Adherence during COVID-19: The role of aging and socio-economics status in shaping drug utilization.

Our study investigates the potential impact that COVID-19 and lockdown restrictions may have had on drug utilization and the role of patient age and education in reshaping it. We focused on patients affected by diabetes mellitus, who are likely to suffer a higher degree of morbidity and mortality due to COVID-19. We used a bi-monthly administrative panel dataset from January 2019 to December 2020 from Liguria (Italy), one of the regions with the highest number of individuals over the age of 65 in Europe. The results demonstrated that, after the initial shock, when patients tried to increase their personal stock of drugs to overcome the risk of possible additional barriers generated by the coronavirus, the hoarding effect almost disappeared. Adherence has drastically reduced during the COVID-19 pandemic and has never reached pre-COVID levels again. Older and poorly educated patients seem to have suffered more from the restrictions imposed by the lockdown and fear of contagion and they may be the ideal target group when considering possible policy interventions to improve adherence.

Introduction

Medication non-adherence is one of the main challenges in the treatment of ailments, especially chronic diseases. It has been estimated that a lack of adherence causes nearly 125,000 deaths, 10% of hospitalizations and costs the strained healthcare system $100–$289 billion a year in the US, wherein approximately 50% of medications for chronic disease are not taken as prescribed (Viswanathan et al., 2012; Wilder et al., 2021). In Europe, the cost of non-adherence is about 125 billion euros per year (Cutler et al., 2018; Khan and Socha-Dietrich, 2018; Pharmaceutical Group of the European Union, 2008) for a death toll of about 200,000, which also accounts for about half of the potentially avoidable cost of inappropriate medication usage (IMS Institute for Healthcare Informatics, 2013).

Non-adherence is more prevalent among chronic patients, particularly in older adults, especially those requiring a complex drug regimen due to a higher number of coexisting diseases (Di Novi et al., 2020). Attempts to explain disparities in therapeutic nonadherence have mostly focused on drug affordability and patient-related factors, such as socioeconomic status (SES), health condition, and type of therapy prescribed (Doshi et al., 2009; Gaynor et al., 2006; Osterberg and Blaschke, 2005; Hughes et al., 2008; Horii et al., 2019). A strand of literature has attempted to study its evolution over time because of innovation (Blankart and Lichtenberg, 2020), generic substitution, (Costa-Font et al., 2014) and other forms of medical intervention (Pagès-Puigdemont et al., 2019).

From a policy point of view, one of the most relevant distinctions is between barriers to access, personal attitudes and patients’ behavior. Barriers to access may be related to both financial constraints and physical barriers. In the US, as in several public healthcare systems across Europe, co-payments are necessary to receive drugs; for chronically-ill patients, they may become a heavy burden. However, some barriers also seem to exist in healthcare systems, such as the Italian one, where drugs are free of charge for low-income and chronically-ill patients (Atella et al., 2017). Physical barriers refer to travel costs (longer travel time or longer distance) in reaching drug providers and “pharmacy desert” that may be important barriers to patients' ability to fill prescriptions even in the absence of economic barriers (Di Novi et al., 2020; Leporatti et al., 2021).1 Finally, adherence to drug therapies depends on individuals’ attitudes toward their health, which is, in turn, determined by age, income, education, and personal traits. For chronic diseases, these aspects are particularly important because drugs should be taken regularly and for a long period of time (usually for the foreseeable future).

The global crisis due to the coronavirus disease 2019 (COVID-19) pandemic has brought significant changes across many aspects of our society. An aspect that has not yet received due attention is the evolution in drug utilization due to COVID-19 as a consequence of the barriers in using healthcare facilities, because of changes in lifestyle, future prospects and as a result of medication management practices that most healthcare systems have implemented to reduce the impact of pandemic events (Clement et al., 2021; Patel et al., 2021; Zhao et al., 2021; Agh et al., 2021; Kardas et al., 2021).

The COVID-19 pandemic may have exacerbated non-adherence, especially during the lockdown periods. In 2020, access to health facilities, general practitioners (GP), and specialists have often been difficult, and this may have deeply compromised adherence, with a consequent possible increase in the number of chronic diseases and/or in the deterioration of patients' conditions in the near future (Atella et al., 2017; Depalo, 2020). Also anxiety and uncertainty regarding the future may have reduced the effort that individuals made in order to be adherent to their therapy (Yeoh et al., 2021).

In Italy, despite the prompt response of the National Health Service (NHS), which has made remote drug prescription (i.e., electronic prescription over the phone sent to the patient through e-mail or SMS) available, the behavior of patients and their attitude toward adherence to therapy may have been influenced by the crisis brought on by the pandemic, especially among the most frail and vulnerable patients; older and lower educated patients may have found more barriers to healthcare access. These groups generally show a reduced level of health and computer literacy, which may have (negatively) affected their access to electronic prescriptions.

Traditional prescription channels required a visit to general practitioners and pharmacies, which has been a risky and difficult activity during the lockdown, especially for older people. Furthermore, social distancing measures, which were often necessary to protect the elderly against the risk of coronavirus, create further difficulties in relying on their usual social network or informal care to perform such activities (Leporatti and Montefiori 2020; Barry and Hughes, 2021).

The objective of this study was to investigate the potential impact that COVID-19 and lockdown restrictions may have had on drug utilization and the role of patient age and education in reshaping it in times of COVID-19. In particular, we focused on two different aspects: stockpiling and therapy adherence where stockpiling measures patients’ concern regarding the possibility of accessing the drug or their fear of shortage, in the context of great uncertainty.2

This study refers to the Ligurian population and focuses on older adult patients (aged 65+) affected by diabetes. Liguria is the region in Italy with the highest concentration of individuals over 65 and anticipates a demographic scenario that, according to Eurostat forecasts, will be prevalent across most European countries in the coming decades (Eurostat, 2020). Since chronic conditions and aging are correlated, Liguria represents a very interesting natural economic laboratory for this study. Older adults are more exposed to the risk of contracting a severe COVID-19 infection; they are more likely to suffer from multiple chronic diseases, which increases the risk of contracting a more acute form of the virus with unfavorable outcomes. Among the chronic pathologies, diabetes (as well as pathologies of the circulatory and respiratory systems) deserves particular interest. Early studies have shown that about 25% of people who went to the hospital with severe COVID-19 infections had diabetes. Those with diabetes were also more likely to have serious complications and to die from the virus. One reason for this is that high blood sugar levels weaken the immune system and hamper its ability to fight off infections (Fang et al., 2020).

To the best of our knowledge, this study is the first attempt to describe how COVID-19 has changed individuals’ behavior in terms of demand for drug therapies. We think that this investigation is relevant for policy makers because, by identifying those categories that are at higher risk of deteriorating their health because of reduced adherence, it offers suitable tools to deal with the ongoing emergencies that threaten the social, economic, and health aspects of the countries coping with the virus.

The remainder of this study is organized as follows. Section 2 describes the data employed in this study while Section 3 presents the empirical model. Section 4 presents the results of this study. The discussion and concluding remarks are provided in Section 5.

Data

We employed a unique ad hoc dataset obtained by linking: (i) the Hospital Discharge Records (HDRs) database containing information about hospitalizations; (ii) the Emergency Department's (ED) registry containing information on all visits for ED services; (iii) the Pharmaceutical Registry (PR) containing information about drug utilization; (iv) the Exemptions database and (v) the Ligurian Population Death Registry. The linkage was possible due to the presence of a unique patient identification code. We specifically considered the period from January 2019 to December 2020.

The HDRs and ED database were used to retrieve patients' demographics and socioeconomic information, that is, sex, date of birth, municipality of residence, marital status, and patients’ educational level. The PR was employed to elicit patients’ drug utilization behavior and medication adherence before and during the spread of the COVID-19 pandemic. The Exemptions database was used to collect information about patients’ health status, specifically whether they suffered from multiple chronic diseases (see Section 3.2). In addition, we employed the Ligurian Population Death Registry to exclude patients who died during the observation period.

We selected patients aged 65 years and over who had diabetes and were being treated with Metformin (Anatomical Therapeutic Chemical – ATC codes equal to A10BA02) during 2018, and who picked up their prescription at least once during the COVID-19 outbreak.3 Metformin is the most common medication worldwide for treating type 2 diabetes mellitus, recommended as first-line therapy by most international and national guidelines including the American Diabetes Association guidelines, the European Association for the Study of Diabetes guidelines (Song, 2016; Schernthaner and Schernthaner, 2020); thanks to its antihyperglycemic effect, it helps to lower blood glucose concentration levels that are higher than normal in diabetes mellitus patients due to an insufficient insulin release or the improper response of cells to insulin (Nasri and Rafieian-Kopaei, 2014).

Each patient was observed for two years over the period of 2019–2020. After conditioning on having no missing value on any dependent variable and/or covariate, the final sample consisted of 8969 patients, observed for 12 bimesters (107,628 observations).4

Empirical model

The empirical model in this study explores the change in drug utilization brought about by the COVID-19 outbreak. This study investigates this occurrence across the following dimensions:

The impact of the COVID-19 pandemic on patients’ adherence.

The pick-ups patterns at the outbreak of COVID-19 and throughout 2020.

From the data (described in detail in Section 2), we derive a bi-monthly measure of adherence (A) and stockpiling (SP) that allows us to define a general equation of change in the pattern of drug utilization (DU)., for patient i in bimester t can be modeled as follows:where D, and H are exogenous predictors, namely patients’ demographics (measured in this model by age, sex and municipality of residence), socioeconomic characteristics (measured in this model by marital status and level of education) and patients’ health conditions unrelated to COVID-19, respectively; while d are five dummy variables identifying the bimester of the year (with January- February as the baseline) which allows encompassing the yearly seasonality of drug utilization; d*y are bimester fixed effects interacted with a dummy variable (y2020) identifying the year 2020. is an individual-specific and time-invariant random component. εit is a time-and individual-specific error term that is uncorrelated across patients and waves. Moreover, it is assumed to be uncorrelated with and strictly exogenous (the explanatory variables are uncorrelated with ε for all t and i).

The key parameters for our study are the marginal effects of d that measure the impact of COVID-19. These interaction effects capture the differences between each bimester of the year 2020 and the average of the corresponding bimester in the previous year which cannot be explained by any of the other observed factors.

Drug prices were not included among the control variables because Metformin, which is used to treat type 2 diabetes mellitus, belongs to the so-called “class A” drug category, which is fully reimbursed by the Italian NHS and dispensed directly through hospital pharmacies (Distribuzione Diretta) or by territorial pharmacies (Distribuzione per conto) (see Folino-Gallo et al., 2008; Leporatti et al., 2021).

We tested the hypothesis that COVID-19 has had heterogeneous effects on patients’ behavior due to age and patients’ education by running the Eq. (1) for age groups and patients’ levels of education separately.

Dependent variables

The indicator of patients’ medication adherence relies on the concept of medication possession rate (MPR). MPR was constructed according to the Italian Medicines Agency's (Agenzia Italiana per il Farmaco [AIFA]) definition, which corresponds to the ratio between the number of dispensed therapy days (calculated on the basis of defined daily dose (DDD)) and the number of days in the time interval of interest.5 , 6 To study patients’ adherence and stockpiling behavior, we exploited the rules set for drug prescriptionsin Italy which, for chronic patients, allow to prescribe a maximum of two-months therapy.

The variable was obtained using a two-step approach:

First, we computed the total number of days of therapy that should be covered in bimester t (relying on the DDD) after identifying the exact date on which the drugs were dispensed.

Second, we checked the time period covered by each patient's drug pick-up. If the drugs picked up in bimester t covered a period that goes beyond the end date of the bimester, the days of therapy that exceed the length of the bimester were recorded as (which presumably covers a part of bimester t + 1). To correctly compute stockpiling for the first bimester (January-February 2019), we have used data on the last bimester of 2018 (November-December 2018) even though this bimester was not included in the period of observation.

Hence, measures, for each patient i in each bimester t (where t = 1, …,12), the excess in days in the format of DDD were picked up and we expect it to increase if patients anticipate barriers/delays to obtain new prescriptions.

Concerning medication adherence , for each patient i in each bimester t (where t = 1, …, 12), we built the following indicator:where denotes the number of days of therapy bought by patient i for the bimester t, denotes the number of days of therapy “inherited” by the previous bimester (t-1) as “stock” and discounts the number of days on therapy that are bought by patient i, in the bimester t, but consumed in the next bimester (t + 1).7

The indicator ranges from 0 to 100. Typically, rates of 80% or more are required for therapy adherence (Tang et al., 2017).8

Independent variables

In our model, we also controlled for a set of individuals’ demographic and socioeconomic characteristics and patients’ health conditions unrelated to the pandemic itself. Table 1 sets out a full description of the variables used in this model.

Table 1

Variables Names and Definitions.

	Variable name	Description
Outcome variables	Adherence Level	Panic buying—adjusted bi-monthly MPR
	Stock Level	Stock level in the bimester
Patient's demographic characteristics	Male	Dummy = 1 if Male
	Age class	Categorical variable reporting the individual age
	Age_class (65–74)	Dummy = 1 if patient's age is in the range 65–74 (baseline category)
	Age_class (74–84)	Dummy = 1 if patient's age in the range 75–84
	Age_class (85+)	Dummy = 1 if patient's age in the range 85+
Patient's socioeconomic characteristics	Marital status	Categorical variable reporting the individual marital status
	Married	Dummy variable = 1 if the patient is married (baseline category)
	Single	Dummy = 1 if the patient is single.
	Divorced	Dummy = 1 if the patient is divorced.
	Widow	Dummy = 1 if the patient is a widow.
	Educational level	Categorical variable reporting the individual educational level
	Primary or No Education	Dummy = 1 if Level of education = "No education” or “Primary Education” (baseline category)
	Lower Secondary Education	Dummy = 1 if Level of education = "Lower Secondary Education” or “Tertiary Education”
	Upper Secondary or Tertiary Education	Dummy = 1 if Level of education = "Upper Secondary”
Patient's clinical characteristics	Comorbidities	Number of comorbidities based on exemption codes
Patient's municipality of residence	Municipality fixed effects	Municipality fixed effects
Seasonal dummies	January-February	Dummy = 1 for bimester January-February (baseline category)
	March-April	Dummy = 1 for bimester March-April
	May-June	Dummy = 1 for COVID-19 bimester May-June 2020
	July-August	Dummy = 1 for bimester July-August
	September-October	Dummy = 1 for bimester September-October
	November-December	Dummy = 1 for bimester November-December
	January -February ## Year 2020	Interaction between bimester January-February and Year 2020
	March-April ## Year 2020	Interaction between bimester March-April and Year 2020
	May-June ## Year 2020	Interaction between bimester May-June and Year 2020
	July-August ## Year 2020	Interaction between bimester July-August and Year 2020
	September-October ## Year 2020	Interaction between bimester September-October and Year 2020
	November-December ## Year 2020	Interaction between bimester November-December and Year 2020

Demographic characteristics are summarized by patients’ sex and age entered as three age classes namely “age class 65–75″ (reference category), “age class 74–84,” and “age class 85+.” In addition, we controlled for the municipality of residence dummies (fixed effects). Marital status and educational level were used to define the socioeconomic status of the patients. Marital status was categorized into four dummy variables, namely: married (reference category), single, divorced, or widowed. The level of education was categorized into three dummy variables, namely: no education or primary school certificate (reference category), lower secondary education, and upper secondary education or tertiary education.9

Finally, we included an indicator of multiple chronic conditions. This information was obtained from the chronic disease registry for medical co-payment exemptions (Exemptions database). Indeed, in order to contain the moral hazard effects, the NHS has envisaged a system of co-payments that are applied to prescription drugs, specialist visits, and diagnostic checks. However, individuals who are over 65, those with an income below a certain threshold level, are entitled, for equity reasons, to be exempted from cost-sharing. Those who suffer from chronic conditions too are exempted from medical co-payment of all healthcare services connected with the disease. Exemption codes are defined at the national level by the Ministry of Health and periodically updated to adequately cover the major chronic conditions affecting the population. For each patient, our variable accounted for the number of exemptions for chronic conditions and was used as an indicator of multiple chronic conditions.

Methodology

As shown in Eq. (2), adherence is a continuous variable restricted to the interval between 0% and 100%; hence, as for patients’ adherence, Eq. (1), is estimated using a fractional regression model assuming a binomial distribution and a Probit Link Function (see Papke and Wooldridge, 1996, 2008).

Stockpiling behavior relies on count data (i.e., whole numbers with a lower bound at 0) whose distribution is skewed to the right. In this last case, we estimated Eq. (1) by applying a negative binomial (NegBin) regression model. NegBin is a modified Poisson regression model that takes into account the overdispersion (due to unobserved heterogeneity), thereby relaxing the restrictive assumption that the variance and mean are equal (Greene, 1994; Sheu et al., 2004; Hilbe, 2014).

We opted for a random-effects model because the fixed-effects model is not suitable for our data, as most of the explanatory variables were time-invariant variables or time-variant variables which were relatively stable over the period of observation.10

Results

Fig. 1 reports the weekly number of Metformin pick-ups by Ligurian patients in 2019 and 2020. As can be seen, in correspondence with the first lockdown (March 2020) there was a peak in the number of drug pick-ups and maintained as so for about three weeks, which was at higher levels than the previous year. In subsequent periods (net of the fluctuations due to seasonality and taken into account in the econometric models proposed) the number of drug pick-ups tended to be lower than that observed in 2019.

Fig. 1

Trend in weekly amount of Metformin picks up by year

Note: Dashed line points out the start of lockdown restrictions.

Table 2a provides the sample means and standard deviations for the variables used in the model (51% male; mean age: 78 years). According to our definitions of adherence, the mean value was equal to 0.57, and around 34% of the sample was highly adherent to their therapy (MPR >80%). Furthermore, patients in each bimester stockpiled drugs for 12 days on average.

Table 2a

Descriptive Statistics (Mean and Standard Deviation).

	Variable name	Mean	Std. Dev	Percentage	25th percentile	75th percentile	Min	Max
Outcome variables	Adherence Level	0.57	0.34		0.29	0.92	0	1
	High Adherence Level (MPR>80%)			34%
	Stockpiling	11.73	18.73		0	18.5	0	85
Patient's demographic characteristics	Male	–	–	51.04%
	Age	78.03	7.03		73.0	83.0	65	102
	Age class
	Age_class (65–74)	–	–	34.15%
	Age_class (74–84)	–	–	46.37%
	Age_class (85+)	–	–	19.48%
Patient's socioeconomic characteristics	Marital status
	Married	–	–	68.25%
	Single	–	–	18.52%
	Divorced	–	–	3.05%
	Widow	–	–	10.18%
	Educational level
	Primary or No Education	–	–	45.91%
	Lower Secondary Education	–	–	37.61%
	Upper Secondary or Tertiary Education	–	–	16.48%
Patient's clinical characteristics	Comorbidities	1.59	1.17		1	2	0	7

Table 2b and Fig. 2 show the average monthly adherence over the period of observation (January 2019–December 2020). Stockpiling showed a peak at the outbreak of the first COVID-19 wave, which was not related to seasonality. Adherence shows a negative trend in 2020; however, there was variability across bimesters that support the opportunity to use a dummy variable associated with each bimester.

Table 2b

Descriptive Statistics (Mean and Standard Deviation).

	Stockpiling					Adherence
Bimester	2019		2020		Difference t-test	2019		2020		Difference t-test
	Mean	Std. Dev	Mean	Std. Dev		Mean	Std. Dev	Mean	Std. Dev
January-February	12.454	20.427	12.648	20.571		0.577	0.33	0.587	0.331	*
March-April	11.804	19.907	13.145	22.121	⁎⁎⁎	0.586	0.334	0.597	0.337	*
May-June	12.53	20.997	12.273	20.309		0.595	0.331	0.557	0.346	⁎⁎⁎
July-August	11.855	20.333	10.905	19.377	⁎⁎⁎	0.587	0.338	0.561	0.343	⁎⁎⁎
September-October	12.313	19.878	11.995	19.694		0.57	0.338	0.546	0.348	⁎⁎⁎
November-December	11.738	19.57	10.781	18.707	⁎⁎⁎	0.576	0.334	0.541	0.352	⁎⁎⁎

Note:.

p < 0.10.

⁎⁎p < 0.05.

p < 0.01.

Fig. 2

Descriptive statistics on adherence and stockpiling.

Table 3 and Fig. 3 show the marginal effects of the RE fractional Probit and NegBin models.11 Specifically, the dependent variable in the first column measures patients’ adherence while the dependent variable in the second one measures their stockpiling behavior. Our results demonstrate that being male and having a higher level of education increases drug utilization: in particular, the rate of adherence for males and those who are highly educated increases of about 3%, while the stockpiling increases by about 11%. In contrast with most of the previous studies (see among others Di Novi et al., 2020), having comorbidities seems to increase the rate of adherence by about 2% and stockpiling by about 6%; living alone is a detriment to both behaviors, specifically: being single reduces the rate adherence by about 2% and stockpiling by about 9%, while being widowed reduces the rate of adherence by about 4% and the stockpiling of about 12%); being older has a negative impact on drug utilization: for the over 85 the adherence rate decreases by about 7% while the stockpiling of about 40%.

Table 3

Results of the random effects regression model on adherence (i.e. fractional model specification) and stockpiling (negative binomial model specification).

	Adherence	Stockpiling
Male	0.026⁎⁎⁎	0.110⁎⁎⁎
	(0.005)	(0.011)
Age (Reference category = 65–74)
Age_class (75–84)	−0.029⁎⁎⁎	−0.170⁎⁎⁎
	(0.005)	(0.012)
Age_class (85+)	−0.070⁎⁎⁎	−0.411⁎⁎⁎
	(0.007)	(0.017)
Marital Status (Reference: Married)
Single	−0.023⁎⁎⁎	−0.091⁎⁎⁎
	(0.007)	(0.015)
Divorced	−0.011	−0.052
	(0.016)	(0.032)
Widow	−0.036⁎⁎⁎	−0.118⁎⁎⁎
	(0.009)	(0.020)
Education (Reference: primary or no education)
Lower Secondary	0.030⁎⁎⁎	0.102⁎⁎⁎
	(0.006)	(0.012)
Upper Secondary or Degree	0.029⁎⁎⁎	0.111⁎⁎⁎
	(0.008)	(0.016)
Comorbidities	0.018⁎⁎⁎	0.061⁎⁎⁎
	(0.002)	(0.005)
January-February # Year 2020	0.013⁎⁎⁎	0.010
	(0.003)	(0.021)
March-April # Year 2020	0.013⁎⁎⁎	0.096⁎⁎⁎
	(0.004)	(0.021)
May-June # Year 2020	−0.036⁎⁎⁎	0.005
	(0.003)	(0.021)
July-August # Year 2020	−0.024⁎⁎⁎	−0.083⁎⁎⁎
	(0.004)	(0.022)
September-October # Year 2020	−0.023⁎⁎⁎	−0.051⁎⁎
	(0.004)	(0.021)
November-December # Year 2020	−0.034⁎⁎⁎	−0.115⁎⁎⁎
	(0.004)	(0.021)
Municipality fixed effects	YES	YES
Bimester fixed effects	YES	YES
Number of observations	107,628	107,628

Note: *p < 0.10.

p < 0.05.

p < 0.01.

Fig. 3

Predictive margins.

The key-parameters to identify the effects of COVID-19 are those of the interaction terms between the six dummy variables for the bimesters with the 2020 dummy variable. As stated before, these interaction effects (reported in the rows from January–February # Year 2020 to November–December # Year 2020) should capture the differences between each bimester of the year 2020 and the average of the corresponding bimester in the previous year.

Concerning the stockpiling behavior, Table 3 shows that the marginal effect of the first bimester of January–February 2020 is not statistically significant, indicating that until the onset of the pandemic (March 2020), there were no unobserved factors distinguishing the pattern of drug stockpiling in the year 2020 from the first bimester of the previous year. More specifically, the COVID-19 shock produced a positive and significant effect on stockpiling, but only in the bimester with the most intense restrictions between March and April of 2020, while the effect of the bimester May-June 2020 was positive but not significantly different from the same bimester of 2019. Stockpiling has decreased since the bimester July-August 2020, and adherence followed a similar pattern, wherein after the March–April semester, all the marginal effects associated with the interaction terms between the remaining bimesters and the year 2020 were negatively signed and strongly significant, which was in line with our expectations. The rapid spread of COVID-19 significantly affected patients’ stockpiling behavior and their adherence in the first months in a positive manner, and then negatively with a more persistent effect. The strongest reductions in patients’ adherence occurred in the bimester May–June 2020 and during the resurgence of COVID-19 cases in the winter bimester (November–December 2020), where we estimated a decreasing effect of more than 3% on patients’ adherence.12 , 13

Fig. 3 reports a graphical description of the change in stockpiling and patients’ adherence between each bimester in the year 2020, relative to the same bimester in the year 2019; from Fig. 3, it is evident that during January–February, there were no differences in the stockpiling behavior in 2020 versus 2019. In the bimester March–April, however, we see a sharp increase in the stockpiling of drugs, the magnitude of which starts to steadily decline over the next 2020 bimesters. The change in patients’ adherence is positive and significant for the first two bimesters of the year 2020 and then adherence too steadily declined following a dynamic that was similar to the stockpiling behavior.

We next examined how age and the patients’ educational level interact with the COVID-19 outbreak. Tables 4 and 5 show the marginal effects of the specification of Eq. (1) for age groups and levels of education, respectively.

Table 4

Results of the random effects regression model on adherence – by age class (i.e. fractional model specification) and stockpiling (negative binomial model specification).

	Adherence			Stockpiling
	65–74	75–84	85+	65–74	75–84	85+
Male	0.025⁎⁎⁎	0.028⁎⁎⁎	0.016	0.113⁎⁎⁎	0.123⁎⁎⁎	0.072⁎⁎
	(0.009)	(0.008)	(0.011)	(0.018)	(0.017)	(0.029)
Marital Status (Reference: Married)
Single	−0.015	−0.029⁎⁎⁎	−0.045⁎⁎⁎	−0.076⁎⁎⁎	−0.058⁎⁎⁎	−0.228⁎⁎⁎
	(0.011)	(0.010)	(0.014)	(0.022)	(0.023)	(0.039)
Divorced	−0.006	−0.021	−0.063	−0.026	−0.050	−0.342⁎⁎⁎
	(0.022)	(0.024)	(0.047)	(0.042)	(0.053)	(0.120)
Widow	−0.025	−0.026*	−0.048⁎⁎⁎	−0.142⁎⁎⁎	−0.029	−0.257⁎⁎⁎
	(0.023)	(0.013)	(0.014)	(0.045)	(0.029)	(0.037)
Education (Reference: Primary or no education)
Lower Secondary	0.044⁎⁎⁎	0.020⁎⁎	0.020	0.138⁎⁎⁎	0.084⁎⁎⁎	0.079⁎⁎
	(0.010)	(0.008)	(0.012)	(0.021)	(0.018)	(0.031)
Upper Secondary or Degree	0.032⁎⁎⁎	0.018	0.016	0.151⁎⁎⁎	0.094⁎⁎⁎	0.050
	(0.012)	(0.012)	(0.020)	(0.024)	(0.025)	(0.049)
Comorbidities	0.020⁎⁎⁎	0.017⁎⁎⁎	0.021⁎⁎⁎	0.060⁎⁎⁎	0.050⁎⁎⁎	0.096⁎⁎⁎
	(0.004)	(0.003)	(0.004)	(0.008)	(0.007)	(0.011)
January-February#Year2020	0.025⁎⁎⁎	0.007	−0.009	0.076⁎⁎	−0.024	−0.094*
	(0.006)	(0.005)	(0.008)	(0.034)	(0.030)	(0.052)
March-April#Year2020	0.022⁎⁎⁎	0.005	0.007	0.116⁎⁎⁎	0.088⁎⁎⁎	0.040
	(0.006)	(0.005)	(0.008)	(0.034)	(0.031)	(0.053)
May-June#Year2020	−0.020⁎⁎⁎	−0.045⁎⁎⁎	−0.054⁎⁎⁎	−0.009	0.014	−0.030
	(0.006)	(0.005)	(0.008)	(0.034)	(0.031)	(0.053)
July-August#Year2020	−0.010	−0.037⁎⁎⁎	−0.027⁎⁎⁎	−0.066*	−0.089⁎⁎⁎	−0.116⁎⁎
	(0.006)	(0.005)	(0.008)	(0.035)	(0.033)	(0.056)
September-October#Year2020	−0.020⁎⁎⁎	−0.027⁎⁎⁎	−0.024⁎⁎⁎	−0.039	−0.053*	−0.099*
	(0.006)	(0.005)	(0.008)	(0.034)	(0.031)	(0.053)
November-December#Year2020	−0.013⁎⁎	−0.041⁎⁎⁎	−0.060⁎⁎⁎	−0.068⁎⁎	−0.137⁎⁎⁎	−0.183⁎⁎⁎
	(0.006)	(0.005)	(0.008)	(0.034)	(0.032)	(0.055)
Municipality fixed effects	YES	YES	YES	YES	YES	YES
Bimester fixed effects	YES	YES	YES	YES	YES	YES
Number of observations	36,781	50,031	20,816	36,781	50,031	20,816

Note:.

p < 0.10.

p < 0.05.

p < 0.01.

Table 5

Results of the random effects regression model on adherence by educational level (i.e. fractional model specification) and stockpiling (negative binomial model specification).

	Adherence			Stockpiling
	Primary or No education	Lower Secondary	Upper Secondary or Degree	Primary or No education	Lower Secondary	Upper Secondary or Degree
Male	0.025⁎⁎⁎	0.030⁎⁎⁎	0.021	0.097⁎⁎⁎	0.115⁎⁎⁎	0.089
	(0.008)	(0.009)	(0.014)	(0.017)	(0.019)	(0.064)
Marital Status (Reference: Married)
Single	−0.026⁎⁎⁎	−0.028⁎⁎	0.004	−0.076⁎⁎⁎	−0.153⁎⁎⁎	0.009
	(0.009)	(0.012)	(0.022)	(0.020)	(0.026)	(0.098)
Divorced	−0.015	−0.033	0.020	−0.009	−0.153⁎⁎⁎	0.220
	(0.030)	(0.023)	(0.031)	(0.063)	(0.049)	(0.137)
Widow	−0.036⁎⁎⁎	−0.029*	−0.059⁎⁎	−0.115⁎⁎⁎	−0.134⁎⁎⁎	−0.430⁎⁎⁎
	(0.013)	(0.015)	(0.026)	(0.029)	(0.032)	(0.121)
Age (Reference category = 65–74)
Age_class (75–84)	−0.017⁎⁎	−0.038⁎⁎⁎	−0.034⁎⁎⁎	−0.138⁎⁎⁎	−0.188⁎⁎⁎	0.088⁎⁎⁎
	(0.008)	(0.008)	(0.012)	(0.019)	(0.019)	(0.017)
Age_class (85+)	−0.058⁎⁎⁎	−0.081⁎⁎⁎	−0.072⁎⁎⁎	−0.385⁎⁎⁎	−0.418⁎⁎⁎	−0.116⁎⁎⁎
	(0.010)	(0.012)	(0.020)	(0.024)	(0.029)	(0.033)
Comorbidities	0.019⁎⁎⁎	0.012⁎⁎⁎	0.027⁎⁎⁎	0.079⁎⁎⁎	0.030⁎⁎⁎	0.108⁎⁎⁎
	(0.003)	(0.004)	(0.005)	(0.007)	(0.008)	(0.026)
January-February#Year2020	0.004	0.016⁎⁎⁎	0.030⁎⁎⁎	−0.010	0.022	0.055⁎⁎⁎
	(0.005)	(0.006)	(0.009)	(0.031)	(0.033)	(0.010)
March-April#Year2020	0.012⁎⁎	0.011*	0.021⁎⁎	0.074⁎⁎	0.115⁎⁎⁎	0.104⁎⁎⁎
	(0.005)	(0.006)	(0.009)	(0.032)	(0.033)	(0.010)
May-June#Year2020	−0.046⁎⁎⁎	−0.029⁎⁎⁎	−0.025⁎⁎⁎	0.044	−0.018	−0.060⁎⁎⁎
	(0.005)	(0.006)	(0.009)	(0.032)	(0.033)	(0.010)
July-August#Year2020	−0.028⁎⁎⁎	−0.026⁎⁎⁎	−0.006	−0.108⁎⁎⁎	−0.080⁎⁎	−0.047⁎⁎⁎
	(0.005)	(0.006)	(0.009)	(0.033)	(0.035)	(0.010)
September-October#Year2020	−0.022⁎⁎⁎	−0.027⁎⁎⁎	−0.015	−0.072⁎⁎	−0.013	−0.033⁎⁎⁎
	(0.005)	(0.006)	(0.009)	(0.032)	(0.034)	(0.010)
November-December#Year2020	−0.035⁎⁎⁎	−0.038⁎⁎⁎	−0.021⁎⁎	−0.123⁎⁎⁎	−0.107⁎⁎⁎	−0.061⁎⁎⁎
	(0.005)	(0.006)	(0.009)	(0.033)	(0.034)	(0.010)
Municipality fixed effects	YES	YES	YES	YES	YES	YES
Bimester fixed effects	YES	YES	YES	YES	YES	YES
Number of observations	49,603	40,544	17,481	49,603	40,544	17,481

Note:.

p < 0.10.

p < 0.05.

p < 0.01.

Table 4 shows that the marginal effects associated with the interaction terms between the first and second bimester (outbreak of COVID-19 in Italy) and the year 2020 were significant and positive for younger classes only. Stockpiling behavior seems to have been prevalent among the age group of 65–84 years, but it has translated into a higher level of adherence only for the “65–74 years” class. This “positive” effect seems to have left room for a negative and more permanent effect in the following bimesters, which registered a lower level of adherence for all age groups, with older age classes (75–84 and 85+) reducing adherence more than other classes.

Table 5 shows that patients who are less educated tend to stockpile less. This group, which was already less adherent to their drug therapy and was also most severely hit by the COVID-19 pandemic, increased the adherence gap. In line with our expectations, the COVID-19 pandemic worsened the situation of patients with lower SES in terms of medication intake. The strongest effect was observed during the second bimester from May–June 2020, when the stockpiling behavior effect, which mainly characterized the first bimester of the pandemic (March-April 2020) also disappeared; between May and June 2020, we estimated that among the less educated individuals, the pandemic reduced the rate of adherence by about 4.6%.

Discussion and conclusions

Drug adherence is an important determinant of outcomes in patients with chronic diseases; non-adherence causes additional healthcare costs. For those who suffer from degenerative ailments such as diabetes, adherence to medications is associated with better control over intermediate risk factors, lower probability of adverse effects, and a better quality of life, as diabetes may degenerate eyesight and other organs.

In the Italian healthcare system, financial barriers have been almost completely abolished: co-payment is sometimes waived on drugs and several diagnostic tests unless patients are exempt because of income or chronic conditions. However, non-financial barriers, such as travel costs in reaching drug providers and “pharmacy desert”, may also be important, as seen in previous studies (see Di Novi et al., 2020; Leporatti et al., 2021). Against this backdrop, the COVID-19 pandemic, which has created an unexpected increase in non-financial barriers due to lockdowns, mobility restrictions and the fear of contagion, may have created barriers to adherence that may differ extensively across populations.

With the outbreak of COVID-19, individuals relied on their primitive instincts and started to stockpile medicines in anticipation of possible barriers to access to the drug or delays in prescriptions (Elek et al., 2021). Stockpiling was not a lower-social-class prerogative, and was more concentrated among the less old age groups. Interestingly, our study demonstrates that after the initial period, patients have reduced this hoarding effect. The reduction in stockpiling during the second wave of COVID-19 may have been driven by two factors: on one hand, by a general decrease in drug purchases (and therefore also in stockpiling) and on the other, due to policies that healthcare providers have swiftly implemented (electronic prescriptions, implementation of digital health and other policies to facilitate citizens' access to medicines) may have reduced patients’ fair of shortages in pharmaceuticals.

The increased uncertainty and the risk of catching the virus seem to have increased patients’ motivation to comply with drug therapy, resulting in a higher level of adherence, but only at the onset of the COVID-19 outbreak. Patients’ adherence declined across the remaining 2020 bimesters and appeared to be negatively affected by the continuing COVID-19 outbreaks compared with 2019 and early 2020. The temporary positive effect of COVID-19 on patients’ adherence made room for a more persistent negative effect, in which widespread illness, quarantine, and social distancing measures exacerbated the barriers to patients’ drug utilization, contributing to the disruption of healthcare and pharmacy access.

Previous studies (see among others Di Novi et al., 2020), show that poor adherence is most prevalent among older adult patients. Our study demonstrates that the negative marginal effect of older age on drug adherence has been further exacerbated by the COVID-19 outbreak, such that the magnitude of the effect has increased over the period from May 2020 to December 2020, especially among those who were over the age of 75. Again, this result can be interpreted as the effect of the additional barriers generated by the measures that were adopted by the Italian government to curb the virus. Arguably, the elderly and those who suffer from chronic diseases (such as diabetes mellitus), with a higher degree of morbidity and mortality due to COVID-19, suffered from additional challenges. The fear of contagion, confinement, and ignorance of the new situation by patients and health professionals might have increased their reluctance to access healthcare facilities and might have affected their drug utilization and therapy adherence.

This specific group is the one that has been the most harshly affected and should be the target group when considering policy interventions to improve adherence. This is an important result from a policy point of view because this population is more fragile and is less adherent to nutrition guidelines for diabetes (Pinto and Braz, 2015); hence, they might suffer from further health problems in the future.

Our study also confirms that education is, in general, one of the most relevant predictors of drug adherence; during the COVID-19 pandemic, the level of adherence among less-educated individuals was more affected than that of highly educated individuals.

Increasing drug adherence to levels similar to those observed before the COVID-19 pandemic should be a priority for regional healthcare systems.

Some limitations must be acknowledged. Firstly, it is important to notice that the study is focused on a specific pathology (diabetes), so our findings may vary for other types of chronic illnesses. In order to support or add information to what is already provided in this paper, future efforts will have to be devoted to investigating other chronic conditions. In addition, this study focuses on a single administrative region (i.e. Liguria). However, given its demographic characteristics and the great burden suffered by Northern Italy due to the COVID-19 pandemic, this region is, in our opinion, of great interest for this type of study because diabetes is more prevalent among older adults.

Finally, we pointed out that the two variables at stake, namely stockpiling and adherence rate, are influenced by health policies implemented both at the national and local levels. However, with the data at our disposal, it is impossible to directly control for these variables. Nonetheless, these variables ultimately affect patient behavior in terms of drug purchase and adherence.

Declaration of Competing Interest

None.

Table A1

Characteristics of patients by missing status.

	Our sample	Missing
Number of patients	8969	18,906
% Male	51.09%	47.50%
Age (mean)	77.95	77.45
Age Class (%)
% 65–74	34.83%	38.32%
% 74–84	45.91%	44.47%
% 85+	19.26%	17.21%
Number of Comorbidities (mean)	1.60	1.40
Adherence Level (mean)	0.578	0.586
Stock Level (mean)	12.00	13.25

Table A2

Toy example–adherence and stockpiling computation.

Bimester	Start of the bimester	End date of the bimester	Num Days of the bimester	Date of first purchase in the bimester	Days of therapy bought in the bimester	End of coverage of bimester purchase	Stockpiling	Adherence
	A	B	C_t=B_t−A_t	D	E	F_t=D_t−E_t	G_t=D_t−E_t	H_t=(E_t−G_t+G_t−1)/C_t
1	10/1/2019	28/2/2019	50	10/1/2019	60	11/03/2019	11	(60–11)/45=1
2	1/3/2019	30/4/2019	61	15/3/2019	70	24/05/2019	24	(70+11–24)/61=0.94
3	1/5/2019	30/6/2019	61	–	0	–	0	(24)/61=0.4
4	1/7/2019	31/08/2019	61	7/8/2019	50	26/09/2019	26	(50–26+0)/61=0.4

Note: The start date of the first bimester period coincides with the date of first purchase in the that bimester