Agreement Between Self-and Partner Reports Obtained by a Self-Administered Questionnaire : Medical and Lifestyle Information

We examined agreement between the subjects' selfand partner-reports of such epidemiological information as medical and family history, smoking and drinking habit and physical activity. Information was obtained by a self-administered questionnaire which was completed by 224 workers (subjects) and by their partners in 1997. Agreement was assessed by calculating kappa statistic, intraclass correlation coefficient (ICC), and per cent agreement. Per cent agreement ranged from a low of 76.2 for general life stress to a high of 98.0 for angina/myocardial infarction and diabetes mellitus as present illness. Kappa values ranged from a low of 0.34 for general life stress to a high of 0.86 for smoking habit. Compared to subjects, their partners tended to report lower level of both exposures (continuous variables) and presence (dichotomous variables). The average kappa was 0.64 for wife-surrogates, whereas 0.53 for husband-surrogates. Overall, our finding suggested that partners could provide acceptable information for the concrete and directly-observable variables (e.g. such present illness as hypertension which required daily medication, or smoking /drinking habit itself), but not so for detailed/subjective variables (e.g. number of cigarettes smoked per day or general life stress) . J Epidemiol, 1999 ; 9 : 183-189

We examined agreement between the subjects' self-and partner-reports of such epidemiological information as medical and family history, smoking and drinking habit and physical activity.Information was obtained by a self-administered questionnaire which was completed by 224 workers (subjects) and by their partners in 1997.Agreement was assessed by calculating kappa statistic, intraclass correlation coefficient (ICC), and per cent agreement.Per cent agreement ranged from a low of 76.2 for general life stress to a high of 98.0 for angina/myocardial infarction and diabetes mellitus as present illness.Kappa values ranged from a low of 0.34 for general life stress to a high of 0.86 for smoking habit.Compared to subjects, their partners tended to report lower level of both exposures (continuous variables) and presence (dichotomous variables).The average kappa was 0.64 for wife-surrogates, whereas 0.53 for husband-surrogates.
Overall, our finding suggested that partners could provide acceptable information for the concrete and directly-observable variables (e.g.such present illness as hypertension which required daily medication, or smoking /drinking habit itself), but not so for detailed/subjective variables (e.g.number of cigarettes smoked per day or general life stress) .J Epidemiol, 1999 ; 9 : 183-189 proxy respondent, agreement, epidemiological information, misclassification, questionnaire In retrospective epidemiological studies, information on such lifestyle behaviors as smoking and drinking habit have commonly been collected from study subjects themselves, using a self-administered questionnaire or by directly interviewing them.
When a subject was deceased/too ill or impaired to obtain available information(such occasions as rapidly advancing fatal cancer, severe neurological diseases or fatal cardiovascular diseases), investigators must rely on information provided by proxy respondents, usually partners or other relatives.In previous case-control studies of lung cancer'', exposure histories were obtained from a surviving partner, children or siblings.
It was, however, uncertain whether proxy respondents could provide acceptable information concerning the study subjects.
In view of conducting retrospective epidemiological studies on fatal or too impaired diseases, it is essential to assess whether information provided by proxy respondents are acceptable or not.According to Gordis 5), most critical issue associated with the use of surrogate respondents might be possible misclassification of subjects and bias which would be related to mortality or presence of disease.
In recent years, attention has been drawn to examine the agreement of information provided by proxy respondents.Agreement of epidemiological information reported between subjects and proxy respondents have been studied by Kolonel et  for continuous variables in order to indicate the proportion of agreement achieved by the respondents beyond the amount that is expected by chance alone.
To evaluate the values of the kappa statistic and ICC, we adopted the same criteria recommended by Landis and Kock23): kappa statistic and ICC greater than 0.75 as excellent agreement, values between 0.40 and 0.74 as good agreement, and values less than 0.40 as poor agreement.
To test whether the level of agreement differed from that expected by chance, we computed z-scores using kappa, ICC and its standard error 24).
To examine whether partners systematically under-or overreport subject's health status and lifestyle compared to subjects' ratings, we used the partner-subject mean difference, which was expressed as the difference in proportions (dichotomous variables)/in mean values (continuous variables) between the subjects' self-and partner-reports 25).Mean difference was calculated by subtracting the proportion or the mean obtained with the subjects information from those obtained with partners information.A mean difference less than zero indicates that partners tended to underestimate the presence or the level of exposures 25).
The computer software SPSS ver.6 for Macintosh was used to perform all analysis.

RESULTS
Table 1 presents demographic characteristics of the subjects and partners.The majority of partners were wives (n=192, 85.7%), followed by husbands (n=32, 14.3%).Mean age tended to be older in the subjects than in partners (46.6*11.0 and 43.7* 9.6, respectively).There was no significant difference in mean age between wife-proxies and husband-proxies.
Table 2 shows percentages of missing information by ques- *Percentages of missing information are the proportion of subject/proxy respondents who could not answer the question to total number of subject proxy respondents tionnaire item according to the relationships between the subjects and the partners.Percentages of missing information were mostly below 10%(range:0-8%), but appeared to be higher in wife-proxies than in husband-proxies for family history and detailed aspects of smoking and drinking habit.Table 3 shows subject-partner agreement for epidemiological information.Exact per cent agreement exceeded 75 percent for all information.Kappa values ranged from a minimum of 0.34 (general life stress) to a maximum of 0.86 (smoking habit).Excellent to good agreement was found for present illness(from kappa=0.66 for diabetes mellitus to 0.81 for angina/myocardial infarction).For family history, kappa values ranged from 0.49 for stroke to 0.65 for subarachnoid hemorrhage.Kappa values were much larger in present illness than in family history.Agreement for general life stress was judged as poor (kappa=0.34).Partner agreement was excellent for smoking habit (kappa=0.86)and moderate for number of cigarettes smoked per day (ICC=0.60)and age at starting smoking (ICC=0.66).Agreement was also moderate for drinking habit (kappa=0.71),and for frequency and consumption pattern.Kappa value for physical exercise was 0.65: indicating moderate agreement.
Table 4 shows the differences in mean values (continuous variables) or proportions (dichotomous variable) between the subjects' self-and partner-reports.Partners tended to underreport the presence for dichotomous variables and the level of exposure for continuous variables.
Significant differences were found for the number of cigarettes smoked per day, drinking habit and amount of sake drinking, though the differences were not large.
Table 5 shows kappa statistic on agreement of information by proxy relationship to the subjects.Kappa values were greater in wife-proxies (0.46-0.84) than in husband-proxies (0.34-0.71); kappa being 0.64 for wife-proxies and 0.53, on an average, respectively.

DISCUSSION
The present study aimed to examine whether the same epidemiological information obtained by a self-administered questionnaire were concordant or not between the subject and partner, but not to examine accuracy of response by partner.Overall, indicated was in this study that partner agreement was good or excellent for most epidemioligical information, though partners tended to underreport the presence for dichotomous variables and the level of exposure for continuous variables .
We observed that percentages of missing information were much higher in wife-proxies than in husband-proxies , likewise in previous reports 18. 19).According to Nelson et al .19), proxy non-response was affected by the topic of the question, the degree of details required, and, most importantly , the relationship of index and proxy respondent to one another.Moreover, +"Go" is a traditional Japanese unit of sake consumption: one "go" is 180 ml.++ One bottle is 663 ml of beer .
when wife/husband doesn't know or was not of interest about detailed aspects of her/his partner's lifestyle, it can be assumed that the differences in the manner of providing a response may also affect the percentages of missing information.
With regard to present illness, our study found that partner agreement was excellent for hypertension and angina/myocardial infarction or good for diabetes mellitus, as previously reported 11.18).In our study, partner agreement for family history was judged to be good for all the 5 chronic diseases questioned.According to Heyman et al. 26), agreement for presence of such chronic diseases as hypertension and diabetes mellitus was high, whereas low for prior short-term /non life-threatening disorders.
Our finding also suggested that information provided by partner could be regarded as acceptable for present illnesses of chronic diseases, but not so for family history.Consistent with other studies 1.6-20), our study found that partner agreement was excellent for smoking and drinking habit itself: being a little higher degree of agreement in smoking habit than in drinking habit.Hatch et al. 16)reported that behav-ior status (e.g.smoker versus non-smoker) was more acceptably reported by partner than amount consumed (e.g.number of cigarettes smoked per day), likewise our study.In general, agreement seemed to be higher for the behaviors which partners could easily/directly observe (e.g.smoking /drinking habit itself), but lower for the behaviors that partners could not directly observe (e.g.numbers of cigarettes smoked per day and amount of alcohol consumed) 1,6.9,11) With regard to physical activity, our study found partner agreement as good, likewise a previous study18).
To our knowledge, few studies have examined partner-subject agreement on general life stress18).Our study found partner agreement for general life stress to be poor.It would generally be difficult for one person (even wife/husband) to guess another person's (husband/wife) general life stress status.Assuming that self-reported status of general life stress was truly correct, much of agreement on general life stress might seemingly be attributable to the degree of cognition of subject's condition by partner.
*For dichotomous variable , the mean differences is the differences in proportions.** Go" is a traditional Japanese unit of sake consumption: one "go" is 180 ml.** One bottle is 663 ml of beer.physical exercise 18), compared to the self reports, though the degree of underreport was more or less.In our study also, partners appeared to underreport medical history, smoking habit, alcohol consumption, and physical exercise, compared to the subject self reports, but almost all of the differences were neither large nor statistically significant.What is epidemiologically meaningful was that the greater the degree of partner's unawareness, the greater the tendency to underestimate subjects' behavior and health items.According to Pickle et al. 30), the most important variable which might affect the proxy agreement was reported to be the relationship between partner and subject which would be influenced by such variables as length of time, in which the subject and proxy respondent lived together, and frequency of proxy's contact with subject.

II
Consistent with other studies 9,13,18,29) our study found that kappa values were much larger in wife-proxies than in husband-proxies for almost all items examined.This might be due to the possibility that wives would watch their husbands' behaviors with much more attention 9).Some studies25,28) indicated that the more easily observed, the more objective and concrete question, a partner's response would be much closer to a subject's report.
In our study higher agreement would be obtained, if some partners were assisted by their subjects when filling the questionnaire, though we asked partners not to be assisted in answering.Judging from the pattern of concordance in response and tendency of underreporting by partners, such assistance might not be frequent in our study, likewise previous studies18,25,29).We could not examine whether blank column not filled by partner was due to merely an omission or any other reason, such as "don't know".We treated blank column as missing information.When the percentage of missing information was higher in partner report than in subject report, reduced sample size due to higher percentage of missing information might result in a decreased statistical power in assessing the agreement of partner report.In our study, however, the percentages of missing information were rather low (less than 10%), and the effect by reduced sample size on the degree of agreement would not be large.An additional methodological issue was that the time lag between the subject self-and partner-report.Our mean time lag was 4-5 weeks on an average.This short time lag was believed not likely to affect the degree of agreement, since subject's lifestyle would not change so quickly.
In our study, we examined the agreement between the subject self-and partner-report, assuming that information from subjects were true.This was simply because there was no objective standard for subjects' lifestyle.
In general, kappa statistic has been used as an index of the degree of agreement.Kappa statistic is, however, reported to be influenced by the prevalence of the characteristics of interest 31), and number of categories 32).According to Thompson et al. 31), kappa value varies with different prevalence of characteristics being measured: kappa being 0.46 for the symptom which has the prevalence of 1% and 0.80 for that having 40%.Kappa value tends to be lower for polytomies than for dichotomies 33), because kappa value is a measure of the frequency of exact agreement rather than a measure of the degree of approximate agreement 32).
In case-control studies, it is necessary to note the relationship between degree of agreement and odds ratio.Thompson et al. 31) indicated that the higher the kappa value, the closer the odds ratio obtained from proxy respondents to that obtained from subjects.This implies that the odds ratio obtained should be carefully interpreted when kappa showing low agreement.
In conclusion, we found that partners could provide acceptable information for the concrete and directly-observable variables (e.g.such present illness as hypertension which required daily medication, or smoking /drinking habit itself), but not so for detailed/subjective variables (e.g.number of cigarettes smoked per day or general life stress).
Restricted to current-smokers.* Restricted to regular-drinkers.

Table 1 .
Age and sex distribution of subject and partners.

Table 2 .
Percentages of missing information* by subject and partner (spouse).

Table 3 .
Subject-partner agreement for epidemiological information.No .ofpairs is often less than 224 because some partners could not answer to all questions.**95% confidence interval . *