Defining and Evaluating Norms for Health and Social Service (HASS) Use

Background The Singaporean demographic transition has increased the proportion of older individuals in the population and the prevalence of multiple chronic health conditions. 1 in 4 Singaporeans aged 40 years and above has at least 1 chronic disease. Worldwide urgency to address chronic conditions is driven by the rapid rise in incidence and also by the associated social and financial costs for the health sector and society.

Often, patients with multiple chronic conditions require more services which are increasingly recognized to include a coordinated mix of clinical and social care. A promising strategy for planning and evaluation of services for an increasingly complex population is population segmentation, where individuals are assigned into groups based on similar health and health-related social needs. Understanding how people in various segments distribute across points of service and in the community more broadly can be used to guide the efficient provision of services. In addition to improved efficiency, meeting otherwise unmet needs would reduce the progression to worse health states and associated high cost medical services such as hospitalization. Through segmentation health care providers, regional health system (RHS) planners, and policy makers will be enabled to develop more person-centric services.

The specific aims and hypotheses of this study are as follows:

Primary Aim 1: To establish norms for high value services for population segments as defined by the SST via a modified-Delphi methodology.

Primary Aim 2: To evaluate the concordance between the types of HASS that patients in each population segment actually utilize versus the types of HASS normatively defined for each population segment over a 3-month period from the point of hospital discharge.

Primary Hypothesis: The concordance between the actual utilization of different types of HASS versus normative HASS is not better than fair.

Secondary Aim: To assess the association between concordance with norm HASS and incidence of adverse outcomes, which includes emergency department, visits, unplanned hospital readmissions and all-cause mortality over a 12-month period from point of hospital discharge for all population segments.

Secondary Hypothesis: Patients with low concordance with norm HASS in their segment will have a higher incidence of adverse outcomes.

To achieve the above aims, the study will be conducted in 3 corresponding phases.

Phase 1: The modified-Delphi methodology will be applied to generate a set of normative HASS for each population segment. This method involves a group of experts who provide individual responses in the questionnaires and re-evaluate their responses subsequently in a group discussion to establish the expert consensus. In addition, this method was chosen for its flexibility in design and is amenable to follow-up interviews, leading to deeper understanding of the research questions.

Population segments will be defined based on the SST in terms of 6 health categories corresponding to the nature of their clinical condition(s) plus a combination of 9 complicating factors which influence the difficulty in managing health conditions and tend to require nursing and social services. Ten specific services based on the proficiency of skills involved will be considered as potentially "high value" from each population segment.

Similar to the RAND approach, each segment will be an "indication," which classifies population in terms of their needs in deciding which services to recommend, and these indications will be grouped for ease of evaluation into "chapters." The research team will identify potential experts from a distribution of relevant disciplines (4 doctors, 4 nurses/allied health professions, and 1 policy maker) in Singapore for the modified-Delphi study. They will be contacted personally to establish their interest to participate in this study, and the process will continue until the required number of panelists in the prescribed distribution are met. Panelists will be provided with a brief description of the method and how it will be applied to this study. The study will consist of an independent round of rating and a group meeting to reconcile the results.

Phase 2: A follow-up study will be conducted to evaluate the concordance between the types of HASS that patients actually utilize 3 months post hospital discharge versus normative HASS defined in Phase 1. Study participants will be recruited from inpatients in the Singapore General Hospital Department of Internal Medicine (SGH DIM) and categorized based on their health care needs using the SST.

The Research Coordinator (RC) will screen patients using the eligibility criteria before inviting them to participate. An Abbreviated Mental Test (AMT) will be administered to determine cognitive capacity to consent. If the patient is deemed unfit, a proxy will then be required to provide informed consent on behalf of the patient.

From past data, approximately 56 patients discharged each day, and 75% are older than 55 years. Assuming a recruitment rate of 30% (12 patients per day), the enrolment period is estimated to be 4 to 5 months.

Baseline data collection: RCs will take the informed consent from the study participants and interview them for socioeconomic and demographic status, health information and prescribed health care services. Study participants will be given a diary to keep track of their health care utilization for 3 months from time of discharge. Doctors who have managed study participants fill out the SST.

Follow-up (3-month) data collection: Study participants will be contacted for follow-up interview by assessors who are blinded to SST categories 3 months after the baseline interview. HASS utilization information over this 3 months period to be obtained by follow-up face-to-face interview, EMR, diary, and the Agency for Integrated Care (AIC) database.

Statistical Analysis Plan The data for aim 2, based on both baseline and follow-up data, can be summarized in a table with (k+1) rows and (k+1) columns. The total number of HASS is k (k=10 for this study). An extra dummy column and an extra dummy row are added to the table to incorporate scenarios where patients do not utilize any high value HASS or utilize HASS that are not deemed high value. Cells in the (k+1)th column represent frequencies of patients who are deemed to be appropriate for some HASS but do not use that service. Similarly, the (k+1)th row represent the cell frequencies corresponding to the patients who use some HASS not deemed high value. So the total number of HASS, for analysis purposes, becomes k+1. nij (i,j=1,2, … k+1) represents the cell frequency (number of patients) corresponding to ith normative HASS and jth actual service: number of patients who have been prescribed ith normative HASS and have used jth actual service. Diagonal elements in the table, nii represent frequencies corresponding to agreement between normative and actual utilization of HASS. Cohen's kappa value is used as the measure concordance between the actual utilization of different types of HASS versus normative HASS.

Therefore, Cohen's kappa can be written as Ka = (p0 - pe)/ (1 - pe), where, overall proportion of observed agreement p0 = (n11 + n22 + …. + nkk + nk+1k+1)/n, is the sum of diagonal entries in the table divided by n; and, overall proportion of chance-expected agreement pe = (n1. x n.1 + n2. x n.2 + … + nk. x n.k + nk+1. x n.k+1)/n2 is the sum of the products of the marginal frequencies divided by n2 .

Based on the primary hypothesis: the concordance between the actual utilization of different types of HASS versus normative HASS is not better than fair; the statistical hypothesis can be written as H0: Ka = 0.41 against H1: Ka < 0.41, where the kappa value of 0.41 denote the lower boundary of the range of moderate concordance. The one sided hypothesis can be tested using the test statistics z = ( K̂a - 0.4)/s.e(K̂a), where K̂a is the estimated value of kappa from the data table and s.e(K̂a) = σ/√n where n is total sample size and s.e denote the standard error. Note that, the statistics z has an asymptotic normal distribution. Based on the data, the null hypothesis is rejected if z < zα and conclude that corresponding concordance is not better than fair, where zα is the (1-α)th quantile of a standard normal distribution.

The proportion of patients using the prescribed services for each of the 10 HASS will also be considered. HASS specific proportion will indicate the measure of agreement within that service.

Sample Size Using the above stated statistical hypothesis and asymptotic normality of the z statistics in a one-sided test with effect size 0.06, σ value as 0.6, type I error as 0.05 and power set to be 0.9, the estimated sample size is 856. Considering 15% dropout during the follow-up, the required sample is 856/0.85 = 1007. To calculate the effect size, the value of kappa was taken to be 0.35 under alternative hypothesis.

Phase 3: The association between normative HASS concordance and incidence of adverse outcomes at 12 months from the day of discharge in each population segment will be assessed by review of information from EMR, National Death Registry, MOH and survey results.

Statistical Analysis Plan In the analysis stage, the effect of the agreement between the normative services and the actual services used by individual patient on the adverse outcomes of that patient will be inferred at each population segment. Separate models for each of the four adverse outcome variables will be considered. As A&E visits and readmission are count variables which can take values 0, 1, 2 …, there may be more patients with '0' value for these two outcome variables. To address this zero-inflated outcome, zero-inflated Poisson regression models will be considered, separately for each of these two outcomes, to see the effect of the agreement between the normative services and the actual services used by individual patient on the adverse outcomes of that patient. The other two adverse outcome variables nursing home placement and death are binary variables 1: yes and 0: no. Two logistic regression models will be used to find out the effect of above mentioned 'agreement' on the two outcome variables separately. In all four regression models, the variable selection procedure will be considered to choose appropriate covariates from the baseline and follow-up data. In each of the four cases, p-values related to the coefficient of the 'agreement' corresponding to the two sided test with null hypothesis that value of the coefficient is zero will be reported, together with confidence intervals of the coefficients. A statistically significant negative value of a coefficient will indicate lower 'agreement' may increase the adverse outcomes.