Recommended method for health services to determine adherence to antiretroviral therapy: a comparison of three models

Abstract

The medication possession ratio (MPR) method is commonly used for the determination of antiretroviral medication adherence. However, different ways of calculating MPR and methodological issues hinder the interpretation of the results and the reproducibility of the method. Thus, this study used three different models of MPR calculation and aimed to identify the one that best represents the situation of patient adherence. The results show that there was a statistically significant difference between the adherence rates determined by the three models, which indicates the need to specify the parameters used for calculation in the MPR method. However, the models individually were found to be related to viral suppression, but none of them had a greater effect than the other in this regard. The model that used residual medication (RM) and a fixed period of analysis allowed for a more precise identification of the number of doses that the patient used when compared to the others. Health services should avoid the application of the model using a variable analysis period. This study found that RM and the period of analysis considered are the main influencing factors in the accuracy of adherence results when the MPR method is used.

Keywords

Antiretroviral adherence medication possession ratio

Introduction

The main goals of antiretroviral therapy (ART) are to reduce the morbidity and mortality of people living with HIV(PLWH), to increase life expectancy and quality of life, and to decrease the chance of virus transmission. In this context, adherence to treatment is one of the greatest challenges in caring for these patients.^1–3 Adherence to ART is a multifactorial and complex process with no consensus regarding the optimal method for its determination. The assessment of ART adherence includes self-reports, pharmacy records, electronic monitoring devices, therapeutic serum level measurements and daily drug registration.^3–12

Pharmacy records are widely used because the data can generally be accessed easily by the health service, incurs a low cost for execution and provides more objective information.^7,13,14 To ensure best operation of this technique, certain aspects should be considered. The locality from which the patient obtains medication is one of them, because the information is more reliable when a single establishment or system records the withdrawal dates of antiretrovirals (ARVs).^13,15 However, when the residual medication (RM) of previous withdrawals is not considered, the patient may eventually be incorrectly classified as non-adherent.¹³ There are different ways of calculating adherence from pharmacy records. One of these is the medication possession ratio (MPR), which measures the amount of time the individual is in possession of ARVs.⁷

Although pharmacy records are correlated to virological and immunological measures, some studies show that patients with high adherence rates do not achieve viral suppression, while others with low adherence maintain an undetectable viral load (VL). These results may suggest, among other reasons, that there are methodological problems in determining adherence using this method.⁷ In this sense, this research aimed to recommend, after assessing three models of adherence calculation by the MPR method, the model that satisfactorily represents the real situation of the patient regarding the use of ARVs.

Methods

This study was carried out at the Testing and Counselling Centre (TCC) located in Curitiba, Paraná State, Brazil. The study was conducted from November 2016 to July 2017. Volunteers, both genders, over 18 years of age, not pregnant and with at least one year and one month of treatment with ARVs, were randomly selected at the TCC during the monthly ARV withdrawal. The results are discussed from the perspective of the statistical difference between the percentages of adherence determined by the three models and their associations with viral suppression, as well as the classification of patients as adherent or not, at distinct cut-off points.

Data collection and adherence measurement

The adherence was determined by using the MPR method. The ARV withdrawal dates of each participant, retrospective over one year, as well as the most recent VL within that period, were obtained from the Logistic Control System of Medicines – a national unified system.

MPR is the ratio between the number of days the patient has medication (numerator) and the number of days of a predetermined period for the study (denominator). The result, when multiplied by 100, indicates the percentage of days that patient would have taken the drug within that time interval. This study used three different ways to calculate MPR¹⁴ that differed by the use (or not) of RM and by the analysis period applied in the denominator. RM can be defined as the doses that the patient had in their possession of a drug withdrawal prior to the first withdrawal within the study period.

The three models exemplified in Table 1 consider a study period from 1 January to 31 December 2017 (365 days). In order to evaluate the presence or absence of RM, the last ARV withdrawal in 2016 is also presented. At each withdrawal, the patient receives medication for 30 days of treatment. Figure 1 shows the similarities between each model.

Figure 1.

Similarities and differences between the three models. NRMF – not used RM in a 365-day fixed analysis period; NRMV – not used RM in a variable analysis period; RMF – used RM in a 365-day fixed analysis period.

Table 1.

Description of the three models of MPR calculation.

MW dates	Model	Analysis period	Total MW in 2017^a	RM	MPR
26 Dec 16 25 Jan 17 26 Feb 17	RMF: used RM in a 365-day Fixed analysis period	365 days (1 Jan 17 to 31 Dec 17)	330 days	25 (1 Jan 17 to 25 Jan 17)	97.30% ([330 + 25]/365) × 100

31 Mar 17 8 May 17 12 Jun 17	NRMV: Not used RM in a variable analysis period	340 days (25 Jan 17 to 31 Dec 17)	330 days	Not considered	97.00% (330/340) × 100

16 Jul 17 14 Aug 17 20 Sep 17 27 Oct 17 25 Nov 17 31 Dec 17	NRMF: Not used RM in a 365-day fixed analysis period	365 days (1 Jan 17 to 31 Dec 17)	330 days	Not considered	90.41% (330/365) × 100

MPR: medication possession ratio; MW: medication withdrawal; RM: residual medication.

^a11 MW × 30 days.

Sample size

The sample was determined based on the number of patients withdrawing ARVs in the TCC in 2016 (4322). Considering a 95% confidence interval, a margin of error of 5%, a non-adherence rate to ART of 30%⁶ and a margin of 10% to compensate for losses, the minimum sample required was 331 patients.

Statistical analysis

To compare the three models, a mixed linear regression model was fitted in which the errors followed a normal distribution, including a random effect to accommodate repeated measures in the same patient.¹⁶ The likelihood ratio test (LRT) was used to evaluate the effect of the three proposed models on the withdrawal record, with a significance level of 5%. The results were reported in terms of the adjusted means, standard error, 95% overall confidence intervals and Wald test to test the contrasts of the variable in the selected model, with the Bonferroni correction.¹⁷

To evaluate the association between the three models and viral suppression, two procedures were performed through multiple logistic regression. In the first, the logistic regression was adjusted for each model separately to assess if they presented statistical significance, demonstrating the results in terms of Akaike Information Criterion (AIC), odds ratios and p-values. Then, it was determined if all models had the same correlation with viral suppression using the Vuong’s LRT¹⁸ for strictly non-nested models, which is based on AIC.

R software¹⁹ was used to perform the analysis, using the Ime4,²⁰ ImerTest²¹ and Nonnest2²² packages (May 2019, Vienna, Austria).

Ethical approval

The study was approved by the Ethics Committee of the Health Sciences Sector of the Federal University of Paraná (protocol 1739379) and the Municipal Health Department of Curitiba (protocol 1620414), Brazil.

Results

A total of 349 volunteers were included in the study. In the sampled population, there was a predominance of men (70.8%) of white ethnicity (69.9%), heterosexual (49%), with no fixed partner (51%) or unmarried (52.2%), without children (57%), and with a mean age of 44.9 years. The majority had more than 11 years of education (77.9%), were Catholic (45.6%) and had more than five years of ARV use (60.2%). Most of the participants used the Unified Health System, provided by the government, for their medical care (59.6%) and 6.3% had a detectable VL in the interview period.

The comparison of the mean adherence rates determined by the three models is presented in Table 2. Although the nominal values between RMF and NRMV were similar, there was a statistically significant difference in the results, as well as between RMF versus NRMF and NRMV versus NRMF, as shown in Table 3.

Table 2.

Comparison of means of adherence between the three models.

	Average adherence (%)	Standard error	CI (95%)
RMF	92.3	0.677	90.0–93.6
NRMV	93.1	0.674	91.8–94.4
NRMF	87.5	0.674	86.2–88.8

CI: confidence interval; NRMF: not used residual medication in a 365-day fixed analysis period; NRMV: not used residual medication in a variable analysis period; RMF: used residual medication in a 365-day fixed analysis period.

Table 3.

Difference between the adjusted means of the three models.

	Difference of the average means	Standard error	CI (95%)	p value^a
RMF – NRMV	−0.8	0.212	[−1.3]–[−0.3]	0.0002
RMF – NRMF	4.7	0.212	4.2–5.2	<0.0001
NRMV – NRMF	5.6	0.201	5.1–6.1	<0.0001

^aWald test.

It is important that patient adherence rates, regardless of the methodology used for calculation, correlate with virological results in order to be relevant. This is demonstrated in Table 4, based on logistic regression considering each model individually. All models were found to be related to viral suppression, but none of them had a greater effect than the other in this regard (Table 5), demonstrated by the p values and because of the small difference in AIC between the two models (Table 4).

Table 4.

Correlation of adherence rates with viral suppression.

	AIC	Estimate	Standard error	Odds ratio	CI (95%)	p value^a
RMF	154.8	0.044	0.011	1.046	1.02–1.07	0.0001
NRMV	160.2	0.038	0.013	1.039	1.01–1.06	0.0024
NRMF	156.9	0.043	0.012	1.044	1.02–1.07	0.0004

AIC: Akaike information criterion; CI: confidence interval; NRMF: not used residual medication in a 365-day fixed analysis period; NRMV: not used residual medication in a variable analysis period; RMF: used residual medication in a 365-day fixed analysis period.

^aWald test.

Table 5.

Difference in correlation with viral suppression.

	Differencein AIC	CI (95%)	p value^a
RMF – NRMV	−5.444	−16.86 to 5.97	0.3497
RMF – NRMF	−2.117	−5.25 to 1.01	0.1848
NRMV – NRMF	3.327	−7.76 to 14.41	0.5563

^aVueng test.

Analysing the differences in adherence obtained by the models, the adherence rates were categorised as ≥95%, ≥90%, ≥85% and ≥80%. The models were compared to each other by evaluating the occurrence of divergence in patient classification with respect to adherence and cut-off points. Table 6 presents these results, but does not include the comparison between the NRMF model and the other two because no patients with disagreement in the classification of adherence were identified.

Table 6.

Comparison between models in relation to the different adherence cut-off points.

		≥95%	≥90%	≥85%	≥80%
RMF→NRMV	T	202	249	274	298
	A/NA	3	2	0	1
	%	1.5	0.8	0	0.3
RMF→NRMF	T	202	249	274	298
	A/NA	96	71	16	21
	%	47.5	28.5	5.8	7.0
NRMV→RMF	T	110	275	304	319
	A/NA	16	18	17	7
	%	14.5	6.5	5.6	2.2
NRMV→NRMF	T	212	263	289	303
	A/NA	108	85	31	26
	%	50.9	32.3	10.7	8.6

A/NA: number of patients considered adherent by the first model and not adherent by the second one, considering each cut-off point; NRMF: not used residual medication in a 365-day fixed analysis period; NRMV: not used residual medication in a variable analysis period; RMF: used residual medication in a 365-day fixed analysis period; T: total number of participants.

It was observed that the divergence between the models was always higher with larger cut-offs, regardless of the comparison being made. We highlight the differences in the classification of adherence between the 90 and 95% cut-off points, especially for the comparison between RMF versus NRMF and NRMV versus NRMF, in which case approximately half of the individuals at the cut-off point of 95% had an erroneous categorisation of adherence.

Discussion

Among the most widely used methods to determine adherence to ART, the use of pharmacy records can be considered the most adequate in relation to the reliability, availability and facility of data collection, since techniques involving self-reports, for example, may overestimate adherence.²³ This study demonstrates that a detailed description of data collection considering withdrawal dates and periods of analysis considered is important to present the results safely. Thus, in the case of analyses that involve the use of MPR, one should describe how the numerator and denominator of the equation were defined.

This study revealed that the three models give statistically different adherence rates (Table 3). Patient adherence when assessed by the RMF and NRMV models resulted in values, on average, 4.7 and 5.6% higher than with the NRMF model, respectively. This means, for example, that for a one-year period of treatment, the adherence rate calculated by NRMV would indicate approximately 20 more days of therapy than by NRMF.

It was observed that NRMV gave a higher average adherence value than the others (Table 2), which can be explained by the fact that the analysis period was less than 365 days in most cases. This is a limiting factor in the use of this model since it considers the period of analysis starting at the date of the first ARV withdrawal within the study range. For example, if the study period is from 1 January to 31 December and drug withdrawals within this range occurred only on 1 August, September, October, November and December (150 tablets in 153 days), the MPR determined by NRMV will be 98% ([150/153] × 100), indicating adequate adherence to ART. However, if the study period of one year (RMF or NRMF models) is considered, adherence will be 41%, which is a better approximation of reality, since the patient remained without medication for seven months. Therefore, the use of variable denominator models must be carefully evaluated on a case-by-case basis so that patients with inadequate adherence are not misclassified as adherent. These outcomes corroborate the findings of Kozma et al.,²⁴ who suggest the use of a fixed denominator for the assessment of adherence to long-term therapies such as ART.

Observing the adherence determined by RMF and NRMF, which differed only in the use of RM, it was expected that the values would be lower with NRMF. Also, if the patient does not have RM, the adherence rate calculated by the two models will be the same. However, the use of NRMF can misclassify individuals as non-adherent when in fact they are adherent. For a cut-off point of 95%, for example, it was observed that approximately half of the patients (47.5%) were mistakenly classified (Table 6), which occurred similarly between NRMV and NRMF; this is a relevant finding when the researcher needs to define which model will be used in a study. Thus, the RMF model indicates adherence to ART more reliably, which corroborates the findings of Boer et al.¹³

It should be emphasised that, in the health services, standardised, cost-effective and routine adherence monitoring is necessary, since adherence to chronic treatment is dynamic and oscillates over time. The identification of patients having difficulties with ART allows for intervention actions by the health team, with a view to promoting an adequate therapeutic efficiency of ARVs.^7,8,23 In this sense, it would be possible to use NRMF in the scope of professional practice, as it is simpler with regard to data collection; it is not necessary to seek withdrawals prior to the observational period. In this situation, given the 4.7% increase in the adherence rate obtained by the NRMF model, the result is close to what would be found by RMF, as recommended in this study (Table 3). Therefore, the use of NRMF would only be feasible with the use of this factor, which was achieved in a TCC with approximately 4000 registered patients and could be extrapolated to other services with similar characteristics.

Regarding the period of analysis considered in the studies, it is noted that the use of short periods for adherence calculation using pharmacy records could give imprecise results regarding the association between adherence and viral suppression.⁸ Studies that contemplate longer time intervals, such as the 12-month period considered in this research, tend to provide results with greater accuracy. It is important to point out that these considerations are relevant for patients who have been on ART for some time before joining a study. For those starting treatment, as long as the study period coincides with the date of the user’s first ARV withdrawal, model selection will not make a significant difference because there will be no RM.

Another consideration that must be made concerns the cut-off point of adherence. Although patient retention in HIV care services is associated with viral suppression, adherence to ARVs is the factor that best relates to this condition.²⁵ Inadequate treatment follow-up is a strong indicator of therapeutic failure, development of drug resistance, disease progression and death.¹⁵ The 95% cut-off value of adherence is the gold standard for older therapies, while studies point out that use of at least 80% of the prescribed doses is necessary to obtain an effective therapeutic response, especially for newer ARVs.^4,15,26 It has been noted that rates of 80% or less do not necessarily mean that the individual has a detectable VL. However, these cases should represent an alert to the health service, since these individuals may be at the limit of virus detection or there may be possible difficulties in the follow-up. For this reason, the cut-off point defined to classify patients as non-adherent should guide the choice of the calculation model to be used. It is enhanced here in the non-recommendation of the NRMF model, mainly for cut-offs above 90%, when the discrepancies are more remarkable (Table 6). It is evident that, with smaller cut-offs, the differences in this classification are less impactful, but no less important to the scientific aspects of the publication of results.

In practice, what is the consequence of misclassifying non-adherents? Especially for the management of the health service, this situation may result in patient rescue procedures that require the time and resources of different health professionals, both of which are generally tight in public services. As a result, there is a waste of resources with patients who are under adequate therapeutic follow-up, leaving those who may actually need assistance without intervention. In addition, individuals with adequate adherence to ART achieve such low VL levels that there is little chance of transmitting the virus. Thus, the correct identification of patients with adherence problems is of social relevance, since this group must be monitored by health professionals and encouraged to return to treatment. It is also emphasised that adequate therapeutic follow-up and clinical monitoring of VL are strategic components of combined HIV prevention, which is designed to make joint use of biomedical, behavioural and structural interventions applied to the individual, their relationships and social groups to which they belong.

It has been shown in this study that the adherence rates obtained by the three models are statistically different, to a greater or lesser degree. The NRMV and NRMF models are simpler in relation to data collection, but the researcher should know that the adherences rates calculated by NRMF will be smaller when compared to the others and may not represent the reality of the studied individuals. Likewise, NRMV may mistakenly indicate individuals as adherent when in fact they are not. In this context, the importance of the use of a scientific methodology with clearly defined validity, safety and precision is pointed out, in order to ensure that the results can be used in the practice of health services with the main objective of adequately identifying patients with adherence problems.

Conclusion

In this study, we identified that the three proposed models showed statistically significant differences in adherence rates, but all of them correlated with viral suppression. The RMF model allowed for a more precise identification of the number of doses that the patient used because it considered RM. In addition, this model reduces the possibility of errors in the classification of adherence. Furthermore, in the RMF model, the time interval previously defined in the study is in fact considered, in contrast to the NRMV model. Thus, application of the NRMV model must be carefully evaluated, given the possibility of mistakenly classifying adherent individuals. Health services should be therefore advised that, when using the MPR method to define adherence rates, the RMF model should be applied, avoiding the use of NRMV in particular.

In studies that use the MPR calculation, it is necessary to describe whether RM was used and the period of analysis considered, since these elements are the main influencing factors regarding the accuracy and safety of the results.

Footnotes

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Camila K Costa

References

Brazil. Ministério da Saúde/Secretaria de Vigilância em Saúde/Departamento de DST/AIDS. Handbook of adherence to treatment for people living with HIV and AIDS, http://bvsms.saude.gov.br/bvs/publicacoes/manual_adesao_tratamento_hiv.pdf (2008, accessed 7 May 2018).

McNabb

Ross

Abriola

, et al. Adherence to highly active antiretroviral therapy predicts virologic outcome at an inner-city human immunodeficiency virus clinic. Clin Infect Dis 2001; 33: 700–705.

Bonolo

Gomes

RRFM

Guimarães

MDC.

Adesão à terapia anti-retroviral (HIV/AIDS): fatores associados e medidas de adesão. Epidemiol Serv Saúde 2007; 16: 261–278.

Brazil. Ministério da Saúde/Secretaria de Vigilância em Saúde/Departamento de DST, Aids e Hepatites Virais. Clinical protocol and therapeutic guidelines for management of HIV infection in adults, http://www.aids.gov.br/pt-br/pub/2013/protocolo-clinico-e-diretrizes-terapeuticas-para-manejo-da-infeccao- pelo-hiv-em-adultos (2017, accessed 23 February 2018).

Brawley

Culos-Reed

SN.

Studying adherence to therapeutics regimens: overview, theories, recommendations. Control Clin Trials 2000; 21: 156S–163S.

Brazil. Ministério da Saúde/Secretaria de Vigilância em Saúde/Departamento de DST, Aids e Hepatites Virais. Adherence to antiretroviral treatment in Brazil: collection of studies from the ATAR Project, https://www.researchgate.net/profile/Maria_das_Gracas_B_Ceccato/publication/284.979896_Adesao_ao_tratamento_antirretroviral_Uma_revisao_sistematica_2004-2009_In_Brasil_Ministerio_da_Saude_Secretaria_de_Vigilancia_em_Saude_Departamento_de_DST_Aids_e_Hepatites_Virais_Adesao_ao_tratamento_/links/57f2f00f08ae8 86b897c00fe/Adesao-ao-tratamento-antirretroviral-Uma-revisao-sistematica-2004-2009-In-Brasil-Ministerio-da-Saude-Secretaria-de-Vigilancia-em-Saude-Departamento- de-DST-Aids-e-Hepatites-Virais-Adesao-ao-tratamento.pdf (2010, accessed 15 December 2017).

McMahon

Jordan

Kelley

, et al. Pharmacy adherence measures to assess adherence to antiretroviral therapy: review of the literature and implications for treatment monitoring. Clin Infect Dis 2011; 52: 493–506.

Henegar

Westreich

Maskew

, et al. Comparison of pharmacy-based measures of adherence to antiretroviral therapy as predictors of virologic failure. AIDS Behav 2015; 19: 612–618.

Berg

Arnsten

JH.

Practical and conceptual challenges in measuring antiretroviral adherence. J Acquir Immune Defic Syndr 2006; 43: S79–S87.

10.

Helena

ETS

Nemes

MIB

ElufNeto

Development and validation of a multidimensional questionnaire assessing non-adherence to medicines. Rev Saúde Pública 2008; 42: 764–767.

11.

Simoni

Kurth

Pearson

, et al. Self-report measures of antiretroviral therapy adherence: a review with recommendations for HIV research and clinical management. AIDS Behav 2006; 10: 227–245.

12.

Steiner

Prochazka

AV.

The assessment of refill compliance using pharmacy records: methods, validity and applications. J Clin Epidemiol 1997; 50: 105–116.

13.

Boer

Prins

Sprangers

, et al. Using different calculations of pharmacy refill adherence to predict virological failure among HIV-infected patients. J Acquir Immune Defic Syndr 2010; 55: 635–640.

14.

Grimes

Andrade

Niemeyer

, et al. Measurement issues in using pharmacy records to calculate adherence to antiretroviral drugs. HIV Clin Trials 2013; 14: 68–74.

15.

Been

Yildiz

Nieuwkerk

, et al. Self-reported adherence and pharmacy refill adherence are both predictive for an undetectable viral load among HIV-infected migrants receiving cART. PLoS One 2017; 12: 1–12.

16.

Faraway

JJ.

Extending the linear model with R: generalized linear, mixed effects and nonparametric regression models. Boca Raton, FL: CRC Press, 2006.

17.

Casella

Berger

RL.

Statistical inference. Pacific Grove, CA: Duxbury Pacific Grove, 2002.

18.

Vuong

QH.

Likelihood ratio tests for model selection and non-nested hypotheses. Econometrica 1989; 57: 307–333.

19.

R Core Team. R: a language and environment for statistical computing. Vienna: R Foundation for Statistical Computing, 2017.

20.

Bates

Machler

Bolker

, et al. Fitting linear mixed-effects models using Ime4. J Stat Softw 2015; 67: 1–48.

21.

Kuznetsova

Brockhoff

Christensen

RHB.

ImerTest Package: tests in linear mixed effects models. J Stat Softw 2017; 82: 1–26.

22.

Merkle

You

Schneider

, et al. Nonnest2: tests of non-nested models. Econometrica 1989; 57: 307–333.

23.

Gutierrez

Sartori

Schmidt

, et al. Measuring adherence to antiretroviral treatment: the role of pharmacy records of drug withdrawals. AIDS Behav 2012; 16: 1482–1490.

24.

Kozma

Dickson

Phillips

, et al. Medication possession ratio: implications of using fixed and variable observation periods in assessing adherence with disease-modifying drugs in patients with multiple sclerosis. Patient Prefer Adherence 2013; 7: 509–516.

25.

Bonacci

Frasca

Swift

, et al. Antiretroviral refill adherence correlates with, but poorly predicts retention in HIV care. AIDS Behav 2016; 20: 1060–1067.

26.

Bezabhe

Chalmers

Bereznicki

, et al. Adherence to antiretroviral therapy and virologic failure: a meta-analysis. Medicine 2016; 95: 1–9.