\tikzset{
  observed/.style={circle, draw},
  partly observed/.style 2 args={draw, fill=#2, path picture={
      \fill[#1, sharp corners] (path picture bounding box.south west) -|
      (path picture bounding box.north east) -- cycle;},
     circle},
    unobserved/.style={draw, circle, fill=gray!40},
    residual/.style={draw, rectangle}
}
\begin{figure}[htbp!]
\centering
\begin{subfigure}[t]{0.48\textwidth}
\centering
\begin{tikzpicture}

  \node[observed] (y) {$Y$};
  \node[unobserved, above=of y] (x) {$X$};
  \node[observed, left=of x] (w) {$W$};
  \node[observed,right=of x] (z) {$Z$};

  \draw[-] (z) to (y);
  \draw[-] (z) -- (x);
  \draw[-] (x) -- (y);
  \draw[-] (x) -- (w);

\end{tikzpicture}
\caption{In \emph{Simulation 1a}, classifications $W$ are conditionally independent of $Y$ so a model using $W$ as a proxy for $X$ has non-differential error. \label{fig:simulation.1a}}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.48\textwidth}
\centering
\begin{tikzpicture}

  \node[observed] (y) {$Y$};
  \node[unobserved, above=of y] (x) {$X$};
  \node[observed, left=of x] (w) {$W$};
  \node[observed,right=of x] (z) {$Z$};

  \draw[-] (z) to (y);
  \draw[-] (z) -- (x);
  \draw[-] (x) -- (y);
  \draw[-] (x) -- (w);

  \draw[-] (x) to (y);

  \draw[-] (w) -- (y);


\end{tikzpicture}
\caption{In \emph{Simulation 1b}, the edge from $W$ to $Y$ signifies that the automatic classifications $W$ are not conditionally independent of $Y$ given $X$, indicating differential error.
\label{fig:simulation.1b}
}
\end{subfigure}
\\
\hfill
\begin{subfigure}[t]{0.48\textwidth}
\centering
\begin{tikzpicture}
  \node[unobserved] (y) {$Y$};

  \node[observed, above=of y] (x) {$X$};
  \node[observed, right=of y] (w) {$W$};


  \node[observed,right=of x] (z) {$Z$};
  \draw[-] (z) to (y);
  \draw[-] (x) -- (y);
  \draw[-] (y) -- (w);
  \draw[-] (x) -- (z);

\end{tikzpicture}
\caption{In \emph{Simulation 2a}, an unbiased classifier measures the outcome. \label{fig:simulation.2a}}
\end{subfigure} \hfill
\begin{subfigure}[t]{0.48\textwidth}
\centering
\begin{tikzpicture}
  \node[unobserved] (y) {$Y$};

  \node[observed={white}{gray!40}, above=of y] (x) {$X$};
  \node[observed, right=of y] (w) {$W$};


  \node[observed,right=of x] (z) {$Z$};
  \draw[-] (x) -- (y);
  \draw[-] (z) -- (w);
  \draw[-] (y) -- (w);
  \draw[-] (x) -- (z);
  \draw[-] (z) -- (y);
\end{tikzpicture}
\caption{In \emph{Simulation 2b}, the edge connecting $W$ and $Z$ signifies that the predictions $W$ are not conditionally independent of $Z$ given $Y$, indicating systematic misclassification. \label{fig:simulation.2b}}
\end{subfigure}
\vspace{1em}
\begin{subfigure}[t]{0.2\textwidth}
\centering
\begin{tikzpicture}
  \matrix [draw, below, font=\small, align=center, column sep=2\pgflinewidth, inner sep=0.4em, outer sep=0em, nodes={align=center, anchor=center}] at (current bounding box.south){
    \node[observed,label=right:observed] {}; \\
    \node[unobserved,label=right:automatically classified]{}; \\
  };
\end{tikzpicture}
\end{subfigure}
\caption{
Bayesnet networks representing the conditional independence structure of our simulations. \label{bayesnets}
}
\end{figure}