Using marginal structural modelling to investigate the cumulative effect of an unconditional tax credit on self-rated health

Pega, Frank

INTRODUCTION: Previous studies assessed short-term effects of financial credits on health in high-income countries using conventional regression analysis, but treatment effects may accumulate long-term and these studies may be biased by time-varying confounders affected by prior treatment (CAPTs). This study assesses the cumulative effect of receiving an unconditional tax credit (Family Tax Credit; FTC) on self-rated health (SRH) in adults in New Zealand using a marginal structural model (MSM) to fully adjust for CAPTs.

METHODS: Seven waves of data (2002-2009) were extracted from the Survey of Family, Income and Employment and restricted to a balanced panel of 6,900 working-age parents in families. The exposure was the total number of years in FTC receipt. The outcome was SRH at wave 7. Conventional linear regression analyses were conducted, unadjusted and adjusted for time-invariant and time-varying confounders for some waves. MSM analysis was used to estimate the average causal treatment effect on SRH at wave 7 adjusted for time-invariant confounders and CAPTs.

RESULTS: Most participants (71.6%) did not receive FTC; 5.3-6.8% of participants received FTC over 1-3 years and 1.8-3.6% over 4-7 years. In conventional regression models, the total number of years in FTC receipt was associated with a small decrease in SRH at wave 7 (β -0.04, 95% confidence interval [CI] -0.05 to -0.02), also after (partial) confounder adjustment (β -0.02, 95% CI -0.03 to -0.01). After controlling for CAPTs using MSM, the effect remained small and negative, but was not significant (β -0.03, 95% CI -0.09 to 0.04).

CONCLUSIONS: The total number of years in receipt of an unconditional tax credit had no discernible cumulative effect on SRH in adults in New Zealand. Effect estimates from conventional regression models unadjusted for CAPTs were comparable to that from an MSM adjusted for CAPTs, suggesting little bias from these variables.