how many words can readers predict?

re: readers predicting and mis-predicting words, here are Stanovich and Stanovich:

Another problem concerns the assumptions that have been made about the properties of contextual information. It is often incorrectly assumed that predicting upcoming words in sentences is a relatively easy and highly accurate activity. Actually, many different empirical studies have indicated that naturalistic text is not that predictable. Alford (1980) found that for a set of moderately long expository passages of text, subjects needed an average of more than four guesses to correctly anticipate upcoming words in the passage (the method of scoring actually makes this a considerable underestimate). Across a variety of subject populations and texts, a reader’s probability of predicting the next word in a passage is usually between .20 and .35 (Aborn, Rubenstein, and Sterling, 1959; Gough, 1983; Miller and Coleman, 1967; Perfetti, Goldman and Hogaboam, 1979; Rubenstein and Aborn, 1958).

Indeed, as Gough (1983) has shown, this figure is highest for function words, and is often quite low for the very words in the passage that carry the most information content.
How research might inform the debate about early reading acquisition
Keith E. Stanovich, Ontario institute for Studies in Education and Paula J. Stanovich, University of Toronto | Journal of Research in Reading, ISSN 0141-0423
Volume 18, Issue 2, 1995, pp 87-105

Here's the abstract from the NYU study, which is quite fascinating (I need to read it again...):

Research in object recognition has tried to distinguish holistic recognition from recognition by parts. One can also guess an object from its context. Words are objects, and how we recognize them is the core question of reading research. Do fast readers rely most on letter-by-letter decoding (i.e., recognition by parts), whole word shape, or sentence context? We manipulated the text to selectively knock out each source of information while sparing the others. Surprisingly, the effects of the knockouts on reading rate reveal a triple dissociation. Each reading process always contributes the same number of words per minute, regardless of whether the other processes are operating.

and, from the text:

The question is: if parts, wholes, and context all play roles in object recognition, do the mental processes associated with them interact? Does impairing one process impair the others as well? Or, alternatively, if we remove one process, will the others continue working, unaffected? To explore this question, we turn to reading.

We want to know how people quickly and effortlessly recognize an object when there are a vast number of possibilities. Ordinary reading demonstrates this amazing human skill. In studying object recognition, reading is one of the few cases where one knows the composition: letters are parts, words are wholes, and sentences provide context. Using reading, we can attempt to isolate and measure the contributions of parts, wholes, and context to the recognition of words as objects.
Parts, Wholes, and Context in Reading: A Triple Dissociation
Denis G. Pelli*, Katharine A. Tillman
PLOS One | August 2007 | Issue 8 | e680

I'm going to knock off for tonight - will try to get some more up tomorrow.

A couple of quick notes, though:

• The study looks at decoding -- reading the words accurately out loud -- not comprehension.  
• "Context" in this study means syntax: the grammatical order of words in the sentence. The authors manipulated context by switching words around. Here's the example they give: contribute others. the of Reading measured - Trying to read this series of words, you have no "context clues" at all, i.e. no syntax to tell you 'a verb has to come next.' 

update 8/9/2012:

from the Eurkalert announcement:

Readers in the study read passages from a Mary Higgins Clark novel. The text was manipulated to selectively knock out each process in turn while retaining the others. Whole word shape was removed by alternating case: “sHe LoOkEd OvEr hEr ShOuLdEr.” To knock out the whole language process, the order of the words was shuffled. To knock out phonics, some of the letters were replaced with others.

Pelli and Tillman’s results show that letter-by-letter decoding, or phonics, is the dominant reading process, accounting for 62 percent of reading speed. However, both holistic word recognition (16 percent) and whole-language processes (22 percent) do contribute substantially to reading speed. Remarkably, the results show that the contributions of these three processes to reading speed are additive. The contribution of each process to reading speed is the same whether the other processes are working or not.

“The contributions made by phonics, holistic word recognition, and whole-language processes are not redundant,” explained Pelli. “These three processes are not working on the same words and, in fact, make contributions to reading speed exclusive of one another.”