Page 10 - IJPS-5-1

Language and self-assessed health in the U.S

is dichotomous. Models include marital status (married vs. not married) and whether the respondent has health insurance
(measured dichotomously). Income in NHANES is represented by annual family income across 17 categories. These were
recorded into approximate terciles and called low, middle, and high. Missing income was included as a separate category.
Education is coded in four categories: Less than high school, high school, some college, and completed college.

2.3 Statistical Analysis

The multivariate analysis employs a multinomial logistic regression model that treats each category of SAH as separate
and unordered. Doing so avoids collapsing meaningfully different responses and allows for the possibility that responses
may not be ordered along a scale with equidistant intervals – a likely outcome of inaccurate translation. In the multinomial
model, each category of the outcome variable is compared against a single baseline. SAH is made up of five categories
(poor, fair, good, very good, and excellent). The category excellent is chosen as the constant baseline, and the regression
procedure compares the likelihood of being in any of the other four categories relative to the baseline, which is relative to
excellent. Results are presented as log-odds ratios which center on zero such that negative coefficients indicate a negative
association or lower likelihood of being in that category relative to the baseline, and positive coefficients indicate a
positive association or higher likelihood. The exponent of the log-odds is the odds ratios. Three regression models are
considered. Model 1 examines the association between survey language and SAH. Model 2 adds citizenship status to
assess whether and how it mediates the association between survey language and SAH. Model 3 adds ethnicity. The -2LL
deviation statistics is used to assess whether the added variables improve predictive capacity. To test for acculturation,
a separate analysis, which only includes non-U.S. born respondents (n=6,457), assesses whether time spent in the U.S.
influences relationships between survey language and SAH by including a measure for years in the U.S. In general, we
report three P-value levels of significance; 0.05<P<0.10; 0.01<P<0.05; and P<0.01. All procedures are run using SPSS
version 25.

3. Results

Table 1 provides distributions for all study variables for the total sample and across citizenship status. The sample consists
of about 13% identifying as Hispanic, although about 60% of non-citizens are Hispanic. About 6% of the surveys were
conducted in Spanish, but about half of the non-citizens completed the survey in Spanish.
Table 2 shows SAH distributions across the three variables of greatest interest in this analysis: Citizenship, ethnicity,
and language of the survey. Chi-square statistics are provided to indicate whether differences in SAH are statistically
significant across categories of citizenship status, ethnicity, and survey language. There are substantial differences across
categories of these measures. The greatest contrast is seen with respect to language. While there is little difference
in percent rating their health as poor, almost 3 times as many Spanish (37.2%) versus English (13.0%) speakers rate
their health as fair, and a greater percentage of those answering the survey in Spanish rate their health as good (40.6%
vs. 33.8%). In contrast, far more English speakers rate their health as very good or excellent. Non-citizens are far more
likely to rate their health as fair (23.3%) than either U.S. (13.2%) born or naturalized citizens (17.3%). Furthermore, U.S.
born citizens are somewhat more likely than others to rate their health as excellent or very good. With respect to ethnicity,
Hispanics are far more likely to rate their health as fair in comparison to those in other categories. White respondents rate
their health best overall, with higher percentages in the very good and excellent category.
Multinomial models are presented in Table 3. While full models included a series of control variables (sex, age, marital
status, insurance status, education, income, and year of survey), only the results that are most pertinent to this analysis
are presented for parsimonious reasons: Language of interview, citizenship status, and ethnicity. The SAH category
“excellent” is the contrast or baseline across all models, and therefore, all coefficients must be interpreted relative to the
contrast category “excellent.” Model 1 indicates that in comparison to English speakers, respondents who take the survey
in Spanish are significantly more likely to rate their health unfavorably. The most striking difference is in the category fair,
which Spanish respondents are much more likely to select (β=+1.086; p<0.01) as opposed to excellent. When citizenship
status is added in Model 2, the propensity among Spanish respondents to rate their health as “fair” becomes even more
pronounced (β=+1.262; p<0.01). Yet this model also demonstrates that, when accounting for survey language, non-
citizens and naturalized citizens are more likely than the U.S. born to rate their health favorably.
When ethnicity is added in Model 3, the effect of survey language persists. Spanish speakers, as opposed to English
speakers, are still far more likely to rate their health as fair (β=+1.187; p<0.01). The effect of citizenship status becomes even
more pronounced than in Model 2. Non-citizens and naturalized citizens are significantly more likely to report favorable
health in comparison to U.S. born when ethnicity is taken into consideration, a strong indication that any relationship

4 International Journal of Population Studies | 2019, Volume 5, Issue 1