R/glm.fit3.R

In emilygoren/BinQuasi: Analyzing Replicated ChIP Sequencing Data Using Quasi-Likelihood

glm.fit3=function (x, y, weights = rep(1, nobs), start = NULL, etastart = NULL, 
                   mustart = NULL, offset = rep(0, nobs), family = gaussian(), 
                   control = list(), intercept = TRUE) 
{
  glm.control=
    function (epsilon = 1e-08, maxit = 25, trace = FALSE, ...) 
    {
      if (!is.numeric(epsilon) || epsilon <= 0) 
        stop("value of 'epsilon' must be > 0")
      if (!is.numeric(maxit) || maxit < 0)  ## allowing zero
        stop("maximum number of iterations must be >= 0")
      list(epsilon = epsilon, maxit = maxit, trace = trace)
    }
  control <- do.call("glm.control", control)
  x <- as.matrix(x)
  xnames <- dimnames(x)[[2L]]
  ynames <- if (is.matrix(y)) 
    rownames(y)
  else names(y)
  conv <- FALSE
  nobs =n<- NROW(y)
  nvars <- ncol(x)
  EMPTY <- nvars == 0
  if (is.null(weights)) 
    weights <- rep.int(1, nobs)
  if (is.null(offset)) 
    offset <- rep.int(0, nobs)
  variance <- family$variance
  linkinv <- family$linkinv
  if (!is.function(variance) || !is.function(linkinv)) 
    stop("'family' argument seems not to be a valid family object", 
         call. = FALSE)
  dev.resids <- family$dev.resids
  aic <- family$aic
  mu.eta <- family$mu.eta
  unless.null <- function(x, if.null) if (is.null(x)) 
    if.null
  else x
  valideta <- unless.null(family$valideta, function(eta) TRUE)
  validmu <- unless.null(family$validmu, function(mu) TRUE)
  if (is.null(mustart)) {
    eval(family$initialize)
  }
  else {
    mukeep <- mustart
    eval(family$initialize)
    mustart <- mukeep
  }
  if (EMPTY) {
    eta <- rep.int(0, nobs) + offset
    if (!valideta(eta)) 
      stop("invalid linear predictor values in empty model", 
           call. = FALSE)
    mu <- linkinv(eta)
    if (!validmu(mu)) 
      stop("invalid fitted means in empty model", call. = FALSE)
    dev <- sum(dev.resids(y, mu, weights))
    w <- ((weights * mu.eta(eta)^2)/variance(mu))^0.5
    residuals <- (y - mu)/mu.eta(eta)
    good <- rep(TRUE, length(residuals))
    boundary <- conv <- TRUE
    coef <- numeric()
    iter <- 0L
  }
  else {
    coefold <- NULL
    eta <- if (!is.null(etastart)) 
      etastart
    else if (!is.null(start)) 
      if (length(start) != nvars) 
        stop(gettextf("length of 'start' should equal %d and correspond to initial coefs for %s", 
                      nvars, paste(deparse(xnames), collapse = ", ")), 
             domain = NA)
    else {
      coefold <- start
      offset + as.vector(if (NCOL(x) == 1L) 
        x * start
        else x %*% start)
    }
    else family$linkfun(mustart)
    
    init.fit=lm.fit(x, eta, offset=offset)
    coefold=init.fit$coefficients
    coefold[is.na(coefold)]=0
    rk=init.fit$rank
    
    mu <- linkinv(eta)
    if (!(validmu(mu) && valideta(eta))) 
      stop("cannot find valid starting values: please specify some", 
           call. = FALSE)
    devold <- sum(dev.resids(y, mu, weights))
    boundary <- conv <- FALSE
    iter=1L
    repeat{
      good <- weights > 0
      varmu <- variance(mu)[good]
      if (any(is.na(varmu))) 
        stop("NAs in V(mu)")
      if (any(varmu == 0)) 
        stop("0s in V(mu)")
      mu.eta.val <- mu.eta(eta)
      if (any(is.na(mu.eta.val[good]))) 
        stop("NAs in d(mu)/d(eta)")
      good <- (weights > 0) & (mu.eta.val != 0)
      if (all(!good)) {
        conv <- FALSE
        warning("no observations informative at iteration ", 
                iter)
        break
      }
      z <- ((eta - offset)[good] + (y - mu)[good]/mu.eta.val[good])
      w <- (sqrt((weights[good] * mu.eta.val[good]^2)/variance(mu)[good]))
      ngoodobs <- as.integer(nobs - sum(!good))
      fit <- lm.fit(x =x[good, , drop = FALSE] * w, y = z * w , singular.ok = TRUE, tol = min(1e-07, control$epsilon/1000))
      
      if(fit$rank != rk){ ## rank should not change during iterations
        fit$rank=fit$qr$rank=rk
        fit$coefficients = qr.coef(fit$qr, z*w)
      }
      fit$coefficients[is.na(fit$coefficients)]=0
      
      if(iter > control$maxit){ ## allows zero maxit 
        conv=FALSE
        coef=coefold
        dev=devold
        break
      }
      
      if (any(!is.finite(fit$coefficients))) {
        conv <- FALSE
        warning(gettextf("non-finite coefficients at iteration %d", 
                         iter), domain = NA)
        break
      }
      if (nobs < fit$rank) 
        stop(gettextf("X matrix has rank %d, but only %d observations", 
                      fit$rank, nobs), domain = NA)
      #start[fit$qr$pivot] <- fit$coefficients
      start <- fit$coefficients
      eta <- drop(x %*% start)
      mu <- linkinv(eta <- eta + offset)
      dev <- sum(dev.resids(y, mu, weights))
      if (control$trace) 
        cat("Deviance =", dev, "Iterations -", iter, 
            "\n")
      boundary <- FALSE
      if (!is.finite(dev)) {
        if (is.null(coefold)) 
          stop("no valid set of coefficients has been found: please supply starting values", 
               call. = FALSE)
        warning("step size truncated due to divergence", 
                call. = FALSE)
        ii <- 1
        while (!is.finite(dev)) {
          if (ii > control$maxit) 
            stop("inner loop 1; cannot correct step size", 
                 call. = FALSE)
          ii <- ii + 1
          start <- (start + coefold)/2
          eta <- drop(x %*% start)
          mu <- linkinv(eta <- eta + offset)
          dev <- sum(dev.resids(y, mu, weights))
        }
        boundary <- TRUE
        if (control$trace) 
          cat("Step halved: new deviance =", dev, "\n")
      }
      if (!(valideta(eta) && validmu(mu))) {
        if (is.null(coefold)) 
          stop("no valid set of coefficients has been found: please supply starting values", 
               call. = FALSE)
        warning("step size truncated: out of bounds", 
                call. = FALSE)
        ii <- 1
        while (!(valideta(eta) && validmu(mu))) {
          if (ii > control$maxit) 
            stop("inner loop 2; cannot correct step size", 
                 call. = FALSE)
          ii <- ii + 1
          start <- (start + coefold)/2
          eta <- drop(x %*% start)
          mu <- linkinv(eta <- eta + offset)
        }
        boundary <- TRUE
        dev <- sum(dev.resids(y, mu, weights))
        if (control$trace) 
          cat("Step halved: new deviance =", dev, "\n")
      }
      if (((dev - devold)/(0.1 + abs(dev)) >= control$epsilon) ) {
        if (is.null(coefold)) 
          stop("no valid set of coefficients has been found: please supply starting values", 
               call. = FALSE)
        warning("step size truncated due to increasing deviance", 
                call. = FALSE)
        ii <- 1
        while ((dev - devold)/(0.1 + abs(dev)) > -control$epsilon) {
          if (ii > control$maxit) 
            break
          ii <- ii + 1
          start <- (start + coefold)/2
          eta <- drop(x %*% start)
          mu <- linkinv(eta <- eta + offset)
          dev <- sum(dev.resids(y, mu, weights))
        }
        if (ii > control$maxit) 
          warning("inner loop 3; cannot correct step size")
        else if (control$trace) 
          cat("Step halved: new deviance =", dev, "\n")
      }
      if (abs(dev - devold)/(0.1 + abs(dev)) < control$epsilon) {
        conv <- TRUE
        coef <- start
        break
      }
      else {
        devold <- dev
        coef <- coefold <- start
      }
      if (iter == control$maxit) break
      iter = iter + 1L
    }
    if (!conv) 
      warning("glm.fit3: algorithm did not converge. Try increasing the maximum iterations", 
              call. = FALSE)
    if (boundary) 
      warning("glm.fit3: algorithm stopped at boundary value", 
              call. = FALSE)
    eps <- 10 * .Machine$double.eps
    if (family$family == "binomial") {
      if (any(mu > 1 - eps) || any(mu < eps)) 
        warning("glm.fit3: fitted probabilities numerically 0 or 1 occurred", 
                call. = FALSE)
    }
    if (family$family == "poisson" || tolower(substr(family$family, 1L, 17L)) == "negative binomial") {
      if (any(mu < eps)) 
        warning("glm.fit3: fitted rates numerically 0 occurred", 
                call. = FALSE)
    }
    
    if (fit$rank < nvars) 
      coef[fit$qr$pivot][seq.int(fit$rank + 1, nvars)] <- NA
    xxnames <- xnames[fit$qr$pivot]
    residuals <- (y - mu)/mu.eta(eta)
    fit$qr$qr <- as.matrix(fit$qr$qr)
    nr <- min(sum(good), nvars)
    if (nr < nvars) {
      Rmat <- diag(nvars)
      Rmat[1L:nr, 1L:nvars] <- fit$qr$qr[1L:nr, 1L:nvars]
    }
    else Rmat <- fit$qr$qr[1L:nvars, 1L:nvars]
    Rmat <- as.matrix(Rmat)
    Rmat[row(Rmat) > col(Rmat)] <- 0
    names(coef) <- xnames
    colnames(fit$qr$qr) <- xxnames
    dimnames(Rmat) <- list(xxnames, xxnames)
  }
  names(residuals) <- ynames
  names(mu) <- ynames
  names(eta) <- ynames
  wt <- rep.int(0, nobs)
  wt[good] <- w^2
  names(wt) <- ynames
  names(weights) <- ynames
  names(y) <- ynames
  if (!EMPTY) 
    names(fit$effects) <- c(xxnames[seq_len(fit$rank)], rep.int("", 
                                                                sum(good) - fit$rank))
  wtdmu <- if (intercept) 
    sum(weights * y)/sum(weights)
  else linkinv(offset)
  nulldev <- sum(dev.resids(y, wtdmu, weights))
  n.ok <- nobs - sum(weights == 0)
  nulldf <- n.ok - as.integer(intercept)
  rank <- if (EMPTY) 
    0
  else fit$rank
  resdf <- n.ok - rank
  aic.model <- aic(y, n, mu, weights, dev) + 2 * rank
  list(coefficients = coef, residuals = residuals, fitted.values = mu, 
       effects = if (!EMPTY) fit$effects, R = if (!EMPTY) Rmat, 
       rank = rank, qr = if (!EMPTY) structure(fit$qr[c("qr", 
                                                        "rank", "qraux", "pivot", "tol")], class = "qr"), 
       family = family, linear.predictors = eta, deviance = dev, 
       aic = aic.model, null.deviance = nulldev, iter = iter, 
       weights = wt, prior.weights = weights, df.residual = resdf, 
       df.null = nulldf, y = y, converged = conv, boundary = boundary)
}