Using administrative health data to identify individuals with intellectual and developmental disabilities: a comparison of algorithms.
Requiring just two doctor visits in administrative data gives the most accurate IDD prevalence for service planning.
01Research in Context
What this study did
Robertson et al. (2013) tested three ways to spot people with intellectual or developmental disabilities inside big health files. They built computer rules that counted doctor visits, hospital stays, and drug claims. Then they ran each rule on the same data to see who got flagged.
The goal was to learn which rule gave a prevalence number that matched published rates. Accurate counts help states plan budgets and services.
What they found
The rule that required at least two doctor visits, with no date cutoff, hit the sweet spot. It gave a prevalence close to textbook figures and kept subgroup sizes in balance. Looser rules over-counted; tighter rules under-counted.
How this fits with other research
Lotfizadeh et al. (2020) tried the same idea for autism and saw the opposite: billing codes alone mis-classified many kids. The difference is the gold standard. E et al. compared to population rates, while D et al. compared to individual ADOS scores. Both warn that raw codes can fool you, but for different reasons.
Vink et al. (2019) also found big swings in counts when they matched census and service registers in Ireland. Their work backs E et al.'s point that the source and the rule matter just as much as the diagnosis.
Cantwell et al. (2014) went one step further and showed that even after you pick an algorithm, state databases still over-represent adults in 24-hour care. So E et al.'s best rule is only the first filter; you still need to check who is missing or over-sampled.
Why it matters
If you help design waiver programs or write grant budgets, you need a head-count you can defend. Ask your data team for the two-visit rule without a look-back window. Then cross-check the age and residential mix against Joanne et al.'s representativeness table. This one-line change in your SQL query can save months of appeals and re-counts later.
Want CEUs on This Topic?
The ABA Clubhouse has 60+ free CEUs — live every Wednesday. Ethics, supervision & clinical topics.
Join Free →Open your state dataset and re-run counts using the two-visit rule; note how age groups shift.
02At a glance
03Original abstract
BACKGROUND: Individuals with intellectual and developmental disabilities (IDD) experience high rates of physical and mental health problems; yet their health care is often inadequate. Information about their characteristics and health services needs is critical for planning efficient and equitable services. A logical source of such information is administrative health data; however, it can be difficult to identify cases with IDD in these data. The purpose of this study is to evaluate three algorithms for case finding of IDD in health administrative data. METHODS: The three algorithms were created following existing approaches in the literature which ranged between maximising sensitivity versus balancing sensitivity and specificity. The broad algorithm required only one IDD service contact across all available data and time periods, the intermediate algorithm added the restriction of a minimum of two physician visits while the narrow algorithm added a further restriction that the time period be limited to 2006 onward. The resulting three cohorts were compared according to socio-demographic and clinical characteristics. Comparisons on different subgroups for a hypothetical population of 50,000 individuals with IDD were also carried out: this information may be relevant for planning specialised treatment or support programmes. RESULTS: The prevalence rates of IDD per 100 were 0.80, 0.52 and 0.18 for the broad, intermediate and narrow algorithms, respectively. Except for 'percentage with psychiatric co-morbidity', the three cohorts had similar characteristics (standardised differences < 0.1). More stringent thresholds increased the percentage of psychiatric co-morbidity and decreased the percentages of women and urban residents in the identified cohorts (standardised differences = 0.12 to 0.46). More concretely, using the narrow algorithm to indirectly estimate the number of individuals with IDD, a practice not uncommon in planning and policy development, classified nearly 7000 more individuals with psychiatric co-morbidities than using the intermediate algorithm. CONCLUSIONS: The prevalence rate produced by the intermediate algorithm most closely approximated the reported literature rate suggesting the value of imposing a two-physician visit minimum but not restricting the time period covered. While the statistical differences among the algorithms were generally minor, differences in the numbers of individuals in specific population subgroups may be important particularly if they have specific service needs. Health administrative data can be useful for broad-based service planning for individuals with IDD and for population level comparisons around their access and quality of care.
Journal of intellectual disability research : JIDR, 2013 · doi:10.1111/jir.12002