The Journal of Nutrition

Nutritional Epidemiology

Validation of a Food Frequency Questionnairefor Assessing Dietary Soy Isoflavone Intakeamong Midlife Chinese Women in Hong Kong1,2

Sieu G. Chan,3,4* Suzanne C. Ho,3 Nancy Kreiger,4 Gerarda Darlington,5 Edward M. Adlaf,4

Kam F. So,3 and Portia Y. Y. Chong3

3Department of Community and Family Medicine, Chinese University of Hong Kong, Hong Kong SAR; 4Department of Public Health

Sciences, University of Toronto, M5G 2L7 Toronto, Canada; and 5Department of Mathematics and Statistics, University of Guelph,

N1G 2W1 Guelph, Canada

Abstract

Substantial evidence suggests that soy isoflavones may protect against certain chronic diseases. This study aims to

assess the reproducibility and validity of a 47-item semiquantitative soy FFQ (SFFQ) designed to measure the usual intake

of soy isoflavones among 145 Chinese mid-life women in Hong Kong. Reproducibility of the SFFQ was assessed by the

stability of dietary intake obtained at baseline (SFFQ0) and at 13-mo follow-up (SFFQ1). Relative validity was measured by

comparing data collected from SFFQ1 with those derived from 23-d, 24-h dietary recalls (DR) collected during the same 1-y

validation period. Isoflavone intake was calculated using analytical values from the Chinese University of Hong Kong Soy

Isoflavone Database. A total of 3,217 DR were successfully obtained. The median (interquartile range) absolute intake

estimated by the SFFQ1 was 0.91 (22.7 to 6.8) mg aglucon equivalents/d higher than the 6.3 (3.7–10.7) mg aglucon

equivalents/d measured by the DR (P , 0.0057; Wilcoxon’s Signed Rank test). Bland-Altman analysis further demonstrated the

presence of significant proportional bias between methods among Cantonese women with above-median intake (Spearman

correlation coefficient; r ¼ 0.44; P ¼ 0.0005). Nonetheless, the intraclass and Pearson correlation coefficients, respectively,

were 0.84 and 0.72 for non-Cantonese and 0.65 and 0.50 for Cantonese, showing moderate to good levels of reproducibility

and validity of the SFFQ (difference between 2 intraclass correlation coefficients, P ¼ 0.09; difference between 2 Pearson

r, P ¼ 0.16). The unadjusted and BMI-adjusted correlations were of similar magnitude. The SFFQ is a reasonably valid

instrument for assessing dietary soy isoflavone exposure in Hong Kong Chinese mid-life women. J. Nutr. 138: 567–573, 2008.

Introduction

The potential role of soy isoflavones in the prevention and ame-lioration of chronic diseases has been a subject of considerableresearch. Isoflavones are heterocyclic phenols structurally sim-ilar to human endogenous estrogens (1) and are found primarilyin soybeans. Many studies have shown that these compoundspossess estrogenic, antiestrogenic, and other biological activities(2). Evidence from epidemiologic studies suggests that soy iso-flavone consumption may reduce the risk of hormonally relatedconditions and diseases (3,4).

Soy is a traditional food among Asians and the variations ofsoy intake would allow studies to assess the relations betweensoy isoflavone exposures and the potential health effects in thesepopulations. The FFQ for estimating the usual intake of anindividual over a specified period is currently the most widely

used dietary assessment method for investigating diet and dis-ease relationship.

Several analytic epidemiological studies recently carried outamong Chinese populations have used FFQ for the assessment ofdietary soy exposure (5–8). However, these FFQ contained a lim-ited number of soy products or were originally developed for theestimation of other nutrients, such as calcium intake (6). More-over, the estimation of soy isoflavone content was based on pub-lished databases from the West, as a local database was not available.

Thus, the development and validation of a FFQ for quanti-fying soy isoflavone exposure would form the basis for furtherresearch into the potential health effects of soy isoflavone intake.This article reports on the development of a semiquantitative soyFFQ (SFFQ)6 for the evaluation of soy isoflavone intake over a12-mo period among Chinese midlife women in Hong Kong andassesses its validity and reproducibility using 23-d, 24-h dietaryrecalls (DR) over the same 12 mo as the reference method.

6 Abbreviations used: CHNS, China Health and Nutrition Survey; CU, the Chinese

University of Hong Kong; DR, dietary recall; HKPOST, Hong Kong Perimenopausal

Osteoporosis Study; ICC, Intraclass correlation coefficient; SFFQ, soy FFQ; SFFQ0,

baseline soy food FFQ; SFFQ1, 13-mo follow-up soy FFQ.

1 Supported by a grant from the Research Grants Council of the Hong Kong

Special Administrative Region, China (project no. CUHK 4047/01M).2 Author disclosures: S. G. Chan, S. C. Ho, N. Kreiger, G. Darlington, E. M. Adlaf,

K. F. So, and P. Y. Y. Chong, no conflicts of interest.

* To whom correspondence should be addressed. E-mail: sieugaen2003@yahoo.

com.

0022-3166/08 $8.00 ª 2008 American Society for Nutrition. 567Manuscript received 10 April 2007. Initial review completed 29 May 2007. Revision accepted 27 November 2007.

by guest on June 10, 2013jn.nutrition.org

Dow

nloaded from

Materials and Methods

Study population

Subjects were participants of the Hong Kong Perimenopausal Osteopo-

rosis Study (HKPOST), an ancillary study to the 1996 cross-sectionalsurvey of ‘‘Women’s Health at Midlife.’’ Descriptions of the study design,

sampling methods, and selection criteria for research participants have

previously been reported (9). Briefly, 1903 women aged 45–55 y were

recruited between June 1994 and March 1996 through random dialingof the directory-listed residential telephone numbers. Eligible partici-

pants from this survey and from the University Family Medicine Clinic

were invited to join the HKPOST. Women aged 45–55 y who were free

from medical conditions and were not taking medications known toaffect bone metabolism were considered eligible. Among the 949 eligible

women identified from these sources, 438 women joined the HKPOST

with a response rate of 46%. Their characteristics were largely similar tothose of the nonparticipants. Assessments of bone mineral density were

made at baseline and at 9-, 18-, 30-, 47-, 74-, and 86-mo follow-up.

In July 2001, 247 members who had completed the HKPOST 47-mo

follow-up study in 1999 were invited by letter followed by telephonecalls to participate in the soy questionnaire validation study. Up to 9

attempts at different times and days of the week were made to reach a

potential participant. Of the 247 HKPOST members, the first 145

women who agreed to participate were enrolled in the validation studybetween September 2001 and February 2002. The validation study was

approved by the Health Sciences I Research Ethics Board of the

University of Toronto and the Ethical Committee of the ChineseUniversity of Hong Kong. Written informed consent of the validation

study was obtained from subjects for their participation in the study.

Dietary assessment

Twenty-four-h DR (reference method). Description of the 24-h DR

collection method has been reported in an earlier article (9). Two non-

consecutive, unannounced 24-h DR, randomly selected within a 1-mo

interval over a 12-mo period, were obtained from all participants fromOctober 2001 to November 2002. Both weekdays and weekend days

were included as recall days to take into account dietary intake pattern

throughout the whole week. Sixteen weekdays and 7 weekend days wereincluded for each participant. The first DR was administered 2 wk after

the baseline soy food frequency questionnaire (SFFQ0) by trained inter-

viewers using the USDA multiple pass technique (10). Details about each

reported food were gathered using standardized food probes listed in theFood Instruction Booklet for the Continuing Survey of Food Intakes by

Individuals 1998 (11).

Development of SFFQ (test method).

Selection of food items. The SFFQ was designed to assess the ha-

bitual dietary intake of soy isoflavones in midlife women. As thesoy isoflavone database for local foods has not yet been developed, the

food items and range of portion sizes selected for the SFFQ were first

developed using dietary soy protein intake data from the China Health

and Nutrition Survey (CHNS) conducted in 1993 (12). Based on 3-d,24-h DR data collected from 2643 female CHNS participants aged 40–

60 y, food items that contributed to 95% of total population intake of

soy protein were identified by contribution analysis (13). Estimates of

the expected contribution of individual soy products were obtainedprimarily from the 1991 Food Composition Table for China (14,15).

The initial SFFQ contained 13 food items. To ensure comprehensiveness

of food items to be included, data from earlier local dietary surveys

were also examined (5,6,8,16). In addition, local ‘‘wet markets’’ andsupermarkets were visited and regional food composition tables (17–

19) and cookbooks were reviewed to identify soy products commonly

consumed by the local Chinese or dialect subgroups in Hong Kong.Thirty-eight additional items were identified and subsequently added to

the initial food list of 13, yielding a SFFQ consisting of 51 soy food

products.

Development of portion size. Respondents chose from a selectionof 3 portion sizes, small, medium, or large, to describe the intake quan-

tity for each of the food items contained in the SFFQ. The quantity of

food for the 3 portion sizes was derived using the 25th, 50th, and 75th

percentiles for the specific food from the 1993 CHNS survey data as

cutoff values (20). The derived portion sizes for foods included in the

SFFQ were reviewed and modified by one coauthor (S. C. Ho) ex-perienced in conducting nutritional epidemiologic studies in the local

population. Color photographs of the 51 foods, each depicted in the

respective portion sizes, were used to assist participants to better esti-

mate the quantity of foods consumed. Participants were asked abouttheir usual intake of the soy food during the 1-y period preceding the

date of interview. Seasonal variation of soy food intake was also taken

into account by asking the participants to report intake frequency during

both in-season and off-season periods for each item. Frequency ofconsumption was recorded as never, or number of times per day, week,

month, or year, and portion size equivalent to either small, medium, or

large.

Pretesting the SFFQ. The draft SFFQ was pretested on 20 women

recruited from women’s health talks or by word of mouth. Using a ret-

rospective, think-aloud technique developed in cognitive science (21),

the first 10 volunteer participants were instructed to think back abouthow they comprehended and interpreted the terms ‘‘soy products’’ and

‘‘seasonal foods.’’ Participants were also asked to explain the recall

strategies used in their estimation of consumption frequency and theamount of 6 soy products (2 solid, liquid, and amorphous) eaten over a

6-mo period prior to the interview. We adjusted the wording of some

questions to improve the ease of understanding. The revised SFFQ was

pretested again on another 10 women of similar age range. No furtherchanges were made and the revised SFFQ was tested on another sample

of 39 women aged 40–60 y to evaluate their ability to answer the ques-

tions and the overall flow of the questionnaire. Participants for this final

stage of the pretesting were identified from a sample of 2000 addressesfrom the list of housing quarters maintained by the Hong Kong Census

and Statistics Department. Soy products not consumed by participants

were excluded (n¼ 4) from the final SFFQ. The final SFFQ contained 47soy products.

Administration of the SFFQ. The 47-item SFFQ was administered in

person by trained interviewers at baseline (SFFQ0). A set of visual aidsincluding a page-numbered food photo album and a set of common

household measures was given to participants at the completion of the

SFFQ0 to take home for reference for the 24-h DR. Another (SFFQ1) was

administered in person at 13 mo, which was 1 mo after the completion ofthe last DR.

Validation procedures. Reproducibility of SFFQ was assessed by thestability of dietary intake obtained at SFFQ0 and at SFFQ1. The relative

validity was assessed by comparing soy isoflavone data collected from

SFFQ1 against the estimated amount from the 23-d, 24-h DR during the

same 1-y validation period.

Statistical analysis

Data processing. We checked the FFQ and DR for completeness priorto food coding and data entry for analysis. Participants with #12 DR

(n ¼ 4) were excluded, leaving a total of 141 participants with 3217 DR

available for analysis.

Isoflavone content of foods. The isoflavone intake for each food was

calculated by multiplying the reported frequency of consumption of

specific food by the isoflavone content of the specified portion. Values forisoflavone content in foods (mg aglucon equivalents/100 g) were ob-

tained from the CU Soy Isoflavone Database (S. Chan, P. Murphy, S. Ho,

N. Kreiger, G. Darlington, K. So, P. Chong, unpublished data). Intake

values were then summed across all foods to obtain the total intake ofisoflavones for each individual.

Statistical methods. Median and interquartile range for isoflavone

intakes were computed from both FFQ and from the means of the 23-d,24-h DR. All other summary statistics are expressed as means 6 SD or

number and percentage, unless otherwise stated. Because the distribu-

tions of isoflavone intakes were right-skewed, the Box-Cox power

568 Chan et al.

by guest on June 10, 2013jn.nutrition.org

Dow

nloaded from

transformation, y ¼ (xl 2 1)/l, was carried out to improve distribution

normality prior to formal analysis. The l values for SFFQ0, SFFQ1, and

DR were 0.3, 0.3, and 0.4, respectively.Proportions were compared with the chi-square test of contingency.

Differences between groups were performed using the Mann-Whitney U

test. Within-subject comparison was examined using the Wilcoxon’s

Signed Rank test.Agreement between the 2 dietary intake methods was further evalu-

ated using the graphical method described by Bland and Altman (22).

With this method, the arithmetic difference in isoflavone intakes between

the 2 methods for each individual was plotted against the mean values ofthe 2 methods. To examine whether the differences between methods

varied systematically with the means obtained from the 2 methods,

Spearman rank correlation procedures were performed. Limits of agree-ment were established defining the range or interval within which most

of the differences between the methods are expected to lie. Because the

Bland-Altman plots indicated the differences between methods were

intake dependent, and log transformation, as suggested by Bland andAltman (23), could not correct the magnitude of differences, 95% limits

of agreement were therefore derived based on the regression method for

nonuniform differences.

The ability of the 2 dietary intake methods to rank subjects into thesame quartile was also examined. Isoflavone values obtained from the

SFFQ1 and through the 24-h DR were divided into quartiles using cut-off

values derived separately for each dietary assessment method. The pro-portion of subjects correctly classified into the same quartile, and mis-

classified into extreme quartiles, was examined.

To evaluate the reproducibility of the SFFQ administered on 2

separate occasions ;13 mo apart, an intraclass correlation coefficient(ICC1,1) was computed using a 1-way random effects ANOVA for a

single measurement (24,25). The ICC, defined as a ratio of between-

person variance to total variance, ranges from 0 to 1. An ICC $0.76

indicates excellent test-retest reliability and an ICC ,0.4 represents poortest-retest reliability (26).

Pearson’s product-moment correlation coefficient was computed to

quantify the extent of linear association between power-transformed

isoflavone intakes derived from the SFFQ1 and the mean values derivedfrom the 23-d, 24-h DR. Partial correlation coefficient adjusted for BMI

was also calculated by multiple linear regression. To correct for the

attenuation of correlation coefficient attributable to within-personvariation in DR, de-attenuated correlation coefficient was calculated

using the formula rc ¼ ro f11[(S2w / S2

B/n)]1/2g(27), where rc is the true

correlation, ro is the observed correlation between power-transformed

isoflavone intakes estimated by the SFFQ1 and 24-h DR, S2w /S2

b is theratio of the within- and between-person variances, and n is the number of

24-h DR per subject. Subgroup correlations were also computed,

stratifying participants by demographic and anthropometric character-

istics. To test the equality of correlations, each subgroup correlationcoefficient was first transformed into a Z score using the Fisher’s r-to-Z

transformation (25). The derived Z scores of 2 and several correlation

coefficients were then compared, respectively, using the Z (28) and chi-square formulas (29).

The Statistical Analysis System for Windows (version 6.12, SAS

Institute) was used to perform all statistical analyses. All tests were

2-sided and P , 0.05 was considered significant.

Results

The baseline demographic characteristics of the participants inthis study were similar to those from the original cohort (Table

1). About one-third had only some primary education or below.The majority were married and approximately one-half werehousewives.

The median intake (interquartile range) of soy isoflavones forSFFQ0, SFFQ1, and 24-h DR were 6.6 (3.4–12.0), 7.7 (3.5–13.0), and 6.3 (3.7–10.7) mg aglucon equivalents/d, respectively.Median isoflavone intakes from the 2 SFFQ administered 1 yapart did not differ (P¼ 0.83; Wilcoxon’s Signed Rank test). The

SFFQ method, however, yielded a higher median daily isoflavoneintake than that derived from the 23-d, 24-h DR. The differencebetween the SFFQ1 and DR medians was 0.91 (22.7 to 6.8) mgaglucon equivalents/d (P , 0.0057; Wilcoxon’s Signed Ranktest).

The Bland-Altman plots for isoflavone intakes for the entiresample (Fig. 1A), Cantonese (Fig. 1B), and non-Cantonese (Fig.1C) showed the individual between-method differences tendedto increase with increasing intake of soy isoflavones, with mag-nitude of arithmetic differences greatest at high intake andsmallest at low intake levels. In addition, the differences betweenmethods were strongly and positively associated with the meansobtained from the 2 methods. The Spearman rank correlationcoefficients for the aforementioned groups were r ¼ 0.40 (P ,

0.0001), r ¼ 0.45 (P ¼ 0.0005), and r ¼ 0.003 (P ¼ 0.9066),respectively. The data suggest the presence of proportional bias(between methods) in the Cantonese dialect group and such biasmay have arisen from 1 or both assessment methods. The 95%limits of agreement derived from the regression method indi-cated ;4% of individual differences were spread above or belowthe lower 95% limit of agreement. Nonetheless, individual datarevealed that the limits of agreement between the 2 methodswere fairly wide.

These findings suggest that the SFFQ might be a usefulmethod for categorizing non-Cantonese according to their in-take of soy isoflavones. To investigate whether the heterosce-dastic distribution observed in Figure 1B would invalidate theuse of the SFFQ in the Cantonese populations, the Cantonesewere categorized according to their intake values below or abovethe absolute median intake of the 2 methods (7.27 mg agluconequivalents), and the relation between the arithmetic differencesand mean intake values obtained from the 2 methods werereexamined. A positive association was observed only amongwomen with above-median intake of soy isoflavones (Spearmanrank correlation coefficient; r ¼ 0.44; P ¼ 0.0005).

In this study, the overall Pearson correlation coefficientbetween the SFFQ1 and 24-h DR for isoflavone intakes of all

TABLE 1 Comparison of study participants andnonparticipants based on the HKPOST cohort1

Characteristics Participants Nonparticipants P 3

n 145 293

Age, y 49.7 6 2.7 50.0 6 2.6 0.14

Weight,3 kg 58.0 6 9.3 58.0 6 9.0 0.82

Height,3 m 1.55 6 0.1 1.54 6 0.1 0.18

BMI,3 kg/m2 24.1 6 3.7 24.4 6 3.6 0.53

Menopausal status, n (%)

Pre- 88 (61) 178 (61) 0.97

Peri- 21 (14) 40 (14)

Post- 36 (25) 75 (26)

Marital status,3 n (%)

Married 132 (92) 277 (95) 0.25

Not married 12 (8) 16 (5)

Occupation,3 n (%)

Housewife 79 (54) 148 (51) 0.75

Working part time 16 (11) 36 (12)

Working full time 50 (34) 108 (37)

1 Values are means 6 SD or n (%).2 Mann-Whitney U test for differences between groups and chi-square test for

categorical variables.3 Variables have missing values.

FFQ validation study 569

by guest on June 10, 2013jn.nutrition.org

Dow

nloaded from

subjects was 0.53 (P , 0.0001) (Table 2). Cross-classificationanalysis further revealed that the observed levels of agreementand disagreement between the 2 methods were higher and lowerthan what might be expected by chance (25 and 12.5%, re-spectively). Thirty-five percent of the participants were classifiedinto similar quartiles and only 4.3% were grossly misclassifiedinto the extreme quartiles. Because 60% of the 24-h DR had

zero consumption of soy isoflavones and Box-Cox transforma-tion could not normalize the distribution of isoflavone values,the ratio of within- to between-person variance in isoflavoneintake derived from the 24-h DR could not be calculated. Failureof the data to follow a normal distribution further precluded thecalculation of the de-attenuated correlation coefficient, whichwhen corrected for attenuation due to day-to-day fluctuations inindividuals would provide a more accurate estimate of thecorrelation between the SFFQ1 and multiple DR.

The reproducibility, as indicated by 1CC between isoflavoneintake assessed by SFFQ at 0 and 13 mo, was 0.68 (Table 2).Except for the non-Cantonese who exhibited a marginally betterreproducibility (1CC1,1 ¼ 0.84) than the Cantonese dialectgroup (1CC1,1 ¼ 0.65; difference between 2 correlation coeffi-cients, P ¼ 0.087), the magnitude of correlation coefficientsdid not differ when stratified by other demographic character-istics.

Discussion

This study examined the relative validity and reproducibility of a47-item SFFQ in estimating the habitual isoflavone intake inmid-life Chinese women. A correlation of 0.53 of isoflavone

FIGURE 1 Difference in isoflavone intake estimated with the SFFQ

and 24-h DR plotted against the mean of isoflavone intake measured

with the 2 methods for the entire study population (n ¼ 141) (A),

Cantonese (n ¼ 120) (B), and non-Cantonese (n ¼ 21) (C). The solid

lines represent the 95% limits of agreement.

TABLE 2 Correlation coefficients between daily mean intakeof isoflavones assessed by SFFQ (SFFQ0 and SFFQ1)and 24-h DR by demographic characteristics1

SFFQ0- SFFQ1 SFFQ1-DR

n ICC n r

All 140 0.68 141 0.53

Age, y

,54.1 35 0.76 35 0.49

54.1–55.2 36 0.58 36 0.47

55.3–57.4 34 0.58 34 0.43

$57.5 35 0.74 36 0.62

Education

Some primary schooling or less 51 0.59 51 0.42

Some high school 57 0.74 58 0.53

High school completion or above 32 0.73 32 0.71

Dialect

Non-Cantonese 21 0.84 21 0.72

Cantonese 119 0.65 120 0.50

BMI, kg/m2

,22.0 35 0.63 35 0.62

22.0–24.1 35 0.78 35 0.44

24.2–26.1 35 0.65 36 0.52

$26.2 35 0.67 35 0.52

Self-rated health

Poor 11 0.73 11 0.76

Fair 73 0.64 73 0.41

Good to excellent 54 0.74 55 0.60

Diet change2

No 99 0.69 100 0.52

Yes 41 0.67 41 0.54

Marital status

Not married 23 0.69 23 0.67

Married 117 0.68 118 0.49

1 Isoflavone intake was power transformed to improve distribution normality.2 Self-reported changes in diet during the 1-y study period. Information was obtained

from SFFQ1.

570 Chan et al.

by guest on June 10, 2013jn.nutrition.org

Dow

nloaded from

intakes derived from the SFFQ and the 23-d, 24-h DR was foundand the results compared favorably with that of other studiesusing self-reported dietary assessment as the reference methods(30–32). The ICC of the 2 SFFQ measured at 1-y intervals (value¼0.68) is also similar to studies that examined reproducibility offood and nutrient intakes within periods of 1–10 y (33) but con-siderably lower than those reported in some recent soy valida-tion studies based on shorter durations. Frankenfeld et al. (34)reported an ICC of 0.98 in a 1-wk study, whereas Kirk et al. (35)reported a 2-wk ICC of 0.89 and Rice et al. (36) reported an ICCof 0.72 for a study over 1–4 mo. The relatively short time in-tervals between the repeated FFQ measurements and thus thelesser variability of dietary intake may explain the fairly highreported correlation values in these studies.

The study found differences in the validity of SFFQ amongthe 2 dialect subgroups. The regularity of food habits andhomogeneity in soy consumption might have contributed toa higher measure of correlation coefficient among the non-Cantonese (9). Nonetheless, this notion of a more consistent dietamong the non-Cantonese was not supported by the 24-h DR,because the variation in isoflavone intake (measured as the semi-interquartile range divided by median) was higher in the non-Cantonese than in the Cantonese. Closer scrutiny revealed thatthe non-Cantonese were better educated than the Cantonese andeducation has previously been shown to be related to betterconsistency of recalls (33).

The DR method provides snapshots of dietary intake atdifferent time points during the reference period. The numberof days of 24-h DR needed to correctly rank individuals varieswith nutrients is a function of the desired correlation betweenobserved and true nutrient intakes and the ratio of within- tobetween-subject variances (37). Our data revealed substantialwithin-person or day-to-day variations in isoflavone intakesamong the study participants. The within- to between-subjectvariance ratio (Sw2/Sb2), obtained by the SAS procedure PROCVARCOMP for non-normally distributed isoflavone data, was9.0. The number of days, therefore, needed to obtain a cor-relation of 0.8 between the 24-h recall-derived mean intake andthe true intake would be 16 d [d¼ (r2/1 2 r2) (Sw2/Sb2)¼ (0.82/1–0.82) (9.0) ¼ 16] (37). Thus, this study has obtained sufficientnumber of 24-h DR from each participant to achieve a rankingof acceptable accuracy.

Our data revealed that the SFFQ yielded a higher medianisoflavone intake value than that obtained from the 24-h DR. Avalidation study by Yamamoto et al. (30) also noted a signif-icantly higher mean absolute intake estimated by the FFQmethod than that obtained from the 24-h DR. Possible expla-nations include an overestimation of the portion sizes assignedto food items in the SFFQ leading to a higher reported amountthan that actually consumed by the participants. A number ofstudies (38,39) have also noted that increasing the number offood items in the FFQ might result in an overestimation of theintake, whereas combining several nutritionally similar fooditems into single items could lead to an underestimation. OurSFFQ, comprising 47 individual food items, is an extensivequestionnaire containing soy products commonly consumed byvarious Chinese dialect groups in Hong Kong. The comprehen-siveness of the SFFQ could thus also give rise to a higherreported consumption of soy isoflavones.

Epidemiological studies of diet and disease relationship seekto compare disease risks across categories of intake. Althoughconsistent over- and underestimation of intake by the FFQmethod would not alter the rankings of individuals along thedistribution of intake, differential misclassification could lead to

inaccurate rankings and erroneous exposure estimates in diet-disease relations (40,41). In this study, despite a moderate posi-tive correlation between the dietary assessment methods, theBland-Altman plot showed that the magnitude of between-method differences was not consistent across the range of intakefor individual participants. Further analysis of the data indicatedthat the degree of agreement was strongly associated with thedialect grouping of the participants and a bias might possiblyexist among the Cantonese dialect group with an above-mediandietary soy intake.

The SFFQ was designed to assess an individual’s usual intakeduring the past year whereas the 24-h DR capture snapshots ofactual intakes at different time points over the reference period.The marked discrepancy between estimates derived from the 2assessment methods in participants with high soy intake couldbe due to the insufficient number of DR or the inability ofparticipants with high intake to accurately recall what they hadconsumed. This finding thus suggests that bias might exist in therisk estimates at certain levels of exposures and caution shouldbe taken in using these results.

Due to limited funding, our study had not incorporated themeasurement of biomarkers. Biomarkers of diet do not rely onthe subject’s reporting and possess measurement errors that arenot likely to be correlated with those of the self-reported dietaryassessment methods (33,42). Lee et al. (43), in a dietaryisoflavone assessment validation study among men in Shanghai,China, compared the validity of a 8-item SFFQ with 4 spot urinesamples collected over a 12-mo period and found a Spearmancorrelation coefficient of 0.48. The ICC reproducibility value forisoflavonoid levels measured in spot urine samples was 0.39.These findings demonstrate that measurements sampled atmultiple time points can yield acceptable reproducibility andurinary isoflavones can serve as a time-integrating biomarkerreflecting longer-term intakes (44). Thus, future validationstudies of self-reported dietary assessment methods should alsoconsider the incorporation of dietary biomarkers to compareand evaluate the validity of the findings.

Our study has several limitations. We had not adjusted theisoflavone intake values for total dietary energy intake, becausethe nutrient composition data for local foods were unavailable.Energy adjustment would presumably improve the correlationcoefficient between estimates from 2 dietary assessment methods(33,45), because it removes variations in nutrient intake dueto different energy requirements among individuals and alsominimizes measurement errors from under- or over-reporting ofenergy intake in FFQ. When the crude validity coefficient in ourstudy was adjusted for BMI (an indirect measure of total energyintake), the unadjusted and BMI-adjusted correlations weresimilar (0.5220 and 0.5239, respectively), suggesting that thevariation in isoflavone intake is not related to variation indietary energy intake.

Also, because of the highly skewed distribution with a highfrequency of zero intake values in the DR, the de-attenuatedcorrelation coefficient was not calculated. Though a high degreeof reproducibility was observed, particularly among the non-Cantonese, the level of agreement between the 2 assessmentmethods varied with a relatively wide scatter of differencesobserved among the Cantonese. The Bland-Altman plot indi-cated that individual differences between methods were depen-dent on intake levels, with better agreement at the lower levelsof intake. Thus, more work is required to establish the presenceof differential bias in dietary intake estimates derived from theSFFQ in population subgroups under investigation. Moreover,as the sample size of the non-Cantonese subjects was rather

FFQ validation study 571

by guest on June 10, 2013jn.nutrition.org

Dow

nloaded from

small, additional research is warranted to examine the effec-tiveness of the SFFQ in larger subgroups of non-Cantonesepopulations.

Despite these limitations, this SFFQ has been validatedagainst 3217 DR with very low attrition from the study par-ticipants. It had acceptable validity and reproducibility and isthus a reasonable instrument for assessing dietary soy iosflavoneexposure in Hong Kong Chinese mid-life women.

Literature Cited

1. Setchell KDR, Adlercreutz H. Mammalian lignans and phyto-oestrogens.Recent studies on their formation, metabolism and biological role inhealth and disease. In: Rowland IA, editor. The role of gut microflora intoxicity and cancer. New York: Academic Press; 1988. p. 315–45.

2. Committee on Toxicity of Chemicals in Food/Consumer Products andthe Environment. COT report: phytoestrogens and health. London:Food Standards Agency; 2003.

3. Cassidy A, Albertazzi P, Lise Nielsen I, Hall W, Williamson G, Tetens I,Atkins S, Cross H, Manios Y, et al. Critical review of health effects ofsoyabean phyto-oestrogens in post-menopausal women. Proc Nutr Soc.2006;65:76–92.

4. Balk E, Chung M, Chew P, Ip S, Raman G, Kupelnick B, Tatsioni A, SunY, Wolk B, et al. Effects of soy on health outcomes. Evidence report/technology assessment no. 126. Prepared by Tufts-New EnglandMedical Center Evidence-Based Practice Center under contract no.290–02–0022. AHRQ publication no. 05–E024–2. Rockville (MD):Agency for Healthcare Research and Quality; August 2005.

5. Ho SC, Woo J, Leung SSF, Sham A, Lam TH, Janus ED. Intake of soyproducts are related to better plasma lipid profiles in the Hong KongChinese population. J Nutr. 2000;130:2590–3.

6. Ho SC, Chan SG, Yi Q, Wong E, Leung PC. Soy intake and themaintenance of peak bone mass in Hong Kong Chinese women. J BoneMiner Res. 2001;16:1363–9.

7. Mei J, Yeung SS, Kung AW. High dietary phytoestrogen intake isassociated with higher bone mineral density in postmenopausal but notpremenopausal women. J Clin Endocrinol Metab. 2001;86:5217–21.

8. Ho SC, Woo J, Lam S, Chen Y, Sham A, Lau J. Soy protein consumptionand bone mass in early postmenopausal Chinese women. OsteoporosInt. 2003;14:835–42.

9. Chan SG, Ho SC, Kreiger N, Darlington G, So KF, Chong PYY. Dietarysources and determinants of soy isoflavone intake among midlifeChinese women in Hong Kong. J Nutr. 2007;137:2451–5.

10. Guenther PM, Cleveland LE, Ingwersen LA. Questionnaire develop-ment and data collection procedures. In: Tippett KS, Cypel YS, editors.Design and operation: the Continuing Survey of Food Intakes byIndividuals and the Diet and Health Knowledge Survey, 1994–96.USDA, Agricultural Research Service, Nationwide Food Surveys reportno. 96–1; 1997 [cited 2005 Aug 31]. Available from: ,http://www.ars.usda.gov/SP2UserFiles/Place/12355000/pdf/DORch_4.pdf..

11. USDA Agricultural Research Service. 1994–6 Continuing Survey of FoodIntakes by Individuals (CSFII): food instruction booklet. 1998 [cited 2005Aug 31]. Available from: http://www.ars.usda.gov/SP2UserFiles/Place/12355000/pdf/FIB.pdf.

12. Zhai F, Guo X, Popkin BM, Ma L, Wang Q, Yu W, Jin S, Ge K. Theevaluation of the 24-hour individual recall method in China. Food NutrBull. 1996;17:154–61.

13. Block G, Dresser CM, Hartman AM, Carroll MD. Nutrient sources inthe American diet: quantitative data from the NHANES II survey. I.Vitamins and minerals. Am J Epidemiol. 1985;122:13–26.

14. Popkin BM, Lu B, Zhai FY. Understanding the nutrition transition:measuring rapid dietary changes in transitional countries. Public HealthNutr. 2002;5:947–53.

15. China CDC. Food Nutrition Components Table. Beijing: People’s Medi-cal Publishing House. 1991.

16. Leung SSF, Ho S, Woo J, Lam TH, Janus ED. Hong Kong Adult DietarySurvey 1995. Hong Kong: Chinese University of Hong Kong; 1997.

17. Wang G, Parpia B, Wen Z. The composition of Chinese foods.Washington, DC: ILSI Press; 1997.

18. Holland B, Welch AA, Unwin ID, Buss DH, Paul AA, Southgate DAT.McCance and Widdowson’s the composition of foods. 5th rev ed.

Cambridge (UK): The Royal Society of Chemistry and Ministry ofAgriculture, Fisheries and Food; 1991.

19. Tsang YS, Fung SW. Diet and health guide and composition of food inSouthern China. Guang Zhou (China): Zhongshan University Press;1991.

20. Carolina Population Center at the University of North Carolina atChapel Hill. China Health and Nutrition Survey. Version current 16August 2005 [cited 2005 Aug 31]. Available from: ,https://www.cpc.unc.edu/projects/china/data/datasets/data_downloads..

21. Sudman S, Bradburn NM, Schwarz N. Thinking about answers: the ap-plication of cognitive processes to survey methodology. San Francisco:Jossey-Bass Publishers; 1996.

22. Bland JM, Altman DG. Statistical methods for assessing agreementbetween two methods of clinical measurement. Lancet. 1986;1:307–10.

23. Bland JM, Altman DG. Measuring agreement in method comparisonstudies. Stat Methods Med Res. 1999;8:135–60.

24. Shrout PE, Fleiss JL. Intraclass correlations: uses in assessing raterreliability. Psychol Bull. 1979;86:420–8.

25. McGraw KO, Wong SP. Forming inferences about some intraclasscorrelation coefficients. Psychol Methods. 1996;1:30–46.

26. Fleiss JL. Statistical methods for rates and proportions. New York:Wiley; 1981.

27. Beaton GH, Milner J, Corey P, McGuire V, Cousins M, Stewart E, deRamos M, Hewitt D, Grambsch PV, et al. Sources of variance in 24-hour dietary recall data: implications for nutrition study design andinterpretation. Am J Clin Nutr. 1979;32:2546–59.

28. Cohen J, Cohen P. Applied multiple regression/correlation analysis forthe behavioral sciences. Second edition. Hillsdale (NJ): LawrenceErlbaum Associates; 1983.

29. Steel RD, Torrie JH, Dickey TA. Principles and practice of statistics: abiomedical approach. New York: McGraw Hill; 1997.

30. Yamamoto S, Sobue T, Sasaki S, Kobayashi M, Arai Y, Uehara M,Adlercreutz H, Watanabe S, Takahashi T, et al. Validity and reproduc-ibility of a self-administered food-frequency questionnaire to assessisoflavone intake in a Japanese population in comparison with dietaryrecords and blood and urine isoflavones. J Nutr. 2001;131:2741–7.

31. Williams AE, Maskarinec G, Hebshi S, Oshiro C, Murphy S, FrankeAA. Validation of a soy questionnaire with repeated dietary recalls andurinary isoflavone assessments over one year. Nutr Cancer. 2003;47:118–25.

32. Huang MH, Harrison GG, Mohamed MM, Gornbein JA, Henning SM,Go VL, Greendale GA. Assessing the accuracy of a food frequencyquestionnaire for estimating usual intake of phytoestrogens. NutrCancer. 2000;37:145–54.

33. Willett W. Nutritional epidemiology. 2nd ed. New York: OxfordUniversity Press; 1998.

34. Frankenfeld CL, Patterson RE, Horner NK, Neuhouser ML, Skor HE,Kalhorn TF, Howald WN, Lampe JW. Validation of a soy food-frequency questionnaire and evaluation of correlates of plasma iso-flavone concentrations in postmenopausal women. Am J Clin Nutr.2003;77:674–80.

35. Kirk P, Patterson RE, Lampe J. Development of a soy food frequencyquestionnaire to estimate isoflavone consumption in US adults. J AmDiet Assoc. 1999;99:558–63.

36. Rice MM, LaCroix AZ, Lampe JW, van Belle G, Kestin M, Sumitani M,Graves AB, Larson EB. Dietary soy isoflavone intake in older JapaneseAmerican women. Public Health Nutr. 2001;4:943–52.

37. Nelson M, Black AE, Morris JA, Cole TJ. Between- and within-subjectvariation in nutrient intake from infancy to old age: estimating thenumber of days required to rank dietary intakes with desired precision.Am J Clin Nutr. 1989;50:155–67.

38. Thompson FE, Kipnis V, Subar AF, Krebs-Smith SM, Kahle LL,Midhune D, Potischman N, Schatzkin A. Evaluation of 2 brief in-struments and a food-frequency questionnaire to estimate daily num-ber of servings of fruit and vegetables. Am J Clin Nutr. 2000;71:1503–10.

39. Smith-Warner SA, Elmer PJ, Fosdick L, Tharp TM, Randall B.Reliability and comparability of three dietary assessment methodsfor estimating fruit and vegetable intakes. Epidemiology. 1997;8:196–201.

40. Beaton GH. Interpretation of results from diet history studies. In:Kohlmeier L, editor. The diet history method. London: Smith-Gordon;1991.

572 Chan et al.

by guest on June 10, 2013jn.nutrition.org

Dow

nloaded from

41. MacIntyre UE, Venter CS, Vorster HH, Steyn HS. A combination ofstatistical methods for the analysis of the relative validation data of thequantitative food frequency questionnaire used in the THUSA study.Transition, Health and Urbanisation in South Africa. Public HealthNutr. 2001;4:45–51.

42. Ocke MC, Kaaks RJ. Biochemical markers as additional measurementsin dietary validity studies: application of the method of triads withexamples from the European Prospective Investigation into Cancer andNutrition. Am J Clin Nutr. 1997;65 Suppl:S1240–5.

43. Lee SA, Wen W, Xiang YB, Barnes S, Liu D, Cai Q, Zheng W, Shu XO.Assessment of dietary isoflavone intake among middle-aged Chinesemen. J Nutr. 2007;137:1011–6.

44. Willett WC. The use of biomarkers in nutritional epidemiology. In: KokFJ, van’t Veer P, editors. Proceedings of the 3rd meeting on nutritionalepidemiology ‘‘biomarkers of dietary exposure.’’ Rotterdam (TheNetherlands): Smith-Gordon; 1991.

45. Flegal KM. Evaluating epidemiologic evidence of the effects of food andnutrient exposures. Am J Clin Nutr. 1999;69 Suppl:S1339–44.

FFQ validation study 573

by guest on June 10, 2013jn.nutrition.org

Dow

nloaded from