To assess factors associated with the therapeutic adherence, it is not sufficient to simply list individual variables or social environmental variables. To obtain systemic research results, a study model is required [21].

For the investigation of the therapeutic adherence, the Health Belief Model, social learning theory, consumer information processing theory, Fishbein's Theory of reasoned action, Multi-attribute attitude model, etc., are used as research models. Among these, the Health Belief Model was initially developed to investigate preventive health behaviors. Nonetheless, recently, it has been used widely to study the therapeutic adherence associated with illness behavior and sick role behavior, and it has been used most widely as a health-related behavior model in the past 30 years [22].

Regarding the Health Belief Model, the value-expectancy theory that is the prerequisite of this model has been accepted widely as the theory explaining the motivation of humans. By establishing the belief or attitude that can be altered as the determining factor of the behavior, it is possible to actually utilize research results. In addition, with the Health Belief Model, data can be collected by diverse methods, the data can include as many variables as the investigator desires, and the number of questions required to measure the effect of the major variables is relatively small. Thus the Health Belief Model is an economical model [23].

In our study, because the study subjects were elderly patients, we kept the number of questions relatively small, and thus made the study economical. Once we selected the Health Belief Model, we could collect data using various methods, including the self-efficacy variable suggested by Bandura [24], which studies the effects on medication drug intake (Fig. 1).

Therapeutic adherence means that the patient observes the medical recommendations, taking the medication, and maintaining a lifestyle as recommended by clinicians [3]. In our study, we focused on medication adherence.

The methods typically used to measure medication adherence include measuring serum drug level, measuring pharmacological marker levels, performing pill counts (pill count method), and patient interview, etc. Among these methods, the adherence assessment method by interview involves directly asking the patient whether they take the drugs. The interview approach is simple and thus this approach was not popular among investigators. Nevertheless, since the approach is readily conducted and other methods assessing adherence have the limited usefulness, such method assessing the adherence has drawn attentions gradually. Methods that assess adherence by interview can be broadly classified into two types of methods. The first approach is to lead the response of the patients by using non-critical and non-threatening questions, and the second approach is to assess the adherence using a structured questionnaire, such as the self-report questionnaire. In our study, we evaluated the medication adherence using the Morisky self-report questionnaire, which was previously validated. The questionnaire is designed validate the assessment of the medication adherence, and in addition, to understand and evaluate the adherence barrier, which is advantageous for clinicians supporting patients who have problems with adherence [25]. In other words, by answering "yes" even to one out of 4 questions, the patients are indicating their non-adherence experience. The self-report questionnaire has the advantage that it is simple and can be conducted in any clinical environment rapidly. This method saves time and improves the validity of the medication adherence [26].

Morisky's self-questionnaire consists of 4 questions: i) Do you ever forget to take your medicine? ii) Are you careless at times about taking your medicine? iii) When you feel better, do you sometimes stop taking your medicine? iv) Sometimes if you feel worse when you take the medicine, do you stop taking it? If the patient answers negatively to just one out of 4 questions, it is defined as 'non-adherence'.

The independent variables to explain the medication adherence of dependent variables include socio-demographic characteristics, lifestyle, the characteristics associated with disease and medication, and variables such as susceptibility, severity, benefit, barrier, cues to action in the Health Belief Model and self-efficacy variable.

Socio-demographic status and lifestyle variables included age, gender, education level, residence, financial level, the type of health security, person paying treatment fees, marital status and subjective health status, meal habits, exercise state, drinking state, and smoking state.

The characteristics associated with disease and medication include the number of currently ailing diseases, the number of medical institutions visited for the current ailment, time since diabetes diagnosis, the number of diabetes drugs, the times of diabetes drug medication per day, the status of the insulin injection, and the condition of drug storage. The total current ailing diseases were classified as cases with only diabetes, 2 diseases, and more than 3 diseases. The time since diabetes diagnosis was classified into 3 groups based on less than 5 years, between 5 and 10 years and more than 10 years. The number of current diabetes medications was classified as either one drug or more than two drugs. The frequency of medication was classified as once or more than two times. The drug storage condition was classified as well storage when the patient stored drug according to instructions of doctors or pharmacists or stored at dry room temperature, and bad storage when the patient stored any drug in a refrigerator, or stored anywhere.

Among variables included in the Health Belief Model, susceptibility was determined by the answer to the question "Do you think that it is possible to develop complications if you do not take diabetes medication as instructed?" Answers were originally ranked on a 5-point scale from "very high" to " not at all". However, there were few answers of "No" and "not at all" so the answers were ultimately divided into 3 scales: "very high", "high", and "lower than moderate". The severity was measured according to answers to the question "Do you think that complications caused by diabetes are very severe?", with the answers graded on a 5-point scale ranging from "very severe" to "not severe at all". These answers were also eventually divided into 3 scales. Benefit was assessed by the question "Do you think that correctly taking your diabetes medication will help to prevent complications?" The answers to this question were divided into 3 scales, as was done with susceptibility. The barrier level was evaluated by the question: "Is it difficult to use hospitals to treat diabetes?" The answers for the barrier level question were divided into 2 categories, either "more than so" or "less than so so". The cue of action was measured by whether the patient knew another patient in the vicinity who had developed diabetic complications. For self-efficacy, the confidence in taking diabetes medication was measured by a 5-point scale, and the answers were classified as "very confident, "confident", and "not so confident".

As independent variables, for the multivariate analysis, we chose the significant variables among general characteristics such as financial level, meal habits, and the state of drug storage. Health security type was related to financial level, and there were only two cases of medical aid at tertiary hospital, so this variable was excluded from the analysis. For the variables of the Health Belief Model, including the susceptibility, severity, and benefit variable, there were fewer than 10 patients who answered 'lower than moderate'. For self-efficacy, only 6 private clinic patients answered 'very high'. When these variables were considered as the same category at bivariate analysis and multiple logistic regression analysis was performed, the model fitness was significant, and the results were similar to those from the bivariate analysis. Nevertheless, the Hosmer and Lemeshow test showed that the significant probability was low, and thus the model was considered to have some problems. As a result, we repeated the analysis after reclassifying susceptibility, severity and benefit variable as "very high" and "lower than high", and reclassifying self-efficacy as "more than high" and "lower than moderate".

In the analysis of medication adherence, any patient answering "no" to all 4 questions was considered to be adherent, and any patient answering "yes" to more than one of the 4 questions was considered to be non-adherent. Then, we compared patients from tertiary hospitals with those from the private clinics.

We performed bivariate and multivariate analysis with medication adherence as the dependent variable, and general characteristics, Health Belief Model variables and the self-efficacy variable as independent variables. For the bivariate analysis, the χ²-test was used. In the multivariate analysis to identify factors that affect the medication adherence, we considered the medication adherence as a dependent variable, and the significant variables among general characteristics, the variables of the Health Belief Model, and self-efficacy as independent variables, and we performed multiple logistic regression analysis.