The observational evaluation of alcoholics' treatments requires a combined analysis of alcoholic behaviour during treatment and of adherence to therapeutic programmes. The application of survival analysis techniques in this setting has been explored in this study. Two hundred and seventy alcoholics admitted to 15 Italian treatment units in a 1-year period were followed-up for 2 years, recording date and length of every recurrence episode and of definitive or transitory interruption of the planned treatment. An extensive use of several survival analysis techniques was made. The length of time between the start of the treatment and the first episode of relapse did not give a reliable measure of frequency of failures. Conversely, the length of time between the start of treatment and withdrawal appeared to be unbiased. The cumulative proportions of treatment-compliant patients (and the corresponding 95% confidence intervals) were 71% (66–76%), 63% (57–69%) and 53% (47–60%) after 6 months, 1 year and 2 years respectively from the start of treatment. Cumulative abstinence duration before withdrawal was significantly and positively associated with the risk of first, of definitive, and of every episode of treatment interruption. This first application of survival analysis techniques to the combined study of alcoholic behaviour and of adherence to treatment can improve our knowledge of treatment evaluation. Our results suggest that compliance to treatment is an objective and versatile outcome measure. Long-term follow-up studies aimed to elucidate the determinants of withdrawal should be performed. Alcoholism treatments.
The design, management, analysis and interpretation of data on the observational monitoring of alcoholics' long-term treatment are rather complex (The Plinius Maior Society, 1994). Crucial issues are the identification of the outcome variables and the high risk of withdrawal of treatment.
Many outcome measures proposed in the literature, such as the ‘cumulative abstinence duration’ (CAD) ( Lehert, 1993 , Plinius Maior Society, 1994 , Whitworth et al., 1996 , Favre et al., 1997 ), are considered as explicit and general treatment objectives. Although abstinence is not necessarily always the only goal of a treatment programme, monitoring abstinence and relapses allows comparison of different therapeutic approaches. This is particularly true under conditions of identical compliance of patients admitted to the treatment programmes. In fact, the unreliability of information on alcohol-intake behaviour of patients who withdraw from therapy limits the possibility of comparing compliant with withdrawn patients and of comparing different approaches under conditions of different compliance.
Compliance, as a proxy measure of the effectiveness of a therapeutic project, does not necessitate an enquiry into alcohol-intake behaviour ( Aricò et al., 1994 ). However, a high degree of compliance does not necessarily reflect the ability of the therapeutic approach to avoid relapse or to maintain abstinence.
There is thus a need to explore more complex procedures, including both alcohol-intake behaviour and compliance with treatment. The aims of our contribution were: (1) to provide guidance to treatment programmes which require self-evaluation, (2) to advance knowledge of the relationship between alcohol-intake behaviour and participation in treatment, (3) to present the use of survival analysis in this setting. For these purposes, we have used data from the ongoing Assessment of Alcoholism Treatment (ASSALT) project. This is an observational prospective study involving 270 alcoholics admitted to 15 alcoholism treatment units throughout Italy in a 1 year period and followed up for 2 years.
A short account on the ASSALT project
The ASSALT project has been designed in the mid-1990s by the Epidemiological Group of the Italian Society of Alcohology (GESIA) with the main aim of identifying prognostic factors associated with alcoholism treatment outcome. Details of the aims, design, phases and instruments of the ASSALT project have been reported elsewhere ( Treatment Evaluation Group of the GESIA, 1994 ). Briefly, each of the 15 alcoholism treatment units (ATUs) contributed to the project, recruiting in a 1-year period all the patients with alcohol-related problems admitted for the first time to the unit. Exclusion criteria were: (1) every diagnosis different from alcohol dependence and/or abuse according to the DSM-III-R diagnostic criteria of the American Psychiatric Association (1987), (2) problems interfering with participation in the therapeutic project, including diagnosis of dementia (severe cerebral degeneration assessed either clinically or instrumentally) or addiction to other substances, when alcohol dependence was a secondary phenomenon, (3) problems interfering with the study protocol, including educational (illiterate status), psychological (documented cognitive deficiency) and/or physical (visual deficit) conditions which implied the inability to understand the questionnaires. Physical, psychological, relational and alcoholism assessments were performed using validated scales. Subsequently, a therapeutic project, either pharmacological treatment, or counselling, or psychotherapy and/or assignment to self-help groups, was defined for each eligible patient. Every treatment programme was scheduled for at least 2 years and required complete abstinence from alcohol. A 2-year follow-up was carried out starting from the beginning of treatment. Finally, every year a physical, psychological, relational and alcoholism assessment was performed using the same instruments administered at entry into the study.
A standardized form reporting information about alcohol-intake behaviour and compliance to treatment was compiled for each patient by the ATU staff every 3 months for 2 years from the start of treatment. Information was collected from the treatment registry of the unit, from other units, from other health services and/or self-help groups, directly from the patient, and from his/her relatives. When a patient had withdrawn from treatment or was lost to follow-up, s/he was contacted by telephone or directly at home 2 years after entry and information was collected directly from the patient, and/or from relatives.
Alcoholism treatment sydney
Days of abstinence, number and duration of the relapse episodes, compliance to, and causes of withdrawal from, treatment were recorded for the whole 2-year follow-up period. A relapse episode was defined as any interruption of abstinence, and its length was measured by the number of continuous days of alcohol consumption from the start of the episode. Transitory and definitive interruptions of treatment were defined as withdrawal for at least 1 month, then respectively followed or not followed by re-establishment of the planned therapeutic project. Reasons for treatment interruption were classified either as refusal to participate in the therapeutic project (withdrawal for refusal) or as other factors independent of the patient's will (e.g. death, hospitalization, imprisonment, institutionalization: unintentional withdrawal).
Four types of outcome variables were independently assessed. The first one is represented by recurrence and the other three by withdrawal from treatment. The specifications on definition of failure, of censoring and of the considered covariates are reported in Table 1 shows their general characteristics at entry into the study and during follow-up. Patients were predominantly men (69%), with a mean age of 44 years and a diagnosis of alcohol dependence (97%). Therapeutic indications included pharmacological treatment and psycho-physical counselling for all the patients, psychotherapy for 158 patients (59%) and/or self-help groups for 211 patients (78%).
One hundred and twenty-nine (48%) and 153 (57%) patients totally abstained from alcohol and totally complied with treatment respectively during the follow-up period. Failures included 141 relapsed patients (52%) and 117 withdrawals for refusal (43%). This last outcome was definitive or transitory for 96 and 21 patients, respectively. Fourteen patients abandoned the treatment for unintentional causes (seven deaths, four hospitalizations, two imprisonments and one institutionalization).
The cumulative proportions of abstainers and of compliant patients during follow-up are shown in Fig. 1 reports the results obtained by fitting Cox's regression model considering the association between CTD categories and the risk of the first relapse. An unexpected significant trend towards a decrease in relative risk at decreased proportion of received treatment was observed. The association between compliance and relapse is also shown in Table 3 reports the results obtained by fitting three Cox's regression models considering the association between CAD categories and the risk of the first episode of withdrawal (first model), of definitive withdrawal (second model) and of each episode of withdrawal (third model). Significant trends towards an increased risk of withdrawal as CAD decreased was observed for the three models. Although statistical hypotheses are not testable for comparisons between the three models, a better goodness-of-fit of the first model was obtained.
Our study shows that: (1) the combined analysis of alcohol-intake behaviour and of compliance to treatment can be useful for monitoring alcoholics during treatment in an observational setting, (2) patients with episodes of recurring alcohol intake during treatment have a higher risk of withdrawal from treatment, (3) survival analysis is an adequate tool in this setting.
The clinician who wants to offer a treatment to his alcoholic patients should have information on the probability that different kinds of failure might occur during treatment. The crucial issue in the monitoring strategy is the a priori choice of the failure events. The characteristics of the patients and of the therapeutic project that might modify the response must be carefully considered. The final goal is to know which is the best treatment for the individual alcoholic patient.
The ASSALT project has been designed as a natural experiment involving the planned observation of the usual therapeutic activity of the alcoholism treatment units who voluntarily agreed to participate in the project (Treatment Evaluation Group of GESIA, 1994). Comparison of the treatments' effectiveness between alcoholism units and/or between different therapeutic approaches was not suitable, however, by the observational design of the study, in the absence of a random allocation of patients into different therapeutic approaches ( Plinius Maior Society, 1994 ). Moreover, the results of the ASSALT project cannot be generalized to the population of alcoholics treated in Italy since a selected group of the more motivated and better-organized alcoholism units has voluntarily participated in the project.
Despite these limitations, the observational approach of the ASSALT project is the most rational tool for characterizing and monitoring alcoholics and for identifying factors associated with the risk of failure of the therapeutic project. However, only questions regarding the choice and the statistical management of the outcome variables were addressed here. Since the observational approach of the ASSALT project can be considered as a realistic simulation of the usual therapeutic activity of an alcoholism unit, these questions are not restricted to this specific project. It has been stated that, since the basis for selection and timing of interventions in observational studies is not precisely specified, sophisticated statistical analyses are necessary to attribute to a specific cause the success (or the failure) of a therapeutic programme ( Bull and Spiegelhalter, 1997 ). Survival analysis is a powerful and versatile tool widely used in several fields to study the occurrence and timing of different kinds of events. Surprisingly, it has found rare applications in the monitoring of alcoholics during treatment ( Fuller and Willford, 1980 , Aricò et al., 1994 ). The main reason why survival analysis seems particularly suitable in this field is that treatment of alcoholics can never be considered wholly successful. In other words, all the observations are necessarily right-censored and survival analysis is particularly suitable in the statistical treatment of censored data.
In this paper, several survival analysis techniques have been used and particular attention has been given to the putative response variables. This is a crucial point in planning prospective observational studies for which survival analysis techniques are suitable. The strict pre-definition of the outcome variables and of their putative determinants should be made on the basis of the characteristics of treatment programmes, but also considering the realistic possibility to collect and to define reliable response data.
In a first approach, the length of time between the start of treatment and first episode of relapse was explored as the cardinal variable. Using this variable as the outcome in a survival analysis model, it is possible to describe the alcohol-intake behaviour of patients over time, rather than only at the end of observation. However, the first episode of relapse is not informative about treatment failure. In fact, an early relapse episode followed by abstinence may be considered as a good result by many clinicians. We observed that patients with lower compliance to treatment paradoxically showed significantly lower risk of relapse in the use of alcohol. One might think that this unexpected result means that shorter interventions could be more effective in obtaining abstinence than intensive treatments. However, all the therapeutic projects considered here were planned as long-term interventions. Another theoretical possibility is that patients with moderate or slight severity of alcohol dependence were less motivated to comply with treatment and more likely to be abstinent after withdrawal. In other words, severity of alcohol dependence might be considered as a confounding variable of the effect of compliance on the risk of relapse. However, since similar results were observed independently from severity of alcohol dependence, this possibility also seems unlikely. We suspect therefore that the unexpected result is explained by the questionable reliability of data on alcohol-intake behaviour of patients withdrawn from the treatment. In fact, while self-reported alcohol-intake behaviours during treatment were always comparable with those referred by operators, only self-reported and unverifiable information was available after withdrawal. Under these conditions, the estimate of the proportion of abstainers over the time can be considered biased and cannot be used as an indicator of success of the therapeutic project. We cannot state the generalizability of this finding in settings different from that considered here, but we suspect that this is a problem common to many multi-centre data-bases.
These considerations justify the second approach used here, where the length of time between start of treatment and its withdrawal was considered as the cardinal variable. The estimate of the cumulative proportion of patients compliant to treatment is realistically unbiased, since all cases of treatment interruption were captured. In correlating this new outcome variable with alcohol-intake behaviour, two problems should be considered. Firstly, since validity of information is questionable after withdrawal from treatment, alcohol-intake behaviour should be considered only in the period in which the patient is in treatment. This recommendation is also justified by the observation that a lower goodness-of-fit of the model was obtained when considering as the outcome the definitive episode of withdrawal, rather than the first one. This could be explained by the fact that periods free from treatment (and therefore characterized by less reliable data on alcohol-intake behaviour) are also included in this case. Secondly, we can reasonably assume that the period of observation is strongly associated with abstinence duration. The longer the period of observation, the more likely the relapse. Thus, variables nominally similar to CAD should be considered as time-dependent covariates. Considering these two aspects in a model of proportional hazards, we observed, as expected, that patients with a lower proportion of abstinence duration during treatment showed a significantly higher risk of withdrawal. These findings suggest that the relationship between recurrence and treatment withdrawal should be considered as a realistic and accurate tool for monitoring treatment.
Survival analysis is able to take into account that some individuals experience more than one failure event during the follow-up period (repeated events analysis). This is the case for patients with more than one withdrawal, each followed by resumption of treatment. In our series, we did not observe a better fitting for the model that considers repeated events in comparison with the conventional model. This is likely to be due to the adjustment for the lack of independence between observations that lead to increasing standard errors of the estimates ( Wei et al., 1989 ). The small number of patients (21) with intermittent treatment is also likely to be involved in this result. However, since intermittence in treatment increases at increased periods of observation, the usefulness of repeated events analysis should be explored for a longer duration of observation.
Alcoholism treatment articles
A crucial problem in the application of survival techniques to analyse data from observational studies is the adjustment for confounding variables. In fact, since patients are not randomly assigned to treatments, several variables associated both with the covariate of interest and with the response variable might confound the observed effect. This typically occurs for the centre, in multi-centre data-bases, but also for several characteristics of the patients, such as demographic and social co-ordinates and initial severity of dependence, that might act as confounders. Proportional hazards models are able to adjust the estimates of interest for the effect of confounding, but several cautions in the choice of the adjustment method are needed.
In conclusion, this first application of survival analysis techniques to the combined study of alcohol-intake behaviour and of adherence to treatment for alcoholics can improve our knowledge on treatment evaluation. The availability of statistical packages allows an easy application of these techniques to the analysis of data. Furthermore, our results suggest that compliance to treatment is an objective and versatile outcome measure in monitoring alcoholics' treatment. The high risk of withdrawal for patients who are not able to maintain abstinence suggests that compliance could be considered as a proxy variable of a treatment's success. Long-term follow-up studies to elucidate the determinants of withdrawal should be performed.
This work has been carried out under the auspices of the Epidemiological Group of the Italian Society of Alcohology (GESIA) through a core grant from Italian Research Counselling (Contract CNR 96.03389.CT04).