The ultimate goals of auditory habilitation programs for children

Transcription

1 Relationships Among Speech Perception, Production, Language, Hearing Loss, and Age in Children With Impaired Hearing Peter J. Blamey Department of Otolaryngology University of Melbourne Australia Julia Z. Sarant Bionic Ear Institute Melbourne, Australia Louise E. Paatsch Johanna G. Barry Catherine P. Bow Department of Otolaryngology University of Melbourne Australia Roger J. Wales Humanities & Social Sciences Faculty La Trobe University Melbourne, Australia Maree Wright Colleen Psarros Kylie Rattigan Rebecca Tooher Children s Cochlear Implant Centre Sydney, Australia Eighty-seven primary-school children with impaired hearing were evaluated using speech perception, production, and language measures over a 3-year period. Forty-seven children with a mean unaided pure-tone-average hearing loss of 106 db HL used a 22-electrode cochlear implant, and 40 with a mean unaided puretone-average hearing loss of 78 db HL were fitted with hearing aids. All children were enrolled in oral/aural habilitation programs, and most attended integrated classes with normally hearing children for part of the time at school. Multiple linear regression was used to describe the relationships among the speech perception, production, and language measures, and the trends over time. Little difference in the level of performance and trends was found for the two groups of children, so the perceptual effect of the implant is equivalent, on average, to an improvement of about 28 db in hearing thresholds. Scores on the Peabody Picture Vocabulary Test (PPVT) and the Clinical Evaluation of Language Fundamentals showed an upward trend at about 60% of the rate for normally hearing children. Rates of improvement for individual children were not correlated significantly with degree of hearing loss. The children showed a wide scatter about the average speech production score of 40% of words correctly produced in spontaneous conversations, with no significant upward trend with age. Scores on the open-set Consonant-Nucleus-Consonant (CNC) monosyllabic word test and the Bench-Kowal-Bamford (BKB) sentence test were strongly related to language level (as measured by an equivalent age on the PPVT) and speech production scores for both auditory-visual and auditory test conditions. After allowing for differences in language, speech perception scores in the auditory test condition showed a slight downward trend over time, which is consistent with the known biological effects of hearing loss on the auditory periphery and brainstem. Speech perception scores in the auditory condition also decreased significantly by about 5% for every 10 db of hearing loss in the hearing aid group. The regression analysis model allows separation of the effects of language, speech production, and hearing levels on speech perception scores so that the effects of habilitation and training in these areas can be observed and/or predicted. The model suggests that most of the children in the study will reach a level of over 90% sentence recognition in the auditory-visual condition when their language becomes equivalent to that of a normally hearing 7-year-old, but they will enter secondary school at age 12 with an average language delay of about 4 or 5 years unless they receive concentrated and effective language training. KEY WORDS: speech, language, hearing-impaired, cochlear prosthesis, hearing aid, children The ultimate goals of auditory habilitation programs for children with impaired hearing include the development of age-appropriate language skills and oral/aural communication skills sufficient for 264 Journal Journal of of Speech, Speech, Language, and and Hearing Hearing Research Vol. Vol April April American Speech-Language-Hearing Association /01/

2 them to function without accommodations at school, and eventually in the wider community (Mecklenburg et al., 1990). The development of these skills should afford hearing-impaired children educational and employment opportunities similar to those of normally hearing children. The achievement of these goals becomes more difficult as the hearing loss of the child increases. In fact, the achievement of these goals may have been impossible for some children with very little residual hearing until the recent development and use of cochlear implants that are capable of providing limited hearing even to totally deaf children (NIH Consensus Statement, 1995). Further improvements in hearing aids and cochlear implants should bring these goals within the reach of increasing proportions of children with impaired hearing. Improvements in devices should also make the habilitation task easier for all hearing-impaired children and their parents and teachers (Galvin, Sarant, & Cowan, 1997). In view of the importance of these outcomes, children using cochlear implants and/or hearing aids need to be assessed regularly with measures of speech perception, language, and speech production designed to provide a comprehensive description of their communication skills under realistic conditions. Closed-set vowel and consonant tests and articulation tests can provide well-controlled information about some aspects of a child s perception and production of speech, but they do not necessarily reflect the functionality of the child s communicative abilities. The authors hold a strong opinion that open-set speech perception tests and conversational speech production measures have greater face validity than closed-set perception tests and articulation tests, although they are undoubtedly more difficult to analyze and control (Andrews & Fey, 1986; Boothroyd, 1995; Lund & Duchan, 1993; Morrison & Shriberg, 1992; Stoel-Gammon, 1989). For a similar reason, perceptual assessments should be carried out using the modality that is most commonly used by the child. For the children in the present study, the usual communication modality is lipreading plus hearing. Perceptual evaluations with hearing alone are included for comparison with other studies, and because the effects of hearing loss are expected to be larger when lipreading is not included. The inclusion of language measures in the assessment of children is not only essential in obtaining a balanced picture of their communication abilities, but is necessary to understand the scores on speech perception and production tests. This paper reports language, speech perception, and speech production data for a group of primary-school children with impaired hearing who were using cochlear implants (CI), hearing aids (HA), or both. The data were collected in the initial evaluations of a longitudinal study designed to follow individual children and relate their rates of progress to factors such as their age, level of residual hearing, and hearing device used. The group data reported here will not allow the study of individual variation, but will provide a cross-sectional view of the variation within the group. Multiple linear regression of the group data will be used to formulate an empirical model of average performance over time, which will be compared with individual performance when the longitudinal study is complete. The model will indicate how close the group was at the time of the evaluation to achieving the goal of age-appropriate language skills and will allow prediction of likely average outcomes at later ages, assuming that group trends will be followed by individual development in later years. Prior large scale investigations of the development of spoken communication in children with impaired hearing include the work of Bench and Bamford (1979) in the United Kingdom and Levitt, McGarr, and Geffner (1987) and Geers, Moog, and Schick (1984) in the United States. Research and practical experience over many years have led to a general view expressed recently by Kent (1997): The importance of auditory information in learning speech is profound. Most individuals born deaf never acquire spoken language as a reliable and successful means of communication (p. 434). Although the latter sentence may be true, its significance needs to be considered in light of the facts that most children with impaired hearing are not profoundly deaf (Graham & Dickins, 1995), and the hearing of profoundly deaf children can now be improved by the use of a cochlear implant to the point where it is equivalent to that of severely hearing-impaired children (Boothroyd & Eran, 1994). Interest in all aspects of speech perception, production, and language in profoundly deaf children using cochlear implants has been renewed since 1990, when the first multichannel implant was approved by the U.S. Food and Drug Administration for clinical use in children aged 2 17 years. Research reports generally show consistent postoperative improvements over time and increasingly positive outcomes as implant speech processors improve and children attain longer periods of auditory experience (e.g., Busby et al., 1989; Dawson, Blamey, Dettman, Barker, & Clark, 1995; Geers & Moog, 1994; Miyamoto et al., 1992; Spencer, Tye-Murray, & Tomblin, 1998; Svirsky, Robbins, Kirk, Pisoni, & Miyamoto, 2000; Tye-Murray, Spencer, & Woodworth, 1995). It is clear from this growing body of research that cochlear implants are of benefit to deaf children. The NIH Consensus Statement on Cochlear Implants in Adults and Children (1995) recommended: comparative research on language development in children with normal hearing, children with hearing impairment who use hearing aids, deaf children with cochlear implants, studies of the Blamey et al.: Speech Perception, Production, Language, and Age 265

3 relationship between the development of speech perception and speech production in cochlear implant users and identification of successful (re)habilitation approaches. (p. 21) These needs statements are just as appropriate for children who use hearing aids as they are for children who use cochlear implants. In the present study, it was necessary to analyze the results for hearing aid users and implant users separately because of the potentially different effects of residual hearing levels and auditory experience in the two groups. However, our expectation was that severely hearing-impaired children using hearing aids and profoundly deaf children using implants would perform similarly on a range of measures. It is already apparent that these groups of children are enrolled in the same types of habilitation and educational programs in Australia. The aims of this study were (a) to describe the relationships among speech perception and production scores, spoken language measures, hearing loss, and age in a group of primary-school children with impaired hearing; (b) to formulate an empirical model of the acquisition of spoken language and speech perception skills in children with impaired hearing; (c) to compare the speech perception, spoken language, and speech production of children using cochlear implants and hearing aids, allowing for differences in age and auditory experience; and (d) to predict the likely outcomes for this group of children in terms of their expected speech perception and language scores at age 12 when they will be entering secondary school. Method Participants The children in the study were chosen according to the following set of criteria: (a) They were attending primary school (ages 4 12 years old) at the beginning of the study; (b) they had a bilateral pure-tone-average hearing loss of at least 40 db HL (ANSI S3.6, 1989; ISO 389, 1985); (c) they used a cochlear implant or hearing aid or both; (d) they spoke English at home as their first language; (e) they were enrolled in an educational program using spoken language without any signing supplement; and (f) they experienced onset of hearing loss before the age of 5. The children in the study were enlisted from four schools in Melbourne and from the Children s Cochlear Implant Centre in Sydney (SCCIC). The particular schools and clinic were chosen for reasons of convenience in the annual collection of data. Most children attended integrated classes with normally hearing children for part of the time at school. Parents of all children who satisfied the above criteria were asked to give permission for their children to participate in the study, and no children were excluded for any reason other than lack of parental consent or noncompliance with the selection criteria above. Table 1 summarizes some of the characteristics of this group of children. Nine children who used a cochlear implant and a hearing aid together have been included in the implant group. The hearing aid users were fitted binaurally. All of the hearing aids used by children in this study were fitted and maintained free of charge by Australian Hearing (formerly National Acoustic Laboratories). The hearing loss figures given in Table 1 are for the unaided pure-tone-average threshold (PTA) in db HL averaged over frequencies of 500, 1000, and 2000 Hz. For the CI group, the PTA refers to the most recent audiogram taken preoperatively in the better ear. Only 3 children in the CI group had PTA less than 90 db HL. For the HA group, the thresholds were taken from the most recent audiogram in the better ear. In the HA group, 3 children had PTA less than 50 db HL, 12 had PTA between 50 and 70 db HL, 11 had PTA between 70 and 90 db HL, and 14 had a profound hearing Table 1. Detailed information on the children. Implant users Hearing aid users Number of children Number with congenital loss Total number of evaluations M SD Range M SD Range Hearing loss (db HL) Onset of hearing loss (years) Age at device fitting (years) Not known Age at evaluation (years) Duration of deafness (years) a Not applicable Auditory experience (years) a a These two variables are defined in the text of the Results section. 266 Journal of Speech, Language, and Hearing Research Vol April 2001

4 loss greater than 90 db HL. For the CI group, onset of hearing loss in Table 1 refers to the age at which the child s hearing loss became profound (i.e., became greater than 90 db HL). For the HA group, onset of hearing loss refers to the age at which the child s hearing loss became greater than 40 db HL. All children in the CI group used the 22- electrode cochlear implant (Clark et al., 1987) manufactured by Cochlear Ltd., together with the SPECTRA wearable speech processor and the SPEAK speech processing strategy (McDermott, McKay, & Vandali, 1992). Children were evaluated annually, on a date close to the anniversary of their first evaluation for this study. Speech Perception Measures The children were assessed with two open-set speech perception tests: the Consonant-Nucleus-Consonant (CNC) monosyllabic word test (Peterson & Lehiste, 1962) and the Bench-Kowal-Bamford (BKB) sentence test (Bench, Doyle, & Greenwood, 1987). The CNC word test consists of phonetically balanced lists of 50 words, which were presented one at a time without a carrier phrase. The BKB sentence test consists of lists of 16 sentences, including 50 key words, with vocabulary, grammar, and sentence length appropriate to the linguistic abilities of most partially hearing 8- to 15-year-old children (Bench & Bamford, 1979). The tests were presented to the children using live voice by an experienced audiologist or teacher at a normal speaking level, speaking about 1 meter from the child in a quiet room. For 56 children, two lists of each test were presented in a situation where the child was able to use lipreading and hearing together (the auditory-visual or AV condition) and two lists in a situation where the child was able to hear but could not use speechreading (the auditory or A condition). The other 31 children, all from the SCCIC, were tested in a single session as part of their annual clinical evaluation by the Centre s clinical staff. These children were tested with only one list of words and one list of sentences in each condition because of the more limited time available. The AV condition was chosen as being most representative of the child s usual communication mode, and the A condition was included for comparison with other studies and to provide a measure likely to be more directly related to the severity of the child s hearing loss. In the A condition, the presenter sat beside the child, who was required to look at a video camera. No screen or hand over the mouth was used, to avoid changing the acoustic characteristics of the sound reaching the hearing aid or implant processor microphone. No items were presented while the child was looking at the presenter, and if the child looked during an item, it was skipped and presented again later in the list. The tests were spread over two or three short sessions on different days. Forty word lists and 20 sentence lists were used in a balanced design to ensure that children were never tested with the same word or sentence list twice within the duration of the study. The children s verbal responses to the words and sentences were recorded on videotape and scored offline. The CNC word test responses were transcribed phonetically by listeners experienced in listening to the speech of children with impaired hearing, using a broad transcription in which each sound was transcribed as the nearest English phoneme without the use of diacritics. CNC words were scored as correct if the phonetic transcript of the response word matched the dictionary pronunciation of the test word (i.e., every phoneme was correct and there were no additional or missing phonemes in the response). A phoneme was scored as correct if the response included the correct phoneme, regardless of position in the word or any additional or incorrect phonemes in the response. The BKB sentence tests were not phonetically transcribed, but response words had to be unambiguously identifiable as matching key words to be scored as correct. A strict criterion was used such that all morphemes in the key word had to be present in the response. For example, a response of walk was not counted as correct if the key word was either walks or walked. To begin with, every list of words and every list of sentences was scored independently by two listeners, and by a third listener if the two scores differed by more than 8%. In a case where three scores were available, individual phonemes and key words were examined and scored as correct only if two or three of the scorers had marked the item correct. After experience had been gained, nearly all pairs of scores were closer than the 8% criterion, and about half of the lists were scored once only to save time. Intertranscriber variability was assessed for 160 lists of BKB sentences and 171 lists of CNC words from the initial round of evaluations, which were scored by two listeners. The root mean squared differences between the transcribers were 5.7% for sentences, 3.0% for CNC phoneme scores, and 6.3% for CNC word scores. These values should be compared with the interlist variability estimated from 126 pairs of BKB sentence lists and 127 pairs of CNC word lists. Each pair of lists was presented in a single test session in the same condition and scored by the same transcriber. The root mean squared difference between the scores for pairs of lists was 10.1% for sentences, 5.9% for phonemes, and 10.8% for words. These values indicate that intertranscriber variability was about half as large as the test-retest variability. Language Measures Two formal language measures and a conversational language sample were used annually with each child. The three measures were chosen to provide general indices Blamey et al.: Speech Perception, Production, Language, and Age 267

5 of language use that could be expected to be relatively independent of the speech perception and production measures. All of the language testing and conversational sampling was carried out in the AV condition, face to face with the child at a comfortable distance of about 1 meter. The formal tests were administered and scored according to the procedures specified in the test manuals without modification. The Peabody Picture Vocabulary Test (PPVT; Dunn & Dunn, 1981, 1997) was used as a measure of receptive vocabulary. It is a closed-set test in which a word is presented and the child is required to select one of four pictures that best represents the meaning of the word. The PPVT-III version was used for children evaluated in Melbourne and the PPVT-R was used in Sydney. Each version of the PPVT has two forms containing different test words. The alternative forms were used in alternate years in a balanced design to avoid specific learning effects for the words contained in the test. The Clinical Evaluation of Language Fundamentals (CELF; Semel, Wiig, & Secord, 1992, 1995) was used as a measure of more general aspects of receptive and expressive language. The CELF has two levels, CELF Preschool for children aged 6 or less, and the CELF-3 for children aged 6 and over. At the first evaluation of each child, both forms of the CELF were used with children aged 6 and over. At subsequent evaluations, the CELF-3 was used if the child had attained an equivalent age of 5 years or more in the previous evaluation. Otherwise, the CELF Preschool was administered, as recommended by the CELF manuals under the heading Out-of-Age-Range Administration. The CELF Preschool consists of three receptive language subtests (Linguistic Concepts, Sentence Structure, and Basic Concepts), which are all closed-set tasks involving questions with pointing responses, and three expressive language subtests (Recalling Sentences in Context, Formulating Labels, and Word Structure) requiring a verbal response following a pictorial and verbal prompt. The CELF-3 has a similar structure, but with more difficult test items. The receptive subtests are Sentence Structure (for ages 6 8 years only), Concepts and Directions, Word Classes, and Semantic Relationships (only for ages 9 years and above). The expressive subtests are Word Structure (for ages 6 8 years only), Formulated Sentences, Recalling Sentences, and Sentence Assembly (only for ages 9 years and above). The formal language tests were presented by an experienced audiologist, teacher, or speech-language pathologist, and videotaped so that scoring could be checked independent of the online scoring that is required for most subtests. Both the PPVT and CELF are normed against children with normal hearing in such a way that raw scores can be expressed as an equivalent age (i.e., the age at which the average score of normally hearing children is equal to the score of the child being evaluated). In this report, all of the data for the PPVT and CELF will be expressed as equivalent ages, because we are particularly concerned with rates of development. The use of equivalent age effectively normalizes the rates of change so that a relatively uniform rate of progress would be expected if the child followed the normal development path. Raw scores of individual subtests, on the other hand, can have a rather nonlinear variation over time because of floor and ceiling effects. The conversational language samples were videotaped during the test sessions with each child. The child s conversational partner was an audiologist, teacher, or speech-language pathologist who had presented the perception and/or formal language tests. The conversation was relatively unstructured and undirected by the conversational partner unless the child was particularly uncommunicative. In this case, a familiar topic, such as the child s family or a sport, was introduced by the partner. The goal was to record a sample of utterances in up to 15 minutes. This was achieved in over 90% of cases, and at least 30 utterances were recorded in every case. A gloss of the first 60 utterances from the conversation was produced from the videotape, and then each word was phonetically transcribed by a speech-language pathologist or linguist who had experience in listening to the speech of children with impaired hearing. The mean number of words transcribed per sample was 286. The orthographic and phonetic transcriptions of the conversations were carried out using the Computer Aided Speech And Language Analysis (CASALA) program (Serry, Blamey, Spain, & James, 1997), which included a phonetic dictionary of Australian English to speed up the phonetic transcription process and which calculated statistics from the transcripts. Mean length of utterance in morphemes was used as a gross measure of syntactic complexity of the child s spontaneous language. Speech Production Measures The transcribed language samples were analyzed in several ways to obtain measures of each child s speech production abilities. Overall intelligibility of the child s speech was assessed by counting the number of unintelligible words. This concept is rather anomalous because anything unintelligible is not a word, and a word is, by definition, a recognizable and meaningful unit of speech. By taking into account the syntactic and semantic context, as well as the duration of unintelligible parts of the child s utterances, the transcriber attempted to mark unintelligible words in the orthographic transcript. The percentages of unintelligible words and words produced without any phonetic error were calculated for each conversational sample. The narrow phonetic 268 Journal of Speech, Language, and Hearing Research Vol April 2001

6 transcription method used included the full set of diacritics of the International Phonetic Alphabet (International Phonetic Association, 1999). This method implies that the correctly produced words were not only phonemically correct, but that each phoneme was correctly articulated with a reasonably high degree of accuracy. For example, hypernasality is a common feature of deaf children s speech. Nasalized vowels were marked by the use of a nasality diacritic, and words mistakenly containing nasalized vowels would not have been included as correctly produced. Intertranscriber agreement for the narrow transcription procedure was estimated by comparing the results of a single conversational sample (containing 1,170 phonemes) transcribed independently by four transcribers. The agreements between pairs of transcribers ranged from 64% to 74%, which is slightly lower than the average of 74% reported by Shriberg and Lof (1991) for consensus teams of transcribers in narrow transcripts of children with mild-to-severe speech delay. Relatively low agreement is to be expected, given that the percentage of phonemes correct for this conversational sample was 72%, compared to an average of 80% for Shriberg and Lof, who found that transcription agreement decreased as the percentage of correct phonemes decreased. Each child s speech was measured at the phoneme level in terms of the percentages of correctly produced consonants and vowels. These percentages excluded the unintelligible parts of the conversations. To be correct, the produced phoneme had to match the target phoneme with no additional diacritic marking. A previous study of the phonetic inventories of children using cochlear implants (Serry & Blamey, 1999) showed that diphthongs were slower to emerge than monophthongs, so these two classes of vowels were analyzed separately. In the case of consonants, the analysis considered only singleton consonants (i.e., not consonant clusters) for a similar reason. Previous experience has shown that consonants occurring within clusters are prone to much higher error rates than singleton consonants through processes such as cluster reduction and deletion. Exclusion of clusters from the consonant scoring in conversational materials is thus likely to produce a measure that is less directly affected by the vocabulary used by the child. Inclusion of clusters would underestimate the consonant production skills of children who use more complex word structures compared to children who use only simple word structures containing few consonant clusters. Statistical Analyses Unless there is a theoretical reason to suppose that experimental variables are related in a nonlinear fashion, it is common practice to use linear regression as a tool to determine whether the variables are statistically related. Linear regression relates a single dependent variable, such as a speech perception score, to one or more independent variables, such as the age and hearing loss of the subject. It is assumed that a constant difference in an independent variable is associated with a constant difference in the dependent variable. For example, the difference in the speech perception scores of a 2-year-old and a 3-year-old is assumed to be equal to the difference in scores between a 6-year-old and a 7-yearold because the age difference is the same. It is also assumed that the effects of the independent variables are additive. Linear regression has been the choice for previous studies of perception scores in adult cochlear implant users (e.g., Blamey et al., 1992; Gantz, Woodworth, Abbas, Knutson, & Tyler, 1993; Summerfield & Marshall, 1995) as well as in children who use cochlear implants (e.g., Dowell, Blamey, & Clark, 1995, 1997; Sarant, Blamey, Dowell, Clark, & Gibson, in press; Staller, Dowell, Beiter, & Brimacombe, 1991). Linear regression will also be used to formulate an empirical model of the data in the present report, but the reader is asked to bear in mind that it is highly unlikely that the relationships among speech perception scores and variables such as hearing thresholds will be linear, because the scores must lie within the restricted range from 0% to 100%. An S-shaped psychometric function is more probable than a straight line. It is also unlikely that the factors examined will have a truly additive effect. For example, totally deaf persons will score close to zero on an open-set word test presented via audition alone, regardless of how long they have been totally deaf or how much auditory training they have had. Therefore, factors relating to a person s auditory experience cannot have an additive effect on auditory perception scores for a totally deaf person. Nonadditivity is sometimes allowed for in multiple regression by introducing interaction terms into the analysis. Interaction terms allow the effect of one regression variable to depend on the value of another. In the example above, the regression line for speech perception against duration of deafness would have zero slope for a totally deaf person because they cannot hear anything, regardless of how long they have been deaf. The slope may be nonzero for people with a partial hearing loss. Thus, the effect of duration of deafness would depend on the amount of residual hearing. Interaction terms were not included in the analyses described below because they are often difficult to interpret. In essence, linear regression can indicate which variables are significantly related to one another in the data and the relative sizes of these effects, but it is unlikely to produce a model that can account for all of the data in a theoretically feasible manner. Despite these limitations, a linear model may provide a useful approximation, leading toward a theoretically plausible nonlinear model. Blamey et al.: Speech Perception, Production, Language, and Age 269

7 Results At the time that these data were analyzed, a total of 87 children had been enrolled in the longitudinal study. The total number of evaluations used in this report is 152. Seventeen children had been evaluated only once, 60 had been evaluated twice, and 5 had been evaluated three times. If the complete data set is used in the analyses, the variability of the trends is minimized, but those children who have been evaluated several times have a greater influence than those who have been evaluated only once. If one uses only the most recent evaluation, or the initial evaluation, the results show similar trends to those of the complete data set. Using multiple observations on the same individuals violates the statistical assumption of independence of observations, potentially leading to an underestimate of the error variance in the analyses and greater than specified Type 1 error. This was corrected in the analyses by using a smaller number of degrees of freedom in the t-test criteria that were used to calculate significance levels. The number of degrees of freedom was based on the number of children instead of the number of observations included in the analysis. If readers are concerned about this aspect of the analysis, they are referred to a previous analysis of the first evaluation for the first 57 children enrolled into the study (Blamey et al., 1998). The trends and the conclusions reached for the smaller data set are highly consistent with those in the present report. Speech Perception Measures Figures 1 4 show each child s individual scores versus age at evaluation for BKB sentences and CNC words in the A and AV conditions. The regression lines show the trends for hearing aid users and implant users separately. These analyses indicate small but statistically significant (p <.05) increases in perception scores with age in the AV condition, but not in the A condition. The percentage of the variance accounted for in every case is less than 10%. The slopes and intercepts from each figure were compared using t tests with standard deviations derived from the regression analyses. The tests indicated no statistically significant differences between the regression lines for the HA and CI user groups in each figure. It is reasonable to expect that perception scores should be related to the amount of residual hearing, especially in the A condition. The reasoning is obvious for the hearing aid users. For the CI users, it is also possible that an ear with more residual hearing will perform better after cochlear implantation because of greater nerve survival or a generally better condition of the cochlea and auditory brainstem. Thus, further regression analysis was carried out using age and PTA as independent variables. Age and PTA were not significantly correlated for the HA group (r =.213, p >.05) or the CI group (r =.295, p >.05). As a result, PTA and age can be used as independent variables in multiple regression analyses to arrive at an estimate of the relative size of their effects. The equations of the regression lines and the proportions of variance they account for (R 2 ) are given in Table 2. The regressions indicate a small positive trend in AV perception scores with increasing age Figure 1. BKB Sentence Perception scores in the auditory-visual condition versus age for children using a cochlear implant (CI) or hearing aid (HA). Figure 2. BKB Sentence Perception scores in the auditory condition versus age for children using a cochlear implant (CI) or hearing aid (HA). 270 Journal of Speech, Language, and Hearing Research Vol April 2001

8 Figure 3. CNC Word scores in the auditory-visual condition versus chronological age for children using a cochlear implant (CI) or hearing aid (HA). Figure 4. CNC Word scores in the auditory condition versus chronological age for children using a cochlear implant (CI) or hearing aid (HA). for both groups and a negative trend in A perception scores with increasing hearing loss for the HA group. The use of PTA as an independent variable increased the percentage of variance accounted for above 10% for four of the HA measures and one of the CI measures. In previous studies of children using cochlear implants (e.g., Dowell, Blamey, & Clark, 1995), the age variable has been split into three components for time periods (a) from birth to onset of hearing loss, (b) from onset of hearing loss to implantation, and (c) from implantation to the time of evaluation. This split is reasonable, because the child s effective hearing levels and auditory experience are usually quite different during these time periods. The three time periods are labeled Onset, Deafness, and Experience, respectively. For the hearing aid group, only two time periods (Onset and Experience) are used in the analyses. In this case, Experience refers to the whole of the period beginning with the onset of hearing loss. The age at first hearing aid fitting is unavailable at the time of writing for most children in the study, Table 2. Regression analyses of auditory-visual (AV) and auditory (A) speech perception scores as a function of age and pure-tone-average hearing loss. Hearing aid group BKB Sentence Score AV = 74.9*** ** age 0.19 PTA R 2 = CNC Word Score AV = 53.2*** * age 0.17 PTA R 2 = CNC Phoneme Score AV = 78.5*** * age 0.07 PTA R 2 = BKB Sentence Score A = 99.5*** * age 0.63*** PTA R 2 = CNC Word Score A = 67.3*** age 0.52*** PTA R 2 = CNC Phoneme Score A = 91.4*** age 0.39*** PTA R 2 = Cochlear implant group BKB Sentence Score AV = 103*** *** age 0.59* PTA R 2 = CNC Word Score AV = 71.7** + 2.6* age 0.26 PTA R 2 = CNC Phoneme Score AV = 91.0*** * age 0.16 PTA R 2 = BKB Sentence Score A = 100*** * age 0.56* PTA R 2 = CNC Word Score A = 57.1* age 0.19 PTA R 2 = CNC Phoneme Score A = 86.6*** age 0.19 PTA R 2 = Note. The scores and the constant coefficients are expressed in percent, the age coefficient is expressed in percent per year, and the PTA coefficient is expressed in percent per decibel of hearing loss (db HL). *p <.05, **p <.01, ***p <.001. Blamey et al.: Speech Perception, Production, Language, and Age 271

9 Table 3. Correlation coefficients for independent variables used in multiple regression analyses. The data for the CI group are in the upper right half of the table, above the diagonal, and the data for the HA group are in the lower left half of the table, below the diagonal. The Deafness variable does not apply to the HA group. Age Onset Deafness Experience PTA Age * 0.728*** Onset Deafness N/A N/A Experience 0.904*** 0.356* N/A 0.458** PTA N/A *p <.05, **p <.01, ***p <.001. so it is not possible to include the period between onset of hearing loss and hearing aid fitting as an independent variable in the analyses. Table 3 shows correlation coefficients between all pairs of variables used in the regression analyses. The correlation between age and experience is high for both CI and HA groups, but this correlation is of no concern, as these two variables will not be used together in any analysis. The correlation between experience and PTA is moderate in the CI group, probably as a result of changing criteria for implantation, which have led to children with greater hearing loss being implanted earlier than children with less profound loss. This correlation means that the regressions for the CI group may not be able to uniquely separate the effects of implant experience from the effects of preoperative thresholds. For example, a difference in the regression coefficient for experience may be balanced to some extent by a difference in the coefficient of PTA. Splitting the age variable into the three components yields the multiple regression results shown in Table 4. In 10 out of 12 cases, splitting the age variable accounted for a significantly greater proportion of the variance. An F test was conducted in the same way as for a step-wise regression. The reduction in the error sum of squares was divided by the difference in the number of parameters in the two analyses, and this mean square value was divided by the mean square error from the analysis in Table 4. These computations give the following F values (in the same order as the analyses in Table 4): 5.73, 6.32, 6.08, 4.25, and 5.25 (all with p <.05), 3.00, 3.11, 3.75 (the latter with p <.05), and 6.78, 6.62, 6.06, and 5.91 (the last four with p <.01). Table 4 shows that postnatal onset of hearing loss had a significant positive association with speech perception scores apart from the CNC words for the CI group. Similarly, the effect of experience was positive with the CI in both conditions and with the HA in the AV condition. The negative effect of PTA was clearer in the CI group once the separate effect of onset of hearing loss had been accounted for. A surprising result was that duration of deafness was not significantly related to perception scores for the CI group. The majority of previous studies for adults have suggested that the duration of deafness is the strongest factor affecting perception scores with the implant (Blamey et al., 1996, Table 3). This aspect of the results will be considered in detail in the Discussion section. Table 4. Regression analyses of perception scores as a function of age at onset of hearing loss, duration of deafness, experience, and pure-tone-average hearing loss. Hearing aid group BKB Sentence Score AV = 72.2*** ** onset * exp 0.15 PTA R 2 = CNC Word Score AV = 49.4*** ** onset * exp 0.12 PTA R 2 = CNC Phoneme Score AV = 76.8*** ** onset * exp 0.05 PTA R 2 = BKB Sentence Score A = 96.1*** ** onset exp 0.59*** PTA R 2 = CNC Word Score A = 64.2*** * onset exp 0.48** PTA R 2 = CNC Phoneme Score A = 89.7*** * onset exp 0.37*** PTA R 2 = Cochlear implant group BKB Sentence Score AV = 116*** * onset deaf *** exp 0.68** PTA R 2 = CNC Word Score AV = 103*** onset 1.47 deaf *** exp 0.52* PTA R 2 = CNC Phoneme Score AV = 105*** onset 0.34 deaf *** exp 0.28* PTA R 2 = BKB Sentence Score A = 114*** * onset deaf ** exp 0.66* PTA R 2 = CNC Word Score A = 81.7** onset 2.37 deaf * exp 0.38 PTA R 2 = CNC Phoneme Score A = 100*** onset 0.82 deaf ** exp 0.30* PTA R 2 = Note. The scores and the constant coefficients are expressed in percent, PTA is expressed in percent per db HL, and the onset, deafness, and experience coefficients are expressed in percent per year. *p <.05, **p <.01, ***p < Journal of Speech, Language, and Hearing Research Vol April 2001

10 Language Measures Figures 5 and 6 show the PPVT and CELF results (with the scores expressed as equivalent ages) plotted against the chronological age at the time of evaluation. These figures show more consistent improvements over time for language than for speech perception scores, although there is still a high degree of variability. Figure 7 Figure 5. PPVT equivalent age versus chronological age for children using a cochlear implant (CI) or hearing aid (HA). The dashed line indicates average results for children with normal hearing. shows the mean length of utterance (MLU) data from the conversational samples, also plotted against age. Equations for the regression lines in the figures are given in Table 5. The increases in language scores with age are significant for all except the MLU measure in the HA group. For the PPVT and CELF regression equations, the constant terms are not significantly different from zero (because one would expect that equivalent language ages should be zero for all children at birth, regardless of hearing status). If the regression is repeated with the constant term set to zero, the rates of improvement in PPVT equivalent ages become 0.65 for the HA group and 0.63 for the CI group. The corresponding rates of Figure 7. Mean length of utterance (MLU) versus chronological age for children using a cochlear implant (CI) or hearing aid (HA). Figure 6. CELF equivalent age versus chronological age for children using a cochlear implant (CI) or hearing aid (HA). The dashed line indicates average results for children with normal hearing. Table 5. Regression analyses of language measures as a function of age. Hearing aid group PPVT equivalent age = *** age r 2 = CELF equivalent age = *** age r 2 = MLU morphemes = 4.16*** age r 2 = Cochlear implant group PPVT equivalent age = *** age r 2 = CELF equivalent age = *** age r 2 = MLU morphemes = 1.87* ** age r 2 = Note. For the PPVT and CELF measures, the equivalent age and the constant coefficients are expressed in years. The regression coefficient for age is dimensionless (years of equivalent age divided by years of chronological age). The MLU measure and the constant terms in the MLU regressions are expressed in morphemes. The coefficients of age in the MLU analyses are expressed in morphemes per year. *p <.05, **p <.01, ***p <.001. Blamey et al.: Speech Perception, Production, Language, and Age 273

11 Table 6. Regression analyses of language measures as a function of age at onset of hearing loss, duration of deafness, experience, and pure-tone-average hearing loss. Hearing aid group PPVT equivalent age = *** onset *** exp PTA R 2 = CELF equivalent age = *** onset *** exp 0.00 PTA R 2 = MLU morphemes = ** onset exp PTA R 2 = Cochlear implant group PPVT equivalent age = * onset *** deaf *** exp 0.04 PTA R 2 = CELF equivalent age = 6.02** *** onset *** deaf *** exp 0.06** PTA R 2 = MLU morphemes = * onset deaf ** exp 0.01 PTA R 2 = Note. The equivalent age and the constant coefficients are expressed in years. The regression coefficients for onset, deafness, and experience are dimensionless (PPVT and CELF regressions) or are expressed in morphemes per year (MLU regressions). The MLU measure and the constant terms in the MLU regressions are expressed in morphemes. The PTA coefficients are expressed in years per db HL (PPVT and CELF regressions) or morphemes per db HL (MLU regressions). *p <.05, **p <.01, ***p <.001. improvement for the CELF are 0.59 and Thus, the children in this study are learning language at about half to two-thirds the normal rate on average. Age accounts for a smaller percentage of the variance for MLU than for PPVT and CELF, and the constant term is significantly different from zero. These differences are due, in part, to the wide variability in the MLU measure and, in part, to a nonlinear relationship between MLU and age. The nonlinearity will be demonstrated later in the Discussion section. There is little difference between the HA and CI groups on any of the language measures. As for the perception measures, it is reasonable to expect that hearing loss will affect the rates of improvement in language. Therefore, the factors PTA, onset, deafness, and experience may also be of interest. In Table 6, 13 of the 15 coefficients for onset time, duration of deafness, and experience are significantly greater than zero, indicating that language is improving during all of these periods, regardless of hearing status and device fitting. Using the standard deviations derived from the regression analyses, t tests were used to compare the coefficients for onset, deafness, and experience with the normal rate of 1 and with one another. The coefficients for onset time are not significantly different from 1, indicating that children are developing language at the normal rate while they have normal hearing. The coefficients for experience and deafness are not significantly different from each other for the PPVT and CELF analyses of the CI group. There are also no significant differences between the experience coefficients for the CI and HA groups in the PPVT, CELF, and MLU analyses. The similarity of these coefficients indicates that children in the two groups were developing the language skills measured by the PPVT, CELF, and MLU tests at roughly equal rates, on average, regardless of the fitting of a hearing aid or cochlear implant. Furthermore, the coefficients for PTA are all very small and only one (for the CELF measure in the CI group) is significantly different from zero, indicating that the level of unaided hearing loss was very weakly related to the rate of language learning after the onset of hearing loss. Speech Production Measures The videotaped conversational samples were transcribed phonetically, as described in the Method section, to yield the proportions of (a) words that were unintelligible; (b) words that were phonetically correct; and (c) monophthongs, diphthongs, and consonants that were phonetically correct. Nine samples from the CI group and two from the HA group could not be transcribed phonetically because of poor-quality recordings. Table 7 gives the mean results for the two groups of children. Table 7. Means (and standard deviations) for the speech production measures derived from recorded conversational samples. Measure CI group HA group Number of samples Unintelligible words 5.8% (6.7%) 2.7% (4.6%) Phonetically correct words 39.4% (13.6%) 40.5% (14.9%) Monophthongs 81.3% (9.1%) 83.3% (10.6%) Diphthongs 74.6% (13.1%) 72.0% (15.6%) Consonants 68.7% (10.5%) 69.0% (10.5%) Note. The values indicate the percentage of words that were unintelligible and the percentage of intelligible words that were correctly produced. The phoneme measures are the percentages of monophthongs, diphthongs, and singleton consonants (excluding consonants in clusters) in intelligible words that were correctly produced. There were no significant differences between the mean values for the CI and HA groups. 274 Journal of Speech, Language, and Hearing Research Vol April 2001

12 Figure 8. Percentage of words produced correctly versus chronological age for children using a cochlear implant (CI) or hearing aid (HA). deafness, experience, and PTA as the independent variables (see Table 8). The analyses suggest that postnatal onset of hearing loss was related positively to speech production for the HA group, and that experience with the implant had a small positive effect in the analyses of consonant production and intelligibility for the CI group. The degree of hearing loss as measured by PTA was not significantly correlated with any of the speech production measures in either group. Very little of the variance in these measures was accounted for by the time and PTA variables. A potential explanation for this result is that most of the speech production skills for these children were acquired at a fairly young age and had not changed very much since. This outcome may have been the result of a reduction in speech production feedback and training from their family and teachers that occurred once they reached a reasonably intelligible standard. Two-sample t tests indicated no significant differences between the two groups of children on any of these measures. The children were mostly intelligible, although the ranges for the measure of unintelligible words were quite wide (0 38% for the CI group and 0 34% for the HA group). None of the speech production measures was correlated significantly with age, so only one example is displayed graphically. Figure 8 shows the percentage of words that were phonetically correct as a function of age at evaluation. As for the perception and language measures, multiple regression analyses were performed with onset, Relationships Among Perception, Production, and Language The group results and analyses presented thus far indicate that speech perception and language are changing over time, and speech perception scores are related significantly to unaided hearing thresholds. Nevertheless, a high proportion of the variance in the regression analyses remains unexplained. It is possible that environmental factors, such as the amount and nature of the habilitation the child receives, the amount of time that hearing aids and implant speech processors are worn, the size and economic status of the child s family, and the child s peer group (or lack of peer group, in some cases), may account for much of the unexplained variance. The nonverbal intelligence of the child may also Table 8. Regression analyses of speech production scores as a function of age at onset of hearing loss, duration of deafness, experience, and pure-tone-average hearing loss. Hearing aid group Unintelligible words = onset exp PTA R 2 = Words correct = 41.9*** ** onset 0.29 exp 0.01 PTA R 2 = Monophthongs = 82.0*** onset 0.54 exp PTA R 2 = Diphthongs = 75.1*** * onset 0.81 exp PTA R 2 = Consonants = 65.8*** ** onset 0.02* exp PTA R 2 = Cochlear implant group Unintelligible words = onset 0.08 deaf 1.68** exp PTA R 2 = Words correct = 52.9** onset 1.39 deaf exp 0.18 PTA R 2 = Monophthongs = 90.7*** 0.39 onset 0.86 deaf exp 0.08 PTA R 2 = Diphthongs = 96.3*** 2.92 onset 1.76 deaf exp 0.20 PTA R 2 = Consonants = 75.2*** onset 1.16 deaf * exp 0.11 PTA R 2 = Note. The scores and the constant coefficients are expressed in percent; the onset, deafness, and experience coefficients are expressed in percent per year; and the PTA coefficients are expressed in percent per db HL. *p <.05, **p <.01, ***p <.001. Blamey et al.: Speech Perception, Production, Language, and Age 275