Page 93 - Effective healthcare cost containment policies Using the Netherlands as a case study - Niek W. Stadhouders

The marginal benefits of healthcare spending in the Netherlands
bias the marginal effect downwards. In addition, morbidity-related QALY loss is not disaggregated to disease category, which may explain differences in disease category thresholds. The approximation of the EQ-5D by the health questionnaires was not validated, and may be imprecise. Also, uncertainty in translation of the EQ-5D to QALY values was not incorporated. The use of QALY-values from the literature assumes comparability, which may be a strong assumption (Gafni and Birch, 1993). Furthermore, we assume that the change in morbidity-related QALY loss is constant over time, while spending-related health shocks may be present.
Our indicator may not capture additional health system outcomes, biasing the threshold upwards (Nixon and Ulmann, 2006). For example, in fertility treatments, reductions in morbidity-related QALY loss and death-related QALY loss may not fully capture all benefits. In these instances, our estimation underestimates the true benefits of health spending. Importantly, effects of spending on future mortality and future gains in patient quality of life are not taken into account. Furthermore, our data do not incorporate all health spending, such as private spending, spending on primary care and municipal health spending. This could bias the marginal effect upward if these types of spending are complementary to hospital spending. However, research suggests no correlation between spending types (de Jong et al., 2016).
One major factor influencing our estimation is the correction for reverse causality using cost in the last year of life. This is not disaggregated to disease category, which may explain differences in disease categories. For example, cost in last year of life of a patient that died due to external causes (e.g. traffic injuries) may be lower than the cost in the last year of life of a cancer patient. This would have little effect on the main estimation, as the translog function is estimating the elasticity for the mean patient group, which by definition also has mean LYoL-costs. However, the marginal effect of specific disease groups may be biased upwards or downwards. Secondly, we disregard the possibility that when patients die at the beginning of the year, not all costs in the last year of life fall into the same year. In stable demographic conditions, this effect can be disregarded, but when mortality is decreasing, this would overcorrect for reverse causality, biasing the marginal effect upwards. Lastly, correcting for the costs in the last year of life could cause censoring bias, as patient groups with negative spending (patient groups with low spending, high mortality and lower-than-average LYoL-costs) cannot be log-transformed. In our analysis, this occurs in less than 1% of patient groups. Nevertheless, future research should take this into account using data sampling and correction methods (Greene, 2005).
In summary, analysis of potential biases is inconclusive on whether the model is overestimating or underestimating the true effect. That not all benefits are included fully
85