Finding the best fit: examining the decision-making of augmentative and alternative communication professionals in the UK using a discrete choice experiment

Objectives Many children with varied disabilities, for example, cerebral palsy, autism, can benefit from augmentative and alternative communication (AAC) systems. However, little is known about professionals’ decision-making when recommending symbol based AAC systems for children. This study examines AAC professionals’ preferences for attributes of AAC systems and how they interact with child characteristics. Design AAC professionals answered a discrete choice experiment survey with AAC system and child-related attributes, where participants chose an AAC system for a child vignette. Setting The survey was administered online in the UK. Participants 155 UK-based AAC professionals were recruited between 20 October 2017 and 4 March 2018. Outcomes The study outcomes were the preferences of AAC professionals’ as quantified using a mixed logit model, with model selection performed using a step-wise procedure and the Bayesian Information Criterion. Results Significant differences were observed in preferences for AAC system attributes, and large interactions were seen between child attributes included in the child vignettes, for example, participants made more ambitious choices for children who were motivated to communicate using AAC, and predicted to progress in skills and abilities. These characteristics were perceived as relatively more important than language ability and previous AAC experience. Conclusions AAC professionals make trade-offs between attributes of AAC systems, and these trade-offs change depending on the characteristics of the child for whom the system is being provided.

The process through which children receive AAC systems can vary 8 9 , but commonly their needs and abilities are assessed by a team of AAC professionals, which may include speech and language therapists, occupational therapists and/or specialist teachers 10 . The team makes recommendations, and a final decision is made with variable input from the child and family, depending on individual circumstances. The large degree of heterogeneity in the population of people who benefit from AAC, and in the systems available, means the assessment and subsequent matching of individual and system is a complex task and unique to each person.
It has been estimated that approximately 1 in 200 children in the UK would benefit from AAC [11][12][13] , although rates of abandonment of AAC systems by children of 30-50% have been observed 14 15 , with the causes of abandonment not well understood. AAC systems can be costly (up to £10,000 for high-tech systems) and require a large amount of professional support 16 . However, they have been suggested to be a cost-effective use of NHS resources 17 .
There is currently a lack of documented evidence for assessment and decision making processes 18 19 , and what does exist is frequently individual case studies [3,17]. AAC practitioners must often make difficult and complex decisions in a complicated, heterogeneous and rapidly evolving environment, balancing the needs of an individual child, and available resources, and take account of the cultural and contextual influences that shape each assessment 20 21 . While there are studies which highlight some important factors in decision making 12 18 22 , available guidelines tend to focus on the organisational structure of services, rather than decision making as such 10 19 This study is part of a wider project examining provision of AAC systems for children entitled I-ASC: Identifying appropriate symbol communication aids for children who are non-speaking. It has several components and has employed a number of research methods with the aim of generating a body of research evidence on current practice, recommendations for best practice, and resources to support AAC professionals in making these decisions. This study aims to contribute quantitative evidence regarding the current decision making rationale of AAC professionals. As yet, there is limited quantitative research in AAC decision making. A study from the current research project ran a Best-worst Scaling (BWS) Case 1 survey 23 , which quantified what AAC professionals considered the most and least important factors related to both children and AAC systems. However, the earlier study did not examine the trade-offs professionals make between different attributes of AAC systems, nor how trade-offs change depending on the characteristics of children.
This study presents a discrete choice experiment (DCE) which seeks to achieve both these aims. Participants were shown a series of vignettes describing hypothetical children and made choices as to which among a set of hypothetical AAC systems they would choose for each child. Analysing the results revealed respondents' preferences for the levels of various AAC system attributes, as well as how those preferences are influenced by the children's characteristics.

Survey development
No stated preference work existed in AAC prior to this research project, meaning a large number of potential attributes and little evidence as to which to include in a DCE. Thus a BWS case 1 study was performed initially and the results used to guide attribute selection for the DCE. Attributes for the BWS study were created using focus groups and interviews with AAC professionals, people who use AAC, their families, and other stakeholders; systematic literature reviews; and input from an expert panel. For more details see section 2 of Webb et al. 23 .
The BWS study produced relative importance scores for 19 child and 18 system attributes given in Appendix B with their rank in terms of relative importance. DCE attributes were selected from these during consensus discussions between authors with expertise in AAC, speech and language therapy, and health economics. The selection criteria were that attributes should: (1)   would not over burden respondents. Consensus was achieved via unstructured discussions until all authors were in agreement. This resulted in four child and five system attributes, listed in Tables 1 and 2 together with non-specialist descriptions.
Broadly speaking, child attributes capture a child's language ability, experience with AAC, attitude/motivation to communicate with AAC, and whether the child is expected to regress, plateau or progress in communication ability. System attributes broadly capture what vocabulary set(s) are preprovided by manufacturers, how many vocabulary items are provided, how they are organised, the type of graphical symbols used, and how consistent the layout of words/symbols are.
A total of 54 vignettes can be formed from the child attributes. Authors with expertise in AAC and speech and language therapy identified and removed 18 vignettes representing unrealistic combinations, leaving 36.
Each participant had three vignettes randomly selected to answer questions about.
Prior experience from the BWS study suggested it would be difficult to recruit a large respondent sample, thus a relatively heavy response burden of 12 choices between three systems was selected. Authors with experience in AAC and speech and language therapy removed 158 unrealistic combinations from the 432 AAC systems which could be formed from the system attributes, leaving 274. A D-efficient survey design was generated using NGene (©ChoiceMetrics) with five blocks, meaning 60 choice tasks in total. Random allocations of block and child vignettes were independent.
The children in the first, second and third vignettes were referred to as Child A, Child B and Child C respectively, and participants made four choices for each with choices for a given child grouped together.
Note that as the sets of AAC systems presented were driven by the experimental design, participants were offered different choices for each child. The order of system attributes was randomised between participants, but consistent within participants and which systems appeared on the left, middle and right of the screen was also randomised. An example question is shown in Appendix A. Participants finally answered questions about themselves and their experiences with AAC (for details, see supplementary online material). The DCE was administered online for ease of recruitment and was tested by five AAC professionals and the wording of some attributes and levels altered. Participants began by confirming they contributed towards AAC decision making for children, and those who did not answered only demographic questions. The precise wording of the question was: "I confirm my work involves assessing children for aided AAC systems and I contribute to the decision making in relation to the language and vocabulary organisation with in AAC systems." During testing it was revealed that some AAC professionals did not have sufficient input into the decision making process in their day-to-day practice for the DCE questions to be meaningful (e.g. occupational therapists specialising in optimising physical access to an AAC system recommended by other members of the team), and this question was designed to filter out such respondents.
Recruitment was carried out using email lists of AAC professionals gathered by the research project as part of prior activities, publically available lists and websites and the professional contacts of authors. The study was also advertised via the mailing list of Communication Matters (www.communicationmatters.org.uk), a UK wide AAC charity, through the project website and online media. Responses were collected between 20/10/17 and 4/3/18.

Analysis
Analysis of participants' choices was grounded in random utility maximisation. In a given choice scenario , participant chooses which of three AAC systems to allocate to child . The utility to participant of allocating AAC system to child in choice scenario is ∈ {1,2,3} = + + where is an alternative specific constant for AAC system , is a vector of dummy variables indicating AAC system levels, is a vector of coefficients which differ across participants and children, and is an error term which varies across choice scenarios and alternatives. The coefficient on level of system attribute , , depends of the characteristics of the child vignette according to where is a constant giving the preference for a system attribute at baseline child levels, is a vector of 0 dummy variables indicating vignette levels and is a vector of coefficients, allowing for heterogeneity in relative preference for AAC system attributes depending on child characteristics.
A full model with all interaction terms includes too many parameters to estimate reliably. Thus parameters were eliminated in a stepwise process and a final preferred mixed logit model identified using the Bayesian Information Criterion. The mixed logit model incorporates participant heterogeneity by allowing AAC system attribute parameters to be random, following a normal distribution with both means and variances depending on child characteristics. For details, see Appendix C.
Models were estimated using the CMC Choice Modelling Centre Code for R version 1.1 24 and all analysis was carried out using R version 3.3.1. Statistical significance was assessed at the 5% level after adjusting for multiple testing using Holm's sequential Bonferroni correction 25 .
Results are presented using a new measure termed relative interaction attribute importance (RIAI) which assesses how big an impact child attributes have on AAC professionals' decision making. It is analogous to relative attribute importance, often used to present DCE results 26 , and it may be calculated either with respect to a single choice object attribute or overall with respect to all choice object attributes. For details, see Appendix D.
Note relative attribute importance is not an appropriate measure of the importance of AAC system attributes here, as their relative importance changes depending on which child AAC professionals are presented with.
Nor is it appropriate to take the mean relative importance over all child vignettes. The set of child vignettes used is not representative of the case mixes seen by AAC professionals: some vignettes may represent children commonly seen, while other vignettes may represent a type of child seldom encountered. Thus averaging over the set of vignettes would not give meaningful insight as to the relative importance of an attribute in AAC professionals' decision making in the real world.

Patient and public involvement
One author (SM) is an AAC user, and one (LM) is the parent of an AAC user, and both were involved in all stages of the research. DCE attributes were developed with impact from AAC stakeholders and the survey tested with AAC professionals as detailed above. Findings from the study and the wider research project have been disseminated to AAC stakeholders and the public at events at the Scottish Parliament (Edinburgh, UK), the Science and Industry Museum (Manchester, UK) and the Houses of Parliament (London, UK).

RESULTS
A total of 172 participants completed the survey, of whom 155 indicated they contributed to decision making regarding AAC systems and answered DCE questions. Summary statistics of their demographics and professional experience are given in Table 3. Most participants were female (~90%) and white. We believe this to be reasonably representative of the population of AAC professionals in the UK. The mean age of DCE participants was around 40 and they had around 10 years experience on average. Around 75% of DCE participants had a speech and language therapy background, with no other background reported by more than 10%. Those who did not answer DCE questions were less likely to have a speech and language therapy background (~50%), with teacher (~20%) and occupational therapist (~30%) more common.
Approximately 30% of the sample worked with all age groups, while 50-60% worked with each of preschool, primary school and secondary school age children. The sample also encountered a wide range of diagnoses, e.g. physical disability (~80% of DCE participants), intellectual disability/developmental delay (~70%) and autism spectrum disorder (~65%).
Turning to DCE responses, respondents chose the left-hand option 37.6% of the time, and the central and   Table 4 contains the results of the final preferred model, with 24 coefficients. The "constant" terms give participants' preferences for AAC system allocation when shown a vignette with all attributes at baseline levels, which represents the most challenging profile that can be formed from the set of child levels. ("Child A/B/C has delayed expressive and receptive language and no previous AAC experience. Child A/B/C does not appear motivated to communicate through any methods and means. Child A/B/C is predicted to progress in skills and abilities (regression).") The interaction terms in the model hence represent how respondents' preferences for AAC systems change if choosing for a vignette presenting less of a challenge on a given child attribute.
For the baseline vignette, vocabulary sets which are fixed or have staged progression are preferred to no preprovided vocabulary. There are no significant differences in preferences between up to 50 and 50-1000 vocab items, but over 1000 items is considered significantly worse. There is no significant preference between visual scene, taxonomic or semantic-syntactic vocabulary organisation, but pragmatic organisation is preferred. There is no preference between graphic representation using photos or pictographs, but text is less preferred than either, and idiographic symbols are considered even worse. Finally, having only some aspects of system layout consistent is less preferred than having all aspects consistent or an idiosyncratic layout.
Compared to this baseline vignette, professionals were much more likely (odds ratio, OR 3.88) to choose systems with staged progression vocabulary sets with staged progression to no pre-installed set if the vignette predicted progress in skills and ability. An intermediate number of vocabulary items (50-1000) became more preferable compared to 50 or fewer for a vignette motivated to communicate using AAC. Over 1000 items became significantly more preferable for vignettes with a variety of characteristics: receptive language exceeding expressive language, an ability to use a range of AAC functions, motivated to communicate using AAC and predicted to progress. Two significant interactions exist between vocabulary organisation and motivation. A vignette with motivation to communicate using AAC became more likely to be allocated a system with taxonomic (OR 2.03), or semantic-syntactic (OR 2.29) organisation compared to visual scene.
Motivation to communicate using AAC also has a large influence on graphic representation. It increases the probability of choosing pictographic symbols (OR 3.88), idiographic symbols (OR 5.31) or text (OR 4.00) rather than photos. However, being predicted to progress makes pictographic symbols less preferable. Figure 1 illustrates the RIAI of child attributes for each system attribute and overall. Consistency of layout is omitted, as there are no interactions for this attribute. Predicted future skills and abilities is the only child attribute to influence preferences for type of vocabulary set. It is also one of only two to influence preferences for graphic representation, although determination and persistence is more important (67% vs. 33%). Determination and persistence is the only child attribute to impact preferences for type of vocabulary organisation. All child attributes have an influence on preferences for vocabulary size, with communication ability with AAC (32%) and determination and persistence (28%) relatively more important than future skills and abilities (22%) and receptive and expressive language (17%). Overall, future skills and abilities has the greatest relative importance (38%), followed by determination and persistence (19%), communication ability with AAC (20%), and receptive and expressive language (12%).

DISCUSSION
This study has demonstrated the feasibility of DCEs as a research tool in AAC. Although some informal feedback was received that participants found the tasks difficult due to not having as much information or as wide a range of options as in real life, the AAC professionals that responded to the survey found the tasks meaningful. The set of AAC system attributes created coherent options and the child vignettes presented meaningful descriptions which conveyed useful information to participants.
This DCE has revealed AAC professionals' priorities when choosing AAC systems for children and shown that these priorities interact with children's, to the extent that for some system attributes their preferences for F o r p e e r r e v i e w o n l y different levels can completely reverse depending which vignette is shown. For example, for the baseline vignette, a system with more than 1000 vocab items is less likely to be chosen then one with less than 50 (OR 0.395). However, for a vignette with a receptive-expressive language gap, can use AAC for a range of functions, is motivated to use AAC and is predicted to progress, a system with more than 1000 vocab items is much more likely to be chosen (OR 22.5).
Overall, motivation to communicate with AAC has the greatest number of interactions with preferences.
Motivation is also more important in terms of RIAI than language ability or previous experience with AAC, although motivation to communicate through non-AAC methods has no bearing on preferences in the final model. Motivation to communicate via AAC tended to drive participants towards what can be regarded as more "ambitious" choices, for example more vocabulary items.
Visual scene vocabulary organisation and graphic representation using photos can both involve items/scenes from an individual's own life and use literal, rather than abstract depictions. Both became less preferred for a vignette motivated to communicate via AAC, in favour of more abstract methods of organisation (taxonomic and semantic-syntactic) and graphic symbols that require more grammar (pictographs, ideographs and text). This may be interpreted as AAC professionals believing that motivated children will be better able to use more complex AAC systems. An alternative and by no means mutually exclusive interpretation is that lack of motivation requires an AAC system involving familiar cues from their everyday environment.
AAC system preferences did not significantly differ between vignettes with skills and abilities predicted to regress or plateau. This may be due to children predicted to regress not being commonly encountered. However, if a child is predicted to progress, this has a large impact on decision making, and future skills and abilities is the highest ranked attribute in terms of RIAI. As with motivation, it leads to more ambitious choices, with more vocab items preferred and pictographs depreciated as graphic symbols compared to ideographs and text. However, unless the vignette featured both predicted progress and motivation to For many vignettes, there are non-linear preferences for vocabulary size. Between 50 and 1000 items was considered better than 50 or fewer for all child vignettes, although the difference was not always significant.
This may indicate that AAC professionals do not wish to limit potential for expression, even for children with low ability and poor prognosis.
Respondents commonly preferred levels of AAC systems that can require a lot of personalisation, e.g.
photographic graphic representation, pragmatic vocabulary organisation or an idiosyncratic layout. Such options require a lot of time and effort on behalf of AAC professionals, highlighting the need in real-world practice to allow sufficient time for AAC system setup. Another implication is that "off the shelf" AAC systems may not generally be suitable for children without alteration. Note that this is not necessarily a criticism of the range of AAC systems available from manufacturers.
The above results are potentially contrasted by preferences for vocabulary sets, where no pre-provided vocabulary set was always considered worse than having pre-provided sets that were either fixed or with staged progression. However, even here, AAC professionals may have intended to customise the provided sets to tailor them to the individual.
Comparing the DCE results with the previous BWS Case 1 study, some similarities may be observed. For example, graphic representation was selected as the wider research project aimed to investigate the properties of graphic symbols, and was the lowest ranked attribute in terms of importance in the BWS to be included in the DCE. In concordance with this finding, if relative importance of AAC system attributes is calculated for each child vignette in the DCE, it is never the most important attribute.
Many differences can also be seen. Language abilities was the most important child attribute in the BWS, yet its RIAI in the DCE is below predicted future abilities, ranked sixth in the BWS. However, differences in results do not necessarily imply contradiction, as the two methodologies do not measure the same thing. The BWS measured the importance of AAC system attributes over the case mix AAC professionals encounter in practice, whereas for the DCE respondents were presented with a specific child vignette.

Limitations
This study has several limitations. As it was not possible to estimate a model with all interactions, results will to a certain extent be sensitive to the model selection strategy. However, the final model was selected using a well-established and widely used selection criterion (BIC). A larger sample size may have allowed robust estimation of more complex models, yet 155 participants represents a large proportion of the population of AAC professionals working in the UK, which is estimated at around 800 (Communication Matters, personal correspondence).
Respondents were more likely to choose AAC systems on the left of the screen and less likely to choose ones on the right, potentially introducing bias in estimated coefficients. However, alternative specific constants were included when modelling responses, and did not provide enough explanatory power to be included in the final model. In addition, the positions in which AAC systems were presented was randomised, mitigating any possible bias.
In some ways, the DCE task does not match how AAC professionals make decisions in practice. Typically, they work together with families and children, as well as part of an AAC team, which could include diverse areas of expertise. They also generally make recommendations, rather than unilaterally choosing a system.
Similarly, the DCE tasks presented one-off static decisions, whereas in reality the process is dynamic, with a Attributes and levels use a mixture of speech and language therapy terms (e.g. receptive and expressive language) and more technological language (e.g. staged progression). This may have made it difficult for respondents from any one speciality to interpret all of them. However, this issue is not limited to the current study, but reflects an ongoing struggle in AAC to establish a common language, given its interdisciplinary nature. In addition, respondents may have been unfamiliar with the generic term ideographic symbols, since only a single commercial set of ideographic symbols is in popular use (Minspeak, © Semantic Compaction Systems, Inc.).
Compared to the real children AAC professionals encounter, vignettes were simple, and lacking in details that would normally be available. A single vignette also represents potentially very different children. For example, the needs of a child who has plateaued in skills and abilities at age five will be very different to a child who has plateaued at age 15. However, this is an inherent limitation of the DCE methodology, and vignettes with a greater number of attributes and levels would have made decisions overly burdensome. In addition, significant interactions between AAC systems and child attributes implies the vignettes were meaningful enough that respondents changed their preferences in response to them, often dramatically.
For a given child vignette, it is only possible to determine relative preferences for system attributes, rather than absolute preferences. Thus it is not possible to tell how suitable a given system is for a given vignette, which is important as some vignettes presented a challenging profile for which it may be hard to find a    A series of vocabulary sets with pre-determined progression through them that simulate language development. E.g. an initial set including just basic words, with subsequent sets introducing more grammatical structure. May be customised.

Consistency of layout (2)
How consistent positions of words/symbols are in system interface, and how consistent navigation to find different symbols is *Consistency of some aspects of layout Words/symbols in multiple categories appear in different positions across categories, but always in the same place in a given category Consistency of all aspects of layout All/nearly all words/symbols always appear in same position in interface Idiosyncratic layout Layout that has been personalised for an individual child Type of vocabulary organisation (5) How words/symbols are organised within the system *Visual scene Interface shows photos, most likely of scenes familiar to the child, with areas of it highlighted to represent words Taxonomic Words/symbols organised according to subject, analogous to nonfiction books in a library Semantic-syntactic Words/symbols organised according to sentence structure, e.g. verbs, nouns, adjectives Pragmatic Words/symbols organised around function in language rather than grammar, e.g. request, mood Size of vocabulary (7) How many words/symbols system can output *Up to 50 vocabulary items Implies only simple communication functions possible 50-1000 vocabulary items Implies combining words/symbols to create grammatical structures More than 1000 vocabulary items Does not imply more complex communication than 50-1000 items, but means a greater load on child's memory Graphic representation (12) Type of symbols used by system     1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  You will be asked a series of questions. Each one has the same format. A brief description of a child will be given, along with three possible choices of aided AAC systems.
q I have read and understood the above and consent to taking part.
I confirm my work involves assessing children for aided AAC systems and I contribute to the decision making in relation to the language and vocabulary organisation within aided AAC systems.

q Yes q No
If yes go to DCE questions.
If no go to a page with the following:-Thank you for your interest in this survey. At present we are only recruiting participants who contribute to decision making in relation to the language and vocabulary organisation within aided AAC for children.
Over the coming 12 months we will be recruiting people with a wider range of AAC experience to test decision making resources we are developing. If you are interested in this aspect of the project or would like to be notified when the free resources are available, there will be an opportunity at the end to submit your email address.
We would still like to ask you a few questions about your experience with AAC to check the representativeness of participants.

Size of vocabulary
The size of the output vocabulary available within the aided AAC system. Child B has receptive language exceeding expressive language and no previous AAC experience. Child B is only motivated to communicate through methods other than symbol communication systems. Child B is predicted to maintain current skills and abilities (plateau).

Size of vocabulary
The size of the output vocabulary available within the aided AAC system. Child C has delayed expressive and receptive language and no previous AAC experience. Child C is only motivated to communicate through methods other than symbol communication systems. Child C is predicted to maintain current skills and abilities (plateau).

Size of vocabulary
The size of the output vocabulary available within the aided AAC system. Child C has delayed expressive and receptive language and no previous AAC experience. Child C is only motivated to communicate through methods other than symbol communication systems. Child C is predicted to maintain current skills and abilities (plateau).

Size of vocabulary
The size of the output vocabulary available within the aided AAC system. Child C has delayed expressive and receptive language and no previous AAC experience. Child C is only motivated to communicate through methods other than symbol communication systems. Child C is predicted to maintain current skills and abilities (plateau).

Size of vocabulary
The size of the output vocabulary available within the aided AAC system. Child C has delayed expressive and receptive language and no previous AAC experience. Child C is only motivated to communicate through methods other than symbol communication systems. Child C is predicted to maintain current skills and abilities (plateau).

Size of vocabulary
The size of the output vocabulary available within the aided AAC system.

End of survey
Thank you for your participation in this survey.
Your responses will contribute to the results of the I-ASC project and support the development of decision making resources for use in AAC assessments.
You can follow the progress of our research project on our website, on Facebook or on Twitter.  Appendix C -Final preferred model selection process A full model with all interaction terms and two alternative specific constants implies 98 parameters, which is too many to reliably estimate given the amount of data collected and given that many interactions are expected to be of very low magnitude. Thus, a strategy was required to identify a suitable model with fewer parameters.
The first stage was estimating a series of stepwise multinomial logit (MNL) models, beginning with a model with all 98 parameters. The parameter with the highest p-value, excluding the constant terms, was 0 eliminated, and a model with 97 parameters was estimated. Then the parameter with the lowest p-value was excluded and a new model run, and so on in an iterative process until only the 12 constant terms remained 0 (one for each non-baseline system level).
The Bayesian Information Criterion (BIC) was used to select the preferred MNL model. This model was then re-estimated as a mixed logit (MIXL) model to account for participant heterogeneity. (The process did not begin by estimating a series of stepwise MIXL models due to the difficulty and greatly increased computational resources required to estimate MIXL models with a large number of parameters.) The \coefficients on system attribute levels were assumed to be drawn from normal distributions with means given by = 0 + and variances given by If p is the number of parameters of the preferred MNL model, then models with between p -3 and p + 3 parameters were re-estimated as MIXL models. The BIC for each MIXL model is given in Error! Reference source not found..
The MIXL model minimising the BIC was chosen as the final preferred model.  decision-making when recommending symbol based AAC systems for children. This study examines AAC professionals' preferences for attributes of AAC systems and how they interact with child characteristics.
Design: AAC professionals answered a discrete choice experiment (DCE) survey with AAC system and childrelated attributes, where participants chose an AAC system for a child vignette.

Setting:
The survey was administered online in the UK.

Outcomes:
The study outcomes were AAC professionals' preferences as quantified using a mixed logit model, with model selection performed using a stepwise procedure and the Bayesian Information Criterion.
Results: Significant differences were observed in preferences for AAC system attributes, and large interactions were seen between child attributes included in the child vignettes, e.g., participants made more ambitious choices for children who were motivated to communicate using AAC, and predicted to progress in skills and abilities. These characteristics were perceived as relatively more important than language ability and previous AAC experience.
Conclusions: AAC professionals make trade-offs between attributes of AAC systems, and these trade-offs change depending on the characteristics of the child for whom the system is being provided.

STRENGTHS AND LIMITATIONS OF THIS STUDY
 This is the first discrete choice experiment, and only the second stated preferences study in the field of augmentative and alternative communication.
 The study used unusual and innovative methodology by (1) using a Best-worst Scaling case 1 study in attribute selection; (2) having AAC system choices be made in the context of a child vignette  In some ways, the discrete choice experiment task differed from how augmentative and alternative professionals make decisions in practice. Major advances in the AAC landscape have occurred in recent years. 11 12 These include technological innovation, for example iPads and eye-tracking, though low-tech systems may still offer the best solution in many cases. 13 14 Another development within services is a greater expectation of participation in all aspects of life for people who use AAC, 11 15-17 coupled with advocacy for the right to communicate. 14  Although there is a lack of robust evidence surrounding the decision-making process, some factors in successful adoption of AAC have been identified. An AAC system is more likely to be adopted by a motivated child 22 with good support from the child's network. 27 33 44 The AAC system must also meet a child's individual needs and circumstances, which will be unique to every child. 14 22 46 A previous study from the current research project investigated the AAC decision-making process using a Best-worst Scaling (BWS) case 1 survey. 45 This method was chosen as it could quantify which of several child and AAC system related factors (37 in total) AAC professionals considered most and least important in decision-making.

INTRODUCTION
The current study sought to complement the previous work by examining fewer factors in more detail using a DCE. 43 It aimed to quantify the clinical judgements and trade-offs AAC professionals make between different attributes of AAC systems, and how those trade-offs change depending on children's characteristics, things not possible using BWS case 1. This is the first DCE carried out in AAC, and there were challenges associated with performing a DCE with a target population of AAC professionals (for details see discussion). Thus, an additional goal was to establish the feasibility of using DCEs as a research tool in AAC. The BWS study produced relative importance scores for 19 child and 18 system attributes given in Appendix A. DCE attributes were selected from these during consensus discussions between authors with expertise in AAC, speech and language therapy, and health economics. The selection criteria were that attributes should: Mirenda. 17 )

Survey development
In summary, the child attributes capture a child's language ability, experience with AAC, attitude/motivation to communicate with AAC, and whether the child is expected to regress, plateau or progress in communication ability. A total of 54 child vignettes were formed from the set of child attributes. Authors with expertise in AAC and speech and language therapy identified and removed 18 child vignettes representing unrealistic combinations, leaving 36.
AAC system attributes broadly captured the vocabulary set(s) provided by manufacturers, vocabulary size and organisation, type of graphic symbols used, and how consistent the navigational layout of words/symbols is when accessing items. It was not stated whether a system was high-tech or low-tech, although certain levels (e.g., vocabulary sets with staged progression) are more common with high-tech systems. Authors with experience in AAC and speech and language therapy removed 158 unrealistic combinations from the 432 AAC systems which could be formed from the system attributes, leaving 274.
Prior experience from the BWS study suggested it would be difficult to recruit a large respondent sample, so to maximise the information captured a relatively heavy response burden of 12 choices between three systems was selected for the DCE. Participants were shown three child vignettes, referred to as Child A, B and C, and made four choices for each child vignette. An example task is shown in Appendix B.
The survey's statistical design (i.e., which levels of the AAC system attributes were presented in each question) was generated using NGene i , with 60 choice tasks split into five blocks. The design sought to maximise Defficiency, a measure of how much information it is possible to extract from survey responses. 49 The survey was piloted by five AAC professionals and consequently the wording of some attributes and levels altered.

Survey administration
The DCE was administered online for ease of recruitment. Recruitment was carried out via AAC professionals email distribution lists (the project's own list and the mailing list of the UK wide charity Communication Matters www.communicationmatters.org.uk). In addition, invitations were sent via publicly available lists and websites, and the professional contacts of authors. Adverts were also placed on the project website and online media. Responses were collected between 20/10/17 and 4/3/18. Ethical approval was received from an NHS Research Ethics Committee (REC reference 6/NW/0165) and informed consent obtained from participants at the start of the survey.
i ©ChoiceMetrics Participants began by confirming they contributed towards AAC decision-making for children, and those who indicated they did not answered only demographic questions. i Three child vignettes and one survey block were randomly allocated to each participant. The order of system attributes was randomised between participants, but consistent within participants, and which systems appeared on the left, middle and right of the screen was also randomised. At the end of the survey, participants answered demographic questions (for details, see Appendix A).

Analysis
Analysis of participants' choices was grounded in random utility theory. This standard approach 50 assumes participants choose the object which maximises their utility. The utility of an object is modelled as depending partly on the object's attributes and partly random, the latter component capturing the influence of all factors not included in the model. In a given choice scenario , participant chooses which of three AAC systems to allocate to child . The utility to participant of allocating AAC system to child in choice ∈ {1,2,3} where is an alternative specific constant for AAC system , is a vector of dummy variables indicating AAC system levels, is a vector of coefficients which differ across participants and children, and is a random error term.
i The precise wording of the question was: "I confirm my work involves assessing children for aided AAC systems and I contribute to the decision making in relation to the language and vocabulary organisation within AAC systems." During testing it was revealed that some AAC professionals did not have sufficient input into the decision making process in their day-to-day practice for the DCE questions to be meaningful (e.g., occupational therapists specialising in optimising physical access to an AAC system recommended by other members of the team), and this question was designed to filter out such respondents. The coefficient on level of system attribute , , depends of the characteristics of the child vignette according to where is a constant giving the preference for a system attribute at baseline child levels, is a vector of 0 dummy variables indicating vignette levels and is a vector of coefficients, allowing for heterogeneity in relative preference for AAC system attributes depending on child characteristics.
A full model with all interaction terms included too many parameters to estimate reliably. Thus, parameters were eliminated in a stepwise process and a final preferred mixed logit model was identified using the Bayesian Information Criterion. The mixed logit model incorporates participant heterogeneity by allowing AAC system attribute parameters to be random, following a normal distribution with both means and variances depending on child characteristics. For details, see Appendix C.
Models were estimated using the CMC Choice Modelling Centre Code for R version 1.1 51 and all analysis was carried out using R version 3.3.1. Statistical significance was assessed at the 5% level after adjusting for multiple testing using Holm's sequential Bonferroni correction. 52 Results are presented using a new measure termed relative interaction attribute importance (RIAI) which assesses how big an impact child attributes have on AAC professionals' decision-making. It is analogous to relative attribute importance, often used to present DCE results, 53 and may be calculated either with respect to a single choice object attribute or overall with respect to all choice object attributes. For a formal definition of RIAI, see Appendix D.

Patient and public involvement
One author (SM) is an AAC user, and one (LM) is the parent of an AAC user, and both were involved in all stages of research development and delivery. A total of 172 participants completed the survey, of whom 155 indicated they contributed to decision-making regarding AAC systems and answered DCE questions. Summary statistics of their demographics and professional experience are given in Table 3. Most participants were female (~90%) and white. We believe this to be reasonably representative of the population of AAC professionals in the UK. i The mean age of DCE participants was around 40, with a range from 24 to 65, and they had 10 years' experience on average of AAC.

RESULTS
Around 75% of DCE participants had a speech and language therapy background, with no other background reported by more than 10%. Those who did not answer DCE questions were less likely to have a speech and language therapy background (~50%), with teacher (~20%) and occupational therapist (~30%) more common.
Approximately 30% of the sample worked with all age groups, while 50-60% worked with pre-school, primary school and secondary school aged children. Participants were asked for the three most common diagnoses encountered in their work, with ~80% stating physical disability, 70% stating intellectual disability/developmental delay and 65% stating autism spectrum.
Turning to DCE responses, respondents chose the left-hand option 37.6% of the time, and the central and right-hand options 33.1% and 29.2% of the time respectively, significantly different from an equal distribution (one sample Kolmogorov-Smirnov p = 0.002). Table 4 contains the results of the final preferred model, with 24 coefficients. Figure 1 illustrates the RIAI of child attributes for each system attribute and overall. The "constant" terms in Table 4 give participants' preferences for AAC system allocation when shown a child vignette with all attributes at baseline levels, which represents what was considered by the researchers as the most challenging profile that can be formed from the set of child attributes. This baseline vignette is as follows: The interaction terms represent how respondents' preferences for AAC systems change if choosing for a child vignette which differs on a given child attribute.

Vocabulary Sets
For the baseline child vignette, vocabulary sets which are fixed or have staged progression were preferred to no pre-installed vocabulary. Only a single child attribute influenced preferences: Professionals were much more likely compared to the baseline to choose systems with staged progression vocabulary sets over no preinstalled set if the child vignette was predicted progress in skills and ability (odds ratio, OR 3.88) ( Table 4).

Consistency of layout,
For the baseline child vignette, consistent layout or an idiosyncratic layout was preferred to only having some aspects of system layout consistent for use, with no interactions with child attributes (Table 4).

Vocabulary organisation
For the baseline child vignette there was no significant preference between visual scene, taxonomic or semantic-syntactic vocabulary organisation, whilst pragmatic organisation was preferred. Two significant interactions exist between vocabulary organisation and motivation. A child vignette with motivation to communicate using AAC became more likely to be allocated a system with taxonomic (OR 2.03), or semanticsyntactic (OR 2.29) organisation compared to visual scene layout ( Table 4).

Size of vocabulary
For the baseline child vignette there were no significant differences in preferences between up to 50 and between 50-1000 vocabulary items, but over 1000 items were considered significantly less appropriate. A mid-size vocabulary (50-1000 items) became more preferable compared to 50 or fewer for a child vignette motivated to communicate using AAC. Over 1000 items became significantly more preferable for child vignettes with each of the following characteristics: receptive language exceeding expressive language, an ability to use a range of AAC functions, motivated to communicate using AAC and predicted to progress  (Table 4). All child attributes influenced preferences for vocabulary size. As measured using RIAI, communication ability with AAC (32%) and determination and persistence (28%) were relatively more important than future skills and abilities (22%) and receptive and expressive language (17%) (Figure 1).

Graphic representation
For the baseline child vignette there was no preference between graphic representation using photos or pictographs, but text was less preferred than either, and idiographic symbols were even less preferred.
Interactions were seen with two child attributes. Motivation to communicate using AAC increased the probability of choosing pictographic symbols (OR 3.88), idiographic symbols (OR 5.31) or text (OR 4.00) rather than photos. However, being predicted to progress made pictographic symbols less preferable (Table   4).

DISCUSSION
This DCE has revealed AAC professionals' priorities when choosing AAC systems for children, and shown that these priorities change when faced with children with different characteristics. This is not unexpected, and in line with previous research showing that AAC professionals recognise the importance of matching an AAC system to an individual person's needs. 22 54 However, this study builds on previous findings by showing the magnitude of preference changes, as for some system attributes their preferences for different levels could completely reverse depending which child vignette was shown. For example, for the baseline child vignette (see Table 4), a system with more than 1000 vocabulary items was less likely to be chosen than one with less than 50 (OR 0.395). However, for a child vignette describing a receptive-expressive language gap, the ability to use AAC for a range of functions, motivation to use AAC and predicted progression, a system with more A key finding was that the attribute of the child's determination and persistence had the greatest number of interactions with preferences and was more important in terms of RIAI than language ability or previous experience with AAC. Specifically, the attribute level motivation to communicate using AAC tended to drive participants towards what can be regarded as more "ambitious" choices, for example more vocabulary items.
It may be that participants believed that motivated children are more likely to succeed with such AAC systems, in line with previous findings that attitude towards AAC, and valuing an AAC system are important factors in successful adoption of AAC. 22 54 Visual scene vocabulary organisation and graphic representation using photos can both involve items/scenes from an individual's own life, and use literal, rather than abstract depictions. Both were less preferred for child vignettes motivated to communicate via AAC. Rather, participants favoured more abstract methods of organisation (taxonomic and semantic-syntactic) and graphic symbols that require more grammar (pictographs, ideographs and text). This may be interpreted as an (unfounded 55 ) belief that motivated children will be better able to use more complex AAC systems. An alternative and by no means mutually exclusive interpretation is that lack of motivation requires an AAC system involving familiar cues from their everyday environment.
Previous studies have also studied how AAC professionals choose graphic symbols for children. 56 For example, Thistle and Wilkinson 33 found that cognitive abilities are an important factor, as did Dada et al. 46 The advantage of a DCE is that the precise interactions between child characteristics and symbol type have been enumerated, showing, e.g., which children were more likely to be given AAC systems with photos, and which were more likely to be given systems with text.AAC system preferences did not significantly differ between child vignettes where their skills and abilities were predicted to regress or plateau. However, if a child was predicted to progress, this had a large impact on professional decision-making, with anticipated future skills and abilities ranked as the highest attribute in terms of RIAI. As with motivation, skills and Photos were still the most preferred aided communication mode unless a child vignette featured both predicted progress and motivation to communicate via AAC. This finding possibly indicates that photos remain a good starting point for a child who is not engaged, regardless of prognosis, and may reflect recommendations that recognise the need to reduce the learning demands of AAC systems for some children. 12 60 Despite unwelcome rates of abandonment, AAC professionals had high expectations of motivated children who were expected to progress, even if their receptive and expressive language were both delayed and they had no previous AAC experience. It has previously been noted that people who use AAC experience an asymmetry between the language they receive and the language they are able to express. 61 One interpretation is that participants wished to minimise asymmetries by choosing text as the expressive output for children they believed could cope with it. These ambitious choices are also encouraging given the greatly increased aspirations for effective societal participation of AAC users. 11  and 1000 items was considered better than 50 or fewer for all child vignettes, although the difference was not always significant. This finding may indicate that participants were mindful of limiting children's potential for expression, even for children with lower cognitive ability and poor prognosis.
Findings suggest that respondents preferred levels of AAC systems that require personalisation, e.g., pragmatic vocabulary organisation or an idiosyncratic layout. This is in line with previous findings that personalisation is important in successful AAC adoption. 28 It indicates that it is not possible to achieve the Many differences to the BWS findings can also be seen. Language abilities were the most important child attribute in the BWS, yet its RIAI in the DCE was below predicted future abilities, ranked sixth in the BWS.
However, differences do not necessarily imply contradiction, as the two methodologies do not measure the same thing. The BWS measured the importance of AAC system attributes over the case mix AAC professionals encounter in practice, whereas for the DCE respondents were presented with a specific child vignette.
Receptive and expressive language had the lowest RIAI overall, with only a single interaction term in the final model. This contrasts with some previous findings that a child's language abilities play a large role in selecting an appropriate AAC System. 13 28 30 37 One possible explanation is that the aspects of language ability which were most relevant were captured in this study by other child attributes, but this remains a question to be addressed by future research.
This study has demonstrated the feasibility of conducting a DCE with a target population of AAC professionals. This is noteworthy given the relative rarity of DCEs studying health professionals' decision-making. For with AAC professionals. The target population in the UK is small, meaning it was uncertain that sufficient participants for a successful study could be recruited. There were also concerns that participants might not find the DCE format acceptable, as they might reject having to make compromises between AAC system attributes in the context of providing a system for a child. Yet despite informal feedback that some respondents found the tasks uncomfortable, many were still willing to complete them. Finally, as interactions between child characteristics and AAC systems are so important, it was necessary to present hypothetical child vignettes, making tasks more complicated than in a typical DCE.
Despite these potential pitfalls, the DCE was successfully carried out, and having demonstrated the feasibility of the method in this area, further DCE studies should be considered in future.

Limitations
This study has several limitations. The sample size was relatively small (155 participants, compared to a median for healthcare DCEs of 401 41 ). However, many studies exist with smaller sample sizes (e.g., Spinks et al. 66 with 35), and it was possible to estimate robust statistical models. Furthermore, it would have been difficult to collect a larger sample, as 155 participants represents a large proportion of the population of AAC professionals in the UK working with children, which is estimated at around 800. i The DCE task may not match how UK AAC professionals make decisions in practice. Typically, many participants have the opportunity to work with families and children, as well as part of an AAC team, which could include diverse areas of clinical and personal expertise. Teams also generally make recommendations, rather than unilaterally choosing a system. However, there is evidence that AAC professionals compare the attributes of AAC systems in everyday practice, 13 and that they make trade-offs between system attributes, 37 akin to DCE tasks. In addition, it is still useful to study the individual decision-making of AAC professionals.
i Communication Matters, personal correspondence The DCE tasks presented one-off static decisions made by a single individual. In reality the decision-making environment is dynamic, with children developing over time, and often having two or more devices over the course of their childhood. These differences are a limiting factor in the external validity of results.
Attributes and levels use a mixture of speech and language therapy terms (e.g., receptive and expressive language) and more AAC specific language (e.g., staged vocabulary progression). This may have made it more difficult for respondents from any one professional speciality to interpret all of them. 44 However, this issue is not limited to the current study, but reflects an ongoing struggle in AAC to establish a common language. 44 In addition, respondents may have been unfamiliar with the generic term ideographic symbols, since only a single commercial set of ideographic symbols is in popular use in the UK (Minspeak i ).
Respondents were more likely to choose AAC systems on the left of the screen and less likely to choose ones on the right. However, the risk of bias was mitigated by allowing for alternative specific constants and randomising the position in which AAC systems were presented.
Compared to the real children AAC professionals encounter, the child vignettes were simple, and lacked information which influences decision-making, such as the child's preferences 33 and contextual factors. 30 However, this is an inherent limitation of the DCE methodology, and vignettes with a greater number of attributes and levels would have made decisions overly burdensome, and therefore were not included.
Significant interactions between AAC systems and child attributes implied that the vignettes were meaningful enough that respondents changed their preferences in response to them, often dramatically.
For a given child vignette, it was only possible to determine relative preferences for system attributes, rather than absolute preferences. Consequently, it is not possible to tell how suitable a given system is for a given i © Semantic Compaction Systems, Inc. child vignette, which is important as some presented a challenging profile for which it may be hard to find a suitable AAC system.

CONCLUSION
A lack of rigorous evidence on how to best assess and provide AAC systems for children has previously been identified, 25 34 44 as well as a gap between research and current practice. 11 In the light of this, the current study's results are encouraging, as it shows AAC professionals following best practice in many areas, for example ensuring AAC systems suit individual needs, and having high expectations for many children.
However, there is still demand from AAC professionals for better support in decision-making, 33 37 and undoubtedly current practice could be improved. The results of this study, together with evidence from the wider research project, have been used to create a heuristic and suite of resources, available at https://iasc.mmu.ac.uk. It is hoped these resources will aid AAC professionals in their clinical practice and help them provide the best possible service for children. Stephane Hess acknowledges additional support by the European Research Council through the consolidator grant 615596-DECISIONS.         A series of vocabulary sets with pre-determined progression through them that simulate language development. E.g. an initial set including just basic words, with subsequent sets introducing more grammatical structure. May be customised.

Consistency of layout (2)
How consistent positions of words/symbols are in system interface, and how consistent navigation to find different symbols is *Consistency of some aspects of layout Words/symbols in multiple categories appear in different positions across categories, but always in the same place in a given category Consistency of all aspects of layout All/nearly all words/symbols always appear in same position in interface Idiosyncratic layout Layout that has been personalised for an individual child Type of vocabulary organisation (5) How words/symbols are organised within the system *Visual scene Interface shows photos, most likely of scenes familiar to the child, with areas of it highlighted to represent words Taxonomic Words/symbols organised according to subject, analogous to nonfiction books in a library Semantic-syntactic Words/symbols organised according to sentence structure, e.g. verbs, nouns, adjectives Pragmatic Words/symbols organised around function in language rather than grammar, e.g. request, mood Size of vocabulary (7) How many words/symbols system can output *Up to 50 vocabulary items Implies only simple communication functions possible 50-1000 vocabulary items Implies combining words/symbols to create grammatical structures More than 1000 vocabulary items Does not imply more complex communication than 50-1000 items, but means a greater load on child's memory Graphic representation (12) Type of symbols used by system *Photos Photographs, possibly of items or environments personal to the child Pictographic symbol set Non-photorealist pictures with specific meanings attached. May be accompanied by text Ideographic symbol system (with rules or encoding) Stylised symbols combined with fixed rules and grammar analogous to Chinese/Japanese characters, e.g. Minspeak Text Text unaccompanied by other symbols Note: * indicates baseline level; numbers in parentheses indicate attributes' rank in relative importance from prior BWS study (reported in Webb et al. 45 ). †Descriptions are not intended as rigorous definitions of AAC terminology, but as a rough guide for the non-AAC specialist reader.   Appendix A -Attributes from best-worst scaling case 1 study  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59    Imagine you had to choose between only these three systems. You should indicate which you would prescribe for the child described. If your preferred option is not available, pick the system from the three options that you think best matches the child's needs. There are no right or wrong answers. It is acknowledged that this may feel uncomfortable for you.
In the survey, there are three different children described. You will be asked four questions about each child (12 questions in total).
In acknowledgement of choices being uncomfortable, after each choice, you will be asked to indicate how well you think that system matches the child's needs.  Your participation in this survey is voluntary. All information is collected anonymously and held in confidence. We hope you complete the survey but you are free to stop responding at any point resulting in your answers will be removed.

Consent
 I have read and understood the above and consent to taking part. I confirm my work involves assessing children for aided AAC systems and I contribute to the decision making in relation to the language and vocabulary organisation within aided AAC systems.

Yes  No
If yes go to DCE questions.
If no go to a page with the following:-Thank you for your interest in this survey. At present we are only recruiting participants who contribute to decision making in relation to the language and vocabulary organisation within aided AAC for children.
Over the coming 12 months we will be recruiting people with a wider range of AAC experience to test decision making resources we are developing. If you are interested in this aspect of the project or would like to be notified when the free resources are available, there will be an opportunity at the end to submit your email address.
We would still like to ask you a few questions about your experience with AAC to check the representativeness of participants.
Then go directly to demographics questionnaire. Child A has delayed expressive and receptive language and is able to use aided AAC for a few communicative functions. Child A is motivated to communicate through symbol communication systems. Child A is predicted to regress in skills and abilities (regression).

Size of vocabulary
The size of the output vocabulary available within the aided AAC system. Child A has delayed expressive and receptive language and is able to use aided AAC for a few communicative functions. Child A is motivated to communicate through symbol communication systems. Child A is predicted to regress in skills and abilities (regression).

Size of vocabulary
The size of the output vocabulary available within the aided AAC system. Child C has delayed expressive and receptive language and no previous AAC experience. Child C is only motivated to communicate through methods other than symbol communication systems. Child C is predicted to maintain current skills and abilities (plateau).

Size of vocabulary
The size of the output vocabulary available within the aided AAC system. Child C has delayed expressive and receptive language and no previous AAC experience. Child C is only motivated to communicate through methods other than symbol communication systems. Child C is predicted to maintain current skills and abilities (plateau).

Size of vocabulary
The size of the output vocabulary available within the aided AAC system. Child C has delayed expressive and receptive language and no previous AAC experience. Child C is only motivated to communicate through methods other than symbol communication systems. Child C is predicted to maintain current skills and abilities (plateau).

Size of vocabulary
The size of the output vocabulary available within the aided AAC system.

End of survey
Thank you for your participation in this survey.
Your responses will contribute to the results of the I-ASC project and support the development of decision making resources for use in AAC assessments.
You can follow the progress of our research project on our website, on Facebook or on Twitter.  Design: AAC professionals answered a discrete choice experiment (DCE) survey with AAC system and childrelated attributes, where participants chose an AAC system for a child vignette.

Setting:
The survey was administered online in the UK.

Outcomes:
The study outcomes were AAC professionals' preferences as quantified using a mixed logit model, with model selection performed using a stepwise procedure and the Bayesian Information Criterion.
Results: Significant differences were observed in preferences for AAC system attributes, and large interactions were seen between child attributes included in the child vignettes, e.g., participants made more ambitious choices for children who were motivated to communicate using AAC, and predicted to progress in skills and abilities. These characteristics were perceived as relatively more important than language ability and previous AAC experience.
Conclusions: AAC professionals make trade-offs between attributes of AAC systems, and these trade-offs change depending on the characteristics of the child for whom the system is being provided.

STRENGTHS AND LIMITATIONS OF THIS STUDY
 This was the first discrete choice experiment (DCE), and only the second stated preference study in the field of augmentative and alternative communication (AAC).
 The study used unusual and innovative methodology by (1) using a Best-worst Scaling case 1 study in attribute selection; (2) having AAC system choices be made in the context of a child vignette  In some ways, the DCE task differed from how augmentative and alternative professionals make decisions in practice.

INTRODUCTION
Many people lack the ability to produce intelligible speech to meet their functional needs for a wide range of reasons, including cerebral palsy, intellectual/developmental delays and autism spectrum disorder. Even Major advances in the AAC landscape have occurred in recent years. 11 12 These include technological innovation, for example iPads and eye-tracking, though low-tech systems may still offer the best solution in many cases. 13 14 Another development within services is a greater expectation of participation in all aspects of life for people who use AAC, 11 15-17 coupled with advocacy for the right to communicate. 14  Although there is a lack of robust evidence surrounding the decision-making process, some factors in successful adoption of AAC have been identified. An AAC system is more likely to be adopted by a motivated child 22 with good support from the child's network. 27 33 44 The AAC system must also meet a child's individual needs and circumstances, which will be unique to every child. 14 22 46 A previous study from the current research project investigated the AAC decision-making process using a Best-worst Scaling (BWS) case 1 survey. 45 This method was chosen as it could quantify which of several child and AAC system related factors (37 in total) AAC professionals considered most and least important in decision-making.
The current study sought to complement the previous work by examining fewer factors in more detail using a DCE. 43 It aimed to quantify the clinical judgements and trade-offs AAC professionals make between different attributes of AAC systems, and how those trade-offs change depending on children's characteristics, things not possible using BWS case 1. This is the first DCE carried out in AAC, and there were challenges associated with performing a DCE with a target population of AAC professionals (for details see discussion). Thus, an additional goal was to establish the feasibility of using DCEs as a research tool in AAC. The BWS study produced relative importance scores for 19 child and 18 system attributes given in Appendix A. DCE attributes were selected from these during consensus discussions between authors with expertise in AAC, speech and language therapy, and health economics. The selection criteria were that attributes should: Mirenda. 17 ) In summary, the child attributes captured a child's language ability, experience with AAC, attitude/motivation to communicate with AAC, and whether the child is expected to regress, plateau or progress in communication ability. A total of 54 child vignettes were formed from the set of child attributes. Authors with expertise in AAC and speech and language therapy identified and removed 18 child vignettes representing unrealistic combinations, leaving 36.

Survey development
AAC system attributes broadly captured the vocabulary set(s) provided by manufacturers, vocabulary size and organisation, type of graphic symbols used, and how consistent the navigational layout of words/symbols is when accessing items. It was not stated whether a system was high-tech or low-tech, although certain levels, e.g., vocabulary sets with staged progression, are more common with high-tech systems. Authors with experience in AAC and speech and language therapy removed 158 unrealistic combinations from the 432 AAC systems which could be formed from the system attributes, leaving 274.
Prior experience from the BWS study suggested it would be difficult to recruit a large respondent sample, so to maximise the information captured a relatively heavy response burden of 12 choices between three systems was selected for the DCE. Participants were shown three child vignettes, referred to as Child A, B and C, and made four choices for each child vignette. An example task is shown in Appendix B.
The survey's statistical design (i.e., which levels of the AAC system attributes were presented in each question) was generated using NGene i , with 60 choice tasks split into five blocks. The design sought to maximise Defficiency, a measure of how much information it is possible to extract from survey responses. 49 The survey was piloted by five AAC professionals and consequently the wording of some attributes and levels altered.

Survey administration
The DCE was administered online for ease of recruitment. Recruitment was carried out via AAC professionals' email distribution lists (the project's own list and the mailing list of the UK wide charity Communication Matters ii ). In addition, invitations were sent via publicly available lists and websites, and the professional contacts of authors. Adverts were also placed on the project website and online media. Responses were collected between 20/10/17 and 4/3/18. Ethical approval was received from an NHS Research Ethics Committee (REC reference 6/NW/0165) and informed consent obtained from participants at the start of the survey.
Participants began by confirming they contributed towards AAC decision-making for children, and those who indicated they did not progressed directly to demographic questions that were at the end of the survey (for i ©ChoiceMetrics ii www.communicationmatters.org.uk details, see Appendix A). i Three child vignettes and one survey block were randomly allocated to each participant. The order of system attributes was randomised between participants, but consistent within participants, and which systems appeared on the left, middle and right of the screen was also randomised.

Analysis
Analysis of participants' choices was grounded in random utility theory. This standard approach 50 assumed participants chose the object which maximised their utility. The utility of an object was modelled as depending partly on the object's attributes and partly random, the latter component capturing the influence of all factors not included in the model. In a given choice scenario , participant chose which of three AAC systems to allocate to child . The utility to participant of allocating AAC system to child in choice ∈ {1,2,3} scenario was

= + +
where was an alternative specific constant for AAC system , was a vector of dummy variables indicating AAC system levels, was a vector of coefficients which differ across participants and children, and was a random error term.
The coefficient on level of system attribute , , depended on the characteristics of the child vignette according to = 0 + i The precise wording of the question was: "I confirm my work involves assessing children for aided AAC systems and I contribute to the decision making in relation to the language and vocabulary organisation within AAC systems." During testing it was revealed that some AAC professionals did not have sufficient input into the decision making process in their day-to-day practice for the DCE questions to be meaningful, e.g., occupational therapists specialising in optimising physical access to an AAC system recommended by other members of the team, and this question was designed to filter out such respondents.  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  60   F  o  r  p  e  e  r  r  e  v  i  e  w  o  n  l  y where was a constant giving the preference for a system attribute at baseline child levels, was a vector 0 of dummy variables indicating vignette levels and was a vector of coefficients, allowing for heterogeneity in relative preference for AAC system attributes depending on child characteristics.
A full model with all interaction terms included too many parameters to estimate reliably. Thus, parameters were eliminated in a stepwise process and a final preferred mixed logit model was identified using the Bayesian Information Criterion (BIC). The mixed logit model incorporated participant heterogeneity by allowing AAC system attribute parameters to be random, following a normal distribution with both means and variances depending on child characteristics. For details, see Appendix C.
Models were estimated using the CMC Choice Modelling Centre Code for R version 1.1 51 and all analysis was carried out using R version 3.3.1. Statistical significance was assessed at the 5% level after adjusting for multiple testing using Holm's sequential Bonferroni correction. 52 Results were presented using a new measure termed relative interaction attribute importance (RIAI) which assessed how big an impact child attributes have on AAC professionals' decision-making. RIAI is analogous to relative attribute importance, often used to present DCE results, 53 and may be calculated either with respect to a single choice object attribute or overall with respect to all choice object attributes. For a formal definition of RIAI, see Appendix D.

Patient and public involvement
One author (SM) is an AAC user, and one (LM) is the parent of an AAC user, and both were involved in all stages of research development and delivery.

RESULTS
A total of 172 participants completed the survey, of whom 155 indicated they contributed to decision-making regarding AAC systems and answered DCE questions. Summary statistics of their demographics and professional experience are given in Table 3. Most participants were female (~90%) and white. We believe  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  60   F  o  r  p  e  e  r  r  e  v  i  e  w  o  n l y this to be reasonably representative of the population of AAC professionals in the UK. i The mean age of DCE participants was around 40, with a range from 24 to 65, and they had on average 10 years' experience of AAC.
Around 75% of DCE participants had a speech and language therapy background, with no other background reported by more than 10%. Those who did not answer DCE questions were less likely to have a speech and language therapy background (~50%), with teacher (~20%) and occupational therapist (~30%) more common.
Approximately 30% of the sample worked with all age groups, while 50-60% worked with pre-school, primary school and secondary school aged children. Participants were asked for the three most common diagnoses encountered in their work, with ~80% stating physical disability, 70% stating intellectual disability/developmental delay and 65% stating autism spectrum disorder.
Turning to DCE responses, respondents chose the left-hand option 37.6% of the time, and the central and right-hand options 33.1% and 29.2% of the time respectively, significantly different from an equal distribution (one sample Kolmogorov-Smirnov p = 0.002). Table 4 contains the results of the final preferred model, with 24 coefficients. Figure 1 illustrates the RIAI of child attributes for each system attribute and overall. The constant terms in Table 4 give participants' preferences for AAC system allocation when shown a child vignette with all attributes at baseline levels. This baseline vignette is as follows: "Child A/B/C has delayed expressive and receptive language and no previous AAC experience. Child A/B/C does not appear motivated to communicate through any methods and means.

Child A/B/C is predicted to regress in skills and abilities (regression)." It represents what was considered by
the researchers as the most challenging profile that can be formed from the set of child attributes.
The interaction terms represent how respondents' preferences for AAC systems changed if choosing for a child vignette which differed on a given child attribute.

Vocabulary sets
i E.g., data from the Health and Care Professionals Council showed speech and language therapists in the UK were 96% female and the Higher Education Statistics Agency found speech and language therapy students in 2017/18 were 79% white. Source: Royal College of Speech and Language Therapists, personal communication. For the baseline child vignette, vocabulary sets which are fixed or have staged progression were preferred to no pre-installed vocabulary. Only a single child attribute influenced preferences: Professionals were much more likely compared to the baseline to choose systems with staged progression vocabulary sets over no preinstalled set if the child vignette was predicted progress in skills and ability (odds ratio, OR 3.88) ( Table 4).

Consistency of layout
For the baseline child vignette, consistent layout or an idiosyncratic layout was preferred to only having some aspects of system layout consistent for use, with no interactions with child attributes (Table 4).

Vocabulary organisation
For the baseline child vignette there was no significant preference between visual scene, taxonomic or semantic-syntactic vocabulary organisation, whilst pragmatic organisation was preferred. There were two significant interactions between vocabulary organisation and motivation. A child vignette with motivation to communicate using AAC became more likely to be allocated a system with taxonomic (OR 2.03), or semanticsyntactic (OR 2.29) organisation compared to visual scene layout ( Table 4).

Size of vocabulary
For the baseline child vignette there were no significant differences in preferences between up to 50 and between 50-1000 vocabulary items, but over 1000 items were considered significantly less appropriate. A mid-size vocabulary (50-1000 items) became more preferable compared to 50 or fewer for a child vignette motivated to communicate using AAC. Over 1000 items became significantly more preferable for child vignettes with each of the following characteristics: receptive language exceeding expressive language, an ability to use a range of AAC functions, motivated to communicate using AAC and predicted to progress (Table 4). All child attributes influenced preferences for vocabulary size. As measured using RIAI, communication ability with AAC (32%) and determination and persistence (28%) were relatively more important than future skills and abilities (22%) and receptive and expressive language (17%) (Figure 1).  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  60   F  o  r  p  e  e  r  r  e  v  i  e  w  o  n  l  y For the baseline child vignette there was no preference between graphic representation using photos or pictographs, but text was less preferred than either, and idiographic symbols were even less preferred.

Graphic representation
Interactions were seen with two child attributes. Motivation to communicate using AAC increased the probability of choosing pictographic symbols (OR 3.88), idiographic symbols (OR 5.31), or text (OR 4.00) rather than photos. However, being predicted to progress made pictographic symbols less preferable (Table   4).

DISCUSSION
This DCE has revealed AAC professionals' priorities when choosing AAC systems for children, and shown that these priorities change when faced with children with different characteristics. That priorities change in this way is not unexpected, and in line with previous research showing that AAC professionals recognise the importance of matching an AAC system to an individual person's needs. 22 54 However, the current study builds on previous findings by showing the magnitude of preference changes, as for some system attributes their preferences for different levels could completely reverse depending which child vignette was shown. For example, for the baseline child vignette (see Table 4), a system with more than 1000 vocabulary items was less likely to be chosen than one with less than 50 (OR 0.395). However, for a child vignette describing a receptive-expressive language gap, the ability to use AAC for a range of functions, motivation to use AAC and predicted progression, a system with more than 1000 vocabulary items was more likely to be chosen (OR   1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  A key finding was that the attribute of the child's determination and persistence had the greatest number of interactions with preferences and was more important in terms of RIAI than language ability or previous experience with AAC. Specifically, the attribute level motivation to communicate using AAC tended to drive participants towards what can be regarded as more ambitious choices, for example more vocabulary items. It may be that participants believed that motivated children are more likely to succeed with such AAC systems, in line with previous findings that attitude towards AAC, and valuing an AAC system are important factors in successful adoption of AAC. 22 54 Visual scene vocabulary organisation and graphic representation using photos can both involve items/scenes from an individual's own life, and use literal, rather than abstract depictions. Both were less preferred for child vignettes motivated to communicate via AAC. Rather, participants favoured more abstract methods of organisation (taxonomic and semantic-syntactic) and graphic symbols that require more grammar (pictographs, ideographs and text). Preferences for abstract methods of organisation and symbols requiring more grammar may be interpreted as an (unfounded 55 ) belief that motivated children will be better able to use more complex AAC systems. An alternative and by no means mutually exclusive interpretation is that lack of motivation requires an AAC system involving familiar cues from their everyday environment.
Previous studies have also studied how AAC professionals choose graphic symbols for children. 56 For example, Thistle and Wilkinson 33 found that cognitive abilities are an important factor, as did Dada et al. 46 The advantage of a DCE was that the precise interactions between child characteristics and symbol type have been enumerated, showing, e.g., which children were more likely to be given AAC systems with photos, and which were more likely to be given systems with text. AAC system preferences did not significantly differ between child vignettes where their skills and abilities were predicted to regress or plateau. However, if a child was predicted to progress, this had a large impact on professional decision-making, with anticipated future skills and abilities ranked as the highest attribute in terms of RIAI. As with motivation, skills and abilities led to more ambitious choices, with more vocabulary items preferred and pictographs depreciated compared to ideographs and text. Such ambitious choices could reflect participants wishing to provide AAC systems that would fulfil the future needs of children who are anticipated to progress, given the large investment that goes into learning a new AAC system. 57-59 With plateau or regression this was less of a concern.
Photos were still the most preferred aided communication mode unless a child vignette featured both predicted progress and motivation to communicate via AAC. This preference for photos possibly indicates that photos remain a good starting point for a child who is not engaged, regardless of prognosis, and may reflect recommendations that recognise the need to reduce the learning demands of AAC systems for some children. 12

60
Despite unwelcome rates of abandonment, AAC professionals had high expectations of motivated children who were expected to progress, even if their receptive and expressive language were both delayed and they had no previous AAC experience. It has previously been noted that people who use AAC experience an asymmetry between the language they receive and the language they are able to express. 61 One interpretation is that participants wished to minimise asymmetries by choosing text as the expressive output for children they believed could cope with it. These ambitious choices are also encouraging given the greatly increased aspirations for effective societal participation of AAC users. 11  and 1000 items was considered better than 50 or fewer for all child vignettes, although the difference was not always significant. Systems with fewer than 50 items being depreciated may indicate that participants were mindful of limiting children's potential for expression, even for children with lower cognitive ability and poor prognosis.
Findings suggest that respondents preferred levels of AAC systems that require personalisation, e.g., pragmatic vocabulary organisation or an idiosyncratic layout, in line with previous findings that personalisation is important in successful AAC adoption. 28 A preference for personalisation indicates that it is not possible to achieve the goal of AAC systems being closely tailored to individuals' needs 14 63 with "offthe-shelf" AAC systems: in other words, some personalisation is always necessary. 64 65 Pre-installed vocabulary sets were always preferred over no pre-provided set, in line with other studies showing that selecting core vocabulary was an important part of AAC professionals' decision-making process. 33 37 Comparing the DCE results with the previous BWS Case 1 study, 45 some similarities may be observed. For example, graphic representation was the lowest ranked attribute in terms of importance in the BWS to be included in the DCE. In concordance with this finding, when the relative importance of AAC system attributes was calculated for each child vignette in the DCE, graphic representation was never the most important attribute. The relative lack of importance ascribed to graphic representation raises debate about the fundamental components of language construction through aided means and suggests much further research is required.
Many differences to the BWS findings can also be seen. Language abilities were the most important child attribute in the BWS, yet its RIAI in the DCE was below predicted future abilities, ranked sixth in the BWS.
However, differences do not necessarily imply contradiction, as the two methodologies did not measure the same thing. The BWS measured the importance of AAC system attributes over the case mix AAC professionals encounter in practice, whereas for the DCE respondents were presented with a specific child vignette.
Receptive and expressive language had the lowest RIAI overall, with only a single interaction term in the final model. This contrasts with some previous findings that a child's language abilities play a large role in selecting an appropriate AAC System. 13 28 30 37 One possible explanation is that the aspects of language ability which were most relevant were captured in this study by other child attributes, but this remains a question to be addressed by future research.  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  60   F  o  r  p  e  e  r  r  e  v  i  e  w  o  n  l  y The current study has demonstrated the feasibility of conducting a DCE with a target population of AAC professionals. The demonstration of feasibility is noteworthy given the relative rarity of DCEs studying health professionals' decision-making. For example in a systematic review 41 of DCEs in health published between 2013 and 2017, only 13% included a sample of health professionals. In addition, there were particular challenges associated with performing a DCE with AAC professionals. The target population in the UK is small, meaning it was uncertain that sufficient participants for a successful study could be recruited. There were also concerns that participants might not find the DCE format acceptable, as they might have rejected having to make compromises between AAC system attributes in the context of providing a system for a child.
Yet despite informal feedback that some respondents found the tasks uncomfortable, many were still willing to complete them. Finally, as interactions between child characteristics and AAC systems are so important, it was necessary to present hypothetical child vignettes, making tasks more complicated than in a typical DCE.
Despite these potential pitfalls, the DCE was successfully carried out, and having demonstrated the feasibility of the method in this area, further DCE studies should be considered in future.

Limitations
The current study has several limitations. The sample size was relatively small (155 participants, compared to a median for healthcare DCEs of 401 41 ). However, many studies exist with smaller sample sizes (e.g., Spinks et al. 66 with 35), and it was possible to estimate robust statistical models. Furthermore, it would have been difficult to collect a larger sample, as 155 participants represented a large proportion of the population of AAC professionals in the UK working with children, which was estimated at around 800. i The DCE task may not match how UK AAC professionals make decisions in practice. Typically, many participants have the opportunity to work with families and children, as well as part of an AAC team, which could include diverse areas of clinical and personal expertise. Teams also generally make recommendations, rather than unilaterally choosing a system. However, there is evidence that AAC professionals compare the i Communication Matters, personal correspondence attributes of AAC systems in everyday practice, 13 and that they make trade-offs between system attributes, 37 akin to DCE tasks. In addition, it is still useful to study the individual decision-making of AAC professionals.
Lynch et al. 30 reported that a wide variety of team structures are used, and the mode of service delivery can have an influence on outcomes. Gathering evidence on individual-level decision-making can thus inform an assessment of how different ways of organising services influence decisions.
The DCE tasks presented one-off static decisions made by a single individual. In reality the decision-making environment is dynamic, with children developing over time, and often having two or more devices over the course of their childhood. These differences are a limiting factor in the external validity of results.
Attributes and levels use a mixture of speech and language therapy terms, e.g., receptive and expressive language, and more AAC specific language, e.g., staged vocabulary progression. Mixing these terms may have made it more difficult for respondents from any one professional speciality to interpret all of them. 44 However, this issue is not limited to the current study, but reflects an ongoing struggle in AAC to establish a common language. 44 In addition, respondents may have been unfamiliar with the generic term ideographic symbols, since only a single commercial set of ideographic symbols is in popular use in the UK (Minspeak i ).
Respondents were more likely to choose AAC systems on the left of the screen and less likely to choose ones on the right. However, the risk of bias was mitigated by allowing for alternative specific constants and randomising the position in which AAC systems were presented.
Compared to the real children AAC professionals encounter, the child vignettes were simple, and lacked information which influences decision-making, such as the child's preferences 33 and contextual factors. 30 However, this is an inherent limitation of the DCE methodology, and vignettes with a greater number of attributes and levels would have made decisions overly burdensome, and therefore were not included.
Significant interactions between AAC systems and child attributes implied that the vignettes were meaningful enough that respondents changed their preferences in response to them, often dramatically.
i © Semantic Compaction Systems, Inc. For a given child vignette, it was only possible to determine relative preferences for system attributes, rather than absolute preferences. Consequently, it is not possible to tell how suitable a given system is for a given child vignette, which is important as some presented a challenging profile for which it may be hard to find a suitable AAC system.

CONCLUSION
A lack of rigorous evidence on how to best assess and provide AAC systems for children has previously been identified, 25 34 44 as well as a gap between research and current practice. 11 In the light of this, the current study's results are encouraging, as it shows AAC professionals following best practice in many areas, for example ensuring AAC systems suit individual needs, and having high expectations for many children.
However, there is still demand from AAC professionals for better support in decision-making, 33 37 and undoubtedly current practice could be improved. The results of the current study, together with evidence from the wider research project, have been used to create a heuristic and suite of resources, available at https://iasc.mmu.ac.uk. It is hoped these resources will aid AAC professionals in their clinical practice and help them provide the best possible service for children.      A series of vocabulary sets with pre-determined progression through them that simulate language development. E.g. an initial set including just basic words, with subsequent sets introducing more grammatical structure. May be customised.
The three AAC systems are described in terms of five characteristics (the systems are identical apart from changes to these five characteristics):-1. Vocabulary sets: Pre-determined vocabulary or language package provided, which can be:-  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  60   F  o  r  p  e  e  r  r  e  v  i  e  w  o  n  l  y 4. Graphic Representation: Primary type of graphic symbol used, which can be:- Photo symbols (i.e. a photo symbol set without rules or encoding)  Pictographic symbols (i.e. a graphic symbol set without rules or encoding)  Ideographic symbols (i.e. a symbol system with rules or encoding)  Graphic symbols with text (i.e. a system with either pictographic or ideographic symbols that incorporates an alphabet for generating text) 5. Consistency of layout: Consistency of layout of symbols on pages, including when navigating through pages to select desired output, which can be:- Inconsistent layout  Somewhat consistent layout  Highly consistent layout Imagine you had to choose between only these three systems. You should indicate which you would prescribe for the child described. If your preferred option is not available, pick the system from the three options that you think best matches the child's needs. There are no right or wrong answers. It is acknowledged that this may feel uncomfortable for you.
In the survey, there are three different children described. You will be asked four questions about each child (12 questions in total).
In acknowledgement of choices being uncomfortable, after each choice, you will be asked to indicate how well you think that system matches the child's needs. ( 1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  60   F  o  r  p  e  e  r  r  e  v  i  e  w  o  n  l  y Your participation in this survey is voluntary. All information is collected anonymously and held in confidence. We hope you complete the survey but you are free to stop responding at any point resulting in your answers will be removed.
 I have read and understood the above and consent to taking part. I confirm my work involves assessing children for aided AAC systems and I contribute to the decision making in relation to the language and vocabulary organisation within aided AAC systems.

Yes  No
If yes go to DCE questions.
If no go to a page with the following:-Thank you for your interest in this survey. At present we are only recruiting participants who contribute to decision making in relation to the language and vocabulary organisation within aided AAC for children.
Over the coming 12 months we will be recruiting people with a wider range of AAC experience to test decision making resources we are developing. If you are interested in this aspect of the project or would like to be notified when the free resources are available, there will be an opportunity at the end to submit your email address.
We would still like to ask you a few questions about your experience with AAC to check the representativeness of participants.

Size of vocabulary
The size of the output vocabulary available within the aided AAC system. Child C has delayed expressive and receptive language and no previous AAC experience. Child C is only motivated to communicate through methods other than symbol communication systems. Child C is predicted to maintain current skills and abilities (plateau).

Size of vocabulary
The size of the output vocabulary available within the aided AAC system.

End of survey
Thank you for your participation in this survey.
Your responses will contribute to the results of the I-ASC project and support the development of decision making resources for use in AAC assessments.
You can follow the progress of our research project on our website, on Facebook or on Twitter. Appendix C -Final preferred model selection process A full model with all interaction terms and two alternative specific constants implies 98 parameters, which is too many to reliably estimate given the amount of data collected and given that many interactions are expected to be of very low magnitude. Thus, a strategy was required to identify a suitable model with fewer parameters.
The first stage was estimating a series of stepwise multinomial logit (MNL) models, beginning with a model with all 98 parameters. The parameter with the highest p-value, excluding the constant terms, was eliminated, and a model with 97 parameters was estimated. Then the parameter with the lowest p-value was excluded and a new model run, and so on in an iterative process until only the 12 constant terms remained (one for each non-baseline system level).
The Bayesian Information Criterion (BIC) was used to select the preferred MNL model. This model was then re-estimated as a mixed logit (MIXL) model to account for participant heterogeneity. (The process did not begin by estimating a series of stepwise MIXL models due to the difficulty and greatly increased computational resources required to estimate MIXL models with a large number of parameters.) The \coefficients on system attribute levels were assumed to be drawn from normal distributions with means given by ̅ = + and variances given by If p is the number of parameters of the preferred MNL model, then models with between p -3 and p + 3 parameters were re-estimated as MIXL models. The BIC for each MIXL model is given in Error! Reference source not found..
The MIXL model minimising the BIC was chosen as the final preferred model.

Study design 4
Present key elements of study design early in the paper Pages 6-7    Cohort study-Report numbers of outcome events or summary measures over time Table 4, Figure 1 Case-control study-Report numbers in each exposure category, or summary measures of exposure Generalisability 21 Discuss the generalisability (external validity) of the study results Pages 16-17 The DCE task may not match how UK AAC professionals make decisions in practice. Typically, many participants have the opportunity to work with families and children, as well as part of an AAC team, which could include diverse areas of clinical and personal expertise. Teams also generally make recommendations, rather than unilaterally choosing a system. However, there is evidence that AAC professionals compare the attributes of AAC systems in everyday practice, 13 and that they make trade-offs between system attributes, 37 akin to DCE tasks. In addition, it is still useful to study the individual decision-making of AAC professionals. Lynch et al. 30 reported that a wide variety of team structures are used, and the mode of service delivery can have an influence on outcomes. Gathering evidence on individual-level decision-making can thus inform an assessment of how different ways of organising services influence decisions. The DCE tasks presented one-off static decisions made by a single individual. In reality the decisionmaking environment is dynamic,  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46 F o r p e e r r e v i e w o n l y 7 with children developing over time, and often having two or more devices over the course of their childhood. These differences are a limiting factor in the external validity of results.

Other information
Funding 22 Give the source of funding and the role of the funders for the present study and, if applicable, for the original study on which the present article is based Page 18 *Give information separately for cases and controls in case-control studies and, if applicable, for exposed and unexposed groups in cohort and cross-sectional studies.  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46