E-Science, the GRID and Statistical Modelling in Social Research Rob Crouchley Collaboratory for Quantitative e-Social Science University of Lancaster - PowerPoint PPT Presentation


PPT – E-Science, the GRID and Statistical Modelling in Social Research Rob Crouchley Collaboratory for Quantitative e-Social Science University of Lancaster PowerPoint presentation | free to view - id: 463168-MGRhZ


The Adobe Flash plugin is needed to view this content

Get the plugin now

View by Category
About This Presentation

E-Science, the GRID and Statistical Modelling in Social Research Rob Crouchley Collaboratory for Quantitative e-Social Science University of Lancaster


E-Science, the GRID and Statistical Modelling in Social Research Rob Crouchley Collaboratory for Quantitative e-Social Science University of Lancaster – PowerPoint PPT presentation

Number of Views:209
Avg rating:3.0/5.0


Write a Comment
User Comments (0)
Transcript and Presenter's Notes

Title: E-Science, the GRID and Statistical Modelling in Social Research Rob Crouchley Collaboratory for Quantitative e-Social Science University of Lancaster

E-Science, the GRID and Statistical Modelling in
Social Research Rob Crouchley Collaboratory
for Quantitative e-Social Science University of
  • The Problem/Motivation Some Background on
    Statistical Methods and Social Research
  • A Solution to part of the Problem? GRID Enabling
    the Analysis of Multiprocess Random Effect
    Response Data
  • Questions.

Part 1. Some Background on Statistical Methods
and Social Research
  • Some Features of Social Science Research
  • Complications
  • A computationally demanding example
  • Sabre and Stata/MP

Some Features of Quantitative Social Science
  • We often want to develop evidence based
    substantive theory. We want to know what
    determines what, e.g. long term unemployment and
    social exclusion
  • And we want to explore the consequences of policy
    changes on individual behaviour, e.g.
    encouragement to stay on at school on educational
    attainment, truancy, and social exclusion
  • Our data sets are often very small (lt10GB)

Some of the Complexities of non experimental data
  • Cluster effects, random and fixed effects
  • Contextual effects
  • Measurement Error
  • Missing data, dropout and selection
  • Parametric Assumptions
  • Endogenous Effects

Some of the Consequent Issues
  • Disentangling the contributions created by the
    different complexities for our results is
    computationally intensive
  • Results really change as our model becomes more
    comprehensive e.g. direct effects change sign,
    other become NS
  • Problems of Large Scale Fixed Effects Analysis,
    sparse matrices
  • To tackle these complexities we could use GRID
    enabled tools, resources and services.

Social Science Research
  • Randomised experiments offer the most powerful
    tool to understand social processes, but outside
    of psychology, they are infeasible, unethical or
    inappropriate (e.g. for instance we can not
    allocate pupils to different levels of
  • Social scientists must therefore rely on
    observational data from longitudinal and other
    surveys e.g. YCS, NCDS, BHPS, The analysis of
    non experimental data involves complications..

Complication 1. Cluster Effects (CE)
  • Most large scale surveys use multi-stage sample
    designs to obtain 'representative' samples this
    procedure often creates cluster effects, e.g.
    BHPS (households), YCS (schools)
  • Pupils in the same class are often more
    behaviourally alike than pupils in different
    classes (even in the same school)

Complication 1. Cluster Effects (CE)
  • Procedures have been developed to model cluster
    effects by means of shared random effects -
    MLwiN, Stata (Gllamm), SAS, AML
  • The estimation of non-identity link (and non
    nested CE) models, e.g. probit, can be
    computationally demanding

Complication 2. Measurement Errors (ME)
  • In observational studies, it is rarely possible
    to measure all relevant covariates accurately,
    e.g. age, educational attainment
  • Ignoring ME can seriously mislead the
    quantification of the link between explanatory
    and response variables
  • ME in one covariate can bias the association
    between other covariates and the response
    variable, even if those other covariates are
    measured without error

Complication 2. Measurement Errors (ME)
  • Also, some important determinants of behaviour
    are either not measured (i.e. omitted) or are
    unmeasurable (e.g. motivation)
  • Repeated measures and longitudinal data provide
    the opportunity to deal with ME in explanatory
    variables, this adds to the computational demands
    of the analysis.

Complication 3. Missing Data, Dropout and
  • All of the major longitudinal data sets available
    to the British social science community, (e.g.
    YCS, BHPS and NCDS), contain missing data and
  • Ignoring this could create bias in the model
    estimated on the data
  • We need to model, as realistically as possible,
    the process by which the observed subjects have
    been retained in the sample, otherwise we will
    not know how much bias is present in our results
  • Also, some sample designs create selection
    effects of their own, e.g. by using a subset of
    locations, or oversampling the poor
  • These add to the computational demands of the

Complication 4. Parametric Assumptions
  • Our statistical tools are assumption rich
  • Parametric linear predictors,
  • Parametric link functions and error structures
  • What if the assumed parametric relationships do
    not hold?
  • BUT - Nonparametric statistical models are
    computationally intensive.

Complication 5. Endogenous effects
  • The curse of endogenous effects, everything seems
    to depend on everything else
  • We need multiprocess models (simultaneous
    equations) to disentangle this complexity, adds
    to computation

Disentangling complexity with existing tools an
  • This is the kind of example that got me
    interested in e-Science.

Disentangling complexity with existing tools an
example endogenous effects
  • The YCS is a multi-stage stratified clustered
    random sample of individuals ages 16-17
  • I use YCS6 which covers young people eligible to
    leave school in 1990-91, who are then observed
    over the 1992-94 period.

Part-time work and truancy are potential
determinants of educational attainment
  • A comprehensive model will allow us to
    disentangle the observable, direct, effects of
    truancy on educational attainment from any
    effects that arise from correlation in the errors
    (unobserved effects).

Educational Attainment
Level of truancy
Part Time Work
Trivariate Ordered Probit Model (Path Diagram)
Independent Errors (ep, et, eq)
Part-time work
Educational Attainment
Independent Errors (ep, et, eq)
  • This model is quick (1-2 seconds) to estimate, 3
    linear predictors
  • - Probit for PT work,
  • - Ordered Probits for Truancy and
  • Qualifications
  • We can use standard software, e.g. Stata.

Correlated Errors
Correlation Structure
Problems and Model Extensions
  • Cant use standard software to fit the model via
  • I used NAG software library, it has special
    routines to evaluate high dimensional
    multivariate normal integrals
  • Even so, this Model can take 2-3 weeks to
    estimate on a P4, 3 linear predictors, 169
    parameters, 8,496 trivariate integrals for each
    function evaluation
  • Results from this model are quite different to
    those estimated under independence e.g. one
    direct effect changes sign, another becomes NS

What is happening?
  • Evaluating lots of 3 dimensional integrals in
    order to compute our likelihood functions is
    computationally demanding
  • We could
  • Try other methods for evaluating integrals such
  • as Gibbs sampling and MCMC,
  • Use approximations
  • Laplace expansions with many terms
  • Pseudo and Quasi Likelihood Methods
  • Estimate fixed effects versions of the models
  • Use Instruments for the endogenous covariates
  • All can be computationally demanding, and each
    approach has its own problems

If we want to go this way, what can we do?
  • Use parallel algorithms on the Grid
  • Use faster Hardware, e.g. HPCx, (also part of the
  • Both

In the education example Ive assumed
  • Particular directions for the direct effects
  • No Non Ignorable dropout in the YCS
  • No School Cluster effects present
  • MVN Error structure
  • Linear predictor, additive function
  • No measurement error in observed covariates
  • We do not yet have the computational power (on
    the GRID) to relax all the assumptions
    simultaneously in this model.

The Grid some Definitions
  • "is the Web on steroids."
  • "is distributed computing across multiple
    administrative domains"
  • Dave Snelling, senior architect of UNICORE
  • provides flexible, secure, coordinated
    resource sharing among dynamic collections of
    individuals, institutions, and resource
  • From The Anatomy of the Grid Enabling Scalable
    Virtual Organizations
  • "enables communities (virtual organizations)
    to share geographically distributed resources as
    they pursue common goals.."

SABRE Software for the Analysis of Binary
Recurrent Events
  • What is it ?
  • Programme for analyising multivariate binary,
    ordinal, count and recurrent events data. Employs
    fast numerical algorithms. Uses Gaussian
    Quadrature and NPMLE for the REs
  • Some typical application areas.
  • Infertility in humans, animal husbandry.
  • Voting, trade union membership, economic activity
    and migration.
  • Absenteeism studies.

(No Transcript)
SABRE Why use it ?
gt6 months
gt1 week
Data is administrative records covering the
duration in employment in the workforce of a
major Australian state government to investigate
the determinants of quits and separations amongst
permanent and temporary workers. NP base line
hazard, quadrature for the REs
An Alternative Stata/MP
What about SABRE and Stata/MP
  • Stata/MP is 1.7 times faster on 2 processors
  • Stata/MP is 2.8 times faster on 4 processors
  • Stata/MP is 4 times faster on 8 processors
  • Sabre can have a bit faster speedup, but the big
    difference is probably the base from which
    Stata/MP starts.
  • Using the previous example on our HPC we could
    have (in minutes)

An empirical analysis of vacancy duration using
micro data from Lancashire Careers Service over
the period 19851992, NP base line hazard,
quadrature for the REs
What have I said so far?
  • That the estimation (via maximum likelihood) of
    some statistical models can be very
    computationally demanding and beyond what you can
    usefully do on your desktop.

Ways of running Sabre on the GRID
  • Directly via the operating system, e.g. Globus
  • Via a Portal, e.g. Science Gateway
  • Via a desktop application, like the tip of an
    iceberg (Im going to concentrate on this for
    the rest of the talk)

Using the Grid Via a Desktop Application
  • Separation of Client and Server Logic
  • Why ?
  • Implementation of Service Logic may change to
    allow for improved algorithms, models or
    scheduling policies and so on
  • However, user interface stays the same!!

Using the Grid Via a Desktop Application
  • Take as an example SABRE
  • Using GROWL Grid Resources on a Workstation
  • 3 Integration of SABRE functionality into
    Statistics Software (R and Stata)

Solution - How
  • Host Sabre as Secure Web Service
  • Service needs to be secure
  • Service needs to be persistent
  • Many services provided via a single host on a
    single port
  • Multiple clients
  • Difficult to do !!
  • Above features easy to host by employing generic
    GROWL server allows the developer to
    concentrate just the service logic (algorithms,
    scheduling etc)

Web services
  • A software system designed to support
    interoperable machine-to-machine interaction over
    a network.
  • It has an interface that is described in a
    machine-processable format such as WSDL.
  • Other systems interact with the Web service in a
    manner prescribed by its interface using
    messages, which may be enclosed in a SOAP
    envelope, or follow a RESTful approach.
  • These messages are typically conveyed using HTTP,
    and normally comprise XML in conjunction with
    other Web-related standards.
  • Software applications written in various
    programming languages and running on various
    platforms can use web services to exchange data
    over computer networks like the Internet in a
    manner similar to inter-process communication on
    a single computer.
  • This interoperability (for example, between Java
    and Python, or Microsoft Windows and Linux
    applications) is due to the use of open
  • OASIS and the W3C are the primary committees
    responsible for the architecture and
    standardization of web services.

First Tier
Second Tier
GROWL Server
Third Tier
Example Using Sabre on a GRID from Stata
  • User gets a Stata plugin (unzip it in the users
    ado directory)
  • This adds some items to the Stata menus
  • And provides a series of dialogue boxes

(No Transcript)
(No Transcript)
  • Could contain lots of other software, e.g. MCMC
    software on the Grid
  • Could use lots of different systems, NGS, NWG, etc

(No Transcript)
(No Transcript)
Authentication required for a Fit
SABRE Availability and Support
  • Web Site http//sabre.lancs.ac.uk
  • Full Command Documentation
  • Tutorials
  • Example Data
  • Publications
  • Downloads
  • SabreR binary R packages including
    documentation (end 06/2006)
  • SabreStata Stata plugin including
    documentation (end 07/2006)
  • Sabre source code

What have I said in part 2
  • .
  • There are beginning to be some tools that can
    make a lot more resources (Grid) available to you
    from within desktop applications.

e-science. lancs.ac.uk/cqess/
Grid Resources on Work Stations
GROWL employs a client/server architecture that
hides the complexity of GRID middleware from the
user. Client access to GROWL employs a secure
(PKI/SSL) connection to a single port on the host
system and clients are authenticated using the
distinguished name extracted from their
certificate. The use of a persistent server to
access grid resources allows all of the service
logic to be hosted by the server, making the
client application, library or plugin extremely
Future developments
  • Course material for the use of Sabre is currently
    being developed.
  • It is planned to launch a Sabre/GROWL service on
    the North West Grid within the coming year. This
    will provide a utility based grid resource.
  • Research into labour markets using Sabre/Growl.
  • SABRE will become available as a plug in for

Further information http//www. sabre.lancs.ac.uk
Middleware for e-Social Science
Development of a parallel, multilevel,
multi-process (OGSA) implementation of SABRE as
an R object to enable the Social Scientists to
disentangle the full stochastic complexity of
socio-economic processes.
SABRE development
GROWL provides a client-side lightweight library
as a plug in to R, providing easy user friendly
access to Grid resources and computational power,
You can watch a more detailed presentation about
Growl by Dan Grose at the NCeSS conference on
line at http//redress.lancs.ac.uk/Workshops/Pre
Version on my PC
  • C\2005-6 laptopfiloes\CQeSS\Oxford
  • Any Questions ?
About PowerShow.com