Abstract
This paper examines the existence of time trends in the infant mortality rates in a number of countries in the twentieth century. We test for the presence of deterministic trends by adopting a linear model for the log-transformed data. Instead of assuming that the error term is a stationary I(0), or alternatively, a non-stationary I(1) process, we allow for the possibility of fractional integration and hence for a much greater degree of flexibility in the dynamic specification of the series. Indeed, once the linear trend is removed, all series appear to be I(d) with 0<d<1, implying long-range dependence. As expected, the time trend coefficients are significantly negative, although of a different magnitude from those obtained assuming integer orders of differentiation.
Notes
1. Note that in Equations (2) and (3) can be more general than a white noise process and it can be specified in terms of a general stationary linear process.
2. See Robinson [Citation31], Doukham et al. [Citation9] and Gil-Alana and Hualde [Citation15] for recent reviews of I(d) models.
3. The Whittle function is an approximation of the likelihood function.
4. Demetrescu et al. [Citation5] use a similar version of the LM test of Robinson [Citation30]’ although specified in the time domain.
5. Other parametric estimation approaches [Citation2,Citation32] were also employed for the empirical analysis producing very similar results to those obtained using the method of Robinson [Citation30].