It doesn't appear that subset is implemented yet:

> normalize(bgd, "loess", subset=spikes$fid)
Normalizing... ^C
Warning message:
In .local(object, ...) :
  Subset not implemented (yet). Returning everything.

So my thought on getting around this would be to extract out the exprs from the background corrected data, normalize with the affy library, reconstruct an ExpressionFeatureSet object with the normalized data, then perform the summarization using oligo's rma function with background and normalize flags set to false. Is this reasonable? I'm not sure, because I don't know off the top of my head if the original feature set object can be modified in place so as to swap in the normalized data, or if I have to break the object down into it's respective features (e.g., phenotype, annotation, and assay data), then generate a new object with the supplied information. Any suggestions?

edit:

I checked the loessNorm function you are referring to, and you are correct that subset is being passed to the functions and being incorporated into the normalization.

## loess normalization helper
loessNormV <- function(v1, v2, subset, span, degree, weights, family){
    x <- (v1+v2)/2
    index <- c(which.min(x), subset, which.max(x))
    xx <- x[index]
    yy <- v1[index]-v2[index]
    aux <- loess(yy ~ xx, span=span, degree=degree,
                 weights=weights, family=family)
    predict(aux, data.frame(xx=x))
}

## loess normalization
## matrices
loessNorm <- function (mat, subset, epsilon=0.01, maxit=1, log.it=TRUE,
                       verbose=TRUE, span=2/3, family.loess="symmetric"){
    II <- nrow(mat)
    J <- ncol(mat)
    if (missing(subset))
        subset <- sample(II, min(c(5000, II)))
    if log.it)
        mat <- log2(mat)

    change <- epsilon + 1
    iter <- 0
    w <- c(0, rep(1, length(subset)), 0)
    while (iter < maxit) {
        iter <- iter + 1
        means <- matrix(0, II, J)
        for (j in 1:(J - 1)) {
            for (k in (j + 1):J) {
                aux <- loessNormV(v1=mat[,j], v2=mat[,k], subset=subset,
                                  span=span, degree=1, weights=w,
                                  family=family.loess)/J
                means[, j] <- means[, j] + aux
                means[, k] <- means[, k] - aux
                if (verbose)
                    message("Done with ", j, " vs ", k, " in iteration ", iter)
            }
        }
        mat <- mat - means
        change <- max(colMeans((means[subset, ])^2))
        if (verbose)
            message(sprintf("Iteration: %02d - Change: %02.8f", iter, change))
    }
    if ((change > epsilon) & (maxit > 1))
        warning("No convergence after ", maxit, " iterations.")
    if log.it) {
        return(2^mat)
    }else{
        return(mat)
    }
}

So the error I'm getting is just caused by a flag in some parent functions, and the normalization is actually performed accordingly where the loess curve is only fit to specified spike in probes. Great.

Appreciate all of the help.

Marty