Quality, Rigour and Usefulness of Free-Text Comments Collected by a Large Population Based Longitudinal Study - ALSWH

While it is common practice for health surveys to include an open-ended question asking for additional comments, the responses to these questions are often not analysed or used by researchers as data. The current project employed an automated semantic program to assess the useability and thematic content of the responses to an open-ended free response item included in the Australian Longitudinal Study on Women’s Health (ALSWH) surveys. The study examined the comments of three cohorts of women, born between 1973–78, 1946–51, and 1921–26, from Survey 1 (in 1996) and Survey 5 (in 2007–2009). Findings revealed important differences in the health status of responders compared to non-responders. Across all three cohorts, and at both time points, women who commented tended to have poorer physical health (except for women aged 82–87) and social functioning, experienced more life events, were less likely to be partnered, and (except for women aged 18–23 years) more likely to have higher levels of education, than women who did not comment. Results for mental health were mixed. The analysis revealed differences between cohorts as well as changes over time. The most common themes to emerge for the 1973–78 cohort were health, time, pregnant and work, for the 1946–51 cohort, the most common themes were health, life, time and work, while for the 1921–26 cohort, the most common themes were husband, health and family. The concepts and frequency of concepts changed from the first to the fifth survey. For women in the 1973–78 cohort, pregnant emerged as a prevalent theme, while eating disappeared. Among women in the 1946–51 cohort, cancer, operation and medication emerged as prevalent themes, while for women in the 1921–26 cohort, the concept children disappeared, while family emerged. This analysis suggests that free-text comments are a valuable data source, suitable for content, thematic and narrative analysis, particularly when collected over time.


Introduction
While many health surveys include an open-ended question to allow participants to provide additional comments and information, the responses to these questions are not often examined beyond surface-level analysis. The sole reliance on numerical survey data has been criticised, with researchers pointing to the interpretive value of descriptive comments [1]. The detail provided by free-text comments may offer an important context for participant responses and reveal issues that cannot be identified using purely quantitative surveys [2]. For example, in a study using Likert Scales and free-text comments to assess quality of life among homeless people, students, and a town population, important discrepancies emerged between the two response methods [3]. While homeless people had similar or better quantitative ratings on several of the health-related measures, their free-text accounts revealed a range of unique difficulties, which appeared to contradict the quantitative results, indicating the value of qualitative comments in evaluating health.
Little is known about the health profile of respondents to openended questions compared to non-respondents in terms of representativeness of responses. It has been argued that respon-dents may not be representative of the population surveyed. In a review of free-text comments from a study of maternity care involving recent mothers, and a longitudinal cohort study of UK medical graduates' careers, Garcia and colleagues [4] found free text respondents were likely to be more articulate, or to have a negative comment to make than non-respondents. These findings suggest that biases may exist in open-ended datasets, however little is known about the more general biases such as demographic factors and overall quality of life.
The potential of free-text responses, particularly in a longitudinal context have increasingly been acknowledged for their utility and narrative potentials. Analysis of free-text data collected over time enables researchers to explore changes in how participants construct meaning across the life course [5,6,7,8,9]. The growing body of literature that has analysed free-text comments begins to suggest that qualitative data sets are of intrinsic value and can be analysed for more than survey evaluation purposes.
The current project employed an automated semantic program to assess the useability and thematic content of the responses to an open-ended free response item included in the Australian Longitudinal Study on Women's Health (ALSWH). These data are well suited to interrogation due to the large number of participants from three broadly representative samples [10].
The aims of this project were to assess the quality, rigour and usefulness of the comments collected by the ALSWH in order to validate the targeted analysis of these comments. Additionally the health status of responders compared to non-responders was assessed.

Methods
This project and the data used in the analysis have received all relevant ethics approvals. Ethics approval was granted by the University of Newcastle (H-076-0795) and the University of Queensland Human Research Ethics Committees (2004000224). Written informed consent was obtained from all participants. This consent procedure was approved by both ethics committees.
The Australian Longitudinal Study on Women's Health (ALSWH) has been collecting postal survey data from three  [11]. In addition to completing quantitative questions on their health and lifestyle, participants are also asked ''Have we missed anything?'' and given the opportunity to answer the question in an open-ended format. Since the study's inception in 1996 the ALSWH has been collecting these comments, often using them to provide further insights into quantitative analysis [11] and more recently for in-depth qualitative analysis [8]. This paper presents examples from two separate time points to highlight the validity of free-text comments as data, Surveys 1 and 5 are presented to exemplify this. Through the use of Leximancer software the researchers were able to provide visual maps of the qualitative data, uncovering the most common themes, words and relationships. This software uses word-association information to elicit emergent concepts from the text [12]. Word frequency and location are used to generate taxonomies which are presented as maps showing the relationships between common concepts. Leximancer software was used to conduct an automatic analysis of the ALSWH qualitative data sets.
This analytical tool enables the generation of a taxonomy which is derived from the data itself, offering an efficient way of conceptually mapping, in this case, large data sets. This type of automatic analysis is grounded and exploratory, as the researcher is removed from making any judgements or preconceptions regarding the coding of the data. Leximancer quantifies large text documents using a classification system of learned lexical concepts rather than just keywords [13].
As this project was exploratory in nature, simply to uncover what women wrote about over the study period, all default settings of Leximancer were maintained with one exception. Leximancer by default reads the data in two-sentence blocks, as generally in common language two sentences presented together contain similar content and topics. However, for this project this setting was changed to analyse one-sentence blocks because the nature of free-text comments is that often the sentences are not continuous in content and are rather spontaneous and short in nature.
Descriptive quantitative analyses (t-tests and chi-squares tests) were additionally undertaken to determine demographic and health-related differences between those women who responded to the free-text item and those who did not respond. Physical and Table 3. 1946-51 Cohort quantitative measures S1 and S5 for those who commented and those who did not comment.

Results
This analysis included data from Survey 1 (in 1996) and Survey 5 (in 2007-2009) of the three ALSWH cohorts. Table 1

Quantitative Results
1973-78 cohort. At Survey 1 (in 1996), those participants who had significantly lower scores across each of the domains assessed on the SF-36 including mental health, general health, physical functioning and social functioning (Table 2), indicating they had poorer health than those who did not comment were worse off across these domains. Participants who also had significantly higher proportional scores on the life events scale, indicating they experienced more life events than those participants who did not. There were no significant differences across demographic measures for the two groups. The results for Survey 5 (in 2009) for the 1973-78 Cohort can also be found in Table 2. At this time there were no significant differences on the Mental Health Index, however all other subscales of the SF-36 were significantly lower for those women who commented. There was also a significant difference detected for education, with those women who commented more likely to have higher levels of education than those women who did not comment. 1946-51 cohort. At Survey 1 (in 1996), those participants who commented had significantly lower scores across all SF-36 domains, with the exception of social functioning. Those who commented also had significantly higher proportional scores on the life events scale. Women who commented were also more likely to be un-partnered, and had higher levels of education than those women who did not comment. Similar results were reported at Survey 5 for all SF-36 domains, life events scores and demographic measures (see Table 3).
1921-26 cohort. At Survey 1 (in 1996), participants who commented had significantly lower scores for Physical Functioning, General Health and Social Functioning scores on the SF-36 compared with women who did not comment. Mental Health Index scores were not significantly different. Significant differences were also detected for higher proportional scores on the life events scale, higher levels of education and partner status. At Survey 5, significant differences were only detected for the Social Functioning score, with women who commented significantly more likely to have lower social functioning scores on the SF-36 compared with women did not write comment. Women who commented were also significantly more likely than women who did not comment to experience a higher proportion of life events, and had a higher level of education (see Table 4). Qualitative Results 1973-78 leximancer analysis results. The data set of the 1973-78 cohort of the ALSWH includes a diverse range of comments regarding the health and life of the participants. As seen in Figure 1, 1973-78 Cohort Survey 1 data, the Leximancer software maps the themes according to frequency of words and connectedness of words to other words i.e. pregnancy and child. As a result the maps create concept circles that are heat mapped. The hot (most frequent and most connected i.e. word association) colours through to cooler colours (i.e. the red, orange through to cooler colours such as green and blue to purple) are able indicate meaning and relationships.
In the 1973-78 data set the hottest coloured theme was Time followed by Health and Family. The size of the circles indicates the interconnectedness of each of the themes. For instance, Time has overlapped with Year (which is not a high frequency colour being Green) however the size highlights the importance of the relationship between Time and Year. So it is clear that through a visual analysis of the Leximancer map the warmth of the colour is important but also the interconnectedness between the circles is important when exploring the comments made by the women.
Within the main theme of Time, key words such as time, months, stress, work, due, past, times, weeks, child, week, depression, relationship derived from the data. The results of this map in comparison to When the 1973-78 Cohort was surveyed approximately 15 years later, the themes and concepts showed transitions and developments. In this map, Figure 2, the main theme is Work, followed by Pregnant, Questions, Months and Survey. This indicates that the women used the free-text space to comment on the survey and the questions in the study. However, this also indicates that issues surrounding Work and Pregnancy are dominating these women's comments, which is dissimilar to the earlier survey where Pregnancy was not a theme.
Within the theme of Work, significant concepts words included work, time, week, home, having, days, months, job.
1946-51 leximancer analysis results. The map of the 1946-51 Cohort data are a rich collection of assorted comments, raising some similar and different issues to the younger cohort. In the first survey of this cohort, the most common theme was Health, followed by Life and Time. Within the theme of Health, concept words found within that theme include health, feel, women, people, better, answers, things, believe, question and etc. (Figure 3). These concepts indicate that the 1946-51 Cohort communicate with the ALSWH researchers about their feelings and health related to the survey items.
It is interesting to note that in survey 5 the main themes, or the hottest colours, have changed slightly. Over time, Work has shifted to become the most prominent theme derived from the data. Work is interlinked with themes such as Husband, Live and Difficult. Within the theme of Work are the concepts, time,health, life, home, mother, full, living, age and job (Figure 4)  1921-26 leximancer analysis results. The map of the 1921-26 Cohort comments is the largest collection of ALSWH comments. As with the other datasets, the 1921-26 Cohort data are rich in diversity of themes, contents and experiences. The participants in this cohort generally wrote longer comments (and often in letter style) and are were likely to write more frequently, that is,. several survey waves.
In the 1921-26 first survey the theme of Husband was the warmest colour. Included in this theme were the concepts life, family, happy, old, live, friends, living, died, mother, people, age, years and full ( Figure 5). It is interesting, that at Survey 5, twelve years after the first survey, the theme Husband remained the hottest theme ( Figure 6).
In 2008, under the theme of Husband, the concepts changed since 1996. These newer concepts include home, week, care, daughter, son, and car. There is a noticeable shift in the types of comments these older women were writing in 2008 in comparison to the 1996 data. For example, the term 'children' has disappeared from this data set in 2008 and been replaced by the emergent term 'family' as these women moved from their mid-seventies into very old age. Discussion Across all cohorts, women who wrote free text comments tended to have poorer physical health and to have higher levels of education. Results for psychological wellbeing were inconsistent. In addition those results for the 1921-26 Cohort at Survey 5 were inconsistent, possibly due to the number of women who have died or withdrawn from the study by Survey 5, due to the high numbers of participants who commented in this cohort by this time.
Examination of the concept maps revealed differences between the three cohorts. The most common themes were for the 1973-78 Cohort, were health, time, pregnant and work, for the 1946-51 Cohort, were health, life, time and work, and for the 1921-26 Cohort, were husband, health and family. All cohorts wrote about health related concepts, as would be expected in a health survey. Common themes included: eating and pain from the 1973-78 Cohort, weight, cancer and medication from the 1946-51 Cohort, and arthritis, eye and pain from the 1921-26 Cohort.
The concepts and frequency of concepts changed from the first to the fifth survey. For women in the 1973-78 Cohort this included the emergence of pregnant and the disappearance of eating as prevalent themes among women moving from their early twenties to their early thirties. For women in the 1946-51 Cohort this included the emergence of cancer, operation and medication as prevalent themes among women moving from their middle forties to their fifties. And for the 1921-26 Cohort, this included the disappearance of the concept children and emergence of the concept family among women moving from their mid seventies into very old age. The maps (see Figures 1,2,3,4,5,6) also illustrate the intersections between concepts and the changes that occur in these intersections over time.

Assessing Quality through the Kitto et al (2008) Framework
In assessing this data the authors have applied a framework explained by Kitto and colleagues [16] which examines qualitative validity. This framework sets some ground rules for assessing quality in qualitative data and was recently published in an edition of the Medical Journal of Australia, applied below [16].
Clarification. This aim of this paper was to reveal via a Leximancer analysis what the women in the ALSWH write about at the end of a survey, the health differences between those who write and those who do not write, and to assess whether or not the data collected is a viable option for research.
Justification. This was an important research question to ask of the ALSWH data as it has never been asked before. Surveys and questionnaires often ask their participants if there is anything else they would like to tell the research team, but, to date, this is a largely untapped source of research data. Often, these data are only used only for quality assurance or evaluation of survey items.
Procedural rigour. All processes of the data collection have been documented and approved by relevant ethical boards. The data used in this analysis have not been edited or changed by researchers, apart from being typed into a database. Leximancer is an automated program based on word count and word association; therefore, this analysis is relatively unbiased and naturally derived from the original data.
Representativeness. This study is a nation-wide longitudinal Australian study. The sample is broadly representative of the population of Australian women in included age groups. Not every participant has written comments; however, many have, as Table 1 details.
Interpretation. This data analysis was conducted by an automated computer program, Leximancer, which consequently has reduced researcher interpretation bias.
Reflexivity and evaluative rigour. This analysis strongly confirmed the views of the researchers, that in fact, qualitative free text comments are a valid source of data, particularly when collected over time. These type of data can be used for content, thematic, narrative and other forms of qualitative analysis and are especially useful when collected over time.
Transferability. The authors conclude that this method can be transferred to similar contexts. As mentioned, the process of collecting free-text comments is common among population and epidemiological studies; this study concludes that these qualitative data can be an asset to these studies and further understandings of particular phenomena.

Contributions and Limitations of this Study
This study was unique in its protocol and analysis. Never before have free text comments collected over a 15 year time period been subjected to a Leximancer analysis. This study has the capacity to encourage other survey based studies to analyse qualitative comments. An important limitation of this study is that only comments written by women who participate in the ALSWH have been analysed, therefore there may be other themes that have not been included in this study, which are of importance to women in Australia who did not write at the back of the ALSWH surveys. A further limitation has been identified by the quantitative analysis. Broadly speaking, those participants who commented were of poorer health, un-partnered and had higher levels of education compared with women who did not comment. Nonetheless, these comments provide valuable insight in to the health, wellbeing and lifestyle of Australian women over time.

Conclusion
This analysis of free text comment from a longitudinal study is novel, and as far as the researchers have found, it has never before been validated that free-text comments collected over time can be used as an effective and justified data source. Free text response offers a rich source of data suitable for content, thematic and narrative analysis, particularly when collected over time.