Markov model
In theory, after the filling of each prescription a patient has to decide whether or not to continue his drug therapy. This means that before filling the next prescription there is a certain probability that a patient will continue drug therapy or discontinue. In time this can result in one or several changes in refill behaviour. The patients' actions, refilling a prescription or not in several fixed periods throughout the study period, can be analysed by a Markov chain model. With the Markov model we describe it is possible to assess the probability of various patterns, instead of just continuous use and/or discontinuation. The patients' refill behaviour is characterized by the filling of at least one ICS prescription within fixed periods (calendar years) during follow-up.
In this particular Markov chain model the states are defined as the years of follow-up. A patient is considered to be in a particular state if he/she has filled at least one ICS prescription in the associated year. If a patient first filled a prescription in a particular yeari, and subsequently in another yearj, and not in the years between yeariand yearj, the patient made a transition from state " yeari" to state " yearj". The state "> 2002" was defined and patients with transitions into it are patients that are not censored before the end of the study (31-12-2002) and of which no assessment of medication use can be performed due to the end of (study) follow-up. The transitions after the end of follow-up are unknown. Transitions from each state are the filling of a prescription in the next calendar year, in one of the other following years or in none of the years while under observation. For each state several transitions from and into another state are possible. As the filling of at least one prescription for an ICS is an inclusion criterion for the cohort all patients are present in the state year93. The first state, 1993, has no transition into it. The number of possible transitions for each state depends on the number of remaining years of follow-up after yeari. The later the year under consideration, the less the number of possible transitions becomes. For example, the first year of follow-up, 1993, has the most number of achievable transitions to another state; 10 and the last year of follow-up, 2002, has the least; 1. Clearly the state ">2002" has no transitions to other states.
As stated before transition probabilities can be assessed for the transition from stateiinto statej, to be more precise yeariinto yearj. To do so, a period of at least two consecutive years is necessary. From the state "2002" patients can only have transitions into ">2002" due to the end of follow-up. A first-order Markov chain model like the one we describe (see appendix) implies that the presence in a particular state only depends on the presence in the directly preceding state but is independent of all former states. This means that the probability of having filled at least one ICS prescription in 1996 and subsequently filling at least one prescription for an ICS in one of the following years, for instance 1997, are conditional to having filled at least one prescription in a preceding year. To have a transition from 1996 to another state, e.g. to refill a prescription in one of the years after 1996, a patient should have filled at least one prescription in one of the years before 1996. The possible transitions into the state, 1996, are 1) year93→year96 (the previous prescription(s) filled in 1993), 2) year94→year96 (the previous prescription(s) filled in 1994), 3) year95→year96, (the previous prescription(s) filled in 1995). E.g. in possibility 2) filling at least one prescription in 1997 given 1996; 1994; 1993, the transition from year96→year97 is conditional only to the transition from year94→year96 regardless of the transition into year94.
The transition matrix gives an overview of the probabilities for all possible transitions throughout the study period. The sum of all possible transition probabilities from a certain state is 1. As the states in this model are defined as subsequent time periods, the transition matrix is a visual description of possible behaviour and change in time. It is therefore possible to detect a change in time of a certain transition probability, e.g. for the filling of the next prescription in the immediately following year.
The occurrence of continuing ICS use can be affected by several factors, possibly leading to confounding. In order to adjust for potential confounding, a stratified analysis according to these factors can be performed. An important confounder is the history of patients at the start of the study. Previous users are "survivors" of the early period of pharmacotherapy and are likely to have a higher probability of continuous use [
23]. Therefore new and previous users were defined based on the use of ICS in 1991 and/or 1992.
Medication gaps
Some of the conceivable patterns of patients' variable prescription refill behaviour are shown in Figure
1. The distribution of patterns with different numbers of cumulative gaps for the follow-up period can be obtained by use of the transition matrix.
Medication gaps were defined as "calendar years without any prescription for an ICS". The occurrence of a gap of one year in the period [1993; 1995] would result in the filling of the next prescription for an ICS in 1995 after 1993, which corresponds to the transition probability
P93→95 (see appendix). The probabilities of cumulative gaps were assessed during the total duration of follow-up (see appendix), for example the probability of a gap of one year
T9 [1993, >2002] -
T8 [1993, >2002], cumulative gap of five years, and so on. Gaps could be in consecutive years or spread over several years. A patient with a cumulative gap duration of two years could have filled prescriptions for ICS in 1993 and for instance subsequently in each year during follow-up, except for 1995 and 2000 (Figure
1b). Another patient could have the same cumulative gap duration with a prescription in each year during follow-up, except for 1995 and 1996 (Figure
1c).