\documentclass[12pt,a4paper]{article}
\usepackage[pdftex,dvipsnames]{color}
\usepackage{graphicx,times}
\usepackage{amsmath}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{algorithm}
\usepackage{algorithmic}
\usepackage[pdfstartview={}]{hyperref}
\textheight=9.5in
\topmargin=-0.5in
\newcommand{\eps}{\varepsilon}
\newcommand{\abso}[1]{\;\mid#1\mid\;}
\renewcommand{\=}{\,=\,}
\newcommand{\+}{\,+\,}
% ----------------------------------------------------------------
\newcommand{\remark}[1]{\par\noindent{\color[named]{ProcessBlue}#1}\par}
\newcommand{\mcc}[2]{\multicolumn{#1}{c}{#2}}
\newcommand{\mcp}[2]{\multicolumn{#1}{c|}{#2}}
\newcommand{\nm}[1]{\textsf{\small #1}}
\newcommand{\nnm}[1]{\textsf{\small\textit{#1}}}
\newcommand{\nmm}[1]{\nnm{#1}}
\newcommand{\R}{{\sf R }}
\newcommand{\sfn}[1]{\textbf{\texttt{#1}}}
\newcommand{\Rn}{{\sf R}}
\newcommand{\rs}{{\sf RSiena}}
\newcommand{\RS}{{\sf RSiena }}
\newcommand{\SI}{{\sf Siena3 }}
\newcommand{\Sn}{{\sf Siena3}}
% no labels in list of references:
\makeatletter
\renewcommand\@biblabel{}
\makeatother

\hyphenation{Snij-ders Duijn DataSpecification dataspecification dependentvariable ModelSpecification}

% centered section headings with a period after the number;
% sans serif fonts for section and subsection headings
\renewcommand{\thesection}{\arabic{section}.}
\renewcommand{\thesubsection}{\thesection\arabic{subsection}}
\makeatletter
 \renewcommand{\section}{\@startsection{section}{1}
                {0pt}{\baselineskip}{0.5\baselineskip}
                {\centering\sffamily} }
 \renewcommand{\subsection}{\@startsection{subsection}{2}
                {0pt}{0.7\baselineskip}{0.3\baselineskip}
                {\sffamily} }
\makeatother

\newcommand{\ts}[1]{\par{\color[named]{Red}TS: #1}\par}

\renewcommand{\baselinestretch}{1.0} %% For line spacing.
\setlength{\parindent}{0pt}
\setlength{\parskip}{1ex plus1ex}
\raggedright
\begin{document}

\title{Maximum likelihood/FRAN}
\author{Ruth Ripley}
\date{}
\maketitle

\centerline{\emph{\today}}
\bigskip
\section{Introduction}

The \R function \nnm{maxlikec} (so named during development when I had other
versions which did not call C!) is the function which controls the maximum
likelihood simulations.  It is used as an argument to the function
\nnm{sienaModelCreate} and its name is stored as the element named \nm{FRAN} of
the model object.  It has three (or 3+) modes: initial, to set up the C++ data
structure, terminal (to tidy up) and ordinary (to do one complete
simulation). (There is an extra mode for using multiple processors, which does
some of the work of an initial call.) It uses C++ functions for most of its
work.

The interface to C++ and to \nnm{robmon} is more or less the same as for
\nnm{simstats0c}. The terminal call is the same as for \nnm{simstats0c} and is
done by the function \nnm{terminateFRAN}. The initial call is process by the
funciton \nnm{initializeFRAN}.

\section{Initial call to maxlikec}



 Call \nnm{initializeFRAN}, which does the data set up in C++ as for
\nnm{simstats0c} plus a call to C++ to initialise the
chains.

Extra details beyond \nnm{simstats0c} processing:
\begin{enumerate}
\item Calculate \nnm{nrunMH} as the multiplication factor (stored on \nnm{x} or
  \nnm{z}) times the sum over dependent variables of the distances by
  wave. Store on \nnm{z}.
\item Store in C++ the parameters needed for maximum likelihood:
\begin{enumerate}
\item Values controlling length of permutation (on the input model)
\begin{description}
\item[maximumPermutationLength]
\item[minimumPermutationLength]
\item[initialPermutationLength]
\end{description}
\item Probabilities of the different MH steps (on the input model)
\begin{description}
\item[insertDiagonalProbability]
\item[cancelDiagonalProbability]
\item[permuteProbability]
\item[insertPermuteProbability]
\item[deletePermuteProbability]
\item[insertRandomMissingProbability]
\item[deleteRandomMissingProbability]
\end{description}
\item Proportion of missing data for each period: calculate from the number of
  missing and non missing entries in each of the dependent variables, which are
  stored as attributes of the group (which is either input or created for
  convenience in \nnm{initializeFRAN}).
\begin{description}
\item[prmin]
\item[prmib]
\end{description}
\item Simple rates flag: TRUE unless any selected effects have type "rate" but
  are not basic rate effects.
\end{enumerate}
\item
\begin{algorithmic}
\IF{this is the first call, (i.e.\ to main process)}
\STATE Set up a minimal chain: see section \ref{sec:minch}.
\STATE Do a pre burnin for each chain: see section \ref{sec:preburnin}
\STATE Do 500 normal MHsteps.
\STATE Return minimal and post burnin chains as lists.
\STATE Store the final chains on the model in C++, along with any intial and
end state differences.
\ELSE
\STATE Copy the post burnin chains from the main process. Add them to
 FRANstore (for use in R, for debugging only?) and store on the model in C++.
\ENDIF
\end{algorithmic}
\end{enumerate}

\subsection{Minimal Chain}
\label{sec:minch}
  \begin{algorithmic}
\FORALL{periods}
\FORALL{variables}
\FORALL{actors}
\IF{network variable}
\FORALL{alters present in out ties for one only of this period or the next}
\STATE create a ministep
\ENDFOR
\ELSE[behavior]
\STATE If the values at either end are different, create enough plus or minus 1
steps to get from one to the other.
\ENDIF
\ENDFOR
\ENDFOR
\STATE In each case ignore structural links or values. We do not need them in
the processing, although they should be added in before returning a chain to
\Rn. TODO: this is not currently done.
\IF {no constraints between the networks}
\STATE Create a chain from the ministeps in random order
\ELSE
\REPEAT
\FORALL{ministeps}
\STATE Try to insert in a random position in the chain
\STATE If this fails, try to insert after any other for the same
actor/alter combination in any network.
\STATE Finally try to insert before any other for the same actor and alter.
\ENDFOR
\UNTIL{done as many loops as dependent variables}
\IF{ any ministeps left}
\STATE stop with error
\ENDIF
\ENDIF
\STATE Initialize the variables for this period
\STATE Calculate the probabilities for the chain (section \ref{sec:chprob})
\ENDFOR
  \end{algorithmic}
 \subsection{Pre-burnin}
\label{sec:preburnin}
  \begin{algorithmic}
\REPEAT
\STATE do an insert diagonal step
\UNTIL{have rejected 5 steps}
\REPEAT
\STATE do an insert permute step with permutation length set to 1
\UNTIL{have rejected 5 steps}
  \end{algorithmic}
 \section{Simulation  call}
\begin{enumerate}
\item Copy over the updated parameters to the Model object.
\item Create a simulation object and set up, from stored values on the Model
  Object:
\begin{description}
\item[simpleRates] flag
\item[prmin] Proportion of missing network data for this period
\item[prmib] Proportion missing behavior data for this period
\item[chain] for this period, terminal chain from last time.
\end{description}
\item Clear out the storage area for this chain.
\item Set up from the input parameters:
\begin{description}
\item[addChainToStore]
\item[needChangeContributions] Set to 1 if either addChainToStore or
  needChangeContributions is 1 (not sure why!)
\item[returnChains] Do we want the final chain returned
\item[returnDataFrame] Return chain, if requested, as data frame rather than
  list.
\item[deriv] do we need to do derivatives (not in phase 2)
\item[nrunMH] Number of MH steps to do
\end{description}
\item Run Initialize method of the simulation object. This sets up the initial
  missing data lists (but no longer the initial state).
\item Reinstate the
  initial state by executing the vector of ministeps defining the differences.
  from the data.
\item If using multiple processes, we need to keep the same process for each
  wave, which happens by default in snow. Just don't try to use load-balancing.
\item Set up the probabilityArray (maybe I could not find anywhere to store
  this, but it also initialises the stores for acceptance and rejection
  statistics).
\item Calculate the chain probabilities (to use the new parameters)
\item Do \nnm{nrunMH} MH steps (section \ref{sec:mh})
\item Set the flags \nnm{needScores} (always set to FALSE while doing the MH
  steps and TRUE at this point) and optionally \nnm{needDerivatives}.
\item Do a final pass of the chain calculating probabilities, scores and
  optionally derivatives.
\item Store the current chain on the model, after converting the initial state
  to a vector of ministeps defining the differences.
\item Calculate the end state differences. The latter are not currently used: if
  they were to be carried over as an initial state for the next wave the chain
  would need correcting. They are returned to \R with the chain.
\item Return:
\begin{description}
\item[scores] These are the simulated targets
\item[derivatives] Returned as a lower triangle of the matrix. (eventually!)
\item[acceptance and rejection statistics] by variable. Permutation steps are
  recorded under the first variable. \nnm{misdat} steps are separated out.
\item[chain] Optional, based on \nnm{returnChains}. Can select format based on
  returnDataFrame or not.
\end{description}
\end{enumerate}

\subsection{MH steps}
\label{sec:mh}
\begin{enumerate}
\item Choose a steptype with probability as in the array
\item Do the step
\item Store whether accepted or rejected
\end{enumerate}
\subsection{Individual Steps}
\subsubsection{Insert Diagonal}
\begin{enumerate}
\item Choose a random ministep in the chain, including the option of the final
  one.
\item Set data to state just before this ministep.
\item Choose variable and actor as for forward simulation.
\item If actor not active, quit (should never occur)
\item Calculate proposal probability
\begin{align*}
p(\tilde{v})&=\text{pr of chain after insert}\\
& = \text{pr (chain length R+1) }
 \prod_{r=1}^{R+1} \left( \text{ pr(var}_r )
\text{pr(actor}_r)\text{pr(choice}_r) \right)\\
p(v)&=\text{pr of chain before insert} \\
 & = \text{pr (chain length R) }
 \prod_{r=1}^{R} \left( \text{ pr(var}_r )
\text{pr(actor}_r)\text{pr(choice}_r) \right)\\
u(\tilde{v}|v) &= \text{pr of proposing this insertion} | \text{not
  there}\\
&=\frac{\text{pr(propose an insert) pr(this var) pr(this actor)}}{(R + 1)}\\
u(v|\tilde{v}) &= \text{pr of proposing deletion} | \text{there}=
\frac{\text{pr(propose a deletion)}}{\text{pr(this diagonal)}}\\
\frac{p(\tilde{v}) u(v|\tilde{v})}
 {p(v) u(\tilde{v}|v)}&=\frac{\text{pr (chain length R+1) }
 \prod_{r=1}^{R+1} \left( \text{ pr(var}_r )
\text{pr(actor}_r)\text{pr(choice}_r)\right) \times}{\text{ (pr chain length R) }
 \prod_{r=1}^{R} \left( \text{ pr(var}_r )
\text{pr(actor}_r)\text{pr(choice}_r) \right)}\\
&\qquad \frac{\text{pr(delete diagonal) pr(this diagonal)}
\times }
{\text{pr(insert diagonal)pr(this position) pr(this var) pr(this actor)} }
\end{align*}
\item The product of probabilities all cancel out except the probability of the
  diagonal step inserted, and further the variable and actor choices cancel out,
  leaving only the choice probability of the diagonal step.
\end{enumerate}
\subsubsection{Cancel Diagonal}
\begin{enumerate}
\item If no diagonal ministeps, quit
\item Find a random diagonal ministep
\item Calculate the proposal probability: as for insert diagonal but upside down.
\item If proposal probability greater than a random uniform, do the deletion
\end{enumerate}
\subsubsection{Permute}
\begin{enumerate}
\item Find a random ministep not equal to the final dummy step.
\item Interval to permute is from here for $c0$ steps, stopping at end.
\item If interval has one or less elements, quit
\item Get a permutation and check if valid
\item If not valid, quit
\item Calculate proposal probability
\begin{align*}
 p(\tilde{v})&=\text{pr of chain after permute}\\
 & = \text{pr (chain length R }
  \prod_{r=1}^{R} \left( \text{ pr(var}_r )
 \text{pr(actor}_r)\text{pr(choice}_r) \right)\\
 p(v)&=\text{pr of chain before permute} \\
  & = \text{pr (chain length R) }
  \prod_{r=1}^{R} \left( \text{ pr(var}_r )
 \text{pr(actor}_r)\text{pr(choice}_r) \right)\\
 u(\tilde{v}|v) &= \text{pr of proposing this permutation} | \text{not
   done}\\
 &=\text{pr(propose a permute) pr(this interval)}\\
u(v|\tilde{v}) &= \text{pr of proposing reverse permutation} | \text{done}\\
 &=
 \text{pr(propose a permute)}\text{pr(this interval)}\\
\frac{p(\tilde{v}) u(v|\tilde{v})}
  {p(v) u(\tilde{v}|v)}&=\frac{ \text{pr (chain length R) }
\prod_{r=1}^{R} \left( \text{ pr(var}_r)
 \text{pr(actor}_r)\text{pr(choice}_r)\right) }
{ \text{pr (chain length R) }\prod_{r=1}^{R} \left( \text{ pr(var}_r )
 \text{pr(actor}_r)\text{pr(choice}_r) \right)}\\
% &\qquad \frac{\text{pr(delete diagonal) pr(this diagonal)}
% \times }
% {\text{pr(insert diagonal)pr(this position) pr(this var) pr(this actor)} }
\end{align*}
\item Calculations of choice probabilities only need to be done for the interval
  permuted.
\item If proposal probability greater than a random uniform, do the permutation.
\end{enumerate}
\subsubsection{Insert missing or CCP and permute}
\begin{enumerate}
\item Select a random real ministep of the chain: \nnm{miniStepA}
\item Get the chain to the state before \nnm{miniStepA}
\item Choose a variable
  using relative rates, calculated as in \nnm{simstatsc}
\item \begin{algorithmic}
\IF [logically here, as our variable and actor choices are separate]
{uponly or downonly for this variable}
\STATE Exit
\ENDIF[Why do we need to exit for missings too?]
\end{algorithmic}
\item  For the variable, choose an actor proportional to their personal lambda.
\item \begin{algorithmic}
\IF [will have rate 0]{actor not active}
\STATE Exit
\ENDIF
\end{algorithmic}
\item Choose change as in forward simulation.
\item Quit if any of the conditions in section \ref{sec:quit} hold
\item \begin{algorithmic}
\IF {chosen change corresponds to a missing value at either end of the
  period}
\STATE set miniStepB to be the dummy at the end.
\ELSE
\STATE set miniStepB to be the place for the matching CCP: chosen
randomly strictly between miniStepA and the next ministep for the same
variable/actor/(alter).
\ENDIF
\end{algorithmic}
\item Find the interval to permute: From miniStepA.next forwards for a maximum
of \nnm{c0} steps. Stop at ministep-just-before-miniStepB if get that far.
\item Shorten the interval by stopping if you find any duplicated non diagonal
  variable/actor/alter combinations.
\item Get a permutation and check whether the resulting chain is valid,
  calculating the chain probabilities as you go.
\item Quit if not valid
\item Otherwise calculate the proposal probability. \\
\sfn{ }\\
For a CCP:
\begin{align*}
p(\tilde{v})&=\text{ pr of chain after insertions}\\
& = \text{pr (chain length R+2) }
 \prod_{r=1}^{R+2} \left( \text{ pr(var}_r )
\text{pr(actor}_r)\text{pr(choice}_r) \right)\\
p(v)&=\text{ pr of chain before insertions}\\
& = \text{ pr (chain length R) }
 \prod_{r=1}^{R} \left( \text{ pr(var}_r )
\text{pr(actor}_r)\text{pr(choice}_r) \right)\\
u(\tilde{v}|v) &= \text{pr of proposing insertions} | \text{not
  there}\\
&=\text{pr(insert permute) pr(these positions) pr(this var)} \times \\
& \qquad  \text{pr(this actor) pr(this choice)}\\
 u(v|\tilde{v}) &= \text{pr of proposing deletions} | \text{there} \\
&=\text{pr(delete permute) pr(not select missing) pr(this CCP)}\\
\frac{p(\tilde{v}) u(v|\tilde{v})}
 {p(v) u(\tilde{v}|v)}&=\frac{\text{pr (chain length R+2) }
 \prod_{r=1}^{R+2} \left( \text{ pr(var}_r )
\text{pr(actor}_r)\text{pr(choice}_r)\right) \times}{\text{ (pr chain length R) }
 \prod_{r=1}^{R} \left( \text{ pr(var}_r )
\text{pr(actor}_r)\text{pr(choice}_r) \right)}\\
&\qquad \frac{\text{pr(delete permute) pr(not select missing) pr(this CCP)}
\times }
{\text{pr(insert permute) pr(these positions) pr(this var)} } \\
& \qquad \frac{1}{\text{pr(this actor) pr(this choice) }}
\end{align*}
For a missing data insertion:
\begin{align*}
p(\tilde{v})&=\text{ pr of chain after insertion}\\
& = \text{ (pr chain length R+1) }
 \prod_{r=1}^{R+1} \left( \text{ pr(var}_r )
\text{pr(actor}_r)\text{pr(choice}_r) \right)\\
p(v)&=\text{ pr of chain before insertions}\\
& = \text{ (pr chain length R) }
 \prod_{r=1}^{R} \left( \text{ pr(var}_r )
\text{pr(actor}_r)\text{pr(choice}_r) \right)\\
u(\tilde{v}|v) &= \text{pr of proposing insertion} | \text{not
  there}\\
&=\text{pr(insert permute) pr(this position) pr(this var)} \times \\
& \qquad  \text{pr(this actor) pr(this choice)}\\
 u(v|\tilde{v}) &= \text{pr of proposing deletions} | \text{there} \\
&=\text{pr(delete permute) pr(select missing) pr(this missing one)}\\
\frac{p(\tilde{v}) u(v|\tilde{v})}
 {p(v) u(\tilde{v}|v)}&=\frac{\text{ (pr chain length R+1) }
 \prod_{r=1}^{R+1} \left( \text{ pr(var}_r )
\text{pr(actor}_r)\text{pr(choice}_r)\right) \times}{\text{ (pr chain length R) }
 \prod_{r=1}^{R} \left( \text{ pr(var}_r )
\text{pr(actor}_r)\text{pr(choice}_r) \right)}\\
&\qquad \frac{\text{pr(delete permute) pr(select missing) pr(this missing)}
\times }
{\text{pr(insert permute) pr(this position) pr(this var)} } \\
& \qquad \frac{1}{\text{pr(this actor) pr(this choice) }}
\end{align*}
The probability that the chain contains $R$ events is either a Poisson with mean
the constant rate parameter:
\begin{equation}\label{eq:like}
\exp(-n\alpha(t_2-t_1))\frac{(n\alpha(t_2-t_1))^R}{R!}
\end{equation}
or approximated by a Gaussian with mean the sum of the reciprocal rates and
variance the sum of the squares of the reciprocal rates.

In each case the product terms will cancel for the part of the chain
before the first insertion and after the final insertion (if there is one).

It is necessary to know the number of CCP's which will exist after the
insertions: the only reliable way I found was to do the insertions and find
out.
\item If this probability is bigger than a random uniform, do the insertion(s).
\end{enumerate}
\subsubsection{Delete missing or CCP and permute}
\begin{enumerate}
\item Decide whether to delete a CCP or a ministep corresponding to a missing
  observation (referred to as a missing ministep in what follows). Do the latter
  with probability \nnm{prmin} + \nnm{prmib} (This does not seem sensible, I
  would suggest using the mean rather than the sum. In an extreme case you could
  have the sum greater than 1).
\item \begin{algorithmic}
\IF {deleting a missing ministep}
\STATE Decide on network or behavior using \nnm{prmin:prmib}.
\STATE Pick a random missing ministep from network or behavior.
\STATE If there are none of this type, quit.
\ELSE
\STATE Pick a random CCP
\ENDIF
\end{algorithmic}Pick a random missing ministep or CCP as appropriate. If the former, if it
  happens to be the last step in the chain, quit.
\item Set ministepA to be the ministep after the random missing or the first of
  the CCP.
\item \begin{algorithmic}
\IF {deleting a missing ministep}
\STATE set miniStepB to be the dummy at the end.
\ELSE
\STATE set miniStepB to be the ministep after the second of the CCP.
\ENDIF
\end{algorithmic}
\item Find the interval to permute: From miniStepA.next forwards for a maximum
of \nnm{c0} steps. Stop at ministep-just-before-miniStepB if get that far.
\item Shorten the interval by stopping if you find any duplicated non diagonal
  variable/actor/alter combinations.
\item Get a permutation and check whether the resulting chain is valid,
  calculating the chain probabilities as you go.
\item Quit if not valid
\item Calculate the proposal probabilities: same as for insert permute but
  upside down.
\item If proposal probability is greater than a random uniform, do the
  deletion(s).
\end{enumerate}
\subsubsection{Insert initial missing}
\begin{enumerate}
\item If no missing values quit.
\item Select a random missing option (variable, actor, plus alter if applicable).
\item
\begin{algorithmic}
\IF {behavior}
\STATE Select up or down at random, but if this goes outside range, try the
other way.
\ENDIF
\end{algorithmic}
\item Select ministepA, a random ministep not after the first occurrence of this
  option, if any. The balancing ministep will be inserted before this ministep.
\item Check that it is valid to alter the initial values of the network by a
  toggle or the reverse of the direction selected, and insert the new ministep
  before  ministepA. Note that you do not need to enforce upOnly and downOnly
  on the change in the initial state.
\item If not valid, and a behavior option, try the reverse change (unless you
  know it won't be accepted because you have already been forced to change it!)
\item If not valid, quit
\item Proposal probability for behavior direction should be set to 1 unless a
  genuine choice was accepted, when it should be 0.5.
\item Calculate the proposal probability:
\begin{align*}
p(\tilde{v})&=\text{ pr of chain after changes}\\
& = \text{ pr(new val of miss item) pr (chain length R+1) } \times\\
& \qquad  \prod_{r=1}^{R+1} \left( \text{ pr(var}_r )
\text{pr(actor}_r)\text{pr(choice}_r) \right)\\
p(v)&=\text{ pr of chain before changes}\\
& = \text{ pr(orig val of miss item)pr (chain length R) }\times \\
& \qquad \prod_{r=1}^{R} \left( \text{ pr(var}_r )
\text{pr(actor}_r)\text{pr(choice}_r) \right)\\
u(\tilde{v}|v) &= \text{pr of proposing changes} | \text{not
  there}\\
&=\text{pr(insert missing) pr(this missing option) pr(direction)
pr(ministep here)} \\
 u(v|\tilde{v}) &= \text{pr of proposing undoing changes} | \text{there} \\
&=\text{pr(delete missing) pr(this missing option)}\\
\frac{p(\tilde{v}) u(v|\tilde{v})}
 {p(v) u(\tilde{v}|v)}&=\frac{\text{ pr (chain length R+1) }
 \prod_{r=1}^{R+1} \left( \text{ pr(var}_r )
\text{pr(actor}_r)\text{pr(choice}_r)\right) \times}{\text{ pr (chain length R) }
 \prod_{r=1}^{R} \left( \text{ pr(var}_r )
\text{pr(actor}_r)\text{pr(choice}_r) \right)}\\
&\qquad \frac{\text{pr(delete missing) pr(new val)} }
{\text{pr(insert missing) pr(direction) pr(ministep here) pr(orig val)} }
\end{align*}
\item Probabilities for the different values of the missing item are set to the
  proportion of non-missing values in the variable at the start of the period
  which are equal to the value in question.
\item The product of probabilities of ministeps will cancel out after ministepA.
\item If proposal probability is greater than a random uniform, do the changes.
\end{enumerate}
\subsubsection{Delete initial missing}
\begin{enumerate}
\item If no missing values, quit
\item Select a random missing option.
\item If there are no ministeps for this option, quit
\item If a behavior option, calculate what the direction probability would have
  been at insertion.
\item Calculate how many places could have been chosen for the position at
  insertion.
\item Check that doing a reverse change to the initial value and deleting the
  ministep is valid.
\item If not valid, exit
\item Calculate proposal probability. Same as insert missing, but upside down.
\item If proposal probability is greater than a random uniform, do the changes.
\end{enumerate}
\subsubsection{quits}
\label{sec:quit}
\begin{algorithmic}
\IF {diagonal}
\STATE Exit
\ENDIF
\IF [we exclude these: need to sort this out ]
{network and structurally fixed link}
\STATE Exit
\ENDIF
\IF [will have rate 0]
{behavior and structurally fixed}
\STATE Exit
\ENDIF
\IF [will be diag or down as not permissible change]
{behavior and goes over end}
\STATE Exit
\ENDIF
\IF
{same variable/actor/(alter) as miniStepA}
\STATE Exit
\ENDIF
\end{algorithmic}
\subsubsection{Update chain probabilities}
\label{sec:chprob}
This reinitialises the data, but no longer the initial state.
\subsection{Calculation of scores and derivatives}
\label{sec:scores}
Scores are as for method of moments. For the rates, set $\tau$ equal to 1 over
the number of real events in the chain. (Not sure about structurals here!)

Using similar notation, with $v_{rk_1k_2}$ the derivative of the score
corresponding to$(\theta_{k_1}, \theta_{k_2})$ in the $r$-th chain,
\begin{align*}
\intertext{for a basic rate parameter, $(\lambda)$, for a
dependent variable, not the total here, just the per actor rate}
v_{rkk} &= \sum_{m=1}^M  \textrm{nbr of non-structurally determined links in the
  chain}/ \lambda_m^2\\
\intertext{for other parameters}
v_{rk_1k_2}&= \sum_{m=1}^M \left [  \left( \sum_{a, \nnm{delta}_a}
  s_{i,a\delta k_a} p_{i,a\delta} \right) \left (
\sum_{ b, \nnm{delta}_b} s_{i,b\delta k_b} p_{i,a\delta} \right) -
\sum_{a, \nnm{delta}}
  s_{i,a\delta k_1}p_{i,a\delta}  s_{i,a\delta k_2}
\right ]\\
\intertext{with derivative matrix}
 D_{jk}&=\frac{1}{R}\sum_{r=1}^R v_{rjk}
\end{align*}

Note that rate effects are uncorrelated with all other effects.


\end{document}
\frac{\exp(p_{rij})}{\sum_k\exp p_{rik}}