Written composition—generating ideas and shaping them into coherent and accurate text—involves a sequence of processes, starting with content retrieval and structuring, and progressing through syntactic planning and spelling retrieval, to the motor planning and execution of keypresses or pen strokes. These processes can be thought of as forming a cascade (Olive, 2014
; Roux et al., 2013
; van Galen, 1991
): processing at a specific level starts as soon as it is provided with input from the preceding (upstream) process. The upstream process is then freed to take new input, processing this in parallel with processing at other levels. It is this “just-in-time” processing (Christiansen & Chater, 2016
) that allows competent writers (communicating non-demanding content) to write fluently with only rare hesitations. Difficulty at any level of this cascade of processes, however, forms a bottleneck that results in production disfluency: the writer’s typing (or pen movement) slows or pauses. Bottlenecks of this nature are more likely to occur when the writer lacks practice in one or more of the component mechanisms (see, for example, Ruthruff et al., 2001
). This will be the case, for example, for adults writing in a language in which they are not fluent (e.g. Chukharev-Hudilainen et al., 2019
) or—the focus of the present study—when the writer is a primary school student.
From a communicational point of view, disfluency in written production is not a problem. Unlike in speech, hesitation during written production, in most contexts, has no effect on readers’ understanding of what is written. Pauses that occur as a result of a child thinking carefully about content are, in fact, welcome. However, unexpectedly long pauses in particular at mid-sentence or within-word locations might be disruptive. Such disfluencies may result from bottlenecks at lower levels in the cascade, as children attempt to retrieve spelling and/or plan the motor actions necessary to form letters on the page or computer screen. Where these lower level transcription processes are the source of disfluency, this may have negative consequences for the quality of the resulting text. If for example a child struggles to spell a word, even if they finally retrieve the correct spelling, this processing difficulty may have consequences for more global features of the text. Fluent output is of course important in any context where the writing task is time constrained. Slower writing will, in many contexts, mean a shorter final product. But there is also the possibility that disfluency in output has knock-on effects for upstream processes. Fluent writing means planning what to say next independently of and without interference from the processing needed to spell and to inscribe1
the words that are currently being written. Christiansen and Chater (2016
) argue that there is a fundamental “now-or-never” bottleneck in human language processing. In the context of written production, this means that delays associated with spelling or with inscription may disrupt the processes responsible for generating and structuring what to say next, i.e. there seems to be important timing issues for optimal text production.
Researchers exploring writing development have frequently suggested that competition for processing resources, in some form, links word-level difficulty with an inability to construct text that is informationally rich and coherently structured (e.g. Berninger et al., 2002
; McCutchen, 1996
; Torrance & Galbraith, 2006
). This has been advanced as an explanation for, for example, the finding that the compositional quality of students’ writing improves as a result of training in handwriting (Alves et al., 2016
) and typing (van Weerdenburg et al., 2019
). The relationship between writers’ transcription skill (skill in spelling and in inscription) and the compositional quality of their text has been explored more directly in a number of studies (e.g. Abbott et al., 2010
; Alves & Limpo, 2015
; Connelly et al., 2005
; Graham et al., 1997
; Jones & Christensen, 1999
; Kim et al., 2015
). In a meta-analysis of studies with children from kindergarten to 12th grade, Kent and Wensek (2016
) found a mean standardized effect of 0.49 for the relationship between spelling-test performance and writing quality (18 effects) and also of 0.49 for the effect of handwriting fluency (17 effects). Feng et al. (2019
) in an overlapping meta-analysis (9 effects) found a mean effect of 0.41 of handwriting fluency on, specifically, substantive quality (i.e. quality of just the content and structure of the completed text, ignoring spelling and handwriting accuracy). Studies have also explored transcription effects on written productivity—the number of words of spontaneous text composed per minute or within a fixed time limit. For example, Graham et al. (1997
) found that productivity in English-speaking upper-primary school students was predicted by accuracy on a spelling-to-dictation test and speed of written alphabet recall. Alves and Limpo (2015
) found similar effects for students writing in Portuguese, a language with a much more transparent orthography (i.e. with more regular sound-to-letter correspondence). Feng et al. (2019
) who found an average effect of handwriting fluency on productivity of 0.53 (12 effects), averaging across a small number of studies, drew similar conclusions. It is worth noting that in most cases the measure attributed to handwriting fluency in all of these studies was speed of handwritten alphabet recall, a measure that includes a memory component and alphabet knowledge in addition to the ability to fluently move the hand.
One possible explanation for these effects is the “process-disruption hypothesis”: Weaker spelling and inscription skills result in production disfluency and this disrupts the writing process in ways that result in damage to text quality. With regards spelling effects, however, there are two reasons why the deficit associated with inaccuracy in spelling-to-dictation may not necessarily mean that the child’s writing processes will be disrupted when they compose full text. First, inaccuracy is not necessarily associated with disfluency. As long as a word is retrieved and produced fluently this will not disrupt upstream processes, regardless of whether it is spelt correctly. Torrance, Rønneberg, et al. (2016
) for example, found that words wrongly spelled in upper secondary students’ written compositions, after statistical control for length and frequency, were produced more, rather than less fluently than words that were spelled correctly. Second, accurately spelled words are not necessarily produced fluently. Fluent spelling is much more likely to be achieved when orthographic retrieval is by a direct (lexical) route (e.g. Delattre et al., 2006
; Martin & Barry, 2012
). However, when writing in a transparent orthography as was the case in the present study, assembling the spelling of a word by phoneme–grapheme translation will typically be successful. Children who do not have the lexical knowledge required to make this direct route possible will therefore nearly always be accurate, but are more likely to be slow. For these reasons, association between children’s accuracy on a spelling task and the quality of their written composition is, at best, only weak evidence that difficulty with spelling disrupts upstream (higher-level) cognitive processing. Child-level factors that independently explain both spelling ability and composition performance probably provide a more parsimonious explanation.
Strong evidence that difficulty with spelling retrieval and/or inscription disrupts higher-level processing requires demonstrating two effects: first, that ability to spell fluently and to inscribe fluently, measured in tasks where these abilities are clearly isolated from other language and literacy-related abilities, predicts word-level fluency when composing full text; second, that word-level fluency when composing a text in turn predicts the compositional quality of the final product.
Word-level fluency here refers to two specific phenomena: the time to initiate the production of a mid-sentence word,2
and the time to initiate production of each letter within a word. In typewriting these can be measured as the time between releasing the spacebar and pressing the character at the start of a new word—henceforth word-initial latency
—or the time between keypresses within a word (within-word latency
). In competent adult writers, word-initial latencies are typically between 250 and 400 ms, and within-word latencies are in the region of 150–250 ms (Chukharev-Hudilainen et al., 2019
; Conijn et al., 2019
). The mental processing underlying word-initial latency and within-word latency is different, but with some overlap. Pre-word processing must include lexical retrieval. In written picture word naming task, performed by adults, frequency affects word-initial latency across a range of languages including Norwegian—the language of the present study—but does appear to affect within-word latencies (Torrance et al., 2018
). Pre-word lexical retrieval includes spelling retrieval—difficult-to-spell words are associated with longer response times in spelling to dictation (e.g. Delattre et al., 2006
). There is some evidence however, that spelling processing is not always complete at inscription onset: within-word latencies also increase for more difficult words, a finding that appears to hold true in both deep and more transparent orthographies and for adults and children (Bonin et al., 2012
; Delattre et al., 2006
; Rønneberg & Torrance, 2019
; Torrance et al., 2018
). Both word-initial and within-word latencies will, of course, also be associated with the motor planning and execution of the keystroke.
Interpretation of these latencies when the writers produce continuous text rather than isolated words is complicated, however, by the fact that at least some of the processing necessary to determine the next keystroke can be conducted in parallel with the output of previous letters or words. This is a direct consequence of the cascading nature of the text production process. So all processing associated with lexical/orthographic retrieval of the next word, for example, does not necessarily occur in the interval after the terminating keystroke of the previous word. Typical means for word-initial and within-word latencies in fact obscure two different distributions of interkey intervals (Roeser et al., 2021
), one associated with intervals dependent only on the time needed to motor plan and execute the next keystroke—processing that takes around 150 ms (Van Waes et al., 2021
)—and a distribution of longer intervals that occur when the writer has to “pause” to engage in additional planning before the keystroke can be output. At a certain point, these longer intervals reach a threshold where they represent processing that is potentially disruptive—i.e. that has the potential to negatively affect the quality of the writer’s output.
A handful of studies have explored the effects of spelling and inscription ability on latency times during composition. In some of these studies latency times above a certain threshold, typically 2 s, are termed pauses. Alves and Limpo (2015
) found that, in both lower and upper-primary school students composing text by hand, the mean number of words between pauses of more than 2 s (a measure referred to as “burst length”) was positively correlated with handwriting fluency (speed of written alphabet recall) and with accuracy in a spelling-to-dictation task (i.e. that students with better handwriting skills and spelling skills paused less, when controlling for the number of words written). Limpo and Alves (2017
) report similar findings in a sample of just 2nd grade writers. Alves et al. (2016
) found that training in handwriting reduces pausing in second grade writers, but that training in spelling does not. Torrance, Rønneberg, et al. (2016
) found longer latencies at both word boundaries and within words, in adolescents with dyslexia (and therefore with weaker spelling skills) composing by keyboard compared to controls. In particular, these students were much more likely to pause within words for periods of 1–2 s, delays that might plausibly have a disruptive effect on upstream processes.
A similarly small number of studies have explored the relationship between pausing during production and the quality of the final text. Grewal and Williams (2018
), in a sample of school age children writing in English as a second or other language, examined effects of mean burst length (with a 2 s pause threshold) and of count and summed duration of pauses over 250 ms on features of the final product, controlling for various child-level factors. They found positive correlations with spelling accuracy and lexical richness, but no relationship with a composite text quality measure. Alves and Limpo (2015
) found that both burst length and mean duration of pauses, again with a 2 s threshold, predicted holistic ratings across grades 2–7 in children writing by hand in Portuguese. The effects were relatively weak after the fourth grade and absent at the sixth grade when children wrote narratives. Asker-Árnason et al., (2010
), in a sample of mainly of secondary school-age students typing narratives, found no relationship between proportion of time on task spent in pauses of greater than 2 s and narrative quality. Deane (2014
) found significant and non-trivial effects of typing fluency on text quality, after control for various product characteristics including spelling accuracy, in a large sample of eighth grade students. The composite nature of the writing fluency measures that they report, however, which aggregated measures including burst length and interkey interval (both subject means and SDs) and deletion counts, makes interpretation difficult. Studies with adult writers have found fewer effects. Medimorec and Risko (2017
) found evidence of a slight tendency for students who pause more before words to write shorter sentences, but few other effects. Conijn (2020
), in a sample of university students writing (proficiently) in English as a second language, found no correlation between subjective text ratings and either word-initial or within-word latencies, or any of a large number of other pause-related measures.
The present study
The aim of our study was to provide a strong test of the hypothesis that difficulty with spelling and/or with inscription disrupts the processing necessary to produce text that is compositionally well formed. Our study sampled sixth grade students writing by typing in a relatively shallow orthography (Norwegian). In this context, we examined the putative causal chain from spelling and inscription ability to production fluency, and from production fluency to the compositional quality of the completed texts. We built on previous research in the following ways. First, our measures of spelling and inscription ability focussed specifically on relevant measures of output fluency, rather than accuracy. For reasons that we have discussed, this is particularly important for assessment of spelling ability. As in previous studies, we assessed spelling ability with a spelling to dictation task, but focussed on speed of retrieval and production rather than accuracy. Second, our analysis of latency times took into account whether these occurred within or before words—the locations likely to be associated with transcription difficulties—or at higher-order text locations. Third, we bring these measures together in a single study that looks at both ability effects on process and process effects on product. For reasons that we have discussed, we believe that these three features are necessary to provide more direct evidence of the process-disruption hypothesis than is, to our knowledge, provided in the existing literature.
Our decision to study typed production in sixth grade students was both expedient and principled. Norwegian children, as in most educational contexts, typically learn to handwrite before they learn to type. There are, however, national curriculum expectations that students achieve a reasonable level of typing proficiency by the time they leave primary school (Ministry of Education and Research, 2019
). As early as third grade, children therefore typically have sufficient typing fluency to create multi-sentence compositions within a reasonable time limit (see, for example, von Koss Torkildsen et al., 2016
), although we anticipated greater variability in skills than might be expected from, for example, a current young adult sample. This, combined with the ease with which writing time course measures can be extracted from typed production, motivated the choice of typing as output medium in our study.
Our study aimed to test the frequently made claims (a) that difficulty with spelling retrieval and/or with the graphomotor skills necessary for inscription leads to slowed or disfluent production at the word level when composing spontaneous text and (b) that this production disfluency then disrupts the processing necessary for planning and structuring text content, and so results in a compositionally inferior final product. We found clear evidence in support of the first of these claims but, at best, weak evidence, for the second. We will discuss each in turn.
Composition fluency was predicted by independent measures of both typing fluency and spelling fluency. Within-word interkeystroke intervals when typing rehearsed sentences from memory (our typing fluency task) predicted within-word interkey intervals when composing. In fact, mean median within-word interkey intervals were only marginally longer during composition than during the typing fluency task. Similarly, word-initial intervals in the typing fluency task predicted word-initial intervals when composing, although in this case latencies were substantially greater when composing, consistent with the considerably greater lexical and orthographic retrieval demands of producing spontaneous text. These effects were after control for response time on the key-finding task, suggesting that ability to construct motor plans that string multiple keypresses together is to some extent distinct from being able to rapidly map letters onto single keys (see also Grabowski, 2008
Key-finding response time also uniquely predicted within-word and word-initial latency when composing. More importantly from a process-disruption perspective, key-finding response time predicted within-word and word-initial pausing during composition: participants who were slower at mapping single letters onto keys were, when composing, more likely to hesitate for intervals of a length that might plausibly be associated with disruption of smooth flow from thought to screen.
Taken together, these findings are robust evidence that typing speed during composition is dependent, in part, on graphomotor skill. They indicate that higher level processes associated with developing content, syntax and spelling do not overdetermine rate of output when composing: keyboard knowledge and ability to string together sequences of keystrokes together explain considerable variance in measures of fluency when writers compose their own text. This finding is not surprising, and is consistent with and extends the findings of previous studies that found correlation between speed of written alphabet recall and composition fluency measures (Alves & Limpo, 2015
; Limpo & Alves, 2017
Previous research has failed to find evidence of a relationship between spelling ability, measured in terms of accuracy in a spelling-to-dictation task, and composition fluency in upper-primary students (Alves & Limpo, 2015
; Graham et al., 1997
). As we discussed in our introduction, however, spelling disfluency rather than spelling inaccuracy is the important factor when considering the possibility that lack of spelling ability might disrupt higher-level processing during composition. We found that fluency in written spelling-to-dictation predicted both word-level interkey interval and the probability of word-level disfluency. Response time on the spelling task predicted word-initial latency and the probability of potentially disruptive word-initial hesitation. Similarly, and interestingly, within-word interkey interval in the spelling task predicted within-word interkey interval when composing and strongly predicted the likelihood of within-word hesitation. It is likely that, once a word has been retrieved, fluent, “just-in-time” written production depends on the ability to then inscribe that word rapidly and effortlessly. The fact that, after control for typing fluency, within-word hesitation increased as a function of spelling fluency suggests that spelling ability affects composition time course and that, at least in a shallow orthography, these effects are not simply associated with word preparation but persist after output of the word has been initiated. This is consistent with findings from experimental written naming studies (Bertram et al., 2015
; Scaltritti et al., 2016
; Torrance et al., 2018
) and studies of spelling to dictation (Rønneberg & Torrance, 2019
), but has not, to our knowledge, previously been shown in spontaneous composition.
Our study therefore provided strong evidence that spelling and inscription fluency, measured independently and directly, and after statistical control for general ability, affects fluency when composing. In particular, lack of ability to spell and inscribe fluently increases the probability of pauses during composition of a duration that might, potentially, result from or in disruption to higher-level processing.
The second question to ask is whether or not these measures of composition fluency predict the quality of the resulting text. Our findings here were much less clear. As might be expected, participants who were less fluent composed shorter texts. Tendency to make spelling errors was associated with a greater tendency to hesitate within a word. The spelling effect is consistent with previous findings that general measures of typing speed during composition correlate negatively with the number of spelling errors in the third grade and seventh grade participants (Asker-Árnason et al., 2010
; von Koss Torkildsen et al., 2016
). Our finding may be due to either or both of writers hesitating within word to retrieve spelling or hesitating because they can see that they have made an error. Better understanding of possible mechanisms behind these effects are possible with more detailed analysis of where writers did and did not hesitate. These analyses are, however, beyond the present scope. The tendency to use a more lexically diverse vocabulary was not associated with a greater tendency to hesitate at the word level. Our results are similar to findings that lexical diversity measured as MTLD was not affected by disfluencies (Medimorec et al., 2017
We did not, however, find strong evidence of a relationship between word-level composition fluency and compositional quality (content and rhetoric) of the completed text. Of 12 possible effects, we found evidence for just one: participants who tended to type more slowly within-word tended to show weaker theme development—their texts tended toward weaker argumentation and tended to lack focus. An increase of 100 ms in median within-word latency (from a sample mean of 323 ms) was associated with a decrease of 0.32 in theme development (scored from 1 to 8). Note, however, that this effect was for within-word typing speed. We found no effects on any of the three quality measures for either of our two disfluency measures, and therefore no evidence that unusually long pauses disrupted higher-level processing.
We see, broadly, three possible explanations for failure in this and previous studies to find effects of processes disfluency on the compositional quality of the resulting text. It may be that there is potential for difficulties with spelling and inscription to interfere with upstream processes, but that in our present sample participants were sufficiently competent for this not to occur. We found clear evidence that children who were less fluent, relative to peers, when performing spelling and inscription tasks were also relatively less fluent when composing. The 25% of children who showed the most within-word disfluency paused for > 1 s before, on average, 1 in 18 keystrokes, and the 25% of children who paused most before words (with a 2 s threshold) did so before 1 in 9 words. It is possible, however, that this degree of hesitancy can be absorbed by the writing-production system without affecting composition quality. Alves and Limpo (2015
) reported sixth grade writers composing a mean of 4.84 words for every pause in excess of 2 s (SD
= 1.91), regardless of the location of pause within the text and found (weak) effects on text quality. Calculating the equivalent statistic for our sample gave M
= 13.2 words (SD
= 10.2). Part of this difference may be due to differences in recording methods. Alves and Limpo’s students wrote by hand, and times for in-air pen movements, between letters and between words, were included when determining pause durations. It is possible, however, that even given this, students in our sample were more fluent, although this is contrary to our expectations: typing, although practiced in Norwegian primary schools, remains the less preferred inscription modality for the majority of classrooms, including those that we sampled in this study.
Second, and more generally, it is possible that struggling with spelling retrieval and/or inscription does not, in fact, interfere with or divert attention from higher-level processes. It could be the case that motor planning or spelling retrieval, even when difficult, do not share processing resources or mechanisms with upstream processes. Third, it is possible that, even if delays at output result in information loss (the now-or-never bottleneck; Christiansen & Chater, 2016
), this loss is recoverable. There is some support for this. Writing, unlike speech, provides an external record of what has just been said. Evidence from eye movement studies suggest that this might be used repeatedly to cue retrieval of what to say next (Torrance, Johansson, et al., 2016
; Torrance, Rønneberg, et al., 2016
), reducing demand on the memory buffers that are an essential part of a cascading text production system (van Galen, 1991
). To our knowledge, the only experimental test of the hypothesis that increased inscription load reduces composition quality gave negative results: across two experiments in which adult competent typists were required to type with both hands and with one hand Medimorec and Risko (2016
) found substantially slowed production in the one-hand condition. However, across a range of text measures they found limited evidence of effects on text quality. Where Medimorec and Risko did find effects in some cases, these pointed toward improved rather than reduced quality in the one-hand condition.
Our study failed to find evidence for a relationship between word-level composition fluency and text quality. This was despite strong evidence composition fluency was predicted by spelling and inscription ability. This is not, of course, to argue that spelling and inscription ability are unrelated to text quality in primary-aged children or persons with writing difficulties. The evidence reviewed by, for example, Feng et al (2019
) points towards a clear association between transcription ability measures and the quality of their written composition. The findings that we present here do not, however, support the frequently made claim that these effects result from process disruption. A more parsimonious explanation for previous findings may simply be that the range of skills required to score well on tasks used to assess transcription ability (spelling to dictation, sentence copying, written alphabet recall), which include reading, short-term memory, and alphabet knowledge alongside orthographic knowledge and motor skill, correlates with the knowledge and skills required to produce well-structured and content-rich text.
In conclusion, therefore, we believe that the present study provides stronger evidence than has previously been available that in upper-primary-aged children the ability to spell and type fluently affects writing time course when composing full text. We did not, however, find strong evidence that fluency in composing affected the compositional quality of the completed text. This is inconsistent with claims made frequently in the existing literature that word-level disfluency when composing text disrupts the processing necessary for developing text content and structure. One possible reason for this is that although there was variation in fluency across our sample, children tended to have sufficient fluency in spelling and typing for this not to be disruptive to upstream processes. Future research could usefully reproduce the measures used in this study with a sample of children at an earlier developmental stage. In the meantime, however, we argue that the process-disruption hypothesis, as a claim that holds true across all writers, is less well supported than is typically assumed.