Saturday , February 24 2018
Latest Article
Home / Blog / Detecting the Health Status of Multiethnic Women in Taiwan
Detecting the Health Status of Multiethnic Women in Taiwan
Asian women portrait in a modern restaurant.

Detecting the Health Status of Multiethnic Women in Taiwan

Title: Detecting the Health Status of Multiethnic Women in Taiwan: A Bayesian

Approach using the SF-36 Questionnaire

Author: Meng-ta Chung

Corresponding author: Meng-ta Chung

Affiliation: Chang Gung University of Science and Technology, Taoyuan, Taiwan




Purpose: The purpose of this study was to estimate a cutoff score using the 36-item Short Form Health Survey to predict the health status for women from different ethnic groups in Taiwan. Method: Total participants comprised 219 Taiwanese, 145 Aborigines and 166 Vietnamese. Applying the single score of the SF-36, a Bayesian logistic model estimated the cutoff score for each ethnic group. In addition to using a single SF-36 score, we also used scores from the 8 subscales of the SF-36 to predict the classification of each participant. A Bayesian discriminant analysis based on Logistic model predicted the likelihood that a participant would be categorized into either a diseased or healthy group. We then calculated the overall accuracy rates to see how accurate the model classified participants. Results: Results from the Bayesian logistic model showed that the cutoff score was 45 for Taiwanese, 30 for Aborigines, 22 for and for Vietnamese. Using the 8 subscales of SF-36, the Bayesian discriminant analysis indicated a satisfactory classification with an accurate rate of 71% for Taiwanese, 65% for Aborigines and 73% for Vietnamese.

Key Words: Women’s Health, Multiethnicity, SF-36, Healthcare, Bayesian


Since the 1970’s the United Nations has been highly concerned with women’s health issues.1 Although women’s health issues have attained higher international prominence in recent decades,2 women are still disadvantaged by having limited access to health related services in many conditions. For example, some of the socio-cultural factors, such as language and culture barriers, prevent women from benefiting quality health services.3 This is especially true for female aborigines and immigrants. The language barrier and culture discrepancy potentially affect their willingness to seek medical help when they have diseases.4 As a result, they are usually relegated into positions where they are less able to have access and control over healthcare resources, causing them to be increasingly vulnerable to suffer health problems.

Taiwan is a multiethnic society, which has become increasingly diverse in its ethnic composition since the 1990s.5 Vietnamese and Aborigines are two major female ethnic groups, in addition to Taiwanese. According to the National Immigration Agency in Taiwan, there were about 91,000 Vietnamese women in Taiwan in 2014, the largest group represented among female immigrants. Despite the increasing number of Vietnamese female immigrants in Taiwan, little attention has been paid to their health requirements.

A similar situation occurs concerning Aboriginal women in Taiwan. According to the Council of Indigenous Peoples, there were approximately 280,000 Aboriginal women in Taiwan in 2014. Health disparities have been known to exist between Aborigines and Taiwanese. Evaluated by life expectancy, mortality rates and the prevalence and incidence of various diseases, the health status of aborigines was in general worse than that of Taiwanese. Astonishingly the aborigines’ life expectancy was on average 10 years less than that of the general population.

Due to existing barriers potentially preventing Aborigines and Vietnamese immigrants from obtaining proper health services, first line health care providers and immigrant agencies in Taiwan may take an active role offering assistance and providing health services with early assessment that reduce morbidity and mortality through the prevention or detection of disease. The short-form general health survey (SF-36), a self-reported screening questionnaire, could be an easy and effective instrument to detect health status6. This research intended to use the SF-36 as a preliminary assessment to detect the health status of multiethnic women in Taiwan. More specifically, we attempted to estimate a cutoff score using the SF-36 questionnaire to predict the health status (healthy or diseased) for women from different ethnic groups in Taiwan. If the SF-36 score of a woman from a certain ethnic group is below the cutoff score, it is very likely that she is at a risk of developing a major disease.



Constructed to examine health status and quality of life,7 the 36-item SF-36 is a generic health status profile measure that consists of 8 subscales: general health (GH), bodily pain (BP), physical functioning (PF), role-physical (RP), mental health (MH), vitality (VT), social functioning (SF), and role-emotional (RE). Each subscale is directly transformed into a 0-100 scale on the assumption that each question carries equal weight. A score ranging from 0 to 100 is assigned for each subscale, with higher scores indicating better health status. Scores from the GH, BP, PF and RP are further summarized into a Physical Component Score (PCS); scores from the rest 4 subscales are additionally aggregated into a Mental Component Score (MCS).


The sample size of this study is 530. The data was part of a larger dataset obtained from Tsai.8 The original data was collected using snowball sampling through referral for a study on healthy literacy of multiethnic women in Taiwan. Among the original participants recruited, 530 were involved in answering the SF-36 questionnaire. The 530 participants comprised 219 Taiwanese, 145 Aborigines and 166 Vietnamese. Participants were required to identify their ethnicity, and they were also asked to provide the information whether they have the following diseases: diabetes, heart diseases, liver diseases, kidney diseases, tuberculosis, and cancers.

Statistical Analysis

Cutoff Score Estimation. Following the scoring scheme advised by Ware et al,9 the score on each subscale was calculated. The score for PCS was the average score of GH, BP, PF and RP. The score for MCS was the average score of MH, VT, SF and RE. The average score of PCS and MCS produced a single composite score. Two cutoff scores were estimated, one using the composite score of PCS and MCS, and the other based on the PCS. An ANOVA was conducted before further analysis to deduce the score disparity between ethnic groups.

A Bayesian logistic model was applied to estimate the cutoff score for each ethnic group. A 95% of probability was adopted to indicate high risk of developing a disease. In other words, if a participant scored below the cutoff score, her probability of having major diseases was greater than 95%. Using to represent the proportion of participants having major diseases, the Bayesian logistic model is as the following: where Y is the health status and X is the SF-36 score. We calculated , obtained a sample for each probability and then estimated the posterior 95% credible interval (Bayesian confidence interval) for each possible score. The prior distribution for each parameter was set to

Although Ware and Kosinski10 argued that PCS and MCS should always be viewed independently, researchers11,12 indicated that PCS and MCS were dependent since mental health and physical health are inseparable, and suggested that it is feasible to use a composite score of PCS and MCS to represent an overall health status, as Kalantar-Zadeh et al13 did in a study on the association among quality of life, hospitalization and mortality in hemodialysis. However, how PCS and MCS interact with each other is not entirely clear. To clarify how the MCS affects the composite cutoff score in predicting health status, we estimated two cutoff scores, using the PCS and using the composite score of PCS and MCS.

Discriminant Analysis. A Bayesian logistic discriminant analysis was used to classify participants into either diseased or healthy group, using the 8 subscales of the SF-36 as predictors. For logistic discrimination with 2 categories (healthy or diseased), the focus is on the ratio of likelihood. That is, if the populations occur at the ratio the logistic model is converted to where G denotes group (healthy=0, diseased=1).

The probability of diseased participants for each ethnic group was incorporated in the prior distribution in the Bayesian procedure in order to reflect the fact that the probability of participants having disease was different in different ethnic group. Probabilities of disease occurrence for Taiwanese, Aborigines and Vietnamese were respectively .317, .386 and .168. If no other information was given, the chance of a woman being classified as diseased was that of her ethnic group. The prior distribution of was assumed to distribute as where p was the probability of having disease for a certain ethnic group. Because the data suggested that the number of healthy subjects was much greater than that of diseased subjects, a constraint was also imposed so that each simulation from had a value greater than 1.

Two points are worth noting when conducting the Bayesian logistic discriminant analysis. First, 1 was coded as diseased and 0 as healthy. Since a higher score in SF-36 indicates a better health status, the estimated coefficients were expected to have negative values. Second, without compromising our findings or altering the final result (i.e. p-values remain the same), scores on the 8 subscale were linearly transformed to a 0-10 scale instead of the original 0-100 scale (see Table 1), so as to avoid the coefficient estimate of each parameter being too small in the logistic regression analysis, an issue that would potentially be misleading.

Setup for Bayesian Analysis. The Bayesian methods for estimating the cutoff score and discriminant analysis were both carried out in OpenBUGS. Two chains were run in the simulation. Each chain ran 60,000 iterations, with the first 10,000 iterations excluded as burn-ins. The initial values of the coefficient were set to 0.01 and 1 for the first and second chain, respectively. The convergence was checked using Gelman-Rubin diagnostic14.

Similarly, OpenBUGS ran two chains for the discriminant analysis, with each running 60,000 iterations. The first 10,000 iterations having been excluded as burn-ins, the program calculated the mean of the remaining 50,000 iterations.

Press’ Q statistics were then used to examine whether the accurate classification rate was statistically significant. For a 95% confidence interval, if the Press’ Q statistic is greater than 3.84, then the classification rate is statistically significant.


Descriptive Statistics

Gelman-Rubin statistics indicated that all of the estimations were converged. The demographic statistics of the 3 ethnic groups and their average scores on SF-36 are shown in Table1.

Cutoff Score Estimation.

Results from the ANOVA suggested that the mean score for at least one ethnic group should be significantly different than others (see Table 2). A post hoc test using the Least Significant Difference indicated that the average score for Taiwanese and Vietnamese were both significantly higher than aborigines (p<.01), and the average scores for Taiwanese and Vietnamese had no statistical difference (p>.05).

On the other hand, results from the logistic regression model disclosed that the estimated composite cutoff scores of the PCS and MCS were 45 for Taiwanese, 30 for Aborigines and 22 for Vietnamese. The estimated cutoff scores using only the PCS were 59 for Taiwanese, 28 for Aborigines and 11 for Vietnamese.

Discriminant Analysis.

Coefficient estimates of the logistic discriminant model were shown in Table 3. The subsequent classification rates were as follows. For Aborigines, the correct classification rate was 65% and the Press’s Q was 12.75 (>3.84). For Taiwanese, the correct classification rate was 71%, and the Press’s Q was 44 (>3.84). For Vietnamese, the correct classification rate was 73%. Press’s Q statistics showed that the correct classification rate was statistically significant for each ethnic group.


Results from the ANOVA indicated no statistical difference of SF-36 scores between Taiwanese and Vietnamese ethnic groups; however an ANOVA was not able to make an inference about health status. That is, different ethnic groups have different proportions of diseased participants. The percentage of Vietnamese women (16.8%) suffering disease was proportionally less than that of Taiwanese women (38.6%). Simply using an ANOVA to compare the SF-36 scores between the two ethnic groups might yield a misleading interpretation that the health status of Taiwanese and Vietnamese held no significant difference. Yet by using a Bayesian logistic regression model that considered health status as the outcome variable, we revealed that the estimated cutoff score were nontrivial between Taiwanese and Vietnamese, either by the single composite score (45 vs. 22) or by the PCS (59 vs. 11). A much lower cutoff score for Vietnamese woman suggested that they were healthier, a result that complied with the well known healthy immigrant phenomenon.

A Bayesian logistic discriminant analysis using the 8 subscales of the SF-36 as predictors was further conducted to see how accurate the health status could be classified and to better understand the structure of the coefficient. The Bayesian logistic discriminant analysis yielded a similar result to that of Mishra et al11, in which PCS and MCS comprised loadings from all subscales, with SF, RE and MH in PCS having a negative weighting in an exploratory factor analysis with orthogonal rotation. Our estimates were correspondent with their results. For Vietnamese, coefficients of the logistic discriminant analysis for PF, RP, BP, SF, RE and GH were all negative. For both Taiwanese and Aborigines, coefficients for PF, RP, BP, GH and VT were negative whereas the coefficients for SF and MH were positive. These positive coefficients for Taiwanese and Aborigines raised concerns, because from the scoring scheme of SF-36, the lower the score the worse the health status. As a result, the positive coefficient represents a negative effect on health status, a reciprocal relationship that suggests better mental health presupposes worse physical health. This influence of MCS was reflected in the cutoff score estimates. Using only PCS and thus excluding the reverse effect from the MCS, the cutoff score for Taiwanese rose to 59 whereas the cutoff score dropped to 11 for Vietnamese.

Although this was intuitively odd as we would expect people with poorer mental health to have poorer health status, some studies15,16 have shown a different tendency. Cancer patients, for instance, have been found experiencing positive changes due to the illness.17 Specifically, a majority of cancer patients frequently reported better relationships with others, altered priorities and life goals, and a greater sense of meaning and appreciation of life.

Vietnamese women in Taiwan generally had poorer mental health than Taiwanese women,18 possibly because immigration was a stressful life experience that lead to cultural differences and conflicts, language barriers, and economic and societal changes, all of which in turn caused changes in values, attitudes and interpersonal relationships. When considering these immigrants usually took the role of care giving, it was not surprising these changes would have negative effects on personal health.18-20

Another finding from the discriminant analysis was that BP was a significant predictor in all ethnic groups. Studies have shown both cross sectional and prospective relationships between bodily pain and various physical and mental aspects of self-reported health and functioning.21-23 HajGhanbari et al24 indicated that people with obstructive pulmonary disease (COPD) report almost 2.5 times greater pain compared to healthy adults. Not only was pain more severe, but it was also more common in people with COPD. Krein et al25 reported that chronic pain was associated with diabetes. Their result showed that even after controlling for general health status and depressive symptoms, chronic pain was a major limiting factor in the performance of self-care behaviors that were important for minimizing diabetes-related complications. Walsh and Sarria26 stated that chronic pain was associated with autosomal dominant polycystic kidney disease and is a significant cause of morbidity. Our finding added more evidence to the literature that bodily pain was associated with and a useful predictor of major diseases.


The health of a society is often tied to the health of its women whose illness or death has far-reaching consequences for their children, families and communities. Estimating a cutoff score for each ethnic group, this research proposed using the SF-36 questionnaire as an early assessment to detect whether a woman from certain ethnic group is at risk of having a major disease. First line health care providers can use the instrument to decide further actions, which will reduce the burden of national health insurance in Taiwan, in that a preventive action can be done before any serious medical condition happens.

Although the SF-36 has been used in many studies, its psychometric property needs further investigation, especially on the reciprocal effects of mental health related subscales. The Bayesian logistic discriminant analysis in the current research showed that coefficients of some of the mental health related subscales (predictors) were positive, a result corresponding to previous studies but contradicting to the concept of the SF-36 questionnaire12. It is suggested that further research is necessary to explore the measurement model using item response theory and refine the scoring scheme to better fit the conceptual model behind the SF-36 questionnaire.


  1. Centre for Social Development, Humanitarian Affairs (United Nations), UNICEF., United Nations Population Fund, United Nations Development Fund for Women, United Nations. Statistical Division & World Health Organization.The World’s Women: Trends and Statistics. United Nations, 2010.
  2. Alexander LL, Alexander W.New dimensions in women’s health. Jones & Bartlett Publishers 2014.
  3. Yang YM, Wang HH, Lee FH, Lin ML, Lin PC. Health Empowerment Among Immigrant Women in Transnational Marriages in Taiwan. Journal of Nursing Scholarship2015;47(2): 135-142.
  4. Todd L, Harvey E, Hoffman-Goetz L. Predicting breast and colon cancer screening among English-as-a-second-language older Chinese immigrant women to Canada.Journal of Cancer Education 2011;26(1):161-169.
  5. Bélanger D, Lee HK, Wang HZ. Ethnic diversity and statistics in East Asia: ‘Foreign brides’ surveys in Taiwan and South Korea. Ethnic and Racial Studies 2010; 33(6):1108-1130.
  6. Lee SYD, Tsai TI, Tsai YW. Accuracy in self-reported health literacy screening: a difference between men and women in Taiwan.BMJ open 2013;3(11):e002928.
  7. Ware JE, Sherbourne CD. The MOS 36-item short-form health survey (SF-36): I. Conceptual framework and item selection.Medical care 1992:473-483.
  8. Tsai SM. Mobile Learning of Health Literacy in Multiethnic Females. Ministry of Science and Technology. 2013
  9. Ware JE, Snow KK, Kosinski M, Gandek B. Manual and interpretation guide.Boston: The Health Institute, New England Medical Center 1993
  10. Ware JE, Kosinski M. Interpreting SF-36 summary health measures: A response. Quality of Life Research 2001;10(5):405–413
  11. Mishra GD, Hockey R, Dobson AJ. A comparison of SF-36 summary measures of physical and mental health for women across the life course.Quality of Life Research 2014;23(5):1515-1521.
  12. Taft C, Karlsson J, Sullivan M. Do SF-36 summary component scores accurately summarize subscale scores?.Quality of Life Research 2001;10(5), 395-404.
  13. Kalantar-Zadeh K, Kopple JD, Block G, Humphreys MH. A malnutrition-inflammation score is correlated with morbidity and mortality in maintenance hemodialysis patients.American journal of kidney diseases 2001;38(6):1251-1263.
  14. Gelman A, Rubin DB. Inference from iterative simulation using multiple sequences.Statistical science 1992:457-472.
  15. Schroevers M, Ranchor AV, Sanderman R. Adjustment to cancer in the 8 years following diagnosis: a longitudinal study comparing cancer survivors with healthy individuals.Social Science & Medicine 2006;63(3):598-610.
  16. Tomich PL, Helgeson VS. Is finding something good in the bad always good? Benefit finding among women with breast cancer. Health Psychol 2004;23:16–23.
  17. Sears SR, Stanton AL, Danoff-Burg S. The yellow brick road and the emerald city: benefit finding, positive reappraisal coping and posttraumatic growth in women with early-stage breast cancer. Health Psychol 2003;22:487–497.
  18. Shu BC, Lung FW, Chen CH. Mental health of female foreign spouses in transnational marriages in southern Taiwan.BMC psychiatry 2011;11(1):4.
  19. Lipson A. Floating Harbor and the good ship Shprintzen.Journal of medical genetics 1991;28(11):807.
  20. Tabora BL, Flaskerud JH. Mental health beliefs, practices, and knowledge of Chinese American immigrant women.Issues in Mental Health Nursing 1997;18(3):173-189.
  21. Elliott AM, Smith BH, Hannaford PC, Smith WC, Chambers WA. The course of chronic pain in the community: results of a 4-year follow-up study.Pain 2002; 99(1):299-307.
  22. Eriksen J, Jensen MK, Sjøgren P, Ekholm O, Rasmussen NK. Epidemiology of chronic non-malignant pain in Denmark.Pain 2003;106(3):221-228.
  23. Bergman S, Jacobsson LT, Herrström P, Petersson IF. Health status as measured by SF-36 reflects changes and predicts outcome in chronic musculoskeletal pain: a 3-year follow up study in the general population.Pain 2004;108(1):115-123.
  24. HajGhanbari B, Holsti L, Road JD, Reid WD. Pain in people with chronic obstructive pulmonary disease (COPD).Respiratory medicine 2012;106(7):998-1005.
  25. Krein SL, Heisler M, Piette JD, Makki F, Kerr EA. The effect of chronic pain on diabetes patients’ self-management.Diabetes Care 2005;28(1):65-70.
  26. Walsh N, Sarria JE. Management of chronic pain in a patient with autosomal dominant polycystic kidney disease by sequential celiac plexus blockade, radiofrequency ablation, and spinal cord stimulation.American Journal of Kidney Diseases 2012;59(6):858-861.


Table 1. Descriptive Statistics

Aboriginals Taiwanese Vietnamese
Mean Age 39 39 31
Marriage % 62.1% 61.9% 83.7%
Job % 37.2% 32.8% 35.8%
Diseased % 38.6% 31.7% 16.8%
PF Score 79 88 83
RP Score 68 87 83
BP Score 70 80 74
GH Score 61 66 64
VT Score 57 60 62
SF Score 73 80 80
RE Score 68 80 84
MH Score 64 66 64
Diseased Score 58 65 62
Healthy Score 74 78 76
Mean Score 68 76 74

Diseased score refers to the average score of those who are diseased. Healthy score is the average score of healthy women.

Table 2. Estimated Cutoff Scores and Results from ANOVA

Taiwanese 45 59 AboriginalsVietnamese <.001.207
Aboriginals 30 28 TaiwaneseVietnamese <.001<.001
Vietnamese 22 11 TaiwaneseAboriginals .207<.001

CS means the estimated cutoff score using the composite score of PCS and MCS. PCS is the estimated cutoff score using the PCS score

Table 3. Coefficients for the Bayesian logistic regression

Aboriginal 2.814 -0.089 -0.004 -0.261* -0.100 -0.118 0.392* -0.155* 0.059
Taiwanese 3.787 -0.285* -0.273* -0.263* -0.328* -0.280 0.436* 0.057 0.658*
Vietnamese 5.333 -0.139 -0.042 -0.034* 0.287 0.252 -0.024 -0.029 -0.432*

* indicates a significant predictor


About hfm

Leave a Reply

Your email address will not be published. Required fields are marked *


You may use these HTML tags and attributes: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>