Japanese Journal of Clinical Oncology

Statistical Methods

To compare the values from the Qx with those from the 3-day dietary records or 24-h recalls, we calculated Pearson correlation coefficients after the nutrient intakes had been logarithmically transformed and adjusted for total energy using the method proposed by Willet (9). Since nutritional measurements are usually treated as categorical variables in epidemiological studies, we also divided both nutrient levels from the Qx and dietary records into quintiles to examine their consistency. In addition, we calculated the intra-class correlation coefficients (ICCs) between the first and second Qxs to evaluate the reproducibility of the Qx after each value was logarithmically transformed and adjusted for total energy (9).

All statistical analyses were performed using PC-SAS version 6.12 (10).

Results

Table 1 indicates that daily caloric intakes from the two Qxs were slightly higher than those estimated from 3-day food records for total males and females. The caloric intakes estimated from the 24-h recalls showed the lowest values. Various nutrients estimated from these methods showed similar relative values i.e. highest in the Qxs, medium in 3-day food records and lowest in the 24-h recalls. Table 2 compares the first Qx with the average values of the food records.

Since the main objectives of the analyses were to assess the validation and reproducibility of the Qx, we show the correlation coefficients among these survey methods for total caloric intakes (Table 3). The correlation coefficient for total caloric intakes between the first Qx and 3-day records for males was higher than that for females (r = 0.38 vs r = 0.25). The correlation coefficients between the first and second Qxs, which indicate the reproducibility of the Qx, were 0.51 for males and 0.29 for females.

Table 3. Correlation matrix of caloric intakes estimated from various records/surveys (Pearson correlation coefficient; n = 58 for males, n = 59 for females)

Male	1.	2.	3.	4.
1. Qxa	1.00
2. 2nd Qxb	0.51	1.00
3. 3-day records	0.38	0.25	1.00
4. 24 h recalls	0.19	0.32	0.44	1.00
Female	1.	2.	3.	4.
1. Qxa	1.00
2. 2nd Qxb	0.29	1.00
3. 3-day records	0.25	0.31	1.00
4. 24 h recalls	-0.02	0.30	0.43	1.00

^aFood frequency questionnaire with portion size information.
^bQx obtained one year later from the Qx.

Table 4. Correlation of various nutrient intakes estimated from Qx^a and 3-day food records by sex (Pearson correlation coefficient)

	Male (n = 58)		Female (n = 59)
	without energy adjustment	with energy adjustment	without energy adjustment	with energy adjustment
Energy	0.38**		0.25
Protein	0.44***	0.45***	0.37**	0.37**
Lipid	0.43***	0.43***	0.41**	0.51***
Carbohydrateb	0.31*	0.51***	0.20	0.29*
Crude fiber	0.45***	0.51***	0.47***	0.66***
Calcium	0.47***	0.51***	0.32***	0.59***
Carotene	0.39**	0.36**	0.45***	0.48***
Vitamin A	0.51***	0.42***	0.34**	0.27*
Vitamin C	0.18	0.21	0.22	0.21
Vitamin E	0.28*	0.29*	0.31*	0.39**
Salt	0.15	0.18	0.18	0.10
Cholesterol	0.46***	0.36**	0.36**	0.31*
Animal fat	0.54***	0.56***	0.38**	0.44***
Fish fat	0.25	0.10	0.35**	0.27*
Vegetable fat	0.31*	0.46***	0.42***	0.59***

*P<0.05, **P<0.01, ***P<0.001.
^aFood frequency questionnaire with portion size information.
^bExcluding fiber.

Correlation coefficients between the first Qx and 3-day food records for various nutrient intakes ranged from 0.15 to 0.54 in males and from 0.18 to 0.47 in females without energy adjustment (Table 4). These correlation coefficients after energy adjustment ranged from 0.10 to 0.56 in males and from 0.10 to 0.66 in females. The nutrients that showed correlation coefficients of 0.30 or less after energy adjustment were vitamin C, vitamin E, salt and fish fat in males and non-fibrous carbohydrate, vitamin A, vitamin C, salt and fish fat in females.

Table 5. Joint classifcation of total caloric intakes assessed by Qx^a and 3-day food records

Male (n = 58)
Qx	3-day food records
	1 (low)	2	3	4	5 (high)	Total
1 (low)	4	4	4	0	0	12
2	3	4	1	3	1	12
3	3	0	3	2	3	11
4	0	3	0	6	3	12
5 (high)	2	1	3	1	4	11
Total	12	12	11	12	11	58
Female (n = 59)
Qx	3-day food records
	1 (low)	2	3	4	5 (high)	Total
1 (low)	2	5	2	1	2	12
2	4	1	4	3	0	12
3	3	2	2	3	2	12
4	1	3	2	2	4	12
5 (high)	2	1	2	3	3	11
Total	12	12	12	12	11	59

^aFood frequency questionnaire with portion size information.

Table 6. Agreement rate (%) of various nutrient intakes estimated from Qx^a and 3-day food records by sex

	Male		Female
	Quintile agreement	Expanded quintile agreementb	Quintile agreement	Expanded quintile agreementb
1. Energy	36.2	55.2	16.9	44.1
2. Protein	32.8	48.3	23.7	50.8
3. Lipid	25.9	43.1	45.8	62.7
4. Carbohydratec	22.4	51.7	23.7	39.0
5. Crude fiber	31.0	43.1	32.2	49.2
6. Calcium	31.0	51.7	32.2	54.2
7. Carotene	32.8	50.0	30.5	52.5
8. Vitamin A	34.5	62.1	40.7	55.9
9. Vitamin C	24.1	43.1	27.1	44.1
10. Vitamin E	27.6	43.1	28.8	45.8
11. Salt	22.4	48.3	25.4	42.4
12. Cholesterol	43.1	63.8	35.6	50.8
13. Animal fat	27.6	46.6	42.4	55.9
14. Fish fat	19.0	39.7	25.4	42.4
15. Vegetable fat	29.3	39.7	32.2	45.8

^aFood frequency questionnaire with portion size information.
^bRatio of the number in `lowest and 2nd lowest', `middle', and `highest and 2nd highest' categories to total number.
^cExcluding fiber.

Among the age groups, the youngest group, 35-44 years old, showed the highest correlation coefficients in general (data not shown).

Table 7. Correlation of various nutrient intakes estimated from Qx^a and 12 one-day food records by sex (Pearson correlation coefficient)

	Male (n = 17)			Female (n = 20)
	without energy adjustment	with energy adjustment	de-attenuated	without energy adjustment	with energy adjustment	de-attenuated
Energy	0.44		0.50	0.49*		0.52
Protein	0.42	0.78***	0.48	0.60**	0.67**	0.63
Lipid	0.13	0.26	0.17	0.50*	0.14	0.58
Carbohydrateb	0.33	0.38	0.39	0.47*	0.24	0.50
Crude fiber	0.66**	0.76***	0.71	0.62**	0.70***	0.65
Calcium	0.76***	0.86***	0.78	0.73***	0.77***	0.75
Carotene	0.40	0.58*	0.50	0.41	0.28	0.46
Vitamin A	0.45	0.47	0.68	0.50*	0.19	0.57
Vitamin C	0.54*	0.55*	0.59	0.42	0.46*	0.45
Vitamin E	0.52*	0.71**	0.71	0.53*	0.34	0.57
Salt	0.17	0.28	0.20	0.52*	0.22	0.56
Cholesterol	0.29	0.18	0.34	0.57**	0.49*	0.66
Animal fat	0.35	0.49*	0.41	0.52*	0.39	0.64
Fish fat	0.62**	0.65**	0.80	0.39	0.30	0.49
Vegetable fat	0.28	0.47	0.49	0.37	0.03	0.48

*P<0.05, **P<0.01, ***P<0.001.
^aFood frequency questionnaire with portion size information.
^bExcluding fiber.

To evaluate the degree of misclassification associated with categorized intakes estimated from the Qx, we examined the joint classification of total caloric intakes from the first Qx and 3-day food records. Of the 12 subjects in the lowest quintile according to the Qx for males, four (36%) were also in the lowest quintile and eight (67%) in the lowest two quintiles according to the 3-day food records (Table 5). Similarly, out of 11 in the highest quintile of the Qx for males, four (36%) and five (45%) were in the highest and the next highest 3-day food record quintiles, respectively. Figures for females were almost identical to those for males. The agreement rates for the quintile analysis were 21/58 = 36% for males and 10/59 = 17% for females. When we combined the lowest and the highest two quintiles, the agreement rates were (15 + 3 + 14)/58 = 55% for males and (12 + 2 + 12)/59 = 44% for females. Similar relationships were observed for other nutrients and are shown in Table 6. The proportion of those which were misclassified into extreme quintiles ranged from 0.0 to 8.6% (average = 4.0%) for males and from 1.7 to 6.8% (average = 3.8%) for females.

Daily caloric intake and other nutrient measurements estimated from the Qx were positively associated with the 12 1-day food records (Table 7). For males, correlation coefficients between the Qx and 12 1-day food records were higher than those between the Qx and 3-day food records in general, particularly when values after energy adjustment were used. For females, energy adjustment did not improve the correlations between the Qx and 12 1-day records but the deattenuation did. Nutrients for which remarkably improved correlations were observed after the deattenuation for both men and women were vitamin A and fish fat. Absolute levels of nutrient intakes from the Qx were somewhat higher than those from the 12 1-day records.

Table 8. Intraclass correlation coefficient of various nutrient intakes estimated from Qx^a and 2nd Qx^b

	Male (n = 58)		Female (n = 59)
	without energy adjustment	with energy adjustment	without energy adjustment	with energy adjustment
Energy	0.53***		0.30*
Protein	0.54***	0.52***	0.33**	0.50***
Lipid	0.65***	0.61***	0.45***	0.57***
Carbohydratec	0.45***	0.65***	0.24*	0.49***
Crude fiber	0.48***	0.65***	0.35**	0.54***
Calcium	0.67***	0.78***	0.50***	0.67***
Carotene	0.49***	0.62***	0.60***	0.58***
Vitamin A	0.11	0.46***	0.51***	0.58***
Vitamin C	0.51***	0.64***	0.18	0.13
Vitamin E	0.55***	0.65***	0.40***	0.57***
Salt	0.55***	0.55***	0.32**	0.48***
Cholesterol	0.67***	0.55***	0.48***	0.36**
Animal fat	0.78***	0.78***	0.59***	0.66***
Fish fat	0.64***	0.53***	0.49***	0.64***
Vegetable fat	0.55***	0.62***	0.41***	0.61***

*P<0.05, **P<0.01, ***P<0.001.
^aFood frequency questionnaire with portion size information.
^bQx obtained one year later from the Qx.
^cExcluding fiber.

Table 9. Correlation of various energy-adjusted nutrient intakes estimated from Qx^a or Qx without portion size information and 3-day records (Pearson correlation coefficient)

	Male (n = 58)		Female (n = 59)
	with portion size	without portion size	with portion size	without portion size
Protein	0.45***	0.32*	0.37**	0.33*
Lipid	0.43***	0.25	0.51***	0.33*
Carbohydrateb	0.51***	0.37**	0.29*	0.28*
Crude fiber	0.23	0.12	0.48***	0.31*
Calcium	0.51***	0.40**	0.66***	0.63***
Carotene	0.36**	0.30*	0.48***	0.47***
Vitamin A	0.42***	0.32*	0.27*	0.23
Vitamin C	0.21	0.07	0.21	0.06
Vitamin E	0.29*	0.15	0.39**	0.29*
Salt	0.18	0.05	0.10	0.09
Cholesterol	0.36**	0.36**	0.31*	0.20
Animal fat	0.56***	0.58***	0.44***	0.32*
Fish fat	0.10	0.14	0.27*	0.24
Vegetable fat	0.46***	0.43***	0.59***	0.48***

*P<0.05, **P<0.01, ***P<0.001.
^aFood frequency questionnaire with portion size information.
^bExcluding fiber.

To evaluate the reproducibility of our questionnaire, we calculated the ICCs between nutrient intakes estimated from the two Qxs. The ICCs were from 0.46 to 0.78 in men and from 0.36 to 0.67 in women, except for vitamin C, when values after the energy adjustment were used (Table 8). The ICCs were relatively low in subjects [ge]65 years in comparison with younger subjects (data not shown).

In our questionnaire, we asked for both frequencies and portion sizes. To determine the effect of using frequencies only, we calculated nutrient intakes using the frequency given by the subject and the `intermediate' or standard portion size. Table 9 shows the correlation coefficients between nutrient intakes estimated from the 3-day food records and those from the Qx and the correlation coefficients without consideration of portion size. The correlation coefficients were notably higher when the subject's portion size was considered for both men and women in general.

Discussion

In Japan, several large-scale prospective studies of cancer are ongoing and some of them are developing food frequency questionnaires to assess individuals' diet quantitatively in the target population (11-13). However, it is difficult to estimate different nutrient daily intakes in Japan because dietary habits vary according to season. Nutrients for which we observed relatively large variations in both men and women were seafood fat, vitamin A, vitamin D and cholesterol. Vegetable fat, vitamin C and salt intakes showed statistically significant seasonal variations in both sexes (high in spring for vitamin C; high in summer for vegetable fat and salt) (8). Also, the number and complexity of food items usually consumed have been increasing dramatically during the last few decades in Japan. The results of this assessment indicate that the Qx for our population-based cohort study is as valid as those reported previously in the USA (9) and Japan (6). For example, the correlation coefficient for protein (r = 0.63) shown in Table 7 for females in our study (A) was higher than those reported by Willet (B: repeat compressed questionnaire vs four 1-week diet records in Nurses' Health Study; r = 0.47) and those reported by Date [C: food frequency questionnaire vs 56 or 63 days' diet records in junior college, mostly female dietitian course students; r = 0.60 (animal protein) and 0.44 (vegetable protein)]. For lipid, r = 0.58 in A, r = 0.53 in B and r = 0.46 (animal fat) and 0.54 (vegetable fat) in C. For carbohydrate, r = 0.50, 0.45 and 0.58 in A, B and C, respectively. For vitamin A, r = 0.57, 0.36 and 0.53, respectively. For vitamin C, r = 0.45, 0.66 and 0.38, respectively. Correlations for crude fiber and cholesterol were 0.65 and 0.66 in A and 0.58 and 0.61 in B.

We also found that portion size information is important to obtain more accurate nutrient intakes. If one has information on portion size which varies little among the study population, nutrient intakes could be estimated from the answers for frequency alone. Our data showed that the Takayama cohort members vary in amounts consumed. Portion size in the questionnaire is one of the important variables for estimating nutrient intake for each subject even if it is roughly categorized. It is eventually suggested that we need portion size information at least in Japan, where studies on portion size have not been conducted sufficiently and cultural backgrounds may be different from those in Western countries (9).

To compare the portion size from the Qx with that from dietary records, we chose 10 items most frequently answered in the Qx and divided the size described in the record into three categories according to the Qx for each item. The agreement rates were as follows: 74.4% for rice, 86.0% for miso soup, 72.6% for milk, 73.4% for egg, 54.0% for cucumber, 60.6% for pickled cucumber or eggplant, 53.8% for tomato, 66.7% for tofu (soy-bean curd), 61.9% for takuan (pickled Japanese radish) and 71.4% for ume-boshi or rakkyo (pickled Japanese apricot or scallions). These figures suggest that portion size indicated in the Qx by the subjects is also valid when we use the dietary record as a `gold standard'.

To determine the true validity of a dietary survey method, we must know the true intake levels of various foods or nutrients. However, there is no accepted gold standard for assessing individual amounts of dietary intake by which one can test the validity of other methods. Therefore, we should compare several methods in which errors occur independently. In the present study we compared the estimated values from several methods such as the Qx with those from 3-day records, 24-h recalls and 12 1-day records, although the errors which occurred in these methods were not completely independent. When we asked the subjects to record their dietary intakes for three continuous days, it was probably difficult for them to have foods specific to the survey for the three days. However, we conducted the survey only in summer and the results may not have represented their dietary habits. On the other hand, it is likely that data from 12 1-day records reflected their dietary habits through a year although we conducted the survey for only one day a month. These may be one of the reasons why we obtained higher correlations between the Qx and 12 1-day records than those between the Qx and 3-day records.

There is a possibility that amount of nutrient intakes estimated from the Qx are higher than those estimated from dietary records because we included several names of dishes as well as food items in the questionnaire. The subjects may have responded to some food items twice when they ate the food in a mixed dish. For example, persons who answered yes to miso soup, which includes tofu (soy-bean curd), may also have responded yes to tofu as a single item. However, we did not observe large differences between the estimated values, except for retinol, carotene and total energy. Our results were consistent with other studies in which amounts of nutrient intakes estimated from food frequency questionnaires tended to yield higher intakes than those estimated from dietary records (14-16).

In conclusion, these results suggest that our Qx can be used to estimate nutrient intakes of each individual in the cohort. The Qx with portion size information has demonstrated both validity and reproducibility among the Japanese population studied.

Acknowledgments

We are grateful to Dr Brian E. Henderson of the University of Southern California and Drs Jean H. Hankin and Laurence N. Kolonel of the Hawaii Cancer Research Center for their useful advice and encouragement. This work was partly supported by grant 06280108 from the Ministry of Education, Culture and Science, Japan.