Data were obtained from the National Health Interview Survey (NHIS) for the baseline year of 2007 and from the Medical Expenditure Panel Survey (MEPS) Household Component for years 2008 and 2009 to incorporate the years of the Great Recession. MEPS survey participants are subsampled from NHIS participants from the prior year, and records of individuals were linked between NHIS and MEPS to allow longitudinal cohort analyses for follow-up years 2008 and 2009. Data linkage files were obtained from the Agency for Healthcare Research and Quality [15]. NHIS and MEPS are both conducted through in-person interview in the participant’s home. NHIS is an on-going mutli-stage probability survey nationally representative of the U.S. civilian, noninstitutionalized population [16]. MEPS, which has a panel design including five interviews over two full calendar years, also provides nationally representative estimates [15].

Due to our interest in working-age adults, we restricted our analyses to adults between the ages of 30 and 62 who were interviewed as Sample Adults in NHIS and sampled into MEPS (n=3,277); a baseline maximum age of 62 allowed respondents to “age up” to 64 and prevented observations of employment exit at age 65, a traditional retirement age, which could have been a potential confounder in our sample. Participants between the ages of 18 and 29 were excluded for three reasons: 1) frequent unstable employment and high prevalence of transitory student status in this age group, 2) comparatively low population prevalence of arthritis, and 3) comparability with prior work which examined 30-64 year olds [4].

NHIS data were obtained from the Sample Adult file with linkage to the Person file for relevant variables [16]; data from MEPS were all from the Household Component Panel 13 Longitudinal file [15]. Response rates for NHIS 2007 participants sampled for MEPS and the MEPS 2008 and 2009 follow up periods were 88.1%, 58.8%, and 55.4%, respectively [17]. The MEPS Panel 13 Household Component was specifically subjected to an experiment in which households were randomized into groups receiving incentive payments of various levels after completion of each interview round. Findings showed that even the highest payment level did not change the demographics of participants across subsequent rounds of participation or when compared to a previous Panel [18]. Examined demographics included: education, race, age group, and employment, marital, and self-rated health status, indicating non-differential loss to follow-up within our sample.

All baseline characteristics were ascertained from the 2007 NHIS. Self-Reported, doctor-diagnosed arthritis was defined as a “yes” response to “Have you ever been told by a doctor or other health professional that you have some form of arthritis, rheumatoid arthritis, gout, lupus, or fibromyalgia?” Sociodemographic characteristics were: age (30-44, 45-54, 55-64 years), sex, race/ethnicity, and education (less than high school, high school graduate, some college, college graduate or higher). Additional personal characteristics were: health insurance (yes/no) and marital status (married/cohabitating, divorced/separated/widowed, never married).

Health characteristics examined were comorbidities, Serious Psychological Distress, and key functional limitations. For comorbidities, root question wording was “Have you ever been told by a doctor or other health professional that you had…” for hypertension, stroke, coronary heart disease, angina, heart attack, any kind of heart condition or heart disease (other than queried separately), cancer, emphysema, and diabetes. Non-Melanoma skin cancer and gestational diabetes were excluded. Current asthma was identified by a “yes” to “do you still have asthma,” and chronic bronchitis was identified with a 12 months recall period for a doctor-diagnosis of the condition. The yes/no question “Do you have trouble seeing, even when wearing glasses or contact lenses?” was used to identify vision trouble. All non-arthritis comorbidities were summed to create a condition count representing the number of non-arthritis chronic conditions: none, 1, 2, ≥3. The Kessler-6 (K6) scale was used to identify Serious Psychological Distress (SPD), using six-questions (items: feeling sad, worthless, nervous, restless, hopeless, and that everything was an effort) with a recall period of 30 days. The K-6 was developed to identify and monitor prevalence and trends in nonspecific SPD at the population-level; reported values were summed to create a total score, and scores of ≥13 identified respondents with SPD [19]. In previous work, the clinical and functional importance of limitations affecting the lower extremity, upper extremity, mobility, and endurance were used to inform the creation of a combined variable to capture the most importance capacities available in the range of functional limitation variables in NHIS [20]. Using the same approach here, we created a key functional limitations variable based on a “very difficult” or “can’t do at all” response to any of: “By yourself, and without using any special equipment, how difficult is it for you to…” “lift or carry something as heavy as 10 pounds such as a full bag of groceries” (upper extremity mobility and strength), “stand or be on your feet for about 2 hours” (overall endurance), and “walk a quarter mile or 3 city blocks” (lower extremity mobility).

Employment status was ascertained at baseline from NHIS and classified as working (employed for pay with a job or business, including self-employed) or not working (all other categories). Over the following two years, 2008-2009, employment status was ascertained from MEPS at five interviews. Employment transitions (stopping or starting work) were also collected in MEPS at each interview as well as timing of transitions to identify the month in which any employment transition occurred.

Distributions (weighted proportions with 95% Confidence Intervals (CI)) of baseline characteristics by employment and arthritis status were generated using SAS 9.3 survey procedures to produce standard errors which accounted for the complex sample design of the survey [21]; adjusted sampling weights as described by Cohen were applied to all estimates to generate nationally representative estimates and to permit longitudinal cohort analyses with NHIS as baseline with 2-years follow-up in MEPS with appropriate adjustment of estimation weights [22]. Estimates with a relative standard error (RSE) of ≥30.0 were suppressed. Estimates with an RSE between 20.0 and 29.9 may be unstable and are flagged in tables and footnotes. Statistically significant differences were conservatively defined as non-overlapping 95% CIs [23].

We used PROC LIFETEST in SAS to estimate Kaplan-Meier survival curves separately for our two outcomes: 1) stopping work for those working at baseline, and 2) starting work for those not working at baseline. The 95% CIs were generated in SUDAAN [24] in order to account for the complex sample survey design not accommodated by SAS for this procedure [25].

The proportional hazards assumption was confirmed graphically [25], and we proceeded to estimate risk factors for each outcome using Cox proportional hazards regression models with hazard ratios (HR) and 95% CI using PROC PHREG [25]. We first estimated unadjusted models for each outcome containing only arthritis status and the work outcome. Next, to explore the relationships between our main exposure of interested (arthritis status) and the potential predictors, we generated paired predictor models for each predictor variable for each work outcome containing arthritis status plus one additional predictor variable (Appendix Table).

Before generating fully adjusted models, we examined Spearman’s correlation coefficients between the predictor variables. Absolute values for correlations ranged from virtually none (0.00) to modest (0.41), with the strongest correlation between key functional limitations and self-rated health. All potential predictors had either consistent (age, function) [5, 10, 12, 13, 26, 27] or ambiguous (sex, race/ethnicity, comorbidity, marital status, education, insurance, psychological characteristics, self-rated health) [10, 12, 13, 26, 28] relationships between arthritis and employment in the literature, so we proceeded with a multistaged approach to allow us to consider all potential predictors. Arthritis status was forced into all models. For each outcome separately, we first fit forward selection models (i.e., significant variables were sequentially added; significance in model building was set at α ≤ 0.05) for all potential predictors except key functional limitations (to avoid potential collinearity with self-rated health). We then repeated the process but exchanged self-rated health for key functional limitations. Next, we fit forward stepwise models for both outcomes with both key functional limitations and self-rated health as candidates in the same models. For good measure, we conducted the entire series of model generations again using backwards stepwise selection (i.e., non-significant variables were sequentially removed). All final models were the same regardless of approach used.

For starting work, we also generated an alternate model which excluded both key functional limitations and self-rated health as candidate variables in order to explore any additional potentially modifiable risk factors for starting work. We conducted the entire forward and backward selection model building process again for starting work. Three of these iterative models are presented in the results.

In order to have a comparison with work entry and exit during the Great Recession, we also estimated the unadjusted Cox proportional hazards regression models with hazard ratios (HR) and 95% CI for each work outcome for the time period immediately preceding the Great Recession using the same methods as described above. The baseline year was 2004 from NHIS with the two year follow-up from years 2005 and 2006 in MEPS.

In 2007, the population prevalence of those 30-64 years old with arthritis was 20.1% (95% CI = 18.0-21.4) and without arthritis was 79.9% (95% CI = 78.6-81.2).¹ At baseline, 78.8% of those without arthritis were working (Table 1), while 21.2% were not working (Table 3). Correspondingly, for those with arthritis, 60.8% were working at baseline (Table 1), and 39.2% were not (Table 3).

Employment was substantially and significantly lower among adults with arthritis compared with those without arthritis at all time points (Figure 1). Employment for both groups declined, particularly in mid- to late-2009, dropping 3.3% percentage points for those without arthritis (baseline = 79.9%, 95% CI = 78.0-81.9; December 2009 = 76.6%, 95% CI = 74.7-78.6) and 6.0% percentage points for those with arthritis (baseline = 61.4%, 95% CI = 56.4-66.5; December 2009 = 55.4%, 95% CI = 50.5-60.4). Despite these declines in employment, these losses were not statistically significant for either group.

Descriptions of selected characteristics for the population employed at baseline are shown in Table 1. Those with arthritis were significantly older and more often female and white compared with those without arthritis. Education and marital status distributions were similar between groups with no significant differences. SPD was low prevalence for both groups. People with arthritis reported ≥3 non-arthritis comorbidities five times more often, fair/poor health more than twice as often, and key functional limitations eight times more often than those without arthritis.

Figure 3 shows the time until work loss among both groups for those employed at baseline. Rate of work loss was significantly greater for those with arthritis compared with those without arthritis starting at month 30 (~July 2009) and continuing through the remainder of the study period. By the end of the study period, 16.8% (95% CI = 14.9-18.9) of those without arthritis and 27.0% (95% CI = 21.8-32.9) of those with arthritis stopped work.

In the unadjusted model, arthritis was significantly associated with 70% greater risk of stopping work (HR = 1.7, 95% CI = 1.3-2.2) (Table 2). Arthritis remained significantly associated with increased risk of stopping work in all paired predictor models (Appendix Table). In the fully adjusted model, arthritis was significantly associated with 50% greater risk of stopping work (HR = 1.5, 95% CI = 1.1-2.0) (Table 2). Other significant predictors were older age, 55-65 years (HR = 1.5, 95% CI = 1.1-2.1), Non-Hispanic Black race/ethnicity (HR = 1.4, 95% CI = 1.1-1.8), high school/GED (HR = 1.4, 95% CI = 1.1-1.9), being divorced/separated/widowed (HR = 1.4, 95% CI = 1.1-1.8), and good self-rated health (HR = 1.4, 95% CI = 1.1-1.9). Interestingly, fair/poor self-rated health did not reach statistical significance for stopping work. Having health insurance was protective for work loss (HR = 0.6, 95% CI = 0.5-0.9).

In the time period immediately preceding the Great Recession, there was no difference in work exit between those with and without arthritis: unadjusted HR = 1.3, 95% CI = 1.0-1.8) for those with arthritis compared with those without arthritis as the referent group. Since arthritis was significantly associated with 70% greater risk of stopping work during the Great Recession, a differential effect of the recession on people with arthritis is suggested. (Appendix Figure 1)

Table 3 shows selected characteristics for the population not employed at baseline. Those with arthritis were significantly older, more often white, and more often divorced/separated/widowed. Proportions of men and women were nearly identical for those with and without arthritis. Those with arthritis reported no comorbidities half as often, 2 comorbidities twice as often, ≥3 comorbidities almost three times as often, SPD ≥3 times as often, fair/poor self-rated health more than twice as often, and key functional limitations more than 4 times as often compared with people without arthritis. Only half as many people with arthritis reported no health insurance compared with those without arthritis.

Figure 4 shows the time until work entry among both groups for those not employed at baseline. Rate of work entry was significantly lower for those with arthritis compared with those without arthritis for most of the study period, from month 11 until month 30 (~December 2007-~July 2009). By the end of the study period, 38.7% (95% CI = 33.8-43.8) of those without arthritis and 24.8% (95% CI = 18.6-32.3) of those with arthritis started work, i.e., a significantly lower proportion among those with arthritis entered work.

In the unadjusted model, arthritis was significantly associated with 40% lower chance of starting work (HR = 0.6, 95% CI = 0.4-0.8). Arthritis remained significantly associated with lower chance of entering work in paired predictor models for sex, race/ethnicity, education, SPD, insurance, and marital status, and was not significant in paired predictor models for age, comorbidities, key functional limitations, or self-rated health (Appendix Table). Three iterative models demonstrating snapshots of the model selection process are presented in Table 4. These models start with adjustment for arthritis, all of the predictors from the paired predictor models in which arthritis remained significant (sex, race/ethnicity, education, SPD, insurance, and marital status) and one predictor, comorbidities, from the paired predictor models in which arthritis was not a significant predictor for work entry. In this first model, the HR for the presence of arthritis continued to demonstrate a negative effect on work entry (HR = 0.9, 95% CI=0.6-1.3), but the estimate was no longer statistically significant. In the second of the 3 models presented in Figure 4, model building has proceeded via the selection process described above to result in a model consisting of: arthritis, age, education comorbidities, and health insurance. In this model, the association of arthritis with starting work is null: HR = 1.0, 95% CI = 0.7-1.6 In the final model presented, adjusting for arthritis, age, sex, key functional limitations, and self-rated health, people with arthritis were more likely to start work, but the estimate was not statistically significant (HR = 1.5, 95% CI = 1.0-2.2) (Table 4). Older age, female sex, and fair/poor self-rated health were all associated with between 40% and 70% lower chance of starting work.

In the time period immediately preceding the Great Recession, there was a difference in work entry between those with and without arthritis: unadjusted HR = 0.5, 95% CI = 0.4-0.7 for those with arthritis compared with those without arthritis as the referent group. However, this relationship was almost identical during the Great Recession (HR = 0.6, 95% CI = 0.4-0.8), so no enhanced effect of the Great Recession on starting work among people with arthritis was demonstrated. (Appendix Figure 2.)