Study participants were part of a stratified, random telephone survey of N
4,013 adult Latinos (aged ≥18 years) designed to produce a statistically representative sample of Latinos in the contiguous United States. Respondents were eligible if they self-identified as Hispanic or Latino (including Mexican, Puerto Rican, Cuban, Dominican, Central or South American). Recent US Census figures indicate that 4.8% Latinos reported not having household telephones.8
Telephone interviews were conducted as part of the Pew Hispanic Center/Robert Wood Johnson Foundation Hispanic/Latino Health survey in summer 2007 and had a response rate of 39.5%.9
This response rate is comparable to other important telephone surveys including: the Centers for Disease Control and Prevention’s, Behavioral Risk Factor Surveillance System (BRFSS), which had a 41.2% response rate in 2004: and the California Health Interview Survey (38.5% in 2004). In accordance with American Association for Public Opinion Research standards, post-survey probability weights were used to adjust for disproportionalities, including non-response bias.10
The final sample for analysis consisted of N
3,899 persons, after excluding 114 participants who refused to give their age or other key demographic variables and who had, in general, more refusals during the interview than other participants. Post-stratification adjustment consisted of a minor rescaling of sampling weights to match the distribution of Latinos in the March 2007 supplement of the Current Population Survey by sex, age, nativity, and education.11
Sampling Frame Telephone area codes and exchanges in the contiguous US were divided into four Latino household incidence strata (very high, high, medium, and low) based on estimates of the proportion of Latino households in each exchange as provided by the GENESYS Sampling System (Marketing System Group; m-s-g.com). Using telephone number listings from InfoUSA and other sources, numbers associated with persons with Latino surnames out of these strata and placed into a fifth stratum: the surname strata. The remaining numbers in the four initial strata subsequently became residual strata containing no telephone numbers associated with a known Latino surname. Separate random samples of telephone numbers were drawn from each of these five strata. Sampling rates in each stratum were designed to minimize the estimated design effect given budget constraints.
Main Outcome All survey respondents were administered a series of eight questions about diabetes awareness and knowledge: 1) symptoms (thirst, urinary and visual problems, and fatigue); 2) risk (family history); 3) prevention (weight control); and 4) treatments (availability of effective treatment and possibility of permanent cure). Factor analysis of the eight items showed a good fit (RMSEA=0.029) for a two factor solution: 1) diabetic symptoms (four items) and 2) risk, prevention and treatments (four items). Reliability for the four diabetic symptoms was higher (0.80) than the four risk, prevention and treatments items (0.39). All items were dichotomized with correct responses assigned a value 1 and all other responses (i.e., incorrect and “don’t know”) a value 0. An additive score of all correct answers was subsequently created. The distribution of the total score (range 0-8) was skewed, and therefore the scores were converted to three categorical diabetes knowledge groups (high [7-8 correct], medium [4-6 correct] and low [≤ 3 correct]). Scores for the low diabetes knowledge group were below the 20th percentile of the total sample.
Self-reported healthcare services utilization was the outcome of interest. We defined having a USHC as meeting the following three criteria: 1) “a place you usually go to when you are sick or need advice about your health”; 2) care was delivered at a doctor’s office (i.e., hospital outpatient clinic, health center, HMO or health community clinic) and not emergency care services; and 3) care was received within the previous 12-months. The 12-month criterion was imposed to capture practice general guidelines for respondents at-risk for or with chronic disorders (e.g., diabetes) and other groups (e.g., older adults). In addition, use of a 12-month criterion reduces recall bias of healthcare visits more than one-year old. This more conservative USHC specification had small effects on the number of respondents meeting three versus
the less conservative and more conventional two criteria listed above.12
Healthcare access factors
Modified Andersen model factors associated with healthcare access were included in our statistical analyses. Predisposing factors included ethnicity, nativity (US- or foreign-born), age (continuous measure) and sex. Five Latino ethnic sub groups (Mexicans, Puerto Ricans, Cuban Americans, Central and South Americans, and other Latinos) were also included. Health need was a self-reported diagnosis of diabetes. It is arguable that Latinos, particularly Mexicans and Puerto Ricans, are at higher risk for diabetes than non-Latino Whites and other ethnic/racial groups and therefore have a need
for diabetes education because they are an at-risk group.3
Enabling factors included education, household income, marital status and insurance coverage. Education was divided into five categories based on years of schooling completed (8 or less; 9-11; high school or equivalent; some college or vocational training; and a college degree or higher). Annual household income was broken into five groups ($14,999 or less; $15,000-24,999; $25,000-34,999; $35,000-59,999; and $60,000 or more). We placed participants in one of three marital status groups (married; divorced, separated or widowed; and never married) and, the presence or absence of health insurance coverage (i.e., employee-based, private or government program) was a dichotomous measure.
Procedures designed for the analysis of complex sample survey data in the Stata software package, version 10.1 (College Station, Texas) were used for all analyses. All statistical estimates were weighted to account for unequal probabilities of selection and post-stratification. Design-based variance estimation methods were used to account for the complex sampling design when computing estimated standard errors.13
First, sample estimates describing demographic characteristics were calculated. Next, generalized ordered logit models were run using the gologit2 command to test the parallel odds constraints and capture any possible differential effects of our predictors on our outcome variable.14
The generalized ordered logit program relaxes the proportional odds assumptions and fits partial proportional odds models, allowing the effects of our predictor variables to vary for different levels of diabetes awareness and knowledge where parallel lines assumptions are unwarranted. The gologit2 procedure is equipped to account for the complex sample design. Both variable specific and global Wald tests indicated that there was no evidence of parallel lines assumptions violations. Therefore, ordered logit models using Stata’s ologit command were deemed acceptable. Ordered logit models are more parsimonious and easier to interpret. We used them to compare diabetes awareness and knowledge between USHC groups while accounting for predisposing factors based on the Andersen model of healthcare access.5
Next, health need was included in the model, and then enabling factors were introduced to the ordered logistic regression model.