Treatment adherence is a key determinant of the success of antiretroviral therapy . However research on adherence has been hampered by the absence of a convenient means of measurement. Various objective methods are available, but each has significant practical limitations: their objectivity and sophistication rises in parallel with cost and difficulty in administration. Patient self-report is, therefore, the principal measure used in studies of adherence to antiretroviral therapy. Various questionnaires have been proposed, most recently that adopted by the Adult AIDS Clinical Trials Group (AACTG) . However, the validity of patients’ responses to such questionnaires is questionable because they have not been compared with objective adherence measures. Data from a small placebo study failed to find a significant correlation between items used in the AACTG questionnaire and electronic adherence measures . Therefore, we determined the validity of responses to our own questionnaire (Medication Adherence Self-Report Inventory: MASRI) compared with two objective measures: MEMS TrackCap (MC, a medication event monitoring system) and pill count (PC). We also determined whether high self-report adherence was associated with a favourable treatment outcome (undetectable plasma viraemia).
The study was performed in a publicly funded specialist clinic where all treatment was free of charge. Subjects were HIV-1-seropositive adults on stable combination antiretroviral therapy dispensed from the clinic's pharmacy. The following were excluded from the study: patients whose medication was administered by another person, those who were not able to read and understand English and those with frank cognitive impairment. Informed consent was obtained.
For each subject, the antiretroviral drug in their combination that presented the greatest barrier to adherence was selected (e.g. higher pill burdens, dietary requirements, more frequent dose intervals). The subject was given a bottle containing this drug closed with a MEMS TrackCap (Medication Event Monitoring Systems, APREX Corporation, Menlo Park California, USA). These are pill bottle caps containing a microprocessor that records the time the bottle is opened as a presumptive dose . Subjects were informed that the study related to their adherence to medication. However, they were not told that the MC recorded their doses. Subjects were asked to dose directly from the MC bottle, use no other container until the end of the study and keep the bottle closed between doses; those who could not agree to this were excluded from the study. Sufficient medication for 6 weeks was dispensed. Subjects returned the MC and unused medication after 1 month and a PC was performed. Without assistance, participants completed the MASRI with reference to the drug that was dispensed with the MC. They were assured that responses were confidential and would not be shared with their physician. Data were retrieved from the MC via a MEMS Communicator and processed using MEMS View 2.61 software (APREX Corporation). Data were extracted from the MC regarding proportion of doses taken over the preceding 3 days and 2 weeks for direct comparison with MASRI responses. MC data are expressed for the preceding month using two indices. First, observed dose events over expected (O/E); this is the number of presumed dosing events recorded on the MC as a proportion of that expected. The range may exceed 100%. Although O/E is the index most often cited, it gives an inflated estimate of adherence if the patient opens the bottle between doses without removing medication or ‘plays’ with the MC. Assuming that repeated opening of the bottle immediately after a presumed dose event does not represent repeat dosing, MEMS View software automatically censors events recorded within 15 min of a previous event. In order to reduce this artefact further O/E was recalculated excluding dose events repeated within 4 h (O/E4h). MC data regarding the timing of doses were expressed as the proportion of doses taken at the exact time a dose was due (± 15 min), within 30 min, 1 or 2 h of the time due and also the proportion less than or more than 2 h late. Plasma HIV viral load was measured by the bDNA 3.0 assay (Chiron, Emeryville, California, USA); the limit of detection was 50 copies/ml HIV RNA.
The MASRI consisted of 12 items with two broad themes. The first section related to the amount of medication actually taken. A permissive statement prefaced this section in order to encourage accurate responses. This read ‘we understand that many people on anti-HIV medication find it very difficult to take it regularly and often miss doses. We won't be surprised if you have missed lots of doses as well. We need to know how many doses you have missed'. Two response formats were used: Likert scales (LS) and a visual analogue scale (VAS).
There were five LS items: questions 1 to 3 were: ‘How many doses of medication did you miss … yesterday?', ‘ … the day before yesterday?’ and ‘ … the day before that (3 days ago)?'. Possible responses were 0, 1, 2, 3 or don't know.
Item 4 asked ‘How many doses of medication have you missed in the 2 weeks before that?’ and possible responses were 0, 1, 2 or more, all of them, or don't know. Patients responding ‘2 or more’ were asked to state ‘roughly how many?'
Item 5 was ‘When was the last time you missed a dose of medication?’ Responses were: today, yesterday, earlier this week, last week, less than a month ago, more than a month ago, never or don't know. For analytical purposes responses of ‘never’ or ‘more than a month ago’ were grouped together, as it was not possible to differentiate them by MC.
The visual analogue scale (VASDOSE) was a VAS for the proportion of doses taken in the preceeding month. The instructions for VASDOSE read: ‘put a cross on the line below at the point showing your best guess about how much medication you have taken in the last month. We would be surprised if this was 100% for most people, e.g. 0% means you have taken no medication; 50% means you have taken half your medication; 100% means you have taken every single dose of medication'. VASDOSE ranged from 0 to 100% in 10% intervals. Participants whose period of observation was less than 25 days were excluded from analyses relating to reported adherence over the preceding month.
The second part of the MASRI addressed the timing of doses. The first four items used the same stem: ‘In the last month, how much of your medication did you take … at the time you were supposed to?’ Items 1–4 then added (1) ‘at the exact time (to within a few minutes)', (2) ‘within half an hour', (3) ‘within 1 hour or (4) ‘within 2 hours of the time you were supposed to?’ Possible responses on an LS were ‘none, very little, less than half, about half, more than half, nearly all or all'.
In the final two items, participants were asked to indicate on a VAS from 0–100% the proportion of doses taken within 2 hours of the correct time (VAS<2h) or more than 2 hours late (VAS2hLATE). The language of MASRI was English with a Flesch reading ease score of 91.3.
The 3 day and 2 week self-report LS items were based on published work [5,6]. VAS items were developed from the concept of asking patients to indicate their ‘percentage adherence’ by selecting the corresponding card . There are no published questionnaires designed to measure dose timing, hence items relating to this were generated by the investigators. MASRI examined a range of time frames around the ideal dosing interval, focusing particularly on the 2 hour interval that, in the clinical practice of the authors, was most meaningful to patients. Before the study, the questionnaire was piloted with a small sample of known patients, and then with a larger group of anonymous patients . Ethical committee approval for the study was obtained.
Parametric data are presented as mean with standard error (SE) and non-parametric data as median with interquartile range (IQR). Differences between two independent group means were tested using unpaired t-test; for paired data, paired t-tests were used. In order to examine discrepancies in responses on LS compared with MC data, LS and MC data were treated as dichotomous variables: fully adherent versus not fully adherent. Agreement between the two data sets was tested using kappa (κ) statistics, which may vary between 0 and 1; values of κ closer to 1 indicate more perfect agreement. The sensitivity and specificity of each item in correctly identifying various levels of adherence within the range of data available were calculated. Receiver operating characteristic (ROC) curves were derived for each of level of adherence tested and the area under these curves (AUC) was calculated. AUC values nearer to 1 are considered more satisfactory and this was tested against the null hypothesis that the true AUC was 0.5. Correlation between parametric quantitative data was tested with Pearson's correlation coefficient (ρ), which may range from 1 to −1. Where r = 0, the two variables are completely independent and show no correlation; values nearer 1 or −1 indicate more perfect positive or negative correlation, respectively. Quantitative data with hypergeometric distribution were normalized using log10 transformation. Spearman's rank correlation (ρ) was used to test association between data with bimodal distribution (viral load). MASRI items relating to number of doses actually taken or missed were converted to percentages of doses expected in order to account for variation in dose frequency between drugs. These were entered into multivariate linear regression models with MC adherence indices as the dependent variable. Regression coefficients are presented with 95% confidence intervals (CI). Items relating to times of doses were excluded from this analysis as they relate to a different concept of adherence. All P values are presented using two-tailed tests. Statistical analysis was performed using SPSS 9.0.0 for Windows software (SPSS Inc., Chicago, Illinois, USA).
Of the 104 patients who were approached for the study between April 1999 and May 2000, three could not give their commitment to use the MC as instructed, two were ineligible because of a prior diagnosis of HIV encephalopathy and one because of low English comprehension. Ninety-eight patients consented to the study but 20 were excluded from analysis: 13 admitted to not using the MC for all their doses; two subjects failed to return the questionnaire; two returned the MC having used it for only 1 week. One patient was lost to follow-up, one lost the MC and one withdrew consent. Of the 78 completing the study, 85.9% were gay males and 5.1% had a history of intravenous drug use. English was the first language of 66.7%. Mean age was 39.9 years (SE, 0.9). AIDS had been diagnosed in 47.4%. The most frequently studied drugs were protease inhibitors (39.7% indinavir, 38.5% nelfinavir, 6.4% saquinavir) followed by nucleoside reverse transcriptase inhibitors (10.2%) then non-nucleoside reverse transcriptase inhibitor (5.2%). These drugs were prescribed once daily in 2.6%, twice daily in 53.8% and three times a day 43.6%. Mean daily pill burden was 12.7 (SE, 0.5); median duration of antiretroviral therapy was 872 days (IQR, 615–1203) and for the current regimen was 642 days (IQR, 450–948). This was the first combination for 59% of subjects. Mean CD4 lymphocyte count was 400 × 106 cells/l (SE, 23). None of these characteristics was associated with adherence. The mean period of observation was 30.2 days (SE, 0.8). Nine participants were monitored for less than 25 days; the mean observation period for the remainder was 31.3 days (SE, 0.8). Mean adherence measured objectively was O/E 92.9% (SE, 1.8%), O/E4h 89.4% (SE, 1.8%) and PC 96.8% (SE, 1.4%); this did not differ significantly between the monitored drugs. Adherence measured by PC was significantly higher than that measured by MC (P < 0.005) and in three subjects PC was between 111 and 151%.
Validity of MASRI dose adherence items
MASRI items about adherence on a particular day or about when a dose was last missed displayed low agreement with MC. This improved when responses for the 3 day items were summated. However the highest level of agreement was for recall of doses missed over the preceding fortnight (Table 1). There was a significant correlation between MC data and the items relating to the last 3 days (summated), fortnight and VASDOSE (Fig. 1 and Table 2). VASDOSE also correlated with PC (r = 0.75;P < 0.001). The items relating to the last 3 days (summated), fortnight (`doses missed over the 2 weeks before') and VASDOSE all displayed a high degree of specificity for non-adherence (see Table 3). While there was no statistically significant difference between the areas under the ROC curves, there was a trend towards higher sensitivity for the VASDOSE item. However the sensitivity of all three items at the lowest level of adherence examined (≤ 80%) was low. Subjects gave the following responses to the item regarding last missed dose: ‘today’ 1.5%, ‘yesterday’ 6.1%, ‘earlier this week’ 12.1%, ‘last week’ 16.7%, ‘less than a month ago’ 9.1%, ‘more than a month ago’ 21.2%, and ‘never’ 33.3%. Agreement with the MC is shown in Table 1. Of the subjects who reported their last missed dose was more than a month ago, 10 indicated < 100% adherence on the VASDOSE question; one subject responded ‘85%’ on VASDOSE (O/E4h, 75.8%). All the remaining subjects scored > 95% on VASDOSE (mean 97.9%; SE, 0.5%) and for five the VASDOSE was ≥ 98%.
Validity of MASRI dose timing items
Respondents indicated on VAS<2h that they had taken a median of 95.7% (IQR, 70–100) of doses within 2 hours of the correct time. After log10 transformation, correlation with the MC was 0.49 (P = 0.001). For VAS2hLATE, the median response was 3.8% (IQR, 0.0–14.0). After log10 transformation, there was no correlation with the MC. However following removal of a solitary outlier (VAS2hLATE 98.0%), a significant association emerged (r = 0.37;P = 0.02). On the LS relating to dose timing, 7.9% reported taking all of their doses exactly on time, 14.5% reported taking them all within 30 min of the exact time, 26.7% within 1 h and 44.1% within 2 h. However, these responses did not agree with MC observations.
Linear regression analysis
Results for univariate analysis are shown in Table 2. Multivariate regression models were constructed using both of the MC adherence indices as dependent variables. In all models, 3 day adherence report was no longer significantly associated with MC. Results for the variables remaining in the models are shown in Table 4. Estimates provided by the full data set may have been inflated by a small number of data points associated with the lowest reported adherence. Therefore, further models were constructed after removing these data points (adherence over last fortnight < 80%, VASDOSE < 60%). Using these more conservative estimates, adherence over the previous fortnight was no longer significantly associated with MC observations. VASDOSE emerged as the only independently associated variable (Table 4).
Association with plasma HIV viral load
Plasma HIV viral load was < 50 copies/ml in 53.2% (IQR, < 50 to 2886 copies/ml). There were significant inverse correlations between viral load and both MC indices of dose adherence (O/E: ρ = −0.37, P = 0.001; O/E4h: ρ = −0.29, P = 0.01;Fig. 2). Examining items within the MASRI there were similar associations with most of the dose adherence items (`last fortnight’ (%): ρ = −0.30, P = 0.010; VASDOSE: ρ = −0.28, P = 0.014 (see Fig. 2); ‘last missed dose': ρ = −0.28, P = 0.017). The mean adherence of those with undetectable viraemia was significantly higher than that of subjects with detectable viraemia for the majority of adherence measures [O/E, 98.6% versus 86.2% (P = 0.003); O/E4h, 93.6% versus 84.3% (P = 0.011); PC, 99.4% versus 93.5% (P = 0.047); ‘last fortnight’ (%), 98.1% versus 92.8% (P = 0.031); VASDOSE, 96.2% versus 89.7% (P (P = 0.012)]. However, there was no significant association between viral load and the 3 day adherence items.
No association was found between viral load and any of the indices of dose-timing accuracy derived from the MC. For example, the mean proportion of doses taken more than 2 h late by subjects with detectable viraemia was 14.4% (SE, 1.8) while for subjects with undetectable viraemia it was 12.9% (SE, 1.7). No association emerged when timing of protease inhibitor drugs or of any individual drug was examined. Neither was there an association between viral load and MASRI items regarding dose timing.
Items within the MASRI relating to whether or not doses were taken showed varying levels of agreement with MC data. Subjects were able to give better estimates of adherence over longer time periods than when asked to recall doses missed 1, 2 or 3 days ago. The strongest independent association with MC data was provided by VASDOSE. Of the dose timing items VAS<2h was most strongly associated with MC data. To our knowledge the MASRI is the only questionnaire on adherence to antiretroviral drugs validated against objective measures. Its use of multiple measures of adherence allows the consistency of an individual patient's responses to be checked.
Adherence to medication is a multifaceted phenomenon the precise definition of which may differ from drug to drug depending on, for example, the fasting or fluid intake requirements of each agent. While undoubtedly important, such considerations are impossible to monitor objectively. We, therefore, chose to examine the performance of patient self-report in relation to the two central questions in adherence: did the patient take the medication and how regular were the dosing intervals? The problems with patient self-report have been well rehearsed: while it is a fairly specific method for detecting non-adherence it is not sufficiently sensitive to exclude it and patient recall remains an obstacle . Nevertheless self-administered questionnaires are economical and convenient. They are likely to remain a cornerstone of the research methodology in this field. However, before introducing any new research instrument, its validity must be assessed: the extent to which the instrument measures what it claims to measure. Validity is a matter of degree rather than an all-or-nothing property .
In the absence of published data comparing responses to different types of question on adherence with objective measures, a variety of question formats were used within the MASRI. Responses by LS and VAS have been shown to be comparable for measurement of quality of life  and hence LS is often preferred in questionnaire design because of ease of interpretation and administration . However the same considerations may not hold for treatment adherence, where responses relate to recall of discrete dosing events rather than attitudes. As regards the dose-timing items, the clinical significance of regular dose timing is unknown. Nevertheless, especially for drugs with rapid clearance, it is biologically plausible that repeatedly delayed doses are likely to promote escape from virological suppression.
We have shown that MASRI items relating to dose adherence had moderate sensitivity and high specificity for detection of adherence below a range of thresholds and there was a trend towards higher sensitivity for the VASDOSE. VASDOSE consistently provided the strongest linear association with MC, even with the most conservative analyses (Table 4). Ten subjects claimed full adherence when asked when they had last missed a dose but indicated < 100% adherence on VASDOSE; the majority of these subjects reported very high adherence on VASDOSE. It is not possible to determine which of the following possibilities might explain these discrepancies: first, the subject may have wished to indicate that they may have actually missed occasional doses; second, subjects may not have been sufficiently careful with the placement of their cross on the VASDOSE and missed the 100% mark; third, they have misunderstood the instructions for this item. Nevertheless, these subjects did consistently report very high adherence across both these items. Although the summated 3 day self-report provided a reasonable estimate of adherence, subjects had poor recall of their dosing behaviour on specific recent days. The objective measures and longer-term self-report items displayed an inverse relationship with plasma HIV viral load. This is consistent with data reported by Paterson et al. . However, we found no association between viral load and either summated 3 day self-report or adherence on individual days.
Reporting of accuracy of dose timing on LS was unreliable. However, use of VAS-based responses resulted in a significant improvement in the accuracy of patient recall of dose timing. For drugs such as protease inhibitors, where the trough plasma concentration is close to the minimum effective concentration, frequent late dosing may permit rounds of viral replication and promote treatment failure. Interestingly, we found no association between even the greatest delays in dosing (doses taken more than 2 h late) and viral load. This suggests that, in clinical practice, ensuring that patients do not miss doses is more important than concerns about timing of doses.
Questionnaires on adherence published to date have adopted two main approaches. Earlier questionnaires aimed to identify attitudes or practical barriers to treatment likely to be associated with non-adherence [13–15]. Recently, a more direct approach has been adopted: subjects are asked about adherence on specific days before the clinic visit [2,6,7,16]. The Morisky scale  is a widely used example of the attitudinal questionnaire. It consists of four questions with dichotomous responses. The range of possible scores is 0 to 4, limiting the variance of the tool. The items make no reference to a specific time period over which adherence is being assessed. Hence the implication is that these items are attempting to identify ‘non-adherent individuals’ rather than identify non-adherent behaviour. This contrasts with current models of adherence, which recognize that the decision to adhere is not once and for all. Rather it is a series of conscious or subconscious decisions about the perceived importance of adherence during the entire dosing period . Moreover, even patients with favourable attitudes to treatment may miss doses accidentally. Attitudinal questionnaires also give no information about how many doses have been missed.
In an early attempt to gauge directly adherence by self-report, Fletcher et al. asked patients to quantify their adherence by choosing a percentage on a card; with serum digoxin level as the gold standard, self-report was more valid than pill count . Using an LS-based 4 week self-report questionnaire, Haubrich et al. described an association between self-report and reductions in viral load . However MC data from Paterson et al. showed that 22% of highly adherent patients experienced virological failure and, conversely, 18% of those displaying the lowest adherence had undetectable viraemia . Therefore, viral load should not be used as a surrogate marker of adherence. In order to reduce incorrect classification of highly adherent patients as non-adherent, the validity of self-report is best tested by direct comparison with MC. Bangsberg et al. described an association between viral load and 3 day self-report, PC and 3 day MC data (adjusted for doses reported taken from another container and for cap events not corresponding to a dose) . However, no direct comparison of the performance of each measure was reported. Golin et al. found that a 7 day self-report did not significantly correlate with MC data  while Wagner and Rabkin  have reported a comparison of 3 day self-report with MC data in patients taking placebo ‘indinavir’ for 2 weeks. Although (perhaps because of the small sample size) they found no statistically significant correlation, it is notable that the correlation coefficient (0.34) was similar to that found in our cohort. Recently Chesney et al. reported the performance of the AACTG Adherence Instrument in a pilot study . The principal adherence measures employed were LS for 2 day self-report, dose timing and recall of last missed dose. The external validity of the instrument was not reported, although construct validity was inferred by association of non-adherence with, for example, higher alcohol consumption and lower adherence self-efficacy. However Paterson et al. have questioned whether such items are sufficiently sensitive to detect adherence of 95% or greater  and, in the present study, such items did not emerge as the most powerful indicators of adherence in comparison with objective data.
The present study has a number of limitations. First, of necessity, the study population was selected by our inclusion criteria to have certain characteristics that would promote higher adherence; indeed subjects who admitted to deviating from our instructions were excluded from the analysis. While this was unavoidable if MC data were to be meaningfully interpreted, it may have skewed the observed behaviour towards higher levels of adherence. Second, participation in the study might have increased subjects’ awareness of their dosing behaviour during the period of observation, artificially improving recall of missed doses. This may have been exacerbated if subjects had been made aware that adherence was being monitored (the effect of electronic monitoring on dosing behaviour is unknown ). We, therefore, took steps to minimize this effect by not alerting subjects to the purpose of the MC. Third, MC data may overestimate adherence. The system is unable to determine whether the patient removed a dose when the bottle was opened and when (if at all) it was actually taken. However, it is unlikely that patients would regularly open the bottle without using the contents . Individuals also occasionally ‘play’ with the MC, opening it many more times than necessary on a given day. In an attempt to minimize these problems, we have provided MC data in a number of formats. In common with our data, previous studies have shown that adherence measured by PC routinely exceeds MC data . This has been ascribed to ‘pill dumping’ (where patients discard excess medication before returning the bottle in order to appear more adherent)  or sharing of medication with friends or family . In the present study, it is also possible that for some subjects PC was higher that MC because, contrary to instructions, doses were removed from the bottle for later consumption. Fourth, notwithstanding the above, the observed adherence in our population was relatively high. This was consistent with the results of our previously reported study using anonymous self-report  and such levels of adherence may perhaps be typical for our clinic population. This may be because of its demographic profile (notably the low prevalence of intravenous drug users). It may also reflect the level of adherence support provided within our clinic, including a specialist adherence nurse and multidisciplinary adherence clinics for patients starting and changing antiretroviral therapy. The performance of the questionnaire in populations with lower adherence needs to be established. Fifth, questionnaire items were generated by the investigators and, therefore, may not fully represent the constructs used by patients in recalling their own adherence. Finally, as all measures of adherence have significant limitations, the MASRI would best be used in tandem with other methods of measuring adherence in future studies. Despite these limitations, we believe our findings are relevant to future research and practice.
The MASRI is the first questionnaire on adherence to antiretroviral therapy to have been validated against an objective measure. Recall of adherence was improved by asking patients to estimate adherence over a longer time period; recall of dosing behaviour on specific recent days was poor. LS provided poor recall of accuracy of dose timing in comparison with VAS items. Dose adherence, but not accuracy of dose timing, was associated with virological outcome. Measurement of adherence to antiretroviral therapy is important in both clinical care and research. This brief questionnaire provides a validated tool for its assessment.
1. Carpenter CCJ, Cooper DA, Fischl MA. et al
. Antiretroviral therapy in adults.Updated recommendations of the International AIDS Society – USA Panel.
JAMA 2000, 283: 381–390.
2. Chesney MA, Ickovics JR, Chambers DB. et al
. Patient Care Committee and Adherence Working Group of the Outcomes Committee of the Adult AIDS Clinical Trials Group (AACTG).Self-reported adherence to antiretroviral medication among participants in HIV clinical trials: the AACTG Adherence instrument.
AIDS Care 2000, 12: 255–266.
3. Wagner GJ, Rabkin JG. Measuring medication adherence: are missed doses reported more accurately then perfect adherence? AIDS Care 2000, 12: 405–408.
4. Cramer JA, Mattson RH, Prevey ML, Scheyer RD, Ouellette VL. How often is medication taken as prescribed? A novel assessment technique. JAMA 1989, 261: 3273–3277.
5. Chesney MA, Hecht FM. Adherence to HIV antiretroviral therapy: an essential element to understanding treatment effect.2nd European Conference on Methods and Results of Social and Behavioural Research on AIDS.
January 1998 [abstract].
6. Hecht FM, Colfax G, Swanson M, Chesney MA . Adherence and effectiveness of protease inhibitors in clinical practice.Fifth Conference on Retroviruses and Opportunistic Infections.
Chicago. February 1998 [abstract 151].
7. Fletcher W, Pappius EM, Harper SJ. Measurement of medication compliance in a clinical setting. Arch Intern Med 1979, 139: 635–638.
8. Walsh JC, Dalton M, Gazzard BG. Adherence to combination antiretroviral therapy assessed by anonymous patient self report. AIDS 1998, 12: 2361–2362.
9. Melkinow J, Kiefe C. Patient compliance and medical research: issues in methodology. J Gen Intern Med 1994, 9: 96–105.
10. Nunnally JC. Psychometric Theory, 2nd
edn. New York: McGraw-Hill; 1978.
11. Jaeschke R, Singer J, Guyatt GH. A comparison of seven-point and visual analogue scales.Data from a randomised trial.
Control Clin Trials 1990, 11: 43–51.
12. Paterson D, Swindells S, Mohr J. et al
. Adherence to protease inhibitor therapy and outcomes in patients with HIV infection
. Ann Intern Med 2000, 133: 21–30.
13. Morisky DE, Green LW, Levine DM. Concurrent and predictive validity of a self-reported measure of medication adherence. Med Care 1986, 24: 67–74.
14. Frick PA, Gal P, Lane TW, Sewell PC. Antiretroviral medication compliance in patients with AIDS. AIDS Patient Care STD 1998, 12: 463–469.
15. Pekovic V, Mayanja H, Vjecha M. et al
. Comparison of three composite compliance indices in a trial of self-administered preventive therapy for tuberculosis in HIV-infected Ugandan adults.Uganda–Case Western Reserve University Research Collaboration.
J Clin Epidemiol 1998, 51: 597–607.
16. Bangsberg DR, Hecht FM, Charlebois ED. et al
. Adherence to protease inhibitors, HIV-1 viral load, and development of drug resistance in an indigent population. AIDS 2000, 14: 357–366.
17. Leventhal H, Diefenbach M, Leventhal EA. Illness cognition: using common sense to understand treatment adherence and affect condition interactions
. Cognit Ther Re 1992, 16: 143–163.
18. Haubrich RH, Little SJ, Currier JS. et al
. The value of patient-reported adherence to antiretroviral therapy in predicting virologic and immunologic response. AIDS 1999, 13: 1099–1107.
19. Golin C, Liu H, Hays R et al.Self-reported adherence to protease inhibitors substantially overestimates an objective measure.Sixth Conference on Retroviruses and Opportunistic Infections.
Chicago, January–February 1999 [abstract 95].
20. Feinstein AR. On white-coat effects and the electronic monitoring of compliance. Arch Intern Med 1990, 150: 1377–1378.
21. Miller LG, Hays RD. Measuring adherence to antiretroviral medications in clinical trials. HIV Clin Trials 2000, 1: 36–46.
22. Gordis L. Conceptual and methodological problems in measuring patient compliance.
In Compliance in Health Care.
Edited by Haynes RB, Taylor DW, Sackett DL. Baltimore: Johns Hopkins University Press; 1979: 23–45.