Decline in Older Persons’ Ability to Recognize Speech in Noise: The Influence of Demographic, Health-Related, Environmental, and Cognitive Factors Marieke Pronk,1 Dorly J. H. Deeg,2,3 Joost M. Festen,1 Jos W. Twisk,2,4 Cas Smits,1 Hannie C. Comijs,3 and Sophia E. Kramer1 Objectives: The first aim was to investigate whether the rate of decline in older persons’ ability to recognize speech in noise over time differs across age and gender. The second aim was to determine extent demographic, health-related, environmental, and cognitive factors influence the change in speech-in-noise recognition over time.
hallmark of ARHL is pronounced sensitivity loss in the high frequencies. ARHL is characterized by reduced hearing sensitivity and speech understanding in noisy environments, slowed central auditory processing, and impaired sound localization (Gates & Mills 2005). For several decades, cross-sectional pure-tone audiometry data are used to estimate age- and gender-specific hearing loss trajectories, and since the 1990s, longitudinal data from large cohort studies are increasingly used for this purpose (e.g., Gates et al. 1990; Morell et al. 1996; Cruickshanks et al. 2003; Kiely et al. 2012). The rate of hearing decline seems to be frequency-specific and dependent on age and gender, although mixed evidence exists. Some studies found faster declines in men across most frequencies up to very old age (Pearson et al. 1995; Chao & Chen 2009), others reported faster frequencyspecific declines in women (Wiley et al. 2008; Kiely et al. 2012), and yet others found no gender differences (Gates et al. 1990; Cruickshanks et al. 2003). Regarding age, in younger old (up to around 70 years) faster hearing declines are generally observed in the high frequencies, while for the older old, this is in the low frequencies (Wiley et al. 2008; Chao & Chen 2009; Kiely et al. 2012). ARHL has long been considered as a natural aging process with associated histological, electrophysiological, and molecular changes in human tissues, causing deficits in hair cells, cochlear neurons, the stria vascularis, and tissues more upstream in the central nervous system (Gates & Mills 2005). Nowadays, ARHL is increasingly recognized as a complex condition the etiology of which is additionally influenced by environmental, healthrelated, and genetic factors (Van Eyken et al. 2007). Noise exposure is the environmental factor most extensively studied. Although excessive noise levels evidently cause hearing loss, it is less clear whether past noise exposure affects the rate of ARHL later in life (Gates et al. 2000; Cruickshanks et al. 2003; Van Eyken et al. 2007; Dubno et al. 2008; Kiely et al. 2012). Smoking and alcohol abuse are other possible risk factors, but mixed results exist: whereas some reported an increased hearing loss risk, others reported none (Rosenhall et al. 1993; Brant et al. 1996; Cruickshanks et al. 1998b; Van Eyken et al. 2007). The same holds for diabetes, cardiovascular disease, and hypertension. Various studies identified them as risk factors for ARHL, but others did not (e.g., Brant et al. 1996; Frisina et al. 2006; Kiely et al. 2012). Most of the aforementioned studies hold important limitations. One is the cross-sectional study design, which limits drawing strong conclusions about causation. The few studies that did use a longitudinal approach mostly failed to find significant risk factors (Dubno et al. 2008; Gopinath et al. 2009, 2010; Mitchell et al. 2009). Another is the study of a certain risk factor in
Design: Data covering 3 to 7 years of follow-up (mean: 4.9 years) of a large sample of the Longitudinal Aging Study Amsterdam were used (n = 1298; 3025 observations; baseline ages: 57 to 93 years). Hearing ability was measured by a digit triplet speech-in-noise test (SNT) yielding a speech reception threshold in noise (SRTn). Multilevel analyses were used to model the change in SRTn over time. First, interaction terms were used to test differences in rate of decline across subgroups. Second, for each of the following factors the authors determined the influence on the change in SRTn: age, gender, educational level, cardiovascular conditions, information processing speed, fluid intelligence, global cognitive functioning, smoking, and alcohol use. This was done by calculating the percentage change in Btime after adding the particular factor to the model. Results: On average, respondents’ SRTn increased (i.e., deteriorated) significantly over time by 0.18 dB signal-to-noise ratio per annum. Rates were accelerated for older ages (Btime = 0.13, 0.14, 0.25, 0.27 for persons who were 57 to 65, 65 to 75, 75 to 85, and 85 to 93 years of age, respectively). Only information processing speed relevantly influenced the change in SRTn over time (17% decrease in Btime). Conclusions: Decline in older persons’ speech-in-noise recognition over time accelerated for older ages. Decline in information processing speed explained a moderate proportion of the SRTn decline. This indicates the relevance of declining cognitive abilities in the ability of older persons to recognize speech in noisy environments. (Ear and Hearing 2013;34;722–732)
INTRODUCTION Hearing loss is one of the most prevalent chronic conditions in old age (Davis 1990; Ries 1994). Several studies described prevalence rates up to 40% in adults 50 years and older (Duijvestijn et al. 1999) and up to 90% in persons more than 80 years of age (Cruickshanks et al. 1998a). Worldwide, adultonset hearing loss is the second leading cause of years lived in disability (World Health Organization 2008). Many studies found significant relationships with poor psychosocial health such as anxiety, depression, and loneliness (e.g., Kramer et al. 2005; Pronk et al. 2011). The bulk of the hearing loss cases can be classified as agerelated hearing loss (ARHL). The traditional audiological Departments of 1Ear Nose Throat/Audiology, 2Epidemiology and Biostatistics, 3Psychiatry/GGZ inGeest, VU University Medical Center, EMGO Institute for Health and Care Research, Amsterdam, the Netherlands; and 4Department of Health Sciences, Faculty of Earth and Life Sciences, VU University Amsterdam, the Netherlands.
0196/0202/13/3406-0722/0 • Ear and Hearing • Copyright © 2013 by Lippincott Williams & Wilkins • Printed in the U.S.A. 722
PRONK ET AL. / EAR & HEARING, VOL. 34, NO. 6, 722–732
isolation. Thus, most studies could not elucidate what the relative contribution of various factors to ARHL was. Last, except for a study by Dubno et al. (2008), all studies used pure-tone thresholds as their hearing loss outcome. Pure-tone audiometry is still considered the gold standard assessment for hearing loss, but its relationship with experienced daily life disability is limited (e.g., Demeester et al. 2012), especially with regard to selfreported speech understanding in noise (Kramer et al. 1996) and during group conversations (Gatehouse & Noble 2004). Trajectories based on the performance on speech-in-noise tests (SNTs) over time have been scarcely examined. Such investigations are especially warranted from the patient’s perspective, as understanding speech in challenging listening situations, such as in noise, is the most frequently reported disability (Kramer et al. 1998). Longitudinal decline in speech perception abilities was examined by Bergman et al. (1976), Møller (1981), Pedersen et al. (1991), Hietanen et al. (2004), Divenyi et al. (2005), and Dubno et al. (2008), but most of these studies were limited by a small (n ≤ 35) sample size of older subjects, included only one birth cohort combined with a relatively short follow-up, or included a non–population-based sample. Moreover, only two (Divenyi et al. 2005; Dubno et al. 2008) included an SNT (the revised SPIN test). Dubno et al. (2008) used data of a population-based sample of 85 older adults (56 to 81 years at baseline) who were measured four times across a 7- to 13-year period. Only marginal or nonsignificant rates of decline over time were reported. No gender differences in the rate of decline appeared, nor did the rate differ across age groups or between those with or without noise history. Divenyi et al. (2005) reported a statistically significant decline in their small sample (n = 29) across 4 to 8.5 years of follow-up, but they did not report the rate of this decline. Speech-based measures may capture central auditory processing functions that decline with increasing age. This concerns functions required to distinguish pitch, loudness, and duration of acoustical signals (Pichora-Fuller & Souza 2003), but also cover more global cognitive functions such as information processing speed (IPS) and working memory (Wingfield & Tun 2007; Akeroyd 2008; Rönnberg et al. 2008; Arlinger et al. 2009). It is generally assumed that the involvement of these neurocognitive functions becomes more pronounced in challenging listening situations such as noisy backgrounds (e.g., Pichora-Fuller & Souza 2003; Koelewijn et al. 2012). Thus, to map functional hearing that goes beyond audibility, mapping trajectories of the ability to recognize speech in noise seems essential. The fact that often correlations between speech-innoise measures and pure-tone measures of 0.4 to 0.7 are found suggests their partly differing underlying constructs (Houtgast & Festen 2008). Divenyi et al. (2005) found that older persons’ performance on speech-recognition tasks declined with increasing age more rapidly than their audiometric performance. Accordingly, they concluded that decline in auditory performance during the later decades of life is due to the degeneration of peripheral auditory structures, but additionally involves a non-peripheral (i.e., central) component. The aim of the present study is to investigate the decline in older persons’ ability to recognize speech in noise over time and the factors influencing this decline. Multilevel modeling will be applied using longitudinal SN data derived from a large Dutch older population-based cohort. First, the mean decline in SN recognition over time will be determined. We will investigate
723
whether the decline differs across age and gender. Second, we will examine to what extent demographic (age, gender, level of education), health-related (cardiovascular conditions), environmental (alcohol use, smoking), and cognitive factors (global cognitive functioning, information processing speed, henceforth IPS, fluid intelligence) influence the decline in speech-in-noise recognition over time and determine their relative contribution.
MATERIALS AND METHODS Sample and Procedures The sample originated from the Longitudinal Aging Study Amsterdam (LASA) cohort (Huisman et al. 2011). LASA is an ongoing cohort study on predictors and consequences of changes in autonomy and wellbeing in an aging population. A random sample of 3107 older persons (aged 55 to 85 years) stratified for age and gender was drawn from the Dutch population for the first LASA measurement in 1992/1993. A second cohort of 1002 respondents (aged 55 to 64 years) was added in 2002 from the same sampling frame as the original cohort from 1992. A followup measurement was conducted every 3 to 4 years. All measurements were performed in the respondent’s home by trained and supervised interviewers. Informed consent was obtained from all respondents. The study was approved by the Medical Ethics Committee of the Vrije Universiteit University Medical Center. For the present study, the original cohort and the second cohort will be referred to as the 1992 cohort and the 2002 cohort, respectively. In 2001/2002 (1992 cohort only), 2005/2006 (both cohorts), and in 2008/2009 (both cohorts), speech-in-noise recognition was measured with an SNT, yielding a speech-reception threshold in noise (SRTn). We selected respondents who had complete SNT data for at least two measurements such that intraindividual change would be sufficiently represented in the statistical models. We tested whether there was selective attrition by comparing age, gender, and SRTn characteristics of those lost to follow-up with those remaining in the sample. Loss to follow-up due to death was regarded as being population-based and was therefore not considered. In general, those lost to follow-up were older and had poorer (higher) SRTns (see Table 1). To identify outliers of the SRTn scores, we linearly regressed the 2005/2006 SRTns on the 2001/2002 SRTns (n = 591), the 2008/2009 SRTns on the 2005/2006 SRTns (n = 1000), and the 2008/2009 SRTns on the 2001/2002 SRTns (n = 136). Thirty-five data points were situated more than 3 SD outside the regression lines and were omitted. The distribution of the respondents and the observations over the measurements is depicted in Figure 1. In total, 1298 respondents provided 3025 SRTn observations. Seven hundred eighty respondents originated from the 1992 cohort, and 518 originated from the 2002 cohort.
Measures In the analyses, the SRTn was used as the dependent variable, and time, demographical, health-related, environmental, and cognitive variables were used as independent variables.
Dependent Variable: SRTn
The SNT was originally developed as a functional hearing screening self-test by telephone (Smits et al. 2004). The test
724
PRONK ET AL. / EAR & HEARING, VOL. 34, NO. 6, 722–732
TABLE 1. Selective attrition between measurements 1998/1999 and 2001/2002 (1992 cohort), 2002 and 2005/2006 (2002 cohort); 2001/2002 and 2005/2006 (1992 cohort); and between 2005/2006 and 2008/2009 (both cohorts) 1998/1999 and 2001/2002 Variable
2002 and 2005/2006
1992 cohort (nattrition = 626) * * —
Age (older) Gender (female) SRTn (higher)
2002 cohort (nattrition = 346) ns ns —
2001/2002 and 2005/2006 1992 cohort (nattrition = 212) * ns *
2005/2006 and 2008/2009 1992 and 2002 cohort (nattrition = 184) ns * *
ns, that is, variable not differing significantly between those lost to follow-up and those remaining in the study sample; —: no SRTn measurement for this cohort in 1998/1999 (1992 cohort) and 2002 (2002 cohort); SRTn, speech reception threshold in noise in dB signal-to-noise ratio. *Variable differing significantly (p < 0.05) between those lost to follow-up and those remaining in the study sample. nattrition, number of respondents lost to follow-up.
determines an individual’s SRTn defined as the signal-to-noise ratio (SNR) in dB corresponding to 50% recognition of digit triplets. The interviewer brought portable testing equipment comprising a telephone, an amplifier, and headphones. At the 2008/2009 measurement, the telephone and amplifier were replaced by a laptop. Before the actual assessment, hearing aids had to be removed and the respondent was instructed to adjust the level of the speech to make sure that signals were 1992 cohort
2002 cohort nobs
TOTAL nobs
727
518
1245
-
0
518
518
x
644
-
0
644
618
-
0
618
1036
3025
nobs t=0 t=3
x -
-
-
t=4 t=7
x
TOTAL nobs
1989
TOTAL nresp
nresp
429
162
53
136
780
TOTAL nresp
518
1298
indicates longitudinal study design reading order SNT data available x
no SNT data available
-
no measurement for this cohort
nobs
number of observations
nresp
number of respondents
t=0
0 years of f.u.; 2001/2002 data (1992 cohort): 2005/2006 data (2002 cohort)
t=3
3 years of f.u.; 2008/2009 data (2002 cohort)
t=4
4 years of f.u.; 2005/2006 data (1992 cohort)
t=7
7 years of f.u.; 2008/2009 data (1992 cohort)
SNT: speech-in-noise test; f.u.: follow-up
Fig. 1. Distribution of the respondents and the observations over the measurements.
sufficiently audible. Subsequently, 23 different monosyllabic digit triplets were presented at different intensity levels against a constant level of stationary background noise. An adaptive up–down procedure was used in which the SNR decreased by 2 dB if the respondent repeated a triplet correctly, and increased by 2 dB after an incorrectly repeated triplet. Different administration and data collection modes caused slight but systematic shifts in the SRTns. The followup scores were therefore corrected. The 2005/2006 shift was estimated by determining age–gender specific SRTn averages at 2001/2002 and 2005/2006. Subsequently, the 2001/2002 averages were linearly regressed on the 2005/2006 averages. After the regression coefficient was set at 1, the constant of the regression equation (i.e., the systematic shift) could be estimated. The same procedure was applied to the 2008/2009 data. The identified shifts were −0.86 dB SNR at 2005/2006, and −0.49 dB SNR at 2008/2009 (both relative to scores at 2001/2002). The SNT has a good validity as indicated by its high correlation (r = 0.87) with the standard Dutch sentences SNT (Plomp & Mimpen 1979). Its test–retest reliability seemed satisfactory in an older subsample from the study by Nachtegaal et al. (2009) (intraclass correlation coefficient; two-way random effects model = 0.70; n = 152; 58 to 82 years). The estimated measurement error is about 1 dB, and is slightly larger for worse SRTns, that is, up to 1.3 dB for SRTns around −1 dB SNR (Smits et al. 2004).
Independent Variables Time in years was included as a continuous variable in the models and was calculated as “time in the study,” taking 2001/2002 (1992 cohort) and 2005/2006 (2002 cohort) as baseline. As such, for the 1992 cohort, the time values included: 0 (2001/2002); 4 (2005/2006), and 7 (2008/2009) years. For the 2002 cohort, these were 0 (2005/2006) and 3 (2008/2009) years. See also Figure 1. Age at baseline was included as a categorical variable (four 10-year groups) because in the first part of the effect analyses (i.e., to determine the change in SRTn over time), the categorical variant fitted the longitudinal SRTn data better than the continuous variant. Gender was included dichotomously. Educational level was assessed by asking the respondent’s highest educational level completed and was categorized into low (uncompleted elementary, elementary, lower vocational), medium (general intermediate, intermediate vocational, general secondary, higher vocational), and high (college and university). Information on cardiovascular conditions was derived from
PRONK ET AL. / EAR & HEARING, VOL. 34, NO. 6, 722–732
self-report and included: incident stroke and myocardial infarction, and presence of claudication, diabetes mellitus, and hypertension. Because diabetes is a chronic condition and the other cardiovascular conditions were assumed to have major and possibly a (life) long impact, the occurrence at one measurement was carried forward to later measurements irrespective of the respondent’s report of it at that time. For example, a respondent who at t = 4 reported having suffered a myocardial infarction between t = 0 and t = 4 would score: t = 0: no myocardial infarction; t = 4 and t = 7: myocardial infarction. As some conditions had rather low prevalence (i.e., stroke, myocardial infarction; claudication), a dichotomous variable was additionally analyzed (one or more of the conditions present: yes/no). Three cognition measures were included: global cognitive functioning, fluid intelligence, and IPS. Global cognitive functioning was measured by the Mini-Mental State Examination (MMSE; Folstein et al. 1975). The MMSE is widely used as a screening instrument for global cognitive dysfunctioning. The MMSE involves recall, orientation, registration, attention, language, and construction. It holds 20 items, and scores range from 0 to 30, higher scores indicating better cognition. Fluid intelligence, defined as the ability to deal with new information, was measured using two subsets of the Raven’s Coloured Progressive Matrices (Raven 1995). Each subset holds 12 items consisting of a picture (matrix) of a pattern from which a section is missing. On the bottom of the page, six possible patterns, including the one fitting, are given. The respondent has to choose the best-fitting pattern. The items increase in difficulty, and so do the two subsets. The total score of the scale ranges from 0 to 24; higher scores indicate better performance. An adapted version of the Coding Task (Savage 1984) was used to assess IPS. Two rows of characters were shown, in which each character in the upper row belonged to a character of the lower row. Two other rows were shown, one of them containing characters and the other one left blank. The respondent had to give (by verbal response) as many as possible correct character combinations within 1 min, using the upper rows as a key. Three trials were performed of which the highest score was used in the present analyses. Smoking was covered by two dichotomous variables: currently smoking (yes/no) and former smoking (former or current smoker versus never smoked). Alcohol use was a categorical variable and defined according to the Netherlands Economic Institute (Reinhard & Rood-Bakker 1998): no use (0 glasses per day), moderate use (1 to 3 [men]/1 to 2 [women] glasses per day), and high/excessive use (≥4 [men]/ ≥3 [women] glasses per day). All independent variables, except gender and education, were considered changeable over time and were included in the models longitudinally, so including values for t = 0, 3, 4, and 7, when available.
Statistical Analyses Descriptive Analyses • Independent samples t tests and oneway analyses of variance were used to compare the mean SRTn scores on t = 0, 3, 4, and 7 years across the different participant groups. The cognition variables were used as dichotomous (MMSE: cut off 24 points) and categorical variables (IPS and fluid intelligence: quartiles) for descriptive purposes in the descriptive analyses; these were analyzed as continuous variables in the effect analyses.
725
Effect Analyses: Change in SRTn Over Time • To examine the change in the SRTn score over time, we performed multilevel analyses (Goldstein 1995). This method accounts for the correlation of observations over time within individuals. Another important advantage is that respondents are allowed to have a different number of repeated measurements; hence no cases are lost due to missing values. This meant that respondents with two or three SRTn scores could be included within the same statistical model. SRTn was included as the dependent variable, and time as the main independent variable. To investigate whether the change in SRTn over time was different for different baseline ages and differed between men and women, interactions were tested. If the interaction (BTime × Gender; BTime × Baseline Age; BTime × Gender ) was significant (p < 0.10), stratified models were pre× Baseline Age sented. In case of significant interaction by baseline age or gender, confounding by gender and baseline age was tested within the baseline age- and gender-stratified models, respectively. In case of significant interaction by baseline age, it was also tested whether SRTn changed in a quadratic fashion for the different age groups. This was done by adding BTime2 and BTime2 × Baseline Age to the model containing BBaseline Age, BTime, and BBaseline Age × Time, and determining whether BTime2 × Baseline Age was significant (p < 0.10). Finally, it was tested whether time of entering the study (i.e., in 1992 or 2002) significantly interacted with time. This way it could be tested whether the SRTn score of respondents entering the study earlier (i.e., in 1992) deteriorated at a different rate than of respondents entering the study later (i.e., in 2002). This was determined for a sample of 445 respondents with baseline ages ranging from 63 to 68 years as only for these respondents the year of entry could differ. Effect Analyses: Factors Influencing the Change in SRTn Over Time • To examine which, and if so, to what extent various factors influence the SRTn change over time, all independent variables were individually added to a model with time (main independent variable) and SRTn (dependent variable). The change in the regression coefficient of time (expressed as percentage change) was determined and was considered relevant if it was at least 10% (Grayson 1987). If multiple factors relevantly influenced the time coefficient, the strongest individual factor was included in the model with time, and subsequently, the second strongest individual relevant factor was added until no further relevant change in the coefficient of time was observed or the statistical significance of the time coefficient exceeded a p value of 0.10. Finally, in a model containing the relevantly influencing factor(s) only, all remaining factors were individually added once more and any additional change in the coefficient of the time variable was assessed. We hereby aimed to determine whether the relevant factors indeed were the primary influencing factors. Statistical Software • Cross-sectional analyses were performed in Statistical Package for the Social Sciences (SPSS, v. 20.0). Multilevel analyses were performed in MLwiN (v. 2.22; Centre for Multilevel Modeling, Institute of Education, London, United Kingdom).
RESULTS Descriptives The total sample (n = 1298) comprised 697 women (54%) and 601 men (46%). Mean baseline age was 70 years (SD = 7). Mean follow-up was 4.9 years. Table 2 shows mean SRTns
Total sample
667 (54)
Women
36 (3)
464 (38)
Yes
880 (71)
−5.93 (2.17)
−5.76 (2.21)
−6.13 (1.92)
0.208
0.050
hallmark of ARHL is pronounced sensitivity loss in the high frequencies. ARHL is characterized by reduced hearing sensitivity and speech understanding in noisy environments, slowed central auditory processing, and impaired sound localization (Gates & Mills 2005). For several decades, cross-sectional pure-tone audiometry data are used to estimate age- and gender-specific hearing loss trajectories, and since the 1990s, longitudinal data from large cohort studies are increasingly used for this purpose (e.g., Gates et al. 1990; Morell et al. 1996; Cruickshanks et al. 2003; Kiely et al. 2012). The rate of hearing decline seems to be frequency-specific and dependent on age and gender, although mixed evidence exists. Some studies found faster declines in men across most frequencies up to very old age (Pearson et al. 1995; Chao & Chen 2009), others reported faster frequencyspecific declines in women (Wiley et al. 2008; Kiely et al. 2012), and yet others found no gender differences (Gates et al. 1990; Cruickshanks et al. 2003). Regarding age, in younger old (up to around 70 years) faster hearing declines are generally observed in the high frequencies, while for the older old, this is in the low frequencies (Wiley et al. 2008; Chao & Chen 2009; Kiely et al. 2012). ARHL has long been considered as a natural aging process with associated histological, electrophysiological, and molecular changes in human tissues, causing deficits in hair cells, cochlear neurons, the stria vascularis, and tissues more upstream in the central nervous system (Gates & Mills 2005). Nowadays, ARHL is increasingly recognized as a complex condition the etiology of which is additionally influenced by environmental, healthrelated, and genetic factors (Van Eyken et al. 2007). Noise exposure is the environmental factor most extensively studied. Although excessive noise levels evidently cause hearing loss, it is less clear whether past noise exposure affects the rate of ARHL later in life (Gates et al. 2000; Cruickshanks et al. 2003; Van Eyken et al. 2007; Dubno et al. 2008; Kiely et al. 2012). Smoking and alcohol abuse are other possible risk factors, but mixed results exist: whereas some reported an increased hearing loss risk, others reported none (Rosenhall et al. 1993; Brant et al. 1996; Cruickshanks et al. 1998b; Van Eyken et al. 2007). The same holds for diabetes, cardiovascular disease, and hypertension. Various studies identified them as risk factors for ARHL, but others did not (e.g., Brant et al. 1996; Frisina et al. 2006; Kiely et al. 2012). Most of the aforementioned studies hold important limitations. One is the cross-sectional study design, which limits drawing strong conclusions about causation. The few studies that did use a longitudinal approach mostly failed to find significant risk factors (Dubno et al. 2008; Gopinath et al. 2009, 2010; Mitchell et al. 2009). Another is the study of a certain risk factor in
Design: Data covering 3 to 7 years of follow-up (mean: 4.9 years) of a large sample of the Longitudinal Aging Study Amsterdam were used (n = 1298; 3025 observations; baseline ages: 57 to 93 years). Hearing ability was measured by a digit triplet speech-in-noise test (SNT) yielding a speech reception threshold in noise (SRTn). Multilevel analyses were used to model the change in SRTn over time. First, interaction terms were used to test differences in rate of decline across subgroups. Second, for each of the following factors the authors determined the influence on the change in SRTn: age, gender, educational level, cardiovascular conditions, information processing speed, fluid intelligence, global cognitive functioning, smoking, and alcohol use. This was done by calculating the percentage change in Btime after adding the particular factor to the model. Results: On average, respondents’ SRTn increased (i.e., deteriorated) significantly over time by 0.18 dB signal-to-noise ratio per annum. Rates were accelerated for older ages (Btime = 0.13, 0.14, 0.25, 0.27 for persons who were 57 to 65, 65 to 75, 75 to 85, and 85 to 93 years of age, respectively). Only information processing speed relevantly influenced the change in SRTn over time (17% decrease in Btime). Conclusions: Decline in older persons’ speech-in-noise recognition over time accelerated for older ages. Decline in information processing speed explained a moderate proportion of the SRTn decline. This indicates the relevance of declining cognitive abilities in the ability of older persons to recognize speech in noisy environments. (Ear and Hearing 2013;34;722–732)
INTRODUCTION Hearing loss is one of the most prevalent chronic conditions in old age (Davis 1990; Ries 1994). Several studies described prevalence rates up to 40% in adults 50 years and older (Duijvestijn et al. 1999) and up to 90% in persons more than 80 years of age (Cruickshanks et al. 1998a). Worldwide, adultonset hearing loss is the second leading cause of years lived in disability (World Health Organization 2008). Many studies found significant relationships with poor psychosocial health such as anxiety, depression, and loneliness (e.g., Kramer et al. 2005; Pronk et al. 2011). The bulk of the hearing loss cases can be classified as agerelated hearing loss (ARHL). The traditional audiological Departments of 1Ear Nose Throat/Audiology, 2Epidemiology and Biostatistics, 3Psychiatry/GGZ inGeest, VU University Medical Center, EMGO Institute for Health and Care Research, Amsterdam, the Netherlands; and 4Department of Health Sciences, Faculty of Earth and Life Sciences, VU University Amsterdam, the Netherlands.
0196/0202/13/3406-0722/0 • Ear and Hearing • Copyright © 2013 by Lippincott Williams & Wilkins • Printed in the U.S.A. 722
PRONK ET AL. / EAR & HEARING, VOL. 34, NO. 6, 722–732
isolation. Thus, most studies could not elucidate what the relative contribution of various factors to ARHL was. Last, except for a study by Dubno et al. (2008), all studies used pure-tone thresholds as their hearing loss outcome. Pure-tone audiometry is still considered the gold standard assessment for hearing loss, but its relationship with experienced daily life disability is limited (e.g., Demeester et al. 2012), especially with regard to selfreported speech understanding in noise (Kramer et al. 1996) and during group conversations (Gatehouse & Noble 2004). Trajectories based on the performance on speech-in-noise tests (SNTs) over time have been scarcely examined. Such investigations are especially warranted from the patient’s perspective, as understanding speech in challenging listening situations, such as in noise, is the most frequently reported disability (Kramer et al. 1998). Longitudinal decline in speech perception abilities was examined by Bergman et al. (1976), Møller (1981), Pedersen et al. (1991), Hietanen et al. (2004), Divenyi et al. (2005), and Dubno et al. (2008), but most of these studies were limited by a small (n ≤ 35) sample size of older subjects, included only one birth cohort combined with a relatively short follow-up, or included a non–population-based sample. Moreover, only two (Divenyi et al. 2005; Dubno et al. 2008) included an SNT (the revised SPIN test). Dubno et al. (2008) used data of a population-based sample of 85 older adults (56 to 81 years at baseline) who were measured four times across a 7- to 13-year period. Only marginal or nonsignificant rates of decline over time were reported. No gender differences in the rate of decline appeared, nor did the rate differ across age groups or between those with or without noise history. Divenyi et al. (2005) reported a statistically significant decline in their small sample (n = 29) across 4 to 8.5 years of follow-up, but they did not report the rate of this decline. Speech-based measures may capture central auditory processing functions that decline with increasing age. This concerns functions required to distinguish pitch, loudness, and duration of acoustical signals (Pichora-Fuller & Souza 2003), but also cover more global cognitive functions such as information processing speed (IPS) and working memory (Wingfield & Tun 2007; Akeroyd 2008; Rönnberg et al. 2008; Arlinger et al. 2009). It is generally assumed that the involvement of these neurocognitive functions becomes more pronounced in challenging listening situations such as noisy backgrounds (e.g., Pichora-Fuller & Souza 2003; Koelewijn et al. 2012). Thus, to map functional hearing that goes beyond audibility, mapping trajectories of the ability to recognize speech in noise seems essential. The fact that often correlations between speech-innoise measures and pure-tone measures of 0.4 to 0.7 are found suggests their partly differing underlying constructs (Houtgast & Festen 2008). Divenyi et al. (2005) found that older persons’ performance on speech-recognition tasks declined with increasing age more rapidly than their audiometric performance. Accordingly, they concluded that decline in auditory performance during the later decades of life is due to the degeneration of peripheral auditory structures, but additionally involves a non-peripheral (i.e., central) component. The aim of the present study is to investigate the decline in older persons’ ability to recognize speech in noise over time and the factors influencing this decline. Multilevel modeling will be applied using longitudinal SN data derived from a large Dutch older population-based cohort. First, the mean decline in SN recognition over time will be determined. We will investigate
723
whether the decline differs across age and gender. Second, we will examine to what extent demographic (age, gender, level of education), health-related (cardiovascular conditions), environmental (alcohol use, smoking), and cognitive factors (global cognitive functioning, information processing speed, henceforth IPS, fluid intelligence) influence the decline in speech-in-noise recognition over time and determine their relative contribution.
MATERIALS AND METHODS Sample and Procedures The sample originated from the Longitudinal Aging Study Amsterdam (LASA) cohort (Huisman et al. 2011). LASA is an ongoing cohort study on predictors and consequences of changes in autonomy and wellbeing in an aging population. A random sample of 3107 older persons (aged 55 to 85 years) stratified for age and gender was drawn from the Dutch population for the first LASA measurement in 1992/1993. A second cohort of 1002 respondents (aged 55 to 64 years) was added in 2002 from the same sampling frame as the original cohort from 1992. A followup measurement was conducted every 3 to 4 years. All measurements were performed in the respondent’s home by trained and supervised interviewers. Informed consent was obtained from all respondents. The study was approved by the Medical Ethics Committee of the Vrije Universiteit University Medical Center. For the present study, the original cohort and the second cohort will be referred to as the 1992 cohort and the 2002 cohort, respectively. In 2001/2002 (1992 cohort only), 2005/2006 (both cohorts), and in 2008/2009 (both cohorts), speech-in-noise recognition was measured with an SNT, yielding a speech-reception threshold in noise (SRTn). We selected respondents who had complete SNT data for at least two measurements such that intraindividual change would be sufficiently represented in the statistical models. We tested whether there was selective attrition by comparing age, gender, and SRTn characteristics of those lost to follow-up with those remaining in the sample. Loss to follow-up due to death was regarded as being population-based and was therefore not considered. In general, those lost to follow-up were older and had poorer (higher) SRTns (see Table 1). To identify outliers of the SRTn scores, we linearly regressed the 2005/2006 SRTns on the 2001/2002 SRTns (n = 591), the 2008/2009 SRTns on the 2005/2006 SRTns (n = 1000), and the 2008/2009 SRTns on the 2001/2002 SRTns (n = 136). Thirty-five data points were situated more than 3 SD outside the regression lines and were omitted. The distribution of the respondents and the observations over the measurements is depicted in Figure 1. In total, 1298 respondents provided 3025 SRTn observations. Seven hundred eighty respondents originated from the 1992 cohort, and 518 originated from the 2002 cohort.
Measures In the analyses, the SRTn was used as the dependent variable, and time, demographical, health-related, environmental, and cognitive variables were used as independent variables.
Dependent Variable: SRTn
The SNT was originally developed as a functional hearing screening self-test by telephone (Smits et al. 2004). The test
724
PRONK ET AL. / EAR & HEARING, VOL. 34, NO. 6, 722–732
TABLE 1. Selective attrition between measurements 1998/1999 and 2001/2002 (1992 cohort), 2002 and 2005/2006 (2002 cohort); 2001/2002 and 2005/2006 (1992 cohort); and between 2005/2006 and 2008/2009 (both cohorts) 1998/1999 and 2001/2002 Variable
2002 and 2005/2006
1992 cohort (nattrition = 626) * * —
Age (older) Gender (female) SRTn (higher)
2002 cohort (nattrition = 346) ns ns —
2001/2002 and 2005/2006 1992 cohort (nattrition = 212) * ns *
2005/2006 and 2008/2009 1992 and 2002 cohort (nattrition = 184) ns * *
ns, that is, variable not differing significantly between those lost to follow-up and those remaining in the study sample; —: no SRTn measurement for this cohort in 1998/1999 (1992 cohort) and 2002 (2002 cohort); SRTn, speech reception threshold in noise in dB signal-to-noise ratio. *Variable differing significantly (p < 0.05) between those lost to follow-up and those remaining in the study sample. nattrition, number of respondents lost to follow-up.
determines an individual’s SRTn defined as the signal-to-noise ratio (SNR) in dB corresponding to 50% recognition of digit triplets. The interviewer brought portable testing equipment comprising a telephone, an amplifier, and headphones. At the 2008/2009 measurement, the telephone and amplifier were replaced by a laptop. Before the actual assessment, hearing aids had to be removed and the respondent was instructed to adjust the level of the speech to make sure that signals were 1992 cohort
2002 cohort nobs
TOTAL nobs
727
518
1245
-
0
518
518
x
644
-
0
644
618
-
0
618
1036
3025
nobs t=0 t=3
x -
-
-
t=4 t=7
x
TOTAL nobs
1989
TOTAL nresp
nresp
429
162
53
136
780
TOTAL nresp
518
1298
indicates longitudinal study design reading order SNT data available x
no SNT data available
-
no measurement for this cohort
nobs
number of observations
nresp
number of respondents
t=0
0 years of f.u.; 2001/2002 data (1992 cohort): 2005/2006 data (2002 cohort)
t=3
3 years of f.u.; 2008/2009 data (2002 cohort)
t=4
4 years of f.u.; 2005/2006 data (1992 cohort)
t=7
7 years of f.u.; 2008/2009 data (1992 cohort)
SNT: speech-in-noise test; f.u.: follow-up
Fig. 1. Distribution of the respondents and the observations over the measurements.
sufficiently audible. Subsequently, 23 different monosyllabic digit triplets were presented at different intensity levels against a constant level of stationary background noise. An adaptive up–down procedure was used in which the SNR decreased by 2 dB if the respondent repeated a triplet correctly, and increased by 2 dB after an incorrectly repeated triplet. Different administration and data collection modes caused slight but systematic shifts in the SRTns. The followup scores were therefore corrected. The 2005/2006 shift was estimated by determining age–gender specific SRTn averages at 2001/2002 and 2005/2006. Subsequently, the 2001/2002 averages were linearly regressed on the 2005/2006 averages. After the regression coefficient was set at 1, the constant of the regression equation (i.e., the systematic shift) could be estimated. The same procedure was applied to the 2008/2009 data. The identified shifts were −0.86 dB SNR at 2005/2006, and −0.49 dB SNR at 2008/2009 (both relative to scores at 2001/2002). The SNT has a good validity as indicated by its high correlation (r = 0.87) with the standard Dutch sentences SNT (Plomp & Mimpen 1979). Its test–retest reliability seemed satisfactory in an older subsample from the study by Nachtegaal et al. (2009) (intraclass correlation coefficient; two-way random effects model = 0.70; n = 152; 58 to 82 years). The estimated measurement error is about 1 dB, and is slightly larger for worse SRTns, that is, up to 1.3 dB for SRTns around −1 dB SNR (Smits et al. 2004).
Independent Variables Time in years was included as a continuous variable in the models and was calculated as “time in the study,” taking 2001/2002 (1992 cohort) and 2005/2006 (2002 cohort) as baseline. As such, for the 1992 cohort, the time values included: 0 (2001/2002); 4 (2005/2006), and 7 (2008/2009) years. For the 2002 cohort, these were 0 (2005/2006) and 3 (2008/2009) years. See also Figure 1. Age at baseline was included as a categorical variable (four 10-year groups) because in the first part of the effect analyses (i.e., to determine the change in SRTn over time), the categorical variant fitted the longitudinal SRTn data better than the continuous variant. Gender was included dichotomously. Educational level was assessed by asking the respondent’s highest educational level completed and was categorized into low (uncompleted elementary, elementary, lower vocational), medium (general intermediate, intermediate vocational, general secondary, higher vocational), and high (college and university). Information on cardiovascular conditions was derived from
PRONK ET AL. / EAR & HEARING, VOL. 34, NO. 6, 722–732
self-report and included: incident stroke and myocardial infarction, and presence of claudication, diabetes mellitus, and hypertension. Because diabetes is a chronic condition and the other cardiovascular conditions were assumed to have major and possibly a (life) long impact, the occurrence at one measurement was carried forward to later measurements irrespective of the respondent’s report of it at that time. For example, a respondent who at t = 4 reported having suffered a myocardial infarction between t = 0 and t = 4 would score: t = 0: no myocardial infarction; t = 4 and t = 7: myocardial infarction. As some conditions had rather low prevalence (i.e., stroke, myocardial infarction; claudication), a dichotomous variable was additionally analyzed (one or more of the conditions present: yes/no). Three cognition measures were included: global cognitive functioning, fluid intelligence, and IPS. Global cognitive functioning was measured by the Mini-Mental State Examination (MMSE; Folstein et al. 1975). The MMSE is widely used as a screening instrument for global cognitive dysfunctioning. The MMSE involves recall, orientation, registration, attention, language, and construction. It holds 20 items, and scores range from 0 to 30, higher scores indicating better cognition. Fluid intelligence, defined as the ability to deal with new information, was measured using two subsets of the Raven’s Coloured Progressive Matrices (Raven 1995). Each subset holds 12 items consisting of a picture (matrix) of a pattern from which a section is missing. On the bottom of the page, six possible patterns, including the one fitting, are given. The respondent has to choose the best-fitting pattern. The items increase in difficulty, and so do the two subsets. The total score of the scale ranges from 0 to 24; higher scores indicate better performance. An adapted version of the Coding Task (Savage 1984) was used to assess IPS. Two rows of characters were shown, in which each character in the upper row belonged to a character of the lower row. Two other rows were shown, one of them containing characters and the other one left blank. The respondent had to give (by verbal response) as many as possible correct character combinations within 1 min, using the upper rows as a key. Three trials were performed of which the highest score was used in the present analyses. Smoking was covered by two dichotomous variables: currently smoking (yes/no) and former smoking (former or current smoker versus never smoked). Alcohol use was a categorical variable and defined according to the Netherlands Economic Institute (Reinhard & Rood-Bakker 1998): no use (0 glasses per day), moderate use (1 to 3 [men]/1 to 2 [women] glasses per day), and high/excessive use (≥4 [men]/ ≥3 [women] glasses per day). All independent variables, except gender and education, were considered changeable over time and were included in the models longitudinally, so including values for t = 0, 3, 4, and 7, when available.
Statistical Analyses Descriptive Analyses • Independent samples t tests and oneway analyses of variance were used to compare the mean SRTn scores on t = 0, 3, 4, and 7 years across the different participant groups. The cognition variables were used as dichotomous (MMSE: cut off 24 points) and categorical variables (IPS and fluid intelligence: quartiles) for descriptive purposes in the descriptive analyses; these were analyzed as continuous variables in the effect analyses.
725
Effect Analyses: Change in SRTn Over Time • To examine the change in the SRTn score over time, we performed multilevel analyses (Goldstein 1995). This method accounts for the correlation of observations over time within individuals. Another important advantage is that respondents are allowed to have a different number of repeated measurements; hence no cases are lost due to missing values. This meant that respondents with two or three SRTn scores could be included within the same statistical model. SRTn was included as the dependent variable, and time as the main independent variable. To investigate whether the change in SRTn over time was different for different baseline ages and differed between men and women, interactions were tested. If the interaction (BTime × Gender; BTime × Baseline Age; BTime × Gender ) was significant (p < 0.10), stratified models were pre× Baseline Age sented. In case of significant interaction by baseline age or gender, confounding by gender and baseline age was tested within the baseline age- and gender-stratified models, respectively. In case of significant interaction by baseline age, it was also tested whether SRTn changed in a quadratic fashion for the different age groups. This was done by adding BTime2 and BTime2 × Baseline Age to the model containing BBaseline Age, BTime, and BBaseline Age × Time, and determining whether BTime2 × Baseline Age was significant (p < 0.10). Finally, it was tested whether time of entering the study (i.e., in 1992 or 2002) significantly interacted with time. This way it could be tested whether the SRTn score of respondents entering the study earlier (i.e., in 1992) deteriorated at a different rate than of respondents entering the study later (i.e., in 2002). This was determined for a sample of 445 respondents with baseline ages ranging from 63 to 68 years as only for these respondents the year of entry could differ. Effect Analyses: Factors Influencing the Change in SRTn Over Time • To examine which, and if so, to what extent various factors influence the SRTn change over time, all independent variables were individually added to a model with time (main independent variable) and SRTn (dependent variable). The change in the regression coefficient of time (expressed as percentage change) was determined and was considered relevant if it was at least 10% (Grayson 1987). If multiple factors relevantly influenced the time coefficient, the strongest individual factor was included in the model with time, and subsequently, the second strongest individual relevant factor was added until no further relevant change in the coefficient of time was observed or the statistical significance of the time coefficient exceeded a p value of 0.10. Finally, in a model containing the relevantly influencing factor(s) only, all remaining factors were individually added once more and any additional change in the coefficient of the time variable was assessed. We hereby aimed to determine whether the relevant factors indeed were the primary influencing factors. Statistical Software • Cross-sectional analyses were performed in Statistical Package for the Social Sciences (SPSS, v. 20.0). Multilevel analyses were performed in MLwiN (v. 2.22; Centre for Multilevel Modeling, Institute of Education, London, United Kingdom).
RESULTS Descriptives The total sample (n = 1298) comprised 697 women (54%) and 601 men (46%). Mean baseline age was 70 years (SD = 7). Mean follow-up was 4.9 years. Table 2 shows mean SRTns
Total sample
667 (54)
Women
36 (3)
464 (38)
Yes
880 (71)
−5.93 (2.17)
−5.76 (2.21)
−6.13 (1.92)
0.208
0.050
