Demographics, health behaviors, and past drug use as predictors of recall accuracy for previous prescription medication use

J Clin Epidemiol Vol. 50, No. 8, pp. 975-980, Copyright 0 1997 Elsevier Science Inc.

0895.4356/97/$17.00 PII SO895.4356(97)00026-7

1997

ELSEVIER

PHARMACOEPIDEMIOLOGY

REPORT

Demographics, Health Behaviors, and Past Drug Use as Predictors of Recall Accuracy for Previous Prescription Medication Use Suzanne L. West, ‘I* David A. Sauitz,’ Gary Koc~,~ Karen L. Sheff,3 Brian L. Strom,4 Harry A. Guess, 1 and Abraham G . Hartzema5 ‘DEPARTMENT OF EPIDEMIOLOGY, AND *DEPARTMENT OF BIOSTATISTICS, SCHOOL OF PLJBLIC HEALTH, UNIVERSITY OF NORTH CAROLINA, CHAPEL HILL, NORTH CAROLINA 27599-7400, ‘SOUTHWEST FOUNDATION FOR BIOMEDICAL RESEARCH, SAN ANTONIO, TEXAS 76902, 4C~~~~~ FOR CLINICAL EPIDEMIOLOGY AND BIOSTATISTICS, DEPARTMENT OF BIOSTATISTICS AND EPIDEMIOLOGY, AND THE DIVISION OF GENERAL INTERNAL MEDICINE OF THE DEPARTMENT OF MEDICINE, UNIVERSITY OF PENNSYLVANIA SCHOOL OF MEDICINE, PHILADELPHIA, PENNSYLVANIA 19104, AND ‘DIVISION OF PHARMACEUTICAL POLICY AND EVALUATIVE SCIENCES, SCHOOL OF PHARMACY, UNIVERSITY OF NORTH CAROLINA, CHAPEL HILL, NORTH CAROLINA 27599

ABSTRACT.

Drug data for pharmacoepidemiologic studies are often ascertained by self-report, but little research has addressed the factors influencing its accuracy. Stratified random sampling was used to select individuals for a study comparing interview data on past prescription drug use with dispensation information from the Group Health Cooperative of Puget Sound pharmacy database. The strata included age, gender, and recency of use. Recall accuracy and its determinants were evaluated for repetitively used non-steroidal anti-inflammatory drugs (NSAIDs), short-term NSAIDs (only a single dispensation), and post-menopausal estrogens. We investigated whether recall accuracy was influenced by education, marital status, race, smoking, alcohol consumption, cumulative drug history, the number of different NSAIDs or estrogens dispensed (both by name and dosage), and the number of dispensations of the drug in question. For repetitively used NSAIDs, recall accuracy was positively associated with the number of NSAID dispensations (the odds of recall were 1.7 [95% confidence interval {CL): 1.3-2.21 times greater for each additional four dispensations of the NSAID), the total number of drugs dispensed, and the number of different NSAIDs dispensed. For estrogen and short-term NSAID use, only higher educational attainment improved recall accuracy: the odds of recall were 4.1 (95% Cl: 1.4-11.7) and 2.1 (95% CI: l.O-

4.7) times greater for those with some college compared with those with only a high school degree, respectively. This

study

demonstrates

that

predictors

of recall

and the repetitiveness of its use. J CLIN EPIDEMIOL KEY WORDS.

Drugs, epidemiologic

accuracy

for previous

50;8:975-980,

methods, questionnaires,

Medication exposures in pharmacoepidemiologic studies can be ascertained by abstracting medical records, administering standardized questionnaires, or using automated databases. Both medical records and self-report are subject to omission errors, and little is known about the factors that influence the availability of drug information in the medical record or recall during the interview [l-3]. Recent evidence [2] suggests that major contributors to recall inaccuracies include the respondent’s age, the type of drug, and recall interval, i.e., the time between when the exposure last occurred and when it is recalled. Gender does not appear to influence underreporting of drug use and respondents rarely overreport drug use. ‘Address for correspondence: Suzanne L. West, Ph.D., Research Triangle Institute, Health and Social Policy Division, 3040 Cornwallis Road, P.O. Box 12194, Research Triangle Park, NC 27709-2194. Accepted for publication on 30 January 1997.

medication

use differ

by the type

of drug

1997. 0 1997 Elsevier Science Inc. recall, reproducibility

of results

The present investigation provides additional information on factors influencing the recall accuracy of repetitively used non-steroidal anti-inflammatory drugs (NSAIDs), short-term NSAIDs (single dispensation in the past 12 years), and post-menopausal estrogens. We compared drug dispensations recorded in the Group Health Cooperative of Puget Sound pharmacy database [4] with drug exposures ascertained in telephone interviews.

MATERIALS

AND METHODS

A stratified, random sample of 50-80-year-old men and women who had been dispensed NSAIDs and/or postmenopausal estrogens (women only) with 12 or more years of continuous enrollment in the Group Health Cooperative of Puget Sound health maintenance organization were selected for this study. Details of the selection procedure, data

S. L. West et al.

collection, and analysis have been described earlier [2]. Briefly, the stratification factors used in sample selection included age, gender, and recency of drug use. A target drug was selected to evaluate recall accuracy. A drug was eligible to be selected as a target drug based on pattern of use (“repetitive” defined as 90 or more days based on dispensation records and “short-term” defined as only one dispensation of only one NSAID in the 12-year study window) and date of last use. For people with only short-term NSAID use, the one NSAID dispensed was the target drug. For repetitively used NSAIDs and estrogens, the individual as well as their target drug were selected using a hierarchical prioritization scheme such that individuals were only eligible for random selection once. However, once selected, the recall accuracy of each eligible drug (NSAID or estrogen) was evaluated. Standardized telephone interviews employing pictorial drug handcards were used to obtain data on past prescription drug use [2]. These data were compared with dispensation information from the Group Health Cooperative of Puget Sound pharmacy database to assess recall accuracy for remembering the target drug name (brand or generic). Previous analyses compared the recall accuracy for repetitively used NSAIDs to that of short-term NSAIDs (single dispensation in the past 12 years) and post-menopausal estrogens, concentrating on age, gender, and recall interval as predictors of accuracy [2]. The present investigation provides additional information on factors influencing the recall accuracy, separately by drug type and repetitiveness of drug use. In the analyses reported here, we examine whether demographic factors (such as education, marital status, and race) or health and behavioral factors (such as current health status, smoking, and alcohol consumption) are associated with recall accuracy after taking the stratification factors into account. We also investigated whether recall accuracy was influenced by past drug consumption, such as the number of different NSAIDs or estrogens dispensed (both by name and dosage); the number of dispensations of the drug in question (as a measure of duration of use where the days’ supply for each dispensation ranged from one to four weeks); or the cumulative drug history over the past 12 years. SUDAAN [5], a statistical software package that accounts for sampling weights and provides standard errors corrected for correlation caused by individual respondents contributing to multiple strata, was used to compute crude recall accuracy proportions for each of the predictors. The SUDAAN program was also used for the logistic regression analyses in order to assess the independent effects of the recall accuracy predictors controlling for several factors simultaneously. Variables were selected for the final logistic regression models if they had odds ratios less than or equal to 0.7 or greater than 1.5 and pvalues of 0.10 or less after controlling for the stratification variables. Considering only the stratification variables and those variables selected for the regressions, all two-way interactions were evaluated to assess model fit.

RESULTS Of the 452 women and 248 men who were selected for this study, 343 women and 177 men were interviewed. This represents a 79% adjusted response rate [2]. Because the study focused on recall accuracy after last use of the drug, 66 people (12.6%) were excluded because they had been dispensed the target drug after sampling. Approximately 19% (n = 98) of the interviewed population were selected for not having used NSAIDs or estrogens. From the results of the previous analysis, the accuracy for recalling the NSAID name was 30% (95% CI: 24-36%) whereas that for recalling the estrogen name was 78% (95% CI: 70-86%). While age, sex, and recall interval were evaluated in the previous analysis, the current analysis focuses on the variables presented in Tables l-3. In these tables, the recall percentages and odds ratios for each predictor of recall accuracy are presented for repetitively used NSAIDs, short-term use of NSAIDs, and post-menopausal estrogens. Recall

of a Repetitively

Used NSAID

Poorer current health status, total number of different prescription medications taken in the past 12 years, number of target NSAID dispensations, and number of different NSAID brands dispensed were positively associated with recall of repetitively used NSAIDs when each was evaluated controlling for the stratification factors (Table 1). Multivariate modeling, controlling for the stratification variables, was used to assess the independent effects of education, current health status, alcohol consumption, marital status, number of target NSAID dispensations, and number of different types of NSAIDs. Little additional confounding was noted for the predictors of interest. Recall

of a Short-Term

NSAlD

Higher educational attainment was predictive of short-term NSAID recall (Table 2). Multivariate modeling, controlling for the stratification variables, was used to assess the independent effects of education, current health status, race, and marital status. Little additional confounding was noted, but there was some suggestion of an interaction between recall interval and education (p = 0.001): a shorter recall interval enhanced the accuracy of recalling the name of a singly dispensed NSAID primarily among those with lower education. However, this interaction may be spurious, driven by sparse counts in some of the strata. Recall

of a Post-Menopausal

Estrogen

A woman’s ability to recall past post-menopausal estrogen use was strongly influenced by education and tended towards being reduced among smokers (p = 0.13) (Table 3). Multivariate modeling, controlling for the stratification variables, was used to assess education, marital status, and smoking status. Only educational attainment continued to

Demographics,

Health

Behaviors,

Drug

977

Use

TABLE

I. Predictors for accuracy a methodologic study conducted Washington State

in recalling the name of a repetitively in 1992 at Group Health Cooperative Total

Predictor Education Some college + SHigh school Marital status Other Married Race Other White Smoking status Current smoker Never or ex-smoker Alcoholic beverage use 2-7 times per week 11 time per week Current health status Very good/Excellent Poor/Fair/Good Total number of past prescriptions 11-34

41 26

35-62 63 or more AlO’ Number

nb

used’ NSAID in of Puget Sound,

% RecallGd

Odds

ratio+

26 (13-40) 47 (26-68)

0.4 (0.1-1.3)

40 (17-63) 30 (17-44)

1.9 (0.5-6.9) Ref

Ref

2 67

3:(24-47)

5 67

43 (-1-87) 33 (21-45)

1.1 (0.2-5.1) Ref

17 55

50 (23-76) 29 (16-41)

Ref

Indeterminate

2.3 (0.6-8.8)

12

13 (-3-28)

0.2 (0.04-1.0)

60

38 (25-50)

Ref

21

27 (10-44) 27 (12-43) 45 (22-68)

Ref 0.8

30 21

(0.2-3.0) 5.9 (1.5-22.7) 1.4 (1.1-1.8)

of target

NSAID

dispensations

26 22 24

2-4 5-9 10+

13 (-1-27) 42 (20-63) 48 (28-69)

44'

Ref

4.4 (0.9-22.0) 6.3 (1.3-30.0) 1.7 (1.3-2.2)

Number l-3

of different

types

of NSAIDs

28 25

4-5 6+ A3i

19

20 (7-33) 37 (17-56) 46 (22-71)

Ref

2.5 (0.7-8.6) 5.2 (1.3-20.9) 3.6 (1.3-9.9)

“Median number of NSAlDs was six. hTotal n # 72 due to missing data. ‘Weighted to account for sampling. d95% confidence interval in parentheses. ‘Odds ratios adjusted for age, gender, and recall interval. ‘Odds ratio for an indicated number of units change on a linear scale

be an important ses: the

odds

predictor of recall

of estrogen were

name

in these

analy-

3.7 (95% CI: 1.3-10.7).

DISCUSSION

Table 4 summarizes the potential predictors of recall accuracy for the names of previously used drugs, combining these new analyses with those previously presented [2]. While the accuracy of recalling the target drug name was much better for post-menopausal estrogens than for NSAIDs, the predictors

of recall

accuracy

varied

by drug

type

(NSAIDs

ver-

sus estrogens) and by the repetitiveness of drug use. There may seem to be a contradiction between our previous report that repetitiveness of drug use does not influence recall accuracy [2] and the current analyses that suggest that drugs used for longer durations are recalled more accurately. The previous report compared recall for persons who had

only one NSAID dispensation in the past 12 years to that for persons who had been dispensed an NSAID for at least 90 days. In the previous analyses, recalling the NSAID name was equally poor for persons who had had only one dispensation of only one NSAID as for those who had multiple, i.e., two or more, dispensations of an NSAID. However, the current analyses refine the previous analyses by evaluating recall accuracy as a function of the absolute number of NSAID dispensations, a more specific analysis of duration of use than the simple dichotomy of one versus two or more dispensations. There has been very limited research on the factors that may influence the recall accuracy for past drug exposures [3]. For long-term recall of estrogen exposures, Goodman et al. [6] reported that Japanese ethnicity, higher educational attainment, and current nonsmoking status were positively associated with recall. Case-control status did not affect the

978

S. L. West

TABLE 2. Predictors in a methodologic Washington State

for study

accuracy conducted

Predictor Education Some college +
in recalling in 1992

et al.

the name of an NSAID dispensed only once at Group Health Cooperative of Puget Sound,

Total

nP

% Recall”’

Odds

ratiobqd

128 59

33 (22-43) 20 (9-32)

2.1 (1.0-4.7)

51

21 (8-34) 30 (21-39)

0.6 (0.2-1.5)

143

8 183

47 (8-86) 27 (19-34)

3.8 (0.6-25.3)

27 167

26 (6-45) 28 (20-37)

Ref

61 134

31 (15-47) 27 (18-35)

Ref

34 (22-47) 22 (12-32)

Ref

57 79 59

16 (4-28) 30 (17-42) 37 (22-53)

AlO’

Ref

Ref

Ref

1.0 (0.3-3.2) 1.4 (0.5-3.7) 1.6 (0.6-3.8)

Ref

2.4 (0.7-8.2) 3.7 (1.0-13.6) 1.3 (0.9-1.9)

“Total n # 195 due to missing data. bWeighted to account for sampling. ‘95% confidence interval in parentheses. dOdds ratios adjusted for age: gender, and recall interval. ‘Odds ratio for an indicated number of units change on a linear scale.

accuracy of reporting nor did marital status or family history of cancer. Similar results were found in the present study. Estrogen recall accuracy was significantly greater in those with higher educational attainment and uninfluenced by marital status. Recall accuracy also appeared to be greater in those who were currently nonor ex-smokers although these results were not significant statistically. Why current smoking status might influence the accuracy with which hormones are recalled is unclear. Exposure to cigarette smoke and/or nicotine may have a physiological effect on the body which impairs memory or more likely, women who are non- or ex-smokers may be more health conscious and may have taken a greater interest in this research. This interest translated into more active searching of memory for the information we requested on past hormone use. For the recall of drug exposures during pregnancy, de Jong et al. [7] found that prescription medications were recalled better than over-the-counter products. Overall accuracy was poor for recalling use of all drugs during pregnancy, even for drug class, ranging from 5% for over-the-counter nasal medications to 91% for general anesthetics. Pooling over NSAIDs and estrogens in the present study, the recall accuracy for the target drug name was 49.6%, compared with 51% for prescription medications in the de Jong et al. study. They reported that higher maternal education, earlier birth order, occurrence of preterm delivery, and lower five-

minute Apgar score were predictive of better recall of drug use during pregnancy. These findings suggest that both interference from other births and the saliency of the birth outcome affected the reporting accuracy of pregnancyrelated drug exposures. In a methodological study using a database of pharmacy dispensations to evaluate recall accuracy and completeness, van den Brandt [8] reported that accuracy improved as the number of drug dispensations increased, a finding similar to the present study for NSAIDs but not for estrogens. Whereas van den Brandt reported that recall was diminished as the total number of past drug exposures increased, we did not find corroborating evidence. A potential explanation for this discrepancy relates to study design. The current study evaluated the recall accuracy of a specific drug while the study by van den Brandt et al. required respondents to accurately and completely report all chronic drug exposures over the past q-year time-period. Consequently, more drug exposures could only make complete reporting less likely. Although there appears to be some agreement between the past studies evaluating the predictors of recall accuracy and those determined by the present study, the reasons for different predictors among the three drug groups studied remains unclear. This difference may result from the indication for drug use, which in turn might be influenced by life-

Demographics,

979

Health Behaviors, Drug Use

TABLE 3. Predictors for accuracy a methodologic study conducted Washington state

in recalling the name of a post-menopausal estrogen in in 1992 at Group Health Cooperative of Puget Sound,

Predictor

Education” Some college + SHigh school Marital status Other Married Race” Other White Smoking status Current smoker Never or exasmoker Alcoholic beverage use 2-7 times per week 51 time per week Current health status Very good/Excellent Poor/Fair/Good Total number of past prescriptions 2-28 29-47 48 or more AlO’ Number of target estrogen dispensations l-2 3-8 9+ 43’ Number of different types of estrogens One only Two 3+ Al’

Total

R=

% Recall”’

Odds

ratio6.-

76 38

85 (76-95) 61 (45-76)

41f (1.4-11.7)

35 81

73 (58-88) 80 (71-90)

0.7 (0.3-1.9) Ref

6 107

100 76 (68-85)

Indeterminate

13 103

60 (32-87) 81 (72-89)

0.4 (0.1-1.3) Ref

30 86

73 (57-90) 80 (70-89)

0.8 (0.3-2.2) Ref

46 69

76 (63-89) 79 (68-90)

0.8 (0.3-2.3) Ref

39 38 39

82 (67-95) 75 (61-89) 78 (63-92)

42 35 39

73 (60-86) 87 (76-98) 76 (60-91)

Ref 2.3 (0.7-8.1) 0.9 (0.3-3.0) 1.0 (0.8-1.1)

42 39 35

72 (59-85) 84 (72-96) 79 (64-94)

Ref 2.3 (0.8-6.7) 1.3 (0.4-4.0) 1.2 (0.6-2.1)

Ef (0.2-2.5) 0.8 (0.2-3.0) 1.1 (0.9-1.4)

Total n # 116 due to missing data. bWeighted to account for sampling. ‘95% confidence interval in parentheses. dOdds ratios adjusted for age and recall interval. ‘Odds ratm for an indicated number of units change on a linear scale

style or income. In a study of the completeness of reporting past chronic conditions, Madow et al. [9] found that completeness was improved for persons with lower educational attainment and poor current health, a finding similar to ours for the predictors of repetitively used NSAID recall. Madow’s findings, combined with de Jong’s and ours, suggest that the significance of the event for which the medications are used influences recall accuracy. A thorough discussion of the present study’s limitations has been reported elsewhere [2]. Briefly, recall accuracy was measured by comparing a respondent’s report of drug use with information available for these individuals from the Group Health Cooperative of Puget Sound pharmacy dispensation database. Drug dispensation, however, does not necessarily reflect compliance. In the present study, compliance was enhanced by requiring two or more consecutive dispensations for repetitively used NSAIDs, where six was

the median number of target NSAIDs actually dispensed. For those individuals who had only one dispensation of only one NSAID, poor compliance may have been reflected in their diminished recall accuracy. However, comparing the overall recall for repetitively used NSAIDs (33%) where compliance was optimized with recall for those with one NSAID (29%), the results are similar. Another limitation of the present study is that the results are generalizable only to persons who are white since more than 90% of our study population was white, which makes the assessment of race as a predictor extremely imprecise. Further, based on the wide confidence intervals for many of the odds ratios, random error calls into question some of the associations that were suggested. In conclusion, our data suggest that recalling the name of a previously used drug is influenced by the drug class as well as the repetitiveness of its use. Based on the results of

S. L. West et al.

980

TABLE 4. A summary of the racy in recalling the names of tified in a methodologic study Health Cooperative of Puget

potential predictors for accupreviously used drugs as idenconducted in 1992 at Group Sound, Washington State

Predictor

Repeat NSAIDs

Females” Advancing age” More recent drug usen Higher education Married White race Current cigarette smoking Weekly alcohol use Very good/excellent current health status Increasing number of drugs used previously Increasing duration of drug use

Increasing number of different drugs in the drug class used

1

One NSAID T ; -

Estrogens N/A c

-

Supported in part by grant HS 06906-01 from the Agency for Health Care Policy and Research in addition to grants from CIBA-GEIGY Colp., Summit, NJ; Merck B Co., Inc., West Point, PA; Pfizer Pharmaceuticals, New York, NY; Robert Wood Johnson Pharrnuceutical Research Institute, Raritan, NJ; and The Upjohn Co., Kafamaroo , MI. The authors would like to thank Dr. Edward Wagner, Director of the Center for Health Studies (CHS) at Group Health Cooperutioe of Puget Sound for access to the databases and for his helpful comments. We would like to thank the CHS staff, especially Cheri Anderson, Mary Verdery, and Alice Fisher for their assistance in collecting the data and programming the data j&s. We are also grateful to Phil Gallagher fur his programming expertise.

1 + 1‘

N/A

1‘

N/A

-

Abbreviations and symbols: N/A = not applicable; NSAIDs = non-steroida1 anti-inflammatory drugs; OR = odds ratio; dash = not predictive of recall accuracy; 1 = decreased recall (OR 5 0.7, p 5 0.10); ? = increased recall (OR 2 1.5, p 5 0.10); + = increased recall (OR = 1.4, p 5 0.10). 4Based on previous analyses [2].

our previous analyses, younger age and shorter recall interval tend to improve recall accuracy for NSAIDs while only a shorter recall interval improved the accuracy of recalling estrogen use. In the present analyses, educational attainment and current health status exhibited inconsistent effects on recall accuracy across the drug groups studied. Since demographic and behavioral characteristics appear to influence recall accuracy, differential distribution of such factors may introduce a spurious heterogeneity of the odds ratio [101,g’lvln g a fa 1se impression that the drug affects population subgroups differently. Thus, the possibility of differential recall accuracy should be considered in the design and conduct of pharmacoepidemiologic studies, especially studies assessing past exposure to NSAIDs. Pharmacy dispensation databases, in spite of their known limitations [3], may be a more accurate source of information on past drug use compared with the expensive, time-consuming, and potentially less accurate ascertainment of self-reported drug data.

References 1. West SL, Strom BL, Freundlich B, Normand E, Koch G, Savitz DA. Completeness of prescription recording in outpatient medical records from a health maintenance organization. J Clin Epidemiol 1994; 47: 165-171. 2. West SL, Savitz DA, Koch G, Strom BL, Guess HA, Hartzema A. Recall accuracy for prescription medications: Self-report compared with database information. Am J Epidemiol 1995; 142: 1103-1112. 3. West SL, Strom BL. Validity of pharmacoepidemiology drug and diagnosis data. In: Strom BL, Ed. Pharmacoepidemiology. 2nd ed. Chichester: John Wiley & Sons Ltd; 1994. 4. Saunders KW, Stergachis AS, Von Korff M. Group Health Cooperative of Puget Sound. In: Strom BL, Ed. Pharmacoepidemiology. 2nd ed. Chichester: John Wiley & Sons; 1994. 5. Shah BV, Bamwell BG, Hunt PN, LaVange LM. SUDAAN User’s Manual. Research Triangle Park, NC: Research Triangle Institute; 1991. 6. Goodman MT, Nomura AM, Wilkens LR, et al. Agreement between interview information and physician records on history of menopausal estrogen use. Am J Epidemiol 1990; 131: 815-825. 7.

De Jong PCMP, Bems MPH, Van Duynhoven YTHP, et al. Recall of medication during pregnancy: Validity and accuracy of an adjusted questionnaire. Pharmacoepidemiol 1995; 4:

23-30. 8. Van den

9.

10.

Brandt

PA,

Petri

of questionnaire

information

Pharm

[Sci]

Weekbl

H, Dorant

E, et al. Comparison

and pharmacy data on drug use.

1991; 13: 91-96. Data on Chronic Conditions Compared with Information Derived from Medical Records. Vi+ tal and Health Statistics, Series 2, No. 23. Washington, DC:

Madow WG. Interview

National Center for Health Statistics; 1967. Greenland S, Robins JM. Confounding and misclassification. Am

J Epidemiol

1985;

122: 495-506.