Estimating unbiased treatmen effects in general practice using instrumental variable analysis


Choice of treatment in general practice is not an exact discipline, and general practitioners may have different preferences for treatments. This treatment variation can in special cases be used to create better evidence using a particular statistical method, instrumental variable analysis (IVA).

Under certain conditions, IVA returns unbiased treatment effect estimates. This makes the method attractive in studies based on healthcare databases that are unbalanced by lack of information on important risk factors. However, when studying endpoints like hospitalization and mortality where patients are included in the study over calendar time, the potential follow-up time obtained from healthcare databases will vary among patients.

There is currently limited theory on how to handle time-to-event outcome in the IVA setup. Furthermore, in order to utilize general practice treatment variation in the IVA setup when conducting primary care research, one hindrance is the missing link between general practitioners and patients containing historical information on which general practitioner a given person is listed with at any given time.

Employing general practice treatment variation in an IVA setup with time-to-event outcome would enable us to study the effect treatment options for bereavement related grief in general practice on psychiatric hospitalization, all-cause hospitalization and mortality.


The aims of this PhD project is threefold:

  • We intend to extend the IVA methodology to time-to-event outcome.
  • We aim to develop, implement and evaluate an algorithm linking general practitioners and patients.
  • The IVA method for time-to-event outcome from the first study and the algorithm from the second study will be applied in a study of the impact of treatment (talk therapy, referral to psychologist and/or antidepressant drugs) of bereavement related grief in general practice on hospitalization for psychiatric conditions, self-induced harm and suicide.


IVA can be used to explore many different research questions in a truly population-based sample of patients from general practice. Inconclusiveness as to whether or not bereaved individuals should be treated constitutes a constant clinical dilemma.

Besides estimating treatment effect in an attempt to improve health among bereaved individuals, a particular vulnerable group of patients in general practice, the third study may potentially contribute to guide decision-making and health care delivery in general practice.

The IVA method for time-to-event outcome in study one will be of interest to any researcher concerned with IVA and time-to-event outcome, and the algorithm of study two will be of interest to anyone studying variation in connection with primary care research in Denmark.