AIDS PATIENT CARE and STDs Volume 20, Number 10, 2006 Mary Ann Liebert, Inc.

A Measurement Model of Medication Adherence to Highly Active Antiretroviral Therapy and Its Relation to Viral Load in HIV-Positive Adults
MARIA M. LLABRE, Ph.D.,1 KATHRYN E. WEAVER, M.S.,1,2 RON E. DURN, Ph.D.,3 MICHAEL H. ANTONI, Ph.D.,1 SHVAWN McPHERSON-BAKER, Pharm.D.,4 and NEIL SCHNEIDERMAN, Ph.D.1

ABSTRACT This study compared a multiple method measurement model of highly active antiretroviral therapy (HAART) adherence with single-method models to determine optimal validity in predicting HIV viral load. Repeated measures of antiretroviral adherence were collected over a 15-month period using three different measurement methods: a self-report questionnaire, an adherence interview item, and electronic medication monitoring. The participants included HIV-positive men and women (n 323) who were currently prescribed HAART. Single-factor models composed of multiple measurements over time were developed for each adherence method and HIV viral load. The three adherence methods were then combined in a second order factor measurement model. Structural equation modeling was used to test the models. Mean adherence, defined as percent of doses taken, was 92%, 90%, and 57% by selfreport, interview, and electronic monitoring, respectively. Reliability of individual measurements of adherence was low. Four or seven assessments were needed to attain acceptable stability, depending on the method. The second-order factor model of adherence fit the data and explained 45% of the variability in HIV viral load. Models including only one method of assessing adherence explained between 20% and 24% of the variability. Models that included both self-report and electronic monitoring optimized predictive validity. Using at least two different methods of adherence measurement, each assessed at multiple times is recommended to derive reliable and valid measurement of medication adherence, which is predictive of biological outcomes such as HIV viral load.

INTRODUCTION

antiretroviral therapy (HAART) has resulted in reductions in AIDS-related morbidity and mortality in the United States1 and made medication adHE ADVENT OF HIGHLY ACTIVE
1Department 2Department

herence a key issue in the management of HIV/AIDS. HAART can lead to reductions in the HIV virus to near undetectable levels; however, suboptimal HAART adherence can have serious consequences for patients and the public alike through the emergence of drug resis-

of Psychology, University of Miami, Miami, Florida. of Psychiatry, University of Illinois-Chicago, Chicago, Illinois. 3California School of Professional Psychology, San Diego, California. 4Department of Microbiology and Immunology, University of Miami, Miami, Florida.

701

702

LLABRE ET AL.

tant HIV strains. HAART has been called the unforgiving regimen, because an extraordinarily high adherence level (90%95% or greater) is necessary for successful virologic control.2,35 Despite the importance of this health behavior, it is estimated that 18% to 73% of HIV-positive persons are nonadherent to HAART.4,68 Poor adherence to antiretroviral medications has been associated with therapeutic failure, disease progression, and viral mutation.9,10 Methods for assessing medication adherence include patient self-report, pill counts, pharmacy refills, medical professionals reports, electronic monitoring devices, direct observation, and blood plasma levels of antiretroviral medications. Reviews of these methods indicate that measuring medication adherence is problematic, because each method introduces its own type of measurement error.9,1113 Selfreport measures may overestimate adherence compared to electronic devices; however, electronic devices may underestimate adherence because of inconsistency in their use1416 or may interfere with methods patients use to improve their adherence, such as the use of pill boxes.17 Both self-report and electronic HAART adherence monitoring methods have been shown to significantly predict HIV viral load with correlations in the 0.20.4 range,1821 with at least three studies reporting stronger effects.14,22,23 Correlations among adherence monitoring methods are similarly variable, with some studies finding statistically significant relationships among varied methods of assessing adherence and others not.14,1820,23,24 The higher correlations are between methods sharing similar sources of error.25 Taken together, these results suggest that a combination of approaches may be necessary to capture adherence behavior. Despite calls to use multiple adherence measures,2,17,26 no study to date has incorporated multiple measures of medication adherence in a latent factor that captures variance common to all measures. This approach allows for an estimate of true adherence, because error uniquely associated with a particular method can be identified. The need for a gold standard (i.e., commonly accepted single measurement procedure) would also be circumvented, as

each measure could be evaluated by examining its contribution to the underlying factor. It should be noted, however, the only measure that guarantees the medication has entered the body involves actually documenting medication blood levels.27 The purpose of this study was to test a measurement model of medication adherence incorporating multiple measurements collected over time with multiple methods in a heterogeneous sample of HIV-positive men and women on HAART. Combining self-report, interview, and electronic monitoring data using structural equation modeling (SEM) allowed us to minimize measurement error and thereby estimate true adherence and its relationship to HIV viral load. Measurement model of adherence Our measurement model of adherence assumes that an observation is made up of three additive components: true adherence, systematic error associated with a specific measurement procedure, and random measurement error associated with a specific measurement at a specific time. By measuring a sample of individuals with different methods multiple times, it is possible to separate these components and use the variability in true adherence to predict viral load, thus eliminating measurement error in the estimate of the effect of adherence on viral load. Although adherence has been defined in a variety of ways, we treated adherence as a continuous variable (percentage of doses taken). The model assumes that in the absence of intervention, true adherence is stable over time but variability in adherence on a specific measure over time reflects random measurement error. There is evidence for this assumption in the literature14,28; however, changes in adherence have been observed in the context of clinical interventions designed to improve medication adherence over both short and long periods of time.29 We conducted preliminary analyses to confirm the absence of systematic change over time in our sample. The model also recognizes that different methods of measuring adherence introduce systematic error or bias that is specific to the measurement procedure.

MEASUREMENT MODEL OF ADHERENCE

703

SEM is suited to address the measurement of adherence and its relation to biologic outcomes such as viral load. In SEM, measurement models (i.e., models incorporating multiple methods and/or time points) may be specified and tested with confirmatory factor analysis. SEM also accommodates models that specify the relationships among latent variables. In this manner, the effect of a predictor may be estimated and tested once measurement error has been removed. MATERIALS AND METHODS Participants Participants were 323 HIV-positive adult men and women from a randomized trial investigating the effects of a 10-week cognitive behavioral stress management (CBSM) intervention on immunologic, health, and psychosocial outcomes. Recruitment took place from 19992004. We combined data from inTABLE 1. SOCIODEMOGRAPHIC
AND

tervention and control participants since intervention effects were not the focus of the present study and both conditions had the same adherence training. Before combining the control and intervention participants, we performed a test of group equivalence comparing the variance/covariance matrices and means between the two groups on all adherence and viral load measures. Results indicated the two groups were comparable. This stringent test of group equivalence provides empirical justification for combining groups. Sociodemographic and disease-related characteristics of the combined sample are shown in Table 1. Adults, 18 to 65 years old on HAART were recruited through clinics, service agencies, community health centers, advertisements, and previous participant referral. Participants were excluded if they were prescribed medications with immunomodulatory effects, or had a history of chemotherapy, whole-body irradiation, or chronic illness associated with immune system changes. Temporary exclusion criteria inOF THE

HIV-RELATED CHARACTERISTICS

SAMPLE 189 134

Gender Men Women Ethnicity Latino/Hispanic Non-Hispanic black Non-Hispanic white Other (American Indian, Asian, or Multiethnic) Employment Status Full-time or part-time Unemployed or disabled Education Less than high school Completed high school Some college 4-year degree or more Income (including public assistance) Less than $5,000 $5,00010,000 $10,00020,000 $20,000 Sexual Orientation Exclusively or predominantly heterosexual Exclusively or predominantly homosexual Bisexual Mean Age (years) Mean time since diagnosis (months) Mean number of daily antiviral pills Mean CD4 count Mean antiretroviral treatment duration (months) SD, standard deviation.

12.7% 60.1% 22.5% 4.7% 25.9% 74.1% 25.3% 26.9% 30.8% 17% 31.5% 32.5% 20.9% 15.1% 47.3% 48.1% 4.6% (SD 8.54) (SD 57.69) (SD 5.31) (SD 273.80) (SD 24)

40.95 92.24 9.22 434.78 28

704

LLABRE ET AL.

cluded antibiotic use for acute infection within the past 2 weeks, changes in the HAART regimen or acute bodily infection during the past month, surgery within the past 3 months, and intravenous drug use within the previous 6 months. All participants were informed of the study purpose and provided informed consent prior to participation. Additional exclusionary criteria, determined at the initial appointment, included: (1) cognitive impairment, (2) inability to read at a sixth grade level, (3) current psychosis, drug or alcohol dependence, and panic disorder, and (4) active suicidality. Study physicians performed a health assessment and blood draw. Of 727 HIV-positive participants who were screened, 188 failed to attend an inperson screening appointment or elected not to participate, and 207 did not meet eligibility criteria. After the interview, participants meeting eligibility criteria completed psychosocial measures and received medication adherence and treatment regimen information from a clinical pharmacist. Participants were provided with MEMS caps (Medication Events Monitoring System; Aprex, Union City, California) and asked to place the cap on the bottle of the most potent HAART agent. This was the protease inhibitor (PI) for 28%, non-nucleoside reverse transcriptase inhibitor (NNRTI) for 26%, and nucleoside reverse transciptase inhibitor (NRTI) for 46% of the sample. Dosage frequency was once per day for 18%, twice per day for 66%, and three or four times per day for 16% of participants. Preliminary analyses indicated no differences in mean rates of adherence for the MEMS as a function of type of medication or dosage frequency. Participants received instruction regarding proper use of MEMS and were asked to refrain from removing more than one dose at a time (i.e., pocket doses). Pocket doses were reported by 4.8% and 9.9% at the second and third assessment, respectively. If a refill was required during the monitoring period, participants were to transfer the cap to the new bottle or refill the bottle only when the cap was opened for regular dose administration. Extra openings were discouraged. Participants were encouraged to contact study staff if the medication regimen changed

or they experienced difficulty using the MEMS. At the completion of the baseline appointment (T1), participants were randomly assigned to either the CBSM intervention or a standard care condition and paid $50. Participants were similarly assessed and paid, approximately 3 months after the initial assessment (T2), and twice again at 6-month intervals (T3 and T4). All participants received medication adherence and treatment information at T1, T2, and T3. Measures MEMS caps contain a pressure-activated microprocessor that records the date, time, and duration of each opening. MEMS data were electronically downloaded using MEMS View (version 2.61; Aprex Corporation). Percent adherence was defined as total doses taken divided by the total doses prescribed multiplied by 100 for three periods: (1) between the first (T1) and second (T2) assessments (approximately 90 days); (2) between T2 and 90 days afterward (T2.5); and (3) between T2.5 and the first follow-up assessment (T3; approximately 90 days after T2.5). MEMS data were excluded when participants refused to use the cap, lost their cap, or when the HAART prescription was discontinued by the provider. Adherence was capped at 100% per day. The Adult AIDS Clinical Trial Group Adherence to Combination Therapy Guide31 (ACTG) assessed self-reported adherence. Participants listed their antiretroviral medications, the number of daily doses, and the number of pills prescribed for each dose. They were asked to think back over the past 4 days, one at a time, and report the number of doses skipped each day. Adherence was calculated as the number of pills taken divided by the number of pills prescribed over the previous 4 days. The Medication Adherence Training Instrument (MATI)32 is a 1-hour initial interview, evaluating participant knowledge about HIV and medications, issues related to side effects, and consequences of noncompliance. Adherence is assessed with the question, If 100% is all of the time and 0% is not at all, over the past month, how much of the time have you taken all of your HIV and AIDS medicines? An ab-

MEASUREMENT MODEL OF ADHERENCE

705

breviated version (1520 minutes) was provided at follow-up study visits (T2 and T3) to assess adherence between visits. Venous blood samples were collected from participants using glass tubes containing sodium heparin (Vacutainer, Cat. #6489, BectonDickinson, Rutherford, NJ). HIV viral load was calculated with an AMPLICOR assay using reverse transcriptase polymerase chain reaction (Roche Laboratories, Nutley, NJ; U.S. #83088) to measure the number of HIV virions per milliliter of peripheral blood plasma. An ultrasensitive assay with a lower limit sensitivity of 50 copies/milliliter was used. Modeling All variables were examined for normality. The distributions of the viral load, ACTG, and MATI were not normal, and they were natural log transformed. Mplus statistical software (version 2.14)33 was used for all analyses. Because all participants in this study received medication adherence counseling, we initially tested a growth model of adherence for each method. The mean and variance associated with the slopes were very close to 0 and nonsignificant. Thus, we were able to rule out a systematic change in adherence over time. Viral load was also tested for systematic change over time, and like adherence, a single factor model proved a better representation of the data. We tested a single factor measurement model separately for each adherence measure (MEMS, MATI, and ACTG) and viral load. In each model, indicators at multiple times were combined to create a latent factor. The purpose of these analyses was to determine the reliability (stability) of each specific method of measuring adherence and of viral load. Each adherence factor was then used to predict the viral load factor to determine validity. Three models were tested, each pairing a different method of assessing adherence with viral load. The MEMS, MATI, and ACTG were then combined into a second-order factor model. This measurement model allowed us to separate the specific error due to measurement method from the true variance of adherence. The final model combined the adherence and viral load

measurement models to estimate the effect of adherence on viral load, independent of specific measurement procedures and random error. In order to determine whether any of the adherence assessment methods could be omitted without loss of predictive power, the second-order factor model was rerun three times, each eliminating a specific method (i.e., MEMS, ACTG, or MATI) from the model. Syntax for all the models tested is available from the corresponding author. To estimate the model parameters with missing data, we used full information maximum likelihood (FIML), an approach based on individual functions from each participants data. Consequently, each person need not have complete data and may contribute a different amount of information to the parameter estimates. This approach provides unbiased parameter estimates when data absence may be predicted from other data in the model,34,35 and performs with greater accuracy and power than older approaches such as listwise deletion. In our study, missing data most commonly resulted from participant attrition across time. In fact, the proportion of complete data systematically decreased over time, as shown on the diagonal of Table 2. To the extent that adherence and viral load values at earlier times predicted those missing at later times, the FIML approach is one of the best available solutions because it can partially recover the missing information by using the information from earlier measurement periods.35 Omitting participants with missing data would have resulted in lower reliability and less power.

RESULTS Table 3 shows the percent of participants in different adherence categories and the mean and standard deviation for each measure at each time. Overall mean adherence, defined as percent of doses taken, was 92%, 90%, and 57% by self-report, interview, and electronic monitoring, respectively. Intercorrelations between adherence and viral load variables are presented in Table 2. The correlations between times for a given measure vary widely, as do the correlations of single adherence measures

706
TABLE 2. Variable 1. MEMS 1 2. MEMS 2 3. MEMS 3 4. ACTG 1 5. ACTG 2 6. ACTG 3 7. ACTG 4 8. MATI 1 9. MATI 2 10. MATI 3 11. VL 1 12. VL 2 13. VL 3 14. VL 4 1 0.60 2 0.37 0.45 3 0.34 0.85 0.44 CORRELATIONS BETWEEN ADHERENCE 4 0.25 0.18 0.12 0.94 5 0.11 0.20 0.19 0.45 0.74 6 0.18 0.25 0.26 0.28 0.33 0.51 7 0.07 0.07 0.09 0.36 0.45 0.19 0.30 8 0.23 0.27 0.25 0.54 0.39 0.26 0.21 0.90
AND

LLABRE ET AL.
VIRAL LOAD INDICATORS 10 0.25 0.29 0.40 0.28 0.31 0.37 0.09 0.45 0.48 0.48 11 0.16 0.26 0.24 0.28 0.24 0.24 0.00 0.31 0.31 0.16 0.96 12 0.23 0.28 0.23 0.25 0.32 0.30 0.06 0.29 0.37 0.15 0.73 0.73 13 0.27 0.38 0.33 0.26 0.17 0.22 0.05 0.18 0.29 0.16 0.77 0.79 0.51 14 0.30 0.29 0.25 0.25 0.36 0.25 0.08 0.22 0.32 0.21 0.66 0.74 0.72 0.31

9 0.27 0.19 0.20 0.43 0.50 0.53 0.36 0.59 0.70

MEMS, Medication Event Monitoring System; ACTG, Adherence to Combination Therapy Guide; MATI, Medication Adherence Training Instrument; VL, HIV viral load. ACTG, MATI, and viral load assessments were collected at baseline (T1) and 3 (T2), 9 (T3), and 15 (T4) months. MEMS assessments were collected from T1 to T2 (MEMS 1), the first 90 days after T2 (MEMS 2), and 90 days after T2 to T3 (MEMS 3). Coverage data is provided on the diagonal.

and viral load. This wide range of values in testretest correlations and prediction of viral load makes it difficult to characterize the reliability and validity of an individual adherence measure using simple correlations. Single factor measurement models The single factor model of viral load with four indicators fit the data with 2 (2) 1.53,
TABLE 3. DESCRIPTIVE STATISTICS
FOR

p 0.46, Comparative Fit Index (CFI) 1.0, root mean squared error of approximation (RMSEA) 0.001. Similarly, the single factor model of ACTG with four indicators fit the data with 2 (2) 0.072, p 0.97, CFI 1.0, RMSEA 0.001. Because there were only three indicators of MEMS and MATI, the single factor model was a saturated model with perfect fit. To estimate reliability, classical theory assumptions of parallel measurement were conADHERENCE
AND

VIRAL LOAD INDICATORS % with adherence 80%90% 8.11 6.19 11.38 10.23 11.73 6.94 7.04 % with adherence 90% 77.36 80.95 79.67 81.82 27.04 37.50 33.10

Adherence indicators ACTG (% doses taken) Time 1 Time 2 Time 3 Time 4 MEMS (% doses taken) Time 1Time 2 Time 2 90 days Time 2 90 Time 3 MATI (% doses taken) Time 1 Time 2 Time 3 Viral load (copies/mL) Time 1 Time 2 Time 3 Time 4

Mean (SD) 92 92 93 93 (18) (20) (18) (20)

% with adherence 70% 7.10 7.62 5.69 5.68 56.12 50.00 57.04

% with adherence 70%80% 7.43 5.24 3.25 2.27 5.10 5.56 2.82

58 (34) 59 (39) 49 (42)

91 (18) 7.12 90 (19) 8.04 91 (18) 5.17 Mean (SD) 16221 (57843) 13833 (36936) 12797 (32318) 19621 (68272)

4.27 9.25 79.36 3.52 10.55 77.89 4.31 10.34 80.17 % With Viral Load 50 copies/mL 41 49 47 36

ACTG, Adherence to Combination Therapy Guide; MEMS, Medication Event Monitoring System; MATI, Medication Adherence Training Instrument; SD, standard deviation.

MEASUREMENT MODEL OF ADHERENCE

707

sidered by fixing the loadings of all indicators to 1 and constraining the error variances equal. The parallel measurement model was an adequate representation of the viral load, ACTG, and MATI data, but not MEMS. For the MEMS the errors between the second and third assessments were correlated and the variance at T1 was allowed to differ from the others. These modifications were theoretically justified, because the second and third MEMS assessments were taken from the same time interval (i.e., between the 3-month and 9-month follow-up appointments). The reliability under the parallel measurement model for a single measurement was 0.73 for viral load, 0.36 for ACTG, 0.51 for MEMS, and 0.53 for MATI. Using the variance components estimated in the measurement model, it is possible to project the number of measurements needed to obtain adequate reliability (specified as 0.8). As shown in Table 4, at least 7 ACTG measurements would be needed to bring reliability to an acceptable level. Four assessments would be needed for the MEMS and MATI. These single factor SEM models confirmed that the multiple measures taken over the 15month period were indicators of a single construct and showed that a single measure of adherence had inadequate reliability. Single factor models predicting viral load The ACTG, MATI, and MEMS latent factors were individually included in separate models predicting the viral load factor. Parallel measurement assumptions were relaxed for these analyses, meaning that the factor loadings and variances were freely estimated. All three models fit the data with minor modifications. The paths from adherence to viral load were statisTABLE 4. Measurement method Reliability True Score Variance [var (T)] Error Variance [var (E)] Number of measurements to obtain reliability of 0.8

tically significant for each of the three factors, with standardized coefficients of 0.48, 0.45, and 0.49 for the ACTG, MATI, and MEMS, respectively. These SEM models allowed the estimation of the predictive validity of each adherence method to viral load when measurement error is controlled by combining the multiple indicators into a factor. Second-order factor model A second-order factor model of adherence combining the ACTG, MATI, and MEMS latent factors was then specified to predict the viral load factor. The second-order factor represents the variability common to all three measures, or true adherence, and controls for method variance specific to each measurement procedure. Because both the MATI and the ACTG are self-report measures with similar biases, their errors were allowed to correlate. This second order factor model of adherence predicting viral load fit the data relatively well [ 2 (70) 110.5, p 0.001, CFI 0.96, RMSEA 0.042], although the 2 was significant. The modification indices specified several correlations between errors that would improve model fit. When these were added to the model specification, the fit improved [ 2 (66) 80.75, p 0.10, CFI 0.99, RMSEA 0.026]; however, the parameter estimates remained relatively the same. Because the error correlations involved the latter measures with more missing data, these correlations may be associated with missingness, and we retained the more parsimonious model. The standardized loadings on the second-order adherence factor were equivalent for the ACTG and MATI (0.672 and 0.678, respectively), and slightly higher for the MEMS (0.75).
THE

MEASUREMENT RELIABILITY UNDER ACTG 0.36 0.689 1.255 7

PARALLEL MEASUREMENT MODEL MEMS 0.51 718.945 704.403 4 MATI 0.53 1.061 0.953 4

The following formula was used to obtain reliability estimates: reliability var (T)/[var (T) var (E)/r], where r the number of assessments. ACTG, Adherence to Combination Therapy Guide; MEMS, Medication Event Monitoring System; MATI, Medication Adherence Training Instrument.

708

LLABRE ET AL.

The correlation between the second-order factor of adherence and the viral load factor was 0.67, larger than any of the simple correlations between individual cross-sectional measurements of adherence and viral load, which ranged between 0.08 and 0.37. The adherence factor predicted 45% of the variability in viral load. Standardized parameter estimates for the final model are shown in Figure 1. This second order SEM model allowed the estimation of the validity of adherence in predicting viral load once measurement error and method bias have been controlled. Omitting one of the self-report methods from the second order factor model yielded parameter estimates comparable to those obtained when all three methods were included. With the ACTG and the MEMS in a second order factor model, the standardized coefficient for predicting viral load was 0.71. With the MATI and MEMS, the standardized coefficient for predicting viral load was 0.62. Although the results are comparable, the ACTG and MEMS combination maximizes the prediction of viral load. Omitting the MEMS resulted in a lower

standardized coefficient for predicting viral load ( 0.49). This coefficient is comparable to what was obtained by using a single method of assessing adherence as reported earlier.

DISCUSSION This study addresses recent calls for research on multiple measurement strategies of adherence assessment.12 We combined indicators of adherence measured by different assessment methods to examine the reliability of each method and the predictive validity of both single and multiple method models. The second order factor model of adherence, combining electronic monitoring, self-report, and interview fit the data and predicted 45% of the variability in viral load, illustrating that using multiple methods, each having a unique source of measurement bias, is preferable to using a single method. However, using multiple methods that share the same sources of error (i.e., two self-report measures) did not improve the ability to predict viral load any better than when

FIG. 1. Standardized parameter estimates of the final model. Large ovals represent latent factors; rectangles represent indicators of each latent factor; small circles labeled E represent residual variance associated with each indicator; and small circles labeled D represent residual variance in the latent factors. All paths and factor loadings are significant (p 0.05. R2 for the latent factor of viral load was 0.45.

MEASUREMENT MODEL OF ADHERENCE

709

using one such method. With any method of assessing adherence, the results clearly indicate that multiple assessments over time are necessary to obtain a reasonable level of reliability, with 4 (MATI and MEMS) or 7 (ACTG) times yielding acceptable values for reliability. Thus, results indicate that medication adherence is best assessed by using both an objective (MEMS) and self-report measure (either ACTG or MATI), each obtained at multiple times. Each HAART adherence method had unique advantages and disadvantages. The ACTG is easy to administer (510 minutes), and the short-time period assessed (4 days) may promote accurate recall of medication-taking behavior.28 However, one may question the representativeness of the short interval and the possible self-presentation bias.12 In contrast, the MATI item, embedded in an interview with a trained professional, covers a 30-day period. Although this method also relies on self-report, the rapport established during the interview may promote accurate reporting of behavior. Although this adherence assessment could be administered without the interview, it is unknown how this would affect its reliability and validity. The MEMS does not rely on self-report or memory and is considered objective; however, the technology is expensive and provides information about bottle opening, not necessarily medication-taking behavior. In addition, the MEMS is potentially sensitive to behavioral characteristics, so for example, the MEMS underestimates adherence if patients remove doses they plan to take later.15,16 Also, some patients find the bulky MEMS cap burdensome and intrusive. Greater compliance and adjustment for pocket doses might result in higher adherence estimates. Other factors thought to influence the MEMS include the complexity of the regimen and the level of activity of the patient. These characteristics may influence the results in any specific study. To the extent that these sources of variance in the MEMS are minimized in a given situation, its validity may approach that obtained by combining it with other methods. The MEMS also has the advantage of gathering specific information about the timing of doses. Although this study did not incorporate this information, it is possible that defining

adherence as the percent of doses taken within the prescribed interval (e.g., 12 hours apart for a twice daily medication) would have improved the predictive power of the MEMS. Future studies are needed to determine the applicability of this framework to alternative assessment strategies such as prescription refill rates, pill counts, and other self-report methods of assessing adherence. Therapeutic drug monitoring of the antiretroviral regimen is another important strategy for assessing medication adherence, but was not available in the current study. While the measures of viral load had reasonable reliability, those of adherence were considerably less stable, with ACTG displaying the lowest reliability. The low reliability for a single measurement of ACTG may reflect greater variability in behavior that spans a short interval of time. Adherence behavior examined over shorter periods may appear less stable because even in the most adherent of patients, the rare missed dose is magnified in an assessment period of 4 days, while the same missed dose would be minimized over a 30day period. The issue of the time frame used for each method of assessing adherence is important to consider when interpreting the recommendations for the number of measures needed to achieve acceptable reliability. Future studies would need to fix the time frame considered by each method for a more direct comparison of their reliability. A limitation to this study is the use of volunteer sample involved in an intervention study. While the sample was heterogeneous with respect to demographic characteristics, it may have been more compliant that the general HIV-positive population. Thus, adherence rates were expected to be higher than those found in the general population. However, our mean rates of adherence for the self-report interview and questionnaire (90%92%) are comparable to others utilizing similar methodologies.21,23,26,36 Another limitation is that all participants received some adherence and treatment regimen information; however, we consider this to be part of standard care. Finally, validity was assessed in terms of the ability of the adherence measures to predict viral load. Thus, our evaluation of the adherence

710

LLABRE ET AL.

measures focused on the relative ordering of individuals in relation to viral load, but did not address the accuracy of the absolute levels of adherence obtained from each measure. The question of which combination of methods is optimal for calculating adherence rates was not addressed here and is worthy of further investigation. Inclusion of viral resistance data would have increased our ability to predict viral load, however these data were not available in the current study. While one could argue that any validity estimate is sample specific, the second-order factor model provides a template that may be replicated in other research contexts. It is this template, rather than the specific factor loadings, that is generalizable to future research. In this manner, reliable and valid adherence estimates may be obtained and used as either predictors or outcomes. Replication across samples and methods would eventually lead to specific recommendations on how different methods of measuring HAART adherence should be weighted. This analysis template is not feasible to implement with individual patients in clinical settings or in all research contexts, however clinicians and researchers are cautioned against relying on single assessments of adherence, as these data suggest that adherence behavior is variable over time and single assessments have limited validity. This study advances our understanding of the measurement of HAART medication adherence, which is critical for both researchers and practitioners working with men and women living with HIV. Given that adherence can explain approximately half of the variability in HIV viral load levels, and sustained viral load suppression leads to improved immune function, enhancing HAART medication adherence is critical.

assistance of graduate students and staff of the intervention study with recruitment and assessment. Most importantly, we are indebted to the participants for providing valuable information. REFERENCES
1. Centers for Disease Control and Prevention. HIV/AIDS Surveillance Report. Atlanta, GA; 2001. 2. Mathews WC, Mar-Tang M, Ballard C, et al. Prevalence, predictors, and outcomes of early adherence after starting or changing antiretroviral therapy. AIDS Patient Care STDs 2002;16:157172. 3. Fogarty L, Roter D, Larson S, et al. Patient adherence to HIV medication regimens: A review of published and abstract reports. Patient Educ Couns 2002;46: 93108. 4. Paterson DL, Swindells S, Mohr J, et al. Adherence to protease inhibitor therapy and outcomes in patients with HIV infection. Ann Intern Med 2000;133:21. 5. Press N, Tyndall MW, Wood E, et al. Virologic and immunologic response, clinical progression, and highly active antiretroviral therapy adherence. J Acquir Immune Defic Syndr 2002;31:S112S117. 6. Catz SL, Kelly JA, Bogart LM, et al. Patterns, correlates, and barriers to medication adherence among persons prescribed new treatments for HIV disease. Health Psychol 2000;19:124133. 7. Ammassari A, Murri R, Pezzotti P, et al. Self-reported symptoms and medication side effects influence adherence to highly active antiretroviral therapy in persons with HIV-infection. J Acquir Immune Defic Syndr 2001;28:445449. 8. Tesoriero J, French T, Weiss L, et al. Stability of adherence to highly active antiretroviral therapy over time among clients enrolled in the treatment adherence demonstration project. J Acquir Immune Defic Syndr 2003;33:484493. 9. Chesney MA, Morin M, Sherr L. Adherence to HIV combination therapy. Soc Sci Med 2000;50:15991605. 10. Perno CF, Ceccherini-Silberstein F, De Luca A, et al. Virologic correlates of adherence to antiretroviral medications and therapuetic failure. J Acquir Immune Defic Syndr 2002;31:S118S122. 11. Kenna LA, Labb L, Barrett JS, Pfister M. Modeling and simulation of adherence: Approaches and applications in therapuetics. AAPS J 2005;7(2):E390E407. 12. Kerr T, Walsh J, Loyd-Smith E, Wood E. Measuring adherence to highly active antiretroviral therapy: Implications for research and practice. Curr HIV/AIDS Rep 2005;2:200205. 13. Paterson DL, Potoski B, Capitano B. Measurement of adherence to antiretroviral medications. J Acquir Immune Defic Syndr 2002;31:S103S106. 14. Arnsten JH, Demas PA, Farzadegan H, et al. Antiretroviral therapy adherence and viral suppression in HIV-infected drug users: Comparison of self-report

ACKNOWLEDGMENTS This project was supported by grants from the National Institute of Mental Health (P01 MH49548 and T32 MH18917). Portions of this research were presented at the American Psychosomatic Society Annual Meeting (Orlando, FL, March 2004). We wish to acknowledge the

MEASUREMENT MODEL OF ADHERENCE

and electronic monitoring. Clin Infect Dis 2001;33: 14171423. Samet JH, Sullivan LM, Traphagen ET, Ickovics JR. Measuring adherence among HIV-infected persons: Is MEMS consummate technology? AIDS Behav 2001;5:2130. Bova CA, Fennie KP, Knafl GJ, et al. Use of electronic monitoring devices to measure antiretroviral adherence: Practical considerations. AIDS Behav 2005;9: 103110. Kalichman, SC, Cain D, Cherry C, et al. Pillboxes and antiretroviral adherence: Prevalence of use, perceived benefits, and implications for electronic medication monitoring devices. AIDS Patient Care STDs 2005;19: 833839. Duong M, Piroth L, Peytavin G, et al. Value of patient self-report and plasma human immunodeficiency virus protease inhibitor level as markers of adherence to antiretroviral therapy: Relationship to virologic response. Clin Infect Dis 2001;33:386392. Howard AA, Arnsten JH, Lo Y, et al. A prospective study of adherence and viral load in a large multicenter cohort of HIV-infected women. AIDS 2002;16: 21752182. Vincke J, Bolton R. Therapy adherence and highly active antiretorviral therapy: Comparison of three sources of information. AIDS Patient Care STDs 2002;16:487495. Walsh JC, Mandalia S, Gazzard BG. Responses to a 1 month self-report on adherence to antiretroviral therapy are consistent with electronic data and virological treatment outcomes. AIDS 2002;16:269277. 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;14:357366. Fletcher CV, Testa MA, Brundage RC, et al. Four measures of antiretroviral medication adherence and virologic response in AIDS clinical trials group study 359. J Acquir Immune Defic Syndr 2005;40:301306. Wagner GJ. Predictors of antiretroviral adherence as measured by self-report, electronic monitoring, and medication diaries. AIDS Patient Care STDs 2002;16: 599608. Bangsberg, DR, Hecht FM, Charlebois ED, Chesney M, Moss, A. Comparing objective measures of adherence to HIV antiretroviral therapy: Electronic medication monitors and unannounced pill counts. AIDS Behav 2001;5:275281.

711
26. Liu H, Golin CE, Miller LG, et al. A comparison study of multiple measures of adherence to HIV protease inhibitors. Ann Intern Med 2001;134:968977. 27. Rakhmanina NY, Van Den Anker, JN, Soldin, SJ. Therapeutic drug monitoring of antiretroviral therapy. AIDS Patient Care STDs 2004; 18: 714. 28. 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;13:0991107. 29. Goujard C, Bernard N, Sohier N, et al. Impact of a patient education program on adherence to HIV medication. J Acquir Immune Defic Syndr 2003;34:191 194. 30. Rigsby MO, Rosen MI, Beauvais JE, et al. Cue-dose training with monetary reinforcement: Pilot study of an antiretroviral adherence intervention. J Gen Intern Med 2000;15:841847. 31. Chesney MA, Ickovics JR, Chambers DB, et al. Selfreported adherence to antiretroviral medications among participants in HIV clinical trials: The AACTG adherence instruments. AIDS Care 2000;12:255266. 32. McPherson-Baker S, Jones D, Duran R, Klimas N, Schneiderman N. Development and implementation of a medication adherence-training instrument (MATI) for people living with HIV. Behav Modif 2005;29: 286317. 33. Muthn LK, Muthn BO. Mplus Users Guide. Los Angeles: Muthen & Muthen, 1998. 34. Schafer JL, Graham JW. Missing data: Our view of the state of the art. Psychol Methods 2002;7:147177. 35. Little RJ, Rubin DB. Statistical Analysis with Missing Data. New York: Wiley & Sons, 1987. 36. Mannheimer S, Friedland G, Matts J, et al. The consistency of adherence to antiretroviral therapy predicts biologic outcomes for human immunodeficiency virus-infected persons in clinical trials. Clin Infect Dis 2002;34:11151121.

15.

16.

17.

18.

19.

20.

21.

22.

23.

24.

25.

Address reprint requests to: Maria M. Llabre, Ph.D. Department of Psychology University of Miami 5665 Ponce de Leon Boulevard Coral Gables, FL 33146 E-mail: mllabre@miami.edu

This article has been cited by: 1. Staci Martin , Deborah K. Elliott-DeSorbo , Sarah Calabrese , Pamela L. Wolters , Gregg Roby , Tara Brennan , Lauren V. Wood . 2009. A Comparison of Adherence Assessment Methods Utilized in the United States: Perspectives of Researchers, HIV-Infected Children, and their CaregiversA Comparison of Adherence Assessment Methods Utilized in the United States: Perspectives of Researchers, HIV-Infected Children, and their Caregivers. AIDS Patient Care and STDs 23:8, 593-601. [Abstract] [Full Text] [PDF] [PDF Plus] 2. Ann E. Deschamps , Sabina De Geest , Anne-Mieke Vandamme , Herman Bobbaers , Willy E. Peetermans , Eric Van Wijngaerden . 2008. Diagnostic Value of Different Adherence Measures Using Electronic Monitoring and Virologic Failure as Reference StandardsDiagnostic Value of Different Adherence Measures Using Electronic Monitoring and Virologic Failure as Reference Standards. AIDS Patient Care and STDs 22:9, 735-743. [Abstract] [PDF] [PDF Plus] 3. MassimoMagnano San Lio, Riccardo Carbini, Paola Germano, Giovanni Guidotti, Sandro Mancinelli, NoorjehanAbdul Magid, Pasquale Narciso, Leonardo Palombi, Elsa Renzi, Ines Zimba, MariaCristina Marazzi. 2008. Evaluating Adherence to Highly Active Antiretroviral Therapy with Use of Pill Counts and Viral Load Measurement in the Drug Resources Enhancement against AIDS and Malnutrition Program in Mozambique. Clinical Infectious Diseases 46:10, 1609-1616. [CrossRef]