Approximating Survivor Regions With Rectangular Cells

Why Approximate Survivor Regions With Cells?

At an interim look, a group sequential procedure continues only for statistic values inside its continuation region. The survivor density describes the distribution of the statistic among paths that have not stopped. To propagate that density to the next look, one integrates over the current survivor region. When the boundary is curved, asymmetric, or inconvenient for direct integration, a grid of rectangular cells provides a transparent numerical approximation.

The grid does not change the statistical stopping rule. It replaces integrals over the continuous state axis with finite sums, while a cell-classification rule decides which parts of the grid are treated as surviving. The approximation is useful only when its two sources of error are understood: integration within cells and classification of cells that intersect a boundary.

Model and Rectangular-Cell Recursion

Assume the statistic has a Gaussian Markov transition between consecutive looks. Specifically, conditional on \(X_{j-1}=x\), suppose \(X_j\) is normal with mean \(m_j(x)\) and variance \(v_j>0\). The transition density is

The Markov assumption means that, given the current statistic, the next-look distribution does not additionally depend on earlier states. Assume also that the continuation regions are specified, the survivor density at the preceding look is available, and the numerical grid covers the region whose mass is being approximated. If the state axis is restricted to a finite interval, mass outside that interval must be accounted for rather than silently treated as zero.

The exact recursion multiplies the previous survivor density by the transition density, integrates over the previous continuation region, and sets the result to zero outside the new continuation region:

To approximate this integral, divide the relevant part of the previous state axis into cells \(A_i\) with widths \(h_i\) and midpoints \(x_i\). Divide the new state axis into cells \(D_\ell\) with widths \(q_\ell\) and midpoints \(y_\ell\). The midpoint rectangular approximation evaluates the integrand at the cell midpoints:

The resulting approximation to survivor mass is another rectangular sum:

Here \(f_j^S(y_\ell)\) is a density, measured as probability per unit of the statistic, whereas \(\widehat f_j^S(y_\ell)q_\ell\) is a probability mass. Confusing these quantities is a common source of errors.

How the Midpoint Approximation Arises

For fixed \(y\), write the integrand as \(g(x)=f_{j-1}^S(x)K_j(y\mid x)\). On a cell of width \(h\) centered at \(x_i\), the midpoint rectangle replaces the cell integral by \(h g(x_i)\). If \(g\) is twice continuously differentiable on the cell and \(|g''(x)|\le M\), Taylor expansion around \(x_i\) gives

The linear Taylor term integrates to zero because the cell is centered at its midpoint; the displayed bound follows by integrating the quadratic remainder bound \(M(x-x_i)^2/2\). Thus, for a smooth integrand and a fixed, correctly classified region, halving a uniform cell width typically reduces this component of the error by about a factor of four. This argument does not cover cells cut by a boundary: the indicator of the continuation region is discontinuous there, so boundary classification can dominate the smooth integration error.

Classifying Cells at a Boundary

For a general continuation region, classify each cell using a stated rule. A simple midpoint rule retains a cell when its midpoint is in the continuation region, and excludes it otherwise. For a two-sided continuation interval \(C_j=(l_j,u_j)\), this means retaining destination cell \(D_\ell\) when \(l_j<y_\ell<u_j\). Whether the interval endpoints are included makes no difference to the continuous model's probability, but the rule for cells near those endpoints still matters.

A midpoint rule can retain a cell that partly lies beyond a stopping boundary, or exclude a cell that partly lies inside. The rectangular sum then assigns the whole cell according to its midpoint. Smaller cells reduce the region affected by this approximation, but do not make a coarse-grid classification exact.

A Destination-Cell Check and Numerical Diagnostics

A useful comparison keeps the source integration as a midpoint sum but integrates the Gaussian transition over each destination cell. If \(D_\ell=[a_\ell,b_\ell]\), a source midpoint \(x_i\) contributes according to the conditional probability

where \(\Phi\) is the standard normal cumulative distribution function. This approach avoids approximating the transition density as constant across a destination cell, although source-cell and boundary-classification errors remain. It is one useful check on a fully rectangular density evaluation.

Several diagnostics help identify a failing approximation:

• Check mass accounting. Survivor mass must be nonnegative and cannot exceed the mass available before the look, apart from small rounding error. Large violations usually indicate a density-versus-mass mistake, a missing cell width, or double counting.
• Check the direction of boundary effects. A cell midpoint just inside a boundary retains the entire cell under the midpoint rule. If the implementation instead drops it, the coding rule and stated method disagree.
• Inspect density scale and units. A density has units reciprocal to the statistic; multiplying by a cell width produces a dimensionless probability mass.
• Repeat with a finer grid. Material changes in survivor mass signal that the original grid is too coarse for the boundary or transition density. Agreement across two grids is reassuring, but does not prove accuracy if both grids mis-handle the same boundary or omitted tail mass.
• Compare destination-cell probabilities. Substantial disagreement with the Gaussian cell-probability calculation suggests that density variation within destination cells is important.

Worked Example: Summing Two Destination Cells

Limitations and Common Pitfall

Rectangular cells are straightforward to audit, but their convenience does not guarantee precision. Accuracy depends on how rapidly the survivor density and transition kernel vary across cells, the location of boundaries relative to cell edges, and whether the grid represents all relevant probability mass. A narrow continuation region or a boundary near a high-density area can make boundary classification especially consequential. The approximation also inherits any error in the previous-look survivor density.

Common mistake: “The cell midpoint is inside the continuation region, so the cell is exactly a survivor cell.” The midpoint rule only assigns the whole rectangle according to one representative point; it does not establish that every point in the cell survives. The correction is to describe the resulting mass as an approximation, inspect cells that straddle boundaries, and assess sensitivity to cell width or a boundary-aware alternative.

Direct integration or simulation can serve as alternatives when cell approximations are inadequate. Cell transition probabilities can improve integration over destination cells; direct numerical integration can represent boundaries more precisely; simulation avoids a state grid but introduces Monte Carlo uncertainty. These methods have different computational costs and error sources, so the choice depends on the required accuracy and the structure of the boundary.