Health-related employer support, recurring pain, and direct insurance costs for a self-insured employer

These data are from a restricted-use dataset consisting of one employer with multiple locations provided under a data use agreement with a national health total population health improvement company. The protocol for this study was approved by the UCLA Institutional Review Board (IRB#11–003195); it was not possible to obtain informed consent from each participant as the data were collected for other purposes and de-identified. The survey data came from participants in 2008. Health Insurance claims and eligibility records were obtained from the data provider from 1/1/2007 to 12/31/2009. Participants were linked in the eligibility, claims, and survey files using an encrypted identification number. Participants were included in the sample if they had worked for their employer for at least 12 months prior to the survey and if they were continuously eligible for health insurance (N=1,570).

Future medical expenditures were measured using administrative claims data that covered inpatient, outpatient, mental health and pharmaceutical insurance claims. The survey was given to participants in January 2008, but the exact dates were not recorded. The date of the start of the aggregation of future medical expenditures was chosen to be January 31st. Claims were aggregated for each eligible individual during the 12-month period after the survey. Prior to being aggregated, claims were adjudicated to eliminate duplicate claims. After adjudication, claims amounts that were still negative (indicating potential payment issues) were dropped from the analysis (less than 0.45% of claims). The expenditure information in the data reflects the amounts paid by the self-insured employer.

Pain was measured by participants’ answers to a three-part question about recurring pain in the past 12 months due to a neck or back condition, knee or leg condition, or other. An affirmative answer to any part was used to construct a dichotomous indicator for pain. Health-related employer support was measured using participants’ dichotomous answers to two questions on whether their employer would be supportive of their efforts to positively change their emotional or physical health. Examples of emotional health improvements were reducing stress, balancing work and home life, dealing with financial concerns, and reducing anxiety or depression. Quitting smoking and losing weight were the examples of physical health improvements. Several additional measures of psychosocial work factors were also used in the analysis: dichotomous indicators for learning or doing interesting things on the job, having fun at work, getting to use individual strengths at work, having enough resources to do the job well, job satisfaction, having a supervisor who created a trusting and open environment, and job security. The measure of whether the employee has enough resources to do their job well was dichotomized from a scale of “strongly agreed”, “agreed”, “disagreed”, or “strongly disagreed”; with agreement answers coded as one and disagreement coded as zero. The measure of job security was adapted from employee responses to a question about whether their employer was increasing, decreasing (job insecurity), or maintaining the size of its workforce–entered as a dichotomous variable indicating that the employer was not downsizing. Typical weekly hours were included as a continuous variable. The largest occupational groups were professional workers (33.10%); clerical or office workers (18.08%); manager, executive, or official workers (12.50%); and service workers (8.66%).

Physical and mental health status was measured by dichotomous indicators for asthma, high blood pressure, high cholesterol, and smoking status, self-reported body mass index (continuous), and an emotional health index (continuous). The emotional health index was based on responses to the following ten items that refer to the day before the survey: smiling or laughter, learning or doing something interesting, being treated with respect, enjoyment, happiness, worry, sadness, anger, stress, and diagnosed with depression. An additional variable, the Charlson Comorbidity Index, was derived from participants’ 2007 claims [39]. Because the sample was relatively healthy, an indicator for having a Charlson Comorbidity Index of at least one was included in lieu of the continuous scores.

Monetary resources were measured using a categorical variable for household monthly income. Household size was unavailable, so the number of children in the household under 18 and an indicator for being married or living with a partner were used to approximate household size and additional potential social support and/or increased opportunity cost of time. The type of health insurance plan was used as a proxy for out-of-pocket prices. Preferred Provider Organization (PPO), Health Maintenance Organization (HMO)-gatekeeper, HMO-Open Access, HMO–Point of Service (POS), and Indemnity were the major plan types available to participants (categorical variable). Gender and age were included in the model as measures of participants’ preferences for using healthcare.

The densities of providers and facilities were taken from the Area Health Resource File: the sum of primary care providers, specialty physicians, and psychiatrists per 10,000 people in the participant’s state and the number of hospital beds per 1,000 people in the state [40]. Indicators for the participants’ census regions were proxies for area input prices and practice patterns.

To appropriately model future medical expenditures, a distribution with many zero values and a skewed distribution, the analysis was conducted using a two-part model after applying the tests suggested in the literature. The tests included Manning and Mullahy suggestions for log transformation and a modified Park test to assess the appropriate family of distributions for the generalized linear model [41-44]. The first part was estimated using a logit model where the dependent variable was equal to one if the worker had any future medical expenditures. The second part of the model (to model the amount of expenditures conditional on having greater than zero expenditures) was a generalized linear model with gamma distribution and log link function.

Average marginal effects were calculated for each covariate for each part of the model, and for the unconditional effect (using both parts of the model). In the first part of the model the average marginal effect is the percentage point difference in the probability of having any expenditure. For categorical variables, the change in the probability of having any expenditure as the variable changes from zero to one was calculated for each observation. The estimated effects were then averaged to get the average marginal effect. For the continuous variables, the average marginal effect was the instantaneous rate of change in the probability of having any expenditure. For the second part of the model, the average marginal effect is the change in expected expenditures conditional on having expenditures greater than zero. Again, the effects were calculated for each observation and then averaged to get the average marginal effect. The unconditional effect included both parts of the model: the probability of having any expenditure multiplied by the amount of the expenditure. The average marginal effects for the complete model represent the unconditional change in expected expenditures for specified changes in the covariates. Confidence intervals were calculated using a Taylor Series expansion (STATA version 12) [45]. Missing values, including “don’t know” answers, were imputed 10 times using a multivariate normal model with categories assigned using the method described by Allison [46-48].