This microsimulation is intended to produce a more accurate estimate of the true prevalence of homelessness. However, the indirect approach to estimating the missing homeless potentially introduces new sources of error while amplifying existing ones. As previously discussed, particular assumptions are made about the timing and length of homeless episodes and the prevalence of homelessness among the out-of-scope population, including those who died or were incarcerated or hospitalised in the 12 months prior to the survey.
There are also several implicit assumptions concerning the GSS sampling strategy that may have been violated. For example, the model assumes no currently homeless
0 50 100 150 200 250 300
Total homeless
Family / friends
0 5 10 15 20 25 30
Caravan / mobile home
Boarding house / hostel
Homeless shelter / refuge
Slept rough / squatted
Number homeless per 10,000 people
Ac c o m m o d a ti o n ty p e
individuals have been inadvertently captured in the survey. Filtering the currently homeless out of the survey is probably straight forward for the street homeless though perhaps more difficult for those staying with family or friends or in longer term supported accommodation that is difficult to distinguish from private housing. Failure to exclude these groups will lead to double counting of homelessness in this model. Further, and as with any estimates based on retrospective data, these estimates rely on accurate respondent recall, understanding and interpretation of the survey questions. In this case, inaccuracies may be amplified where, for example, respondents overestimate the length of their most recent homelessness episode. This is because the ๐ values are larger for longer past episodes, meaning that over-reporting of episode length will lead to an overestimate of the missing homeless.
One way to validate the estimates is to compare them against administrative data from Australiaโs homelessness services system. The GSS asks whether respondents sought assistance from housing and homelessness services during their most recent episode:
โDid you seek assistance from services such as these? โข Housing service providers
โข Crisis accommodation/supported accommodation for the homeless (e.g. Shelter, womenโs refuge etc.)โ ABS (2015, p.100)
Other types of services provided in response categories include mental health services, church or community organisations, health services, members of parliament, hospitals and the police. Respondents were allowed to provide multiple responses.
These responses are used to predict the probability of seeking help from housing and homelessness services, ๐โ๐๐๐, by age, sex, episode duration and timing of most recent
episode among those homeless in the last 10 years:
log ( ๐โ๐๐๐ 1 โ ๐โ๐๐๐
) = ๐ฝ0+ ๐ฝ1. ๐๐๐ + ๐ฝ2. ๐๐๐2+ ๐ฝ3. ๐๐๐3+ ๐ฝ4. max(0, ๐๐๐ โ ๐๐)3
Age is again modelled with a cubic spline with knots at 20, 30, 40 and 50 years. Interactions were tested between each set of variables, with only an interaction between age and sex retained in the final model. The resulting probabilities are multiplied by the estimated number of adults homeless in the last 12 months to derive an estimate of the number of homeless people to seek assistance from housing and homelessness providers during the 2013-14 financial year.
Of the annual homeless population, 20 per cent are predicted to have sought assistance from housing and homelessness services in 2013-14. This amounts to 98,200 adults (s.e. 15,200). This prediction is compared to the equivalent figure available in the
Specialist Homelessness Service Collection (SHSC), the national data collection for government-funded homelessness service providers in Australia (AIHW 2018). In 2013-14, 197,400 adults sought or received services. Of these, 93,400 were recorded as either homeless at first or final presentation or at some stage during a support period. There were another 33,900 adults whose homelessness status was not recorded. This may have been because an initial contact was not maintained or followed up. The remainder were considered at risk of homelessness, perhaps because they faced a threat of housing or accommodation loss (for example, an eviction notice or release from institutional care) which was resolved before homelessness occurred. One means of imputing the homelessness of the 33,900 adults with an unrecorded status is to assume that this population has the same rate of homelessness as those who did not receive support services but provided enough information to have a status inferred. This rate is approximately 12 per cent, which gives an estimate of an extra 4,000 homeless adults and 97,300 in total. This is reasonably close to the model prediction (98,200). On face value, this suggests a good degree of accuracy in the model estimates.
Caution is warranted for at least three reasons. Firstly, this comparison understates the true size of the error in the total homeless population. That only one-in-five people are predicted to seek help suggests the prediction error for the total homeless population could
be five times larger than the error among those who seek help. Although, in this case, the error for the total population remains modest (4,100 people), this is sensitive to changes in the parameters and assumptions of the model and the imputation of administrative data. Secondly, the validity of the comparison to administrative data rests on a common definition and measurement of homelessness. The two sources use the same conceptual definition and categories of homelessness. However, differences in measurement no doubt arise given the GSS asks respondents to self-report past experiences with specifically worded questions while the SHSC measurement is based on administrative processing of data collected and reported by service providers and intake systems. Thirdly, the GSS questionnaire asks whether respondents sought help from housing or homelessness services. It is not known how these responses concord with actual presentations to services which are likely to be the product of referrals from other services, outreach in which services providers seek out clients and family and group presentations, in addition to self-referrals. In view of these points, the comparison to administrative data gives the model estimates plausibility as opposed to a high degree of certainty.
Validating the estimates for the different accommodation types is more difficult. One possibility is to compare estimates of shelter and refuge use to administrative data. In 2013-14, 59,400 adults were recorded in the SHSC as having received accommodation support (AIHW 2018), 2.5 times larger than the number predicted by the model to have stayed in homeless shelters and refuges (23,500, s.e. 7,900). A large portion of this difference is likely explained by the broader range of accommodation provided in Australiaโs homelessness services system. While the GSS focuses on accommodation in shelters and refuges, the homelessness services system includes longer-term and semi-permanent accommodation including transitional housing. The model estimates also better conform to Wright and Devineโs (1995) argument that point-in-time ratios of street to sheltered homelessness are typically greater than one than if the shelter/refuge count was indeed substantially higher. On the other hand, the ratio of the average point-in-time rate of shelter/refuge use to the annual rate
predicted by the model (0.36) is lower than previously published estimates (Culhane et al. 1994). This suggests that the model predicts relatively long homelessness episodes among shelter users.
Model results can also be compared against Census counts and the Journeys Home
survey. This is shown in Figure 3.5. To provide a reasonable basis for comparison against the point-in-time Census counts, the model estimates are of those at the time of the GSS. The
Journeys Home estimates are calculated by weighting the number of respondents who were homeless at wave four so that the population estimates are equal to the model prediction of the total number of homeless adults at the time of the GSS by sex, age group and duration of episode. Wave four is chosen because the survey period covered a similar time of year to the GSS (March-May 2013) but early enough that episode durations are not affected by censoring.
Perhaps the most striking finding is that the model predicts a substantially larger prevalence of staying with family and friends than the two most recent Census estimates. Although this may result from model overestimates, it is also likely to reflect a longstanding issue in estimating this form of homelessness in the Australian Census (Chamberlain and Mackenzie 1999). The weighted Journeys Home estimates also suggest it is much more prevalent than the Census suggests, though not as much as in the GSS model. For the other accommodation forms, the model appears to underestimate the other forms. In saying that, note that a) Journeys Home represents a highly disadvantaged population, so will perhaps overestimate these three categories, particularly rough sleeping; b) as with the SHSC, the Census shelter count includes all forms of accommodation support, not just shelters and refuges; and c) unlike in the Census and Journeys Home, GSS respondents are asked to self-report staying in different accommodation types specifically because they did not have a permanent place to live. In other words, people may stay in different accommodation forms without necessarily considering it temporary. Nevertheless, these comparisons suggest a high degree of uncertainty worthy of further research.
Source: authorโs calculations from ABS (2012c, 2015, 2018); Wooden et al. (2012); Scutella et al. (2017)
Figure 3.5 Comparison of GSS model predictions to Census and Journeys Home estimates
Conclusion
Homelessness is difficult to measure. Just as with point-in-time estimates, annual predictions are subject to substantial uncertainty. Estimates of homelessness in different accommodation types has proved particularly challenging. The Jackknife standard errors give a sense of the sampling error, however the potential for non-sampling error appears large. As discussed throughout the chapter, there are a number of possible sources of error, including those related to sampling, sample sizes and recall. Sampling is a particular issue for measuring such a phenomenon that is reasonably rare in the general population, highly episodic and highly elusive, requiring data that is both targeted and able to be generalised. Combining different datasets provides the ability to overcome some of the limitations inherent in each, though this requires uniformity not only in how homelessness is defined but also operationalised and measured. To the extent these were managed in this study, notable shortcomings include the omission of homelessness among children, emigrants who left the country and those who
0 10 20 30 40 50 60 70 Family/friends Caravan/mobile home Boarding house/hostel Homeless shelter/refuge Slept rough/squatted Hom e le s s p e r 1 0 ,0 0 0 a d u lts Accommodation type
Model estimates Weighted Journeys Home
died, were institutionalised or moved out-of-scope before the end of year, as well as in other increasingly recognised categories of homelessness such as household crowding (ABS 2012a). Future data collection in this area ought to consider how to better integrate the design, timing and measurement aspects of point-in-time, household survey and administrative data instruments.
In the meantime, the annual rates estimated in this study are valuable. The comparison to administrative data suggests the magnitudes of these estimates are plausible. Even if not highly precise, they point to certain truths about homelessness that are sometimes suppressed in traditional measures. Firstly, homelessness is more prevalent than typically measured, affecting a larger cross-section of the population. Secondly, homelessness is more diverse in length and severity with a larger population experiencing temporary and episodic homelessness than point-in-time estimates indicate. In this chapter, the effects on estimates by age, sex and accommodation type were reasonably modest, though associations between annual homelessness and personal characteristics and attributes remain a topic for future research. Thirdly, staying with family/friends and in various forms of marginal and sub-market accommodation is as common if not substantially more so than street and sheltered homelessness. While it is a long running topic of debate whether these constitute homelessness (Rossi 1989; Chamberlain and Mackenzie 1992; Koebel and Murray 1999), it is noteworthy in the GSS questionnaire that survey respondents self-report staying in these forms of accommodation specifically because they do โnot have a permanent place to liveโ (ABS 2015, p.97) and for reasons related to previous housing loss and homelessness. Whether considered โhomelessโ or not, interpersonal housing support is evidently a common means through which young adults, in particular, manage a lack of permanent, independent housing.
These points offer theoretical and practical insights into the nature of homelessness. Consideration of the greater prevalence and diversity of homelessness gives rise to the hypothesis that social and economic structures expose a larger population to housing loss and
homelessness than generally understood. The experience for many individuals and families however is relatively transitory, perhaps owing to the personal, interpersonal and institutional resources they utilise to avoid or escape the most severe and long-lasting consequences of homelessness including chronic โrough sleepingโ (Piliavin et al. 1993; Wong and Piliavin 1997; Shinn et al. 1998; OโDonnell 2019). In particular, people with well-developed support networks draw on these supports to a very large extent in managing housing loss, family breakdown and economic crises. For many, this may be a positive mechanism, providing a stepping stone back into stable housing. For others though, it may be a pathway to deeper housing deprivation and homelessness.