channel width. This is because most legiti- mate sources have a spectral width that is signiﬁcantly larger than the channel width of a spectral-line image, so sources that have a small channel width are likely to be noise peaks or interference. For example, the thinnest galaxy in the HIPASS Catalogue Meyer et al. (2004), J1336-29 has a velocity width of 30.4km/s 1 compared to the WALLABY survey, which will have a spectral res- olution of 4km/s Koribalski & Staveley-Smith (2009). The framework allows for more complex conﬁrmation techniques to be added. The conﬁrmed sources and their parameters are written to the output catalogue.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	PGSF can scale to search larger images, up to datasets that are 231 − 1 elements along each side. The number of processes to be used by the source ﬁnder has a upper limit equal to the number of voxels in the image, and a lower limit set by memory limits. Each process has a copy of its assigned portion of the image, including the halo data, which is (dx + Hx)(dy + Hy)(dz + Hz) elements per process. A second array data structure of the same size is used to store the ﬁltered image, and optionally a third data structure to store the weights information of the image. Additional memory is used when calculating the statistics, as mentioned above, and a variable amount of memory is needed to store the source detections. The accuracy and computational performance of this program is measured in the next section.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	that the program was working correctly. The second set of tests shows the accuracy of the source ﬁnder for dif- ferent sources. Finally, the third set of tests analyse the suitability of the program for processing large datasets. The primary machine used to test the program was the Epic@Murdoch supercomputer. This system has 800 nodes, each possessing two six-core Intel Xeon X5660 CPUS, 24GB of RAM and a QLogic IBA7322 QDR In- ﬁniband interconnect. The MPI library used was Open- MPI version 1.6.3.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	The correctness of the program was examined using unit testing. These tests do not concern the overall accu- racy of the program. Rather, they show that the frame- work functions work correctly. Each function used in PGSF was individually tested for correctness. The cor- rectness for the program as a whole was demonstrated by running the source ﬁnder on a 2GB simulated data cube and comparing the output to the expected results, which were obtained by executing a single-threaded im- plementation of the program on the same image. The results were found to be identical.
Processed_A_Framework_for_HI_Spectral_Source_Finding_Using_D.txt	The accuracy of PGSF was demonstrated by execut- ing the source ﬁnder on a pair of simulated data cubes, one containing point sources and one containing ex- tended sources. These are the images used in Popping et al. (2012). Point sources are objects whose spatial extent is smaller than the resolution of the instrument used to observe them, whilst extended sources are those who are spatially larger than the resolution of the in- strument. In practice, both types of sources have spec- tral sizes signiﬁcantly larger than the channel width of the telescope. This test used a series of ﬁlters that are 1, 3, 5 and 9 pixels wide in the spatial dimensions and 9, 17, 33, 65 and 129 channels wide in the spectral di- mension. For the data cube used, these ﬁlters are 10”, 30”, 50” and 90” in spatial size, and 659kHz, 1.24MHz, 2.42MHz, 4.76MHz and 9.45MHz in spectral size. In all, 20 diﬀerent ﬁlters were used with a total of 29, 348 ﬁl- ter elements between them. For each ﬁlter, a threshold equal to four standard deviations above the mean is ap- plied to the ﬁltered data. As in Whiting & Humphreys (2012), the mean and standard deviation for a voxel are calculated from a range 101 voxels wide in right ascension and declination, and a single voxel wide in frequency. Detected voxels were merged into the same object if they were within 5 voxels spatially and 80 fre- quency channels of another voxel in that object. The ﬁnal selection of voxels were those that were above the threshold for 13 or more ﬁlters, and had a spectral width of at least 10 channels.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	Adaptive Markov Chain Monte Carlo Forward Simulation for Statistical Analysis in Epidemic Modelling of Human Papillomavirus.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	We develop a Bayesian statistical model and estimation methodology based on Forward Pro- jection Adaptive Markov chain Monte Carlo in order to perform the calibration of a high- dimensional non-linear system of Ordinary Differential Equations representing an epidemic model for Human Papillomavirus types 6 and 11 (HPV-6, HPV-11). The model is compart- mental and involves stratiﬁcation by age, gender and sexual activity-group. Developing this model and a means to calibrate it efﬁciently is relevant since HPV is a very multi-typed and common sexually transmitted infection with more than 100 types currently known. The two types studied in this paper, types 6 and 11, are causing about 90% of anogenital warts.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	We extend the development of a sexual mixing matrix for the population, based on a formu- lation ﬁrst suggested by Garnett and Anderson. In particular we consider a stochastic mixing matrix framework which allows us to jointly estimate unknown attributes and parameters of the mixing matrix along with the parameters involved in the calibration of the HPV epidemic model. This matrix describes the sexual interactions between members of the population un- der study and relies on several quantities which are a priori unknown. The Bayesian model developed allows one to estimate jointly the HPV-6 and HPV-11 epidemic model parameters such as the probability of transmission, HPV incubation period, duration of infection, duration of genital warts treatment, duration of immunity, the probability of seroconversion, per gen- der, age-group and sexual activity-group, as well as unknown sexual mixing matrix parameters related to assortativity.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	Finally, we explore the ability of an extension to the class of adaptive Markov chain Monte Carlo algorithms to incorporate a forward projection simulation strategy for the ordinary dif- ferential equation state trajectories. Efﬁcient exploration of the Bayesian posterior distribu- tion developed for the ODE parameters provides a challenge for any Markov chain sampling methodology, hence the interest in adaptive Markov chain methods. We conclude with simula- tion studies on synthetic and actual data from studies undertaken recently in Australia.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	The human papillomaviruses (HPV) are a family of small DNA viruses that preferentially infect differentiating epithelial cells of the skin and mucosae. More than 100 HPV genotypes have thus far been identiﬁed, classiﬁed according to their tissue tropism (mucosal or cutaneous) and oncogenic potential (high or low). About 40 HPV types are known to infect the mucosae, including those of the anogenital and oral tracts, and 13-18 of these are considered to be oncogenic (high-risk) on the basis of their association with malignancies. Low-risk HPV types are associated with benign lesions such as genital warts and low-grade intraepithelial neoplasias of the cervix [41, 55]. Sexual contact is the primary mode of transmission [12] and HPV is the most common sexually transmitted infection in the world. HPV is known to be the causal factor in the vast majority of cervical cancer cases and is also implicated in a proportion of other anogenital cancers and cancers of the head and neck. The overall burden of disease attributed to HPV, both cancers (as much as 5.2% of incident cancers worldwide) and benign lesions such as genital warts, is considerable [39].
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	Two vaccines have been developed and shown through clinical trials to be highly effective in the prevention of precancerous lesions and persistent infection due to an important subset of HPV types [44, 54]. The quadrivalent vaccine (Gardasil) protects against high-risk HPV types 16 and 18 that are associated with 70-75% of cervical cancers, and against low-risk HPV types 6 and 11 that cause more than 90% of genital warts. The bivalent vaccine (Cervarix) provides protection against HPV types 16 and 18 only. Both vaccines have been licensed in more than 100 countries and publicly funded national immunisation programmes have commenced in some of these including Australia (Gardasil) [19].
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	National immunisation programmes are costly and decisions regarding their implementation are generally made on the basis of health-economic evaluations. In regard to HPV, these decisions are complicated by two related factors: 1) HPV is a sexually transmitted infection; and 2) only a small proportion of infections do not resolve and can lead to cancer many years or decades subsequent to acquisition. Both of these factors have generally been addressed by employing models to estimate the long-term impact of vaccination on the incidence of HPV-related disease so that the costs and beneﬁts can be calculated. However, it is the former of these factors that is of particular relevance in the context of this study. Because HPV is an infectious disease, the rate of transmission in a population, commonly referred to as the ”force of infection”, is a function of the prevalence of the infection in the population at any given time [1]. Furthermore, the beneﬁt of vaccination that confers immunity to infection (immunisation) is not conﬁned only to those directly immunised— ”unvaccinated individuals” (who remain susceptible to infection) enjoy a degree of indirect protec- tion because their risk of exposure is reduced through a diminishment in circulating virus. This indirect beneﬁt of vaccination is referred to as ”herd immunity” [20]. In order to model the impact of vaccination on the course of the HPV epidemic over time, in a manner that captures the herd immunity effect, we must use dynamic transmission models [9, 16]. Failure to do so can result in an underestimation of the potential beneﬁt of vaccination.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	models have traditionally been formulated in a deterministic framework as systems of differential equations (ordinary or partial) [1, 35, 59]. With the increasing power of personal desktop computers and access to high performance computing facilities, the use of agent-based stochastic modelling approaches has become more prominent (examples for HPV include [11, 57]). The latter approach is particularly useful for low prevalence infections where there is a possibility of extinction and/or where it is necessary to capture events that occur at the level of the individual (e.g., tracing and treating sexual partners of infected individuals). However, for their computational efﬁciency, ana- lytical tractability and ability to provide mechanistic insights to epidemic dynamics, deterministic ordinary differential equation (ODE) models are often preferred, particularly for endemic infec- tions such as HPV [47]. We have previously developed a deterministic single-type transmission model for HPV-16 [45] and more recently multi-type models for HPV types 6, 11, 16 and 18 in order to evaluate the potential impact of vaccination. In this paper we discuss a novel formulation of a model for HPV-6 and -11 that incorporates Australian data on genital warts incidence and type-speciﬁc seroprevalence.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	In this study, we develop a Bayesian statistical model and estimation methodology to perform the calibration of a high-dimensional non-linear system of ordinary differential equations repre- senting an epidemic model for HPV types 6 and 11. While the health and economic consequences of HPV-16 and -18 are more serious than for types 6 and 11 (hence their inclusion in both currently available vaccines), a model for types 6 and 11 was chosen for this study because of the availabil- ity of Australian genital warts incidence [43] and seroprevalence [42] data that could be used for calibration.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	Despite extensive study, there remains considerable uncertainty regarding aspects of the natural history of HPV and the patterns of sexual behaviour that underpin transmission [46, 58]. Further- more, many studies have not been designed with transmission models in mind, so the processes and phenomena they measure cannot always be applied directly and/or their interpretation in the context of transmission is not clear. Areas of uncertainty that present particular challenges for mod- elling include interpretation of vaccine efﬁcacy from clinical trials in the context of transmission, the duration of infectiousness (as opposed to the duration of detectability using currently available tests), the duration and nature of naturally acquired immunity, the relationship between seropositiv- ity and immunity, and the probability of transmission on sexual contact. Some of these are difﬁcult to measure at a population level for practical and/or ethical reasons. We demonstrate a Bayesian statistical methodology that will address these uncertainties in the estimation and calibration of the ODE epidemic model we have developed. This methodology allows us to statistically quantify the extent of the uncertainty in model outcomes that are derived from uncertainty in the inputs, and the contribution of uncertainty in individual parameters to the uncertainty in the outcomes [51, 31]. It can also be used to predict the impact of vaccination on a population.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	Standard MCMC algorithms that do not incorporate adaptation often require a degree of ”tun- ing” of the parameters controlling the algorithms performance. This is typically performed by off-line simulations to assess performance of the mixing of the resulting Markov chain followed by numerical investigation of the convergence rates to stationarity of the chain for different algorithmic settings of the proposal distribution. For example, the widely used variant of the Metropolis Hast- ings algorithm, the Random Walk Metropolis algorithm has mixing performance that is controlled through speciﬁcation of the Markov chain proposal distributions covariance matrix. Tuning this matrix for optimal performance can be computationally expensive and inefﬁcient. Optimal perfor- mance of an MCMC algorithm is typically either speciﬁed by the convergence rate of the Markov chain to stationarity or through the related quantity, the acceptance probability of the rejection step in the MCMC algorithm. In this regard, theoretically optimal results have been derived for several classes of statistical models, which now act as guides for more complicated sampling problems (see discussions in [50]).
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	In this paper, the ODE HPV epidemic model is constructed on a high dimensional space both in the parameters of the model and also in the latent ODE state trajectories solved for at each dis- crete time point in the ”forward projection” ODE solver. This high dimensionality in the posterior parameter space provides a signiﬁcant challenge for standard MCMC algorithms with respect to the design of an efﬁcient proposal mechanism for the Markov chain. In particular, in the model considered in this paper the fact that we incorporate a forward projection stage for the ODE solver adds additional complications in the design of the proposal. Therefore, it is desirable to automate this proposal construction for the MCMC sampler, avoiding computationally expensive tuning pro- cesses. Hence, we develop an adaptive version of the Random Walk Metropolis algorithm, coupled with forward projection. The incorporation of an adaptive proposal mechanism in an Markov chain Monte Carlo algorithm has been demonstrated to improve the performance of the sampling algo- rithm relative to standard MCMC approaches, see reviews of several examples of this improvement in ??. This improvement is achieved by learning on-line the structure of the Markov chain proposal distribution in an automated fashion, avoiding tuning of the MCMC proposal mechanism.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	In this paper we formulated a model for the HPV types 6 and 11 which adequately covers all aspects of the disease transmission and treats the incorporated seropositivity as a state associated with an individual’s recovery. Our aim was to ensure the model structure is minimalistic yet sufﬁcient for testing whether this particular interpretation of seropositivity agrees with the available data the model is calibrated to.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	Using this deterministic ODE model we undertake three key tasks: construction of a robust statistical modelling framework under a Bayesian paradigm to perform calibration of this coupled ODE transmission model with extensions to a stochastic population mixing matrix formulation; de- velopment of an automated statistical estimation methodology, based on a modiﬁcation to adaptive MCMC to incorporate forward projection for the ODE, that will provide a robust means of perform- ing calibration and statistical analysis of the calibration performance; and a statistical methodology to study vaccination responses based on the posterior predictive distribution we estimated via adap- tive MCMC.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	We perform detailed studies on synthetic data to assess the properties of the model and adaptive MCMC forward projection methodology. This is followed by assessment of the calibration perfor- mance on real data collected from Australian sources (genital warts incidence [43], and HPV-6 and -11 type-speciﬁc seroprevalence [42]).
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	We consider a dynamic transmission model for HPV types 6 and 11 (Figure 1). This model is compartmental, such that the entire population is viewed as being distributed between a set of non-overlapping groups (”compartments”) representing the stages of disease progression. The compartments are speciﬁed as presented in Table 1. The model is intended to describe how the number of people in each compartment changes over time. For example, members of the sus- ceptible population ’move’ from S to I as they become infected, and members of the recovered seropositive population ’move’ from P to S as they lose their immunity.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	model the recovered/seropositive state corresponds to those people who have developed de- tectable antibodies (using currently available tests) to to HPV-6/-11 through an immunogenic response. Conversely, the recovered/seronegative state corresponds to those who have recov- ered and are immune to reinfection but have not developed detectable antibodies. In general, seropositivity can serve as a long-lasting marker of ongoing or prior infection, though not a particularly reliable one in the case of HPV as only a proportion of those exposed to in- fection develop detectable antibodies [42]. Furthermore, there is some evidence supporting an alternative notion of seropositiviy to the one we have assumed, whereby seropositivity in only a marker of previous infection and is not correlated with protection against reinfection (see [56]). Such a perspective would lead to a more complicated model structure and we therefore do not focus on it in this paper. However, the methodology we develop can easily be extended to this context.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	This term is obtained by dividing the total number of individuals leaving the oldest age-group (group 9) each year on reaching age 60, (Sg,s,9 + Ig,s,9 + Gg,s,9 + Pg,s,9 + Ng,s,9)/5, evenly between 2 genders and 4 sexual activity-groups. To every g and s is added Sg,s,1 according to the previously deﬁned distribution of the population across risk-groups (see Table 9). The implementation of aging is a mechanism for people to enter and leave the sexually active population continuously and is necessary to propagate the effect of vaccination: we must ensure that vaccination of individuals in a particular age-group will later contribute to the number of vaccinated in older age-groups.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	Each of the equations (2)-(6) describes the change in the number of individuals that occurs during a small time period as the sum of the number of individuals entering this compartment from other compartments and those leaving the compartment. Discussions on construction of com- partmental disease transmission models are presented in [35] and [59]. Consider, for example, compartment G (Figure 1): we can calculate the change in the number of individuals in this com- partment during a small interval of time by adding up the individuals entering the compartment during this time interval (the infected who developed genital warts, γgIg,s,a and the aging mem- bers of G from age-group a − 1, Gg,s,a/5 ) and subtracting the number that leave the compartment (recovered who go either to P or N, rgGg,s,a and the aging members of G moving to age-group a + 1, Gg,s,a/5). In so doing, we will obtain γgIg,s,a + Gg,s,(a−1)/5 − rgGg,s,a − Gg,s,a/5 which is the right-hand side of equation (4).
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	It is necessary to point out the crucial role of the force of infection in our model (Table 2). This is the only non-constant coefﬁcient we have to deal with, which introduces non-linearity into the system (2)-(6). It depends on the age and risk-group speciﬁc patterns of sexual behaviour represented by a matrix usually known as a ’sexual mixing’ matrix [1].
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	In the following subsection we provide a concise description of the construction of the sexual mixing matrix. For complete details we refer the reader to Appendix 1 and an associated research paper [24].
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	In this section, we discuss the model choice, the construction and extensions we made in this paper to the development of a statistical model for the sexual interaction of members of the population, as deﬁned by the sexual mixing matrix. We consider a Sexually Transmitted Infection (STI) transmis- sion model which has many features in common with other STI models. For example the models for gonorrhea in [26] or those developed for HIV in [24] and [23], which describe patterns of mix- ing between age and sexual activity-groups with respect to HIV in heterosexual communities. Like these other models, our approach relies on certain assumptions about the way individuals form their sexual partnerships. This partnership formation process is commonly referred to as ”sexual mixing”. A simpliﬁed model for sexual mixing can be described via a sexual mixing matrix as described in [1], and for which examples can be found in [24], [22], [25] and [27].
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	to the personal nature of such surveys it is understood that their results are to be taken cautiously. In our model we will assume that two of the parameters specifying the sexual mixing matrix are unknown and should be jointly estimated with the ODE model parameters.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	In our heterosexual model formulation, the mixing matrix is a (2 × 4 × 4 × 9 × 9) dimensional matrix comprised of the product terms cg,s,s′,a,a′ρg,s,s′,a,a′, where cg,s,s′,a,a′ is the mean per capita an- nual rate at which an individual of gender g from a risk or activity-group s and age-group a acquires new sexual partners of the opposite gender g′ from a risk-group s′ and age-group a′; ρg,s,s′,a,a′ is the conditional probability that an individual of gender g from sexual activity-group s and age-group a acquires a sexual partner of the opposite gender g′ from sexual activity-group s′ and age-group a′. It is clear that estimation of all of these parameters is an almost insurmountable statistical challenge, which is one of the reasons why these parameters are often taken as ﬁxed in any given calibration study of STI transmission models. There are two broad approaches one could pursue, the ﬁrst, given we are working in a Bayesian modelling framework in this paper would involve prior elicita- tion for these population parameters based on expert opinions of annual interaction rates that would be reasonably understood by medical practitioners in sexual health clinics, sexual health workers and social workers in regions in which respondents were recorded. The other alternative involves re-parameterizing aspects of this matrix, simplifying it signiﬁcantly, allowing one to account for the uncertainty associated with speciﬁcation of this matrix in an appropriate simpliﬁed stochastic model. This would involve ﬁnding suitable factors common to aspects of this matrix that could instead be taken as stochastic and estimated in the model calibration, which in turn allow one to derive each element of the sexual mixing matrix. Most importantly, the framework we develop and present for the estimation and calibration of the transmission model is general enough to be used for either of these approaches and any degree of unknown parameters in the sexual mixing matrix and any parameterisation deemed suitable for a given population study.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	where βg is the transmission probability per partnership, i.e. the probability that a susceptible person of gender g will become infected due to a partnership with an infected person of gender g′. As discussed in [24], we can easily incorporate into speciﬁcation of (7) several parameters allowing us to control to what extent the matrix should reﬂect assortative partnership formation patterns, i.e. how likely are individuals to ﬁnd sexual partners among ’similar’ individuals, thereby altering the properties of the matrix, making it more or less dense. For example, we can vary the extent to which older males prefer to have younger female sexual partners, or, possibly, a tendency of older females to choose younger male sexual partners. It is also necessary to be able to adjust for or account for the readiness of each gender to compromise with the wishes of the opposite gender. This last point becomes signiﬁcant whenever supply of individuals of one gender does not meet the demand for sexual partners from the individuals of the opposite gender.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	In this paper, we will treat the degrees of assortativity by age and sexual activity-groups, ob- served to have a noticeable effect on the model calibration, as uncertain, while the rest of the sexual mixing matrix parameters will be ﬁxed. We do not assume that all of the parameters specifying this matrix are unknown since we want to keep the total number of parameters in our model as low as possible.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	While a matrix with these features cannot encompass all the complexities of human sexual mixing, it certainly enables us to explore various relatively plausible mixing scenarios. In addition it is a manageable model formulation and this framework has been widely used by STI modellers (for examples of its use in HPV models see [7, 37, 45, 17]).
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	In constructing a statistical model for the HPV epidemic model we treat the parameters of the non-linear system of ODE equations describing the epidemic as unknown random variables. In addition, we treat some of the parameters of the sexual mixing matrix associated with assortativity as unknown random variables to be estimated jointly with the ODE epidemic model parameters. We formulate a Bayesian model for this system with the prior speciﬁcations given in Table 3. Furthermore, we treat the ODE system as purely deterministic, conditional on a set of system parameters. In other words, the trajectories of the latent states in the ODE system over time are conditionally, deterministically speciﬁed by the system of ODEs. Therefore, we do not derive a system of stochastic differential equations as has been done in the Pharmaco-Kinetic/Pharmaco- dynamics literature (see [14] or [15]). Instead our focus lies purely in the ”calibration” of this ODE epidemic model to an observed set of data based on observations which are formed by a transformation of the state of the ODE system and observed in noise. We detail the process of forward simulation we utilise to obtain the state of the non-linear ODE system at any given time point, conditional on the model parameters deﬁned in Section 3.1.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	The states of the system at time points t ∈ {1, . . . , T } are denoted by vectors Xg,a,s(1 : T ) = [Xg,a,s(1), . . . , Xg,a,s(T )] , which generically represent the dynamic evolution of the ODE sys- tem, where for a given gender, age and sexual activity-group at time t the state encompasses the following components Xg,a,s(t) = [Sg,a,s(t), Ig,a,s(t), Gg,a,s(t), Pg,a,s(t), Ng,a,s(t)]. Therefore, con- ditional on a set of parameters in the ODE system, we can iterate the ODE system forward in time using an ODE solver to obtain the states for each population group at time t. Note, the times t, for which the system is solved, will generally be of a much ﬁner granularity to those at which observa- tions are collected in a population study. This set of times should include, for the minimum subset of times for which the system is solved, the observation times.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	In our estimation procedure we considered several different solvers, and noticed that the choice of solver can have a signiﬁcant effect on the accuracy and on the efﬁciency of the computations undertaken in the statistical estimation. In particular, depending on the parametrization and devel- opment of the relationships between states or compartments of the model and the parametrization of the mixing matrix, one can obtain non-stiff or stiff systems of equations. In such settings repeated application of generic solvers will be computationally inefﬁcient.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	where t is an independent variable, x is a vector of state variables to be solved for and g(x, t) is a function of t and x.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	An important distinction between these solvers is that each of them being of explicit, implicit or hybrid type is particularly efﬁcient for a given degree of stiffness of the system. The universal solver ﬁrst estimates how stiff the system is and then employs an appropriate solver to handle it. It is pertinent to note that in the case where (8) can be solved explicitly a modiﬁcation of the fourth order Merson’s method combined with a ﬁrst order multi-stage method will be applied. Should (8) be treated implicitly, an L-stable fourth order (5, 2)-method with an option to ﬁx the Jacobian will be used.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	We chose the dodesol solver not only because of its convenient Fortran interface, for which the remainder of the model and code was developed, but also efﬁcient performance which we com- pared with that of the two standard MATLAB solvers ode45 and ode15s. The solver dodesol was especially superior to those provided by MATLAB solvers ode45 and ode15s when the stiffness of our ODE system increased. However, it should be pointed out that dodesol speed can be considerably affected by the selection of its minimal time step and relative error threshold. To obtain the simulation results discussed in this paper we used the minimal step size 1 × 10−10, the initial step size 1 × 10−6 and the relative error tolerance 1 × 10−3. Also, we did not make use of the option to pre-specify the Jacobian provided by the solver.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	We also note, that when specifying the prior distributions and their hyper-parameters (i.e. pa- rameters describing the shape of distributions assigned to the prior distribution parameters) we utilise data reported in the literature from previous studies of HPV transmission to inform some ap- propriate prior speciﬁcations (see Table 4). Importantly, these data are completely different to the real data sets that were studied in the present paper, based on different populations, different time periods and different locations, hence we have not used our observation data twice. Furthermore, where data were not informative on particular parameters, uninformative prior speciﬁcations were utilised. As is usually the case with interpreting any reported HPV data, it is appropriate to consider the reported values of observations and calibrations as nothing more than an estimation or a trial with its own particular limitations and proneness to error. This means we can safely consider values outside of the reported conﬁdence intervals to be plausible in our study. Using (4) we specify the priors for each of the HPV models (Table 5).
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	The prior parameters for seroprevalence observation errors denoted by AY had hyper-prior pa- rameters given by (kAm, θAm) = (2, 2) and the diagonal element σ of the incidence observation error covariance matrix was distributed according to (kσ, θσ) = (2, 5). Finally, the priors for the sexual mixing matrix parameters associated with assortativity had hyper-priors speciﬁed according to (αǫa, βǫa) = (0.5, 0.7) and (αǫr , βǫr) = (0.5, 0.7).
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	Having speciﬁed the prior structure of the model we present the observation model, likelihood and details of the actual data studied. In particular, we note that conditional on a particular realisa- tion of the model parameters in Table 3, the latent unobserved state trajectories of the system are deterministic, non-analytic solutions to the system of ODEs in Equations (6). These solution tra- jectories are transformed according to age, gender and risk-group to produce a set of results that can be directly observed in noise in the population under study. In the following section we will describe and detail these transformations and the associated statistical assumptions made on the observation error. In doing so we will also specify models suitable for situations of disequilibrium of the disease states within the population such as occurs after a serious outbreak, a vaccination or community education awareness program, etc., and models for situations in which the disease in the population under study can be considered to be at equilibrium.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	In this section we ﬁrst describe the observations that are utilised in the calibration of the ODE epidemic model and the sexual mixing matrix. Importantly, the real observations utilised in this paper, based on HPV-6/-11 data collected in Australia, will be treated as taken from a population at endemic equilibrium.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	Then we present the resulting likelihood model utilised, followed by derivation of the poste- rior distribution. We ﬁrst note that the general likelihood model will be presented in which ob- servations of the transformed state trajectories are known at a ﬁxed set of discrete time points t ∈ {1, 2, . . . , T }. In practice, these discrete time points may be unevenly spaced and this would not affect the estimation procedure to be developed.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	samples to 2005 Australian midyear population estimates by age. Incidence data were estimated from the Bettering the Evaluation of Care and Health (BEACH) cross sectional database. In par- ticular, the annualised new case consultation rate stratiﬁed by gender and age from April 2000 to September 2006 was analysed and normalised to the corresponding 2004 Australian population. In total, 639 consultations related to genital warts were registered and of these roughly 35% were new cases. BEACH did not capture the people who attended sexual health clinics to seek treatment for genital warts. However, it was estimated that for every person visiting a GP there was 0.298 persons visiting a sexual health clinic, so the genital warts cases managed in such clinics were ac- counted for by multiplying the incidence rates obtained from the GP data by 1.298. The data are summarised in Tables 6 and 7.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	Generally, observations can be considered in this model as arriving at irregularly spaced inter- vals and generally not at all times t ∈ {1, . . . , T }. For simplicity, we will assume that a full panel of observations for all age-groups, activity-groups and genders is available at each observation time, though this assumption can also be relaxed under the model developed in this paper. We will also assume for simplicity that since we can solve for the state of the non-linear system at any speciﬁed time point, conditional on a set of model parameters. We will always have the observation time corresponding to one of the time points t ∈ {1, . . . , T }.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	We model the observation vector as the result of a vector function of the non-linear ODE sys- tem state Xg,a,s,t at each time t, denoted by Og,a,s(t) = [Dg,a,s(t), Yg,a,s(t)] which contains the observation Dg,a,s(t) representing incidence at time t for a given gender, age-group and sexual activity-group as well as Yg,a,s(t), which represents the seroprevalence in the given category.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	which is a function of the state vector at time t and the model parameters. The multiplier 1000 appears here because we compare the simulated incidence with the real incidence data reported per 1000 individuals [43].
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	Note, we have obtained the state vectors for the system Xg,a,s(t) used in the evaluation of the likelihood model via forward simulation, conditional on the model parameters of the non-linear ODE model, ensuring that the solution points at which we evaluated the states of the system at least corresponded to the observation times t ∈ {1, . . . , T } and generally would be a ﬁner grid than these time points. This was achieved using a specialised ODE solver as described previously in Section 3.1.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	The resulting posterior distribution therefore has 16 model parameters. At each time point t, the forward simulated ODE state vectors X(t) ∈ R360, due to 2 sexes, 9 age-groups, 4 activity-groups and 5 compartments. Hence, the total dimension explored in the simulation performed in the results section is 360 × T , where we set T = 120.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	Next we demonstrate how to perform inference under this Bayesian model formulation. In particular we are interested in the Bayesian point estimators corresponding to the posterior mean (MMSE) and the posterior mode (MAP) estimates, as well as the distribution properties of the pos- terior. To explore these we must resort to a numerical procedure to draw samples from the posterior distribution. The class of methods we consider in this paper involves combining the forward sim- ulation procedure for solving the non-linear ODE system given a set of model parameters, with an adaptive Markov chain Monte Carlo (AdMCMC) algorithm to propose a new set of model parame- ters given past history of proposed parameter vectors. The details of the AdMCMC methodologies considered are presented in Section 4.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	In this section we present details for the adaptive Markov chain Monte Carlo (AdMCMC) algo- rithm that will be combined with the Forward simulation algorithm in order to sample from the posterior distribution given in Equation (15). In particular, we ﬁrst introduce the background of an AdMCMC algorithm, before detailing the adaptation strategy we will explore in this paper based on the adaptation rules developed in [6, 49].
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	In essence, we ﬁrst construct a MCMC proposal distribution to sample the static posterior pa- rameters, denoted by vector Θ = [TR, WIP, GWT, DAI, PSC, DI, Σ, AY , EP Sa, EP Sr]. Then, conditional on these model parameters proposed, we obtain the state trajectories for the ODE HPV epidemic model, given by Xg,a,s(1 : T ) = [Xg,a,s(1), . . . , Xg,a,s(T )], which are generated us- ing forward simulation. The MCMC proposal distribution is parameterised by parameter vector Φ. The idea of adaptive MCMC methods is to learn appropriate values for Φ recursively utilising the previous samples of the Markov chain that have been accepted under the MCMC accept-reject mechanism. This is achieved on-line, adapting according to the support of the posterior distribution, thereby allowing the Markov chain to discover and explore the regions of the posterior distribution that have the most mass. Through this on-line adaptive learning mechanism the Markov chain proposal distribution can signiﬁcantly improve the acceptance rate of the Markov chain, enabling efﬁcient mixing and improving the samples obtained from the posterior.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	where we evaluate this acceptance probability utilising the expressions developed in Section 3.2, Section 3.2 and Section 4.1. These steps are repeated for j ∈ {1, . . . , J}.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	To complete the speciﬁcation of the methodology utilised, we present the internal adaptation strat- egy we considered in this paper based on the adaptive Metropolis algorithm detailed in [49]. This is a variant of the approach proposed in [30] which develops a Random Walk Metropolis-Hastings (RWMH) that estimates the global covariance structure from the past samples.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	and distributed as shown in Table 5. Note that DIm and DIf are omitted because we decided to assume a life-long immunity for all individuals in the population. This simpliﬁcation can easily be removed and would not modify our estimation procedure or models developed.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	With an initial population size of 10,000 we simulated trajectories of system (2)-(6) over annual time steps for 120 years (t ∈ [0, 120]) and every 10 years we calculated the synthetically generated ”true” state and a noisy set of observations for each age-group under the speciﬁed statistical models in Equation 10. These ODE state trajectories and the observations were taken as the ”true” syn- thetic data. Next we ran 100,000 iterations of the AdMCMC Forward Projection sampler to obtain samples from the posterior distribution conditional on these synthetically generated observations. In Figure 2 we present the posterior estimates for the ODE epidemic static model parameters (calibration performance) under the Bayesian model constructed in Section 3. We illustrate this performance for the model parameters that describe the rates of transmission between states as well as those that quantify the statistical uncertainty we model in the sexual mixing matrix. Each of these parameters is estimated based on observations generated for the number of new diag- noses (incidence) and seroprevalence utilising prior speciﬁcations derived from previous literature studies. The important features of this synthetic study is that we can illustrate that our Adaptive MCMC Forward Projection sampling methodology is able to generate samples from the resulting high dimensional posterior efﬁciently and that the estimated posterior parameter MMSE values and 95% posterior C.I. contain, in all cases, the ”true” model parameters utilised to generate the data. These results demonstrate that the AdMCMC Forward project estimation procedure we developed is working accurately in this controlled synthetic study. This is also conﬁrmed by the estimation of state trajectories. An example of estimated state trajectories for the total number of males in every disease state is shown in Figure 3. For all simulated time steps and for each disease state we observe that the ”true” trajectory is contained within the 95% posterior conﬁdence interval for the aggregated trajectory and, in addition, the estimated mean trajectory is in agreement with the ”true” aggregated trajectory.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	Figure 2: ”True” model parameters (gray circles) for synthetic study versus estimated model pa- rameters (light gray circles) and error bars (95% posterior C.I.) for the model corresponding to the combined HPV-6 and -11 case.
Processed_Adaptive_Markov_Chain_Monte_Carlo_Forward_Simulati.txt	In this section we undertook statistical estimation and performed calibration of the ODE model to the real data observations for HPV-6, HPV-11 and HPV-6 and -11 combined. We maintained the assumption that the duration of immunity is life-long for all individuals, since we found that as immunity was chosen closer to permanent the calibration became consistently better. This was measured with respect to the predictive performance of the posterior MMSE and associated pos- terior 95% C.I. obtained from the model for each age-group and gender for seroprevalence and number of new diagnoses (incidence).

Subsets and Splits

No saved queries yet

Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.