Estimation of census month prevalence ratesEstimation of prevalence in the 2010 national psychosis survey was by means of a two-phase survey. This involves the use of a brief and easy to administer screening instrument at the first phase. Participants are differentially sampled for a detailed and more accurate interview at the second phase based on screening status.28 Typically, a much larger proportion of screen positive participants is interviewed than of screen negative participants.
For prevalence estimation, data from screen negative participants are critical. Unless it can be assumed that the screen has perfect sensitivity and thus the prevalence of the condition of interest in screen negatives is zero, ignoring this group will lead to the under-estimation of prevalence. Conversely, ignoring the relative sampling frequencies of screen positive and screen negative participants will lead to overestimation, as second phase interviews are enriched with those more likely to meet diagnostic criteria.
There are a number of methods in use for prevalence estimation in two-phase surveys.28, 36 The application of sampling weights derived from phase 1 to phase 2 data is the most widely used method and was used in this report. This method is known as Horvitz-Thompson inverse probability weighting.36
Participants were classified according to sex and age strata and screen status within sites. Phase 2 sampling weights were calculated according to the ratio of the number of members of the census population in each 'cell' relative to the number interviewed. For example, a sampling weight of 10 implies that each phase 2 interviewee in a particular stratum with a particular screening status at a site represents 10 comparable members of the phase 1 census. The phase 2 sample was designed to recruit equal numbers of men and women and equal numbers into younger (18-34) and older (35-64) strata. Weights were based on actual strata frequencies – empirical weights – rather according to design weights as this has been demonstrated to yield results that are more accurate.
At some sites, no screen negative individuals in particular strata were interviewed leading to an undefined weight. These members of the census were not 'represented' by anyone at phase 2. In these circumstances, the stratum was combined with the same sex age-adjacent stratum or strata at that site.
Prevalence in the census population within each stratum at each site was estimated as the weighted proportion of persons meeting diagnostic criteria. The statistical package Stata/IC version 10.1 was used for estimation. As the number of cells in the design was large and small numbers could lead to unstable individual estimates, a logistic model was fitted to the data with diagnostic status predicted by site, sex and age group. This yielded results highly consistent with the approach using cell-based proportions.
Top of page
Estimation of one-month population prevalenceFrom the estimated revalence and size of each stratum at each site, the numbers of people meeting criteria in that cell could be calculated. The corresponding resident population was estimated from data provided by the Australian Bureau of Statistics for each catchment. Population numbers for 2010 were not available at the time of calculation so growth rates for each stratum in the whole Australian population were applied to 2009 population estimates.
Prevalence for each stratum at each site is the estimated number of people in the census group divided by estimated catchment population of the stratum.
Prevalences from sites were combined by weighting each value by the proportion of the population across all catchments represented by each site. Weighting was applied separately for each stratum. This assumes that the sites are either a random or a representative sample of sites nationally. This yielded the strata specific prevalences reported in table 2-1.
National numbers of people meeting criteria were derived from these prevalence values and the corresponding population size.
Where prevalences are reported aggregating strata over sex or age groups, adjustments were made so that the estimates correspond to the age and, where appropriate, sex distribution of the Australian population aged 18 to 64.
Estimation of 12-month prevalence ratesPhase 1 of the survey included the enumeration of people in each catchment who were screen positive for psychosis and, while not in contact with public specialised mental health services during the census month, had been in contact with these services in the prior 11 months. Individuals in this category were eligible for recruitment to Phase 2, during which the formal diagnosis could be confirmed by the diagnostic interview.
Unlike the one-month prevalence estimates, the enumeration process could not, by definition, yield screen negative individuals and so 11-month prevalence estimates are based only on screen positive individuals and calibrated for false positives. Apart from this difference, prevalence rates and estimated numbers were estimated using the same methods as for one-month values.
The 12-month prevalence rates and estimated numbers reported in table 2-2 aggregate values for the 11 months prior to the census month with the one-month values reported in table 2-1.
Top of page
Estimation of prevalence in non-government organisationsThe relatively smaller numbers of participants ascertained from non-government organisations who were interviewed precluded construction of a separate set of weights for this group.
A set of weights for the combined groups of those presenting at mainstream mental health services and non-government organisations was developed and the logistic model described above was fitted to the interview data with the addition of an indicator of non-government organisation status. Prevalence was estimated for members of the census from non-government organisations using this model and all other statistics were then derived in a manner identical to that used for attendees of mainstream mental health services.
Possible sources of uncertainty and bias in estimatesThere are a number of sources of potential uncertainty or imprecision in the estimates reported. These arise from normal sampling variation and any errors in population estimates. In addition, the Diagnostic Interview for Psychosis – Diagnostic Module (DIP-DM)4, developed for the first psychosis survey in 1997-98, was used to determine the diagnosis of a psychotic illness using formal ICD-10 criteria. While this diagnostic instrument has been validated and translated into eight languages for use internationally for research and clinical purposes, it is possible that a small number of people were misclassified. Further, some people attending public mental health services within a catchment site may have been missed, resulting in an underestimation of prevalence at the site.
Of particular concern are participant refusal and the inability to interview some very ill patients. This may have biased prevalences downward, particularly if refusal was associated with a higher likelihood of meeting diagnostic criteria. In addition, this may also impact on the results particularly in relation to course of illness and functioning.
The aggregate effect of these factors is difficult to estimate. Beyond straightforward sampling variation, factors that might bias estimates can generally be seen to be likely to be negligible or to result in underestimation of the prevalence of psychotic illnesses.