Using predicted Spanish preference to target bilingual mailings in a mail survey with telephone follow‐up

Abstract

Objective

Spanish‐preferring Medicare beneficiaries are underrepresented in national patient experience surveys. We test a method for improving their representation via higher response rates.

Data Sources/Study Setting

2009‐2010 Medicare CAHPS surveys; Medicare population.

Study Design

We used surname and address to predict Spanish‐language preference for a national sample of 177 139 beneficiaries. We randomized half of the 10 000 non‐Puerto Rico beneficiaries with the highest predicted probabilities of Spanish preference (>10 percent) to bilingual mailings (intervention) and half to standard English‐only mailings (control).

Data Collection

Medicare CAHPS Survey data were collected through mail surveys with telephone follow‐up of nonrespondents.

Principal Findings

Mail response rate was higher for intervention (28.7 percent) than control (23.9 percent) (P < 0.0001); phone response rates among mail nonrespondents were similar in intervention and control arms (15.8 percent vs 15.7 percent, P = 0.90). Targeted bilingual mailings induced 6.5 percent of those who would not have responded to respond by mail and 54.0 percent of those who would have responded in English to respond in Spanish. Beneficiaries with greater Spanish probabilities showed greater increases in response rates, a higher proportion of responses in Spanish, and lower control response rates among.

Conclusions

Targeted bilingual mailing of mixed‐mode surveys using commonly available surname and address information can efficiently increase representation of this underrepresented group.

1. INTRODUCTION

Health researchers and social scientists use surveys with national probability samples to track population health care experiences and disparities between subgroups. As the United States becomes more ethnically and linguistically diverse,1 it is becoming more important to accurately estimate characteristics and experiences of those for whom English is not the preferred language. Surveys may underrepresent those unable to respond in English because of the high cost of administering surveys in multiple languages. Linguistically isolated Spanish‐language‐speaking households had the lowest percentage of response by mail on the American Community Survey.2 Research also suggests that bilingual (Spanish‐English) patients give fewer “don't know” responses to diagnostic surveys if the survey is completed in Spanish instead of English.3, 4

Spanish is the second most used language in the United States; the 2010 American Community Survey estimates that it is the primary language spoken at home for 37 million people, almost half of whom have limited English proficiency.5 Here, we describe and evaluate an effort to efficiently collect information about the health care experiences of Spanish‐preferring Medicare beneficiaries with limited English proficiency (SPLEP). These beneficiaries, who are entitled to Medicare coverage due to age (65 or older) and/or disability, often have linguistic and health literacy barriers and report poorer care than other Medicare beneficiaries; for instance, they are immunized at half the rates of English‐speaking non‐Hispanic Whites.6

The Centers for Medicare & Medicaid Services (CMS) implements the Consumer Assessments of Healthcare Providers and Systems (CAHPS^®) surveys to measure beneficiaries’ perceptions of care and services received through Medicare managed care plans and fee‐for‐service insurance.7, 8 As the largest nationally representative survey regarding the health care experiences of U.S. seniors and disabled adults, the Medicare CAHPS Survey (MCAHPS) provides a unique opportunity to monitor the health care experiences of seniors and disabled adults.

However, SPLEP are underrepresented in MCAHPS and similar surveys because of low response rates.9, 10, 11, 12, 13 While more than half of those in the MCAHPS sample who are listed administratively as non‐Hispanic White complete a survey, less than one‐third of those identified as Hispanic or Asian do.13, 14 Compared to beneficiaries of the same race/ethnicity who answer in English, beneficiaries who prefer non‐English languages (eg, Spanish, Chinese) have lower scores on CAHPS composites.14, 15, 16, 17, 18 Increasing response rates for those who prefer languages other than English may improve the representativeness of seniors particularly vulnerable to poor care experiences.

“Double‐stuffed” mailing of the survey—including both a Spanish and an English copy in the same envelope—increases mailing and printing costs by 35 percent.19 Due to the large total sample size, it was not feasible to send the Spanish‐language 2010 MCAHPS survey to beneficiaries outside of Puerto Rico with the rare exception (<1 percent) of beneficiaries who called a toll‐free number on the bilingual prenotification letter but refused the offer to complete the survey on the telephone in Spanish immediately. Nonrespondents to the survey mailings were followed up by telephone and offered the choice to complete the survey on the phone in English or Spanish. Almost all (>99 percent) responses in Spanish from the U.S. mainland occurred during telephone follow‐up. Increasing Spanish‐language mail responses has the potential to improve the measurement of patient experience both by increasing the representation of underrepresented groups and by making the mix of mail and phone survey modes for Spanish‐preferring respondents, and hence their survey mode effects, more similar to those of English‐preferring respondents.20, 21 Increasing the representation of Spanish speakers is superior to nonresponse weighting for their underrepresentation, in that it results in greater precision and requires fewer assumptions.

In this study, we report on an experiment to evaluate the response rate effects of mailing a bilingual survey package to targeted Medicare beneficiaries when cost makes it infeasible to implement 100 percent bilingual mailing. Targeting is based on an innovative approach to identify beneficiaries with the highest probabilities of being Spanish‐preferring based on information from CMS administrative data, prior MCAHPS survey data, and U.S. Census data.

1.1. CMS administrative data, race/ethnicity, and Spanish‐language preference

Centers for Medicare & Medicaid Services administrative data contain information on beneficiaries’ age, gender, and other limited demographic information, including race/ethnicity. The standard CMS administrative race/ethnicity variable is derived from Social Security Administration (SSA) records. For persons assigned a Social Security number prior to 1980, race/ethnicity categories were limited to “Black,” “White,” or “Other.” Starting in 1980, Social Security forms used race/ethnicity categories that conform to current Office of Management and Budget standards.22 Persons who apply for a replacement Social Security card have an opportunity to update their race/ethnicity reported in the SSA database, and this information is transferred to CMS. Because of how data were collected prior to 1980, CMS administrative records are known to substantially underidentify non‐Black minority groups, including Hispanics.23, 24, 25, 26, 27

In the absence of direct race/ethnicity information, the Institute of Medicine (now the National Academy of Medicine) recommends the use of indirect estimation to monitor health disparities in care as a bridging strategy.28 Surname analyses and geocoding have been used to infer race/ethnicity when it is otherwise unavailable. Surname analyses use dichotomous dictionaries (eg, Hispanic, not Hispanic) to identify Hispanics29, 30, 31, 32; although useful, these surname lists do not fully utilize the information that surnames might convey regarding race/ethnicity because dichotomization loses information about relative probabilities of being Hispanic given one's surname. Geocoding links an individual's address to a Census measure of their neighborhood's racial/ethnic composition and uses that measure as a basis for inferring the individual's race/ethnicity. Because Hispanics tend to live in less‐segregated neighborhoods than some other racial/ethnic groups (eg, Blacks),33, 34 geocoding alone cannot accurately identify members of these minority groups.

One indirect estimation method recommended by the Institute of Medicine28 combines probabilities based on residential racial/ethnic information with probabilities based on the latest surname list, both from U.S. Census data, to produce a vector of probabilities that a person is Hispanic, non‐Hispanic White, Black, Asian/Pacific Islander, or Multiracial.26 This Bayesian Improved Surname and Geocoding (BISG) method achieves a weighted average of 93 percent concordance (area under the curve for predicted probabilities vs observed outcome; 95 percent for Hispanic; 94 percent for Asian/Pacific Islander; 93 percent for Black and White) compared to self‐report.26 The BISG approach has been adapted to improve race/ethnicity information from CMS administrative files.35 In this paper, we adapt that Medicare‐specific version of the BISG method (the Medicare BISG or MBISG) to predict Spanish‐language preference.

2. METHODS AND RESULTS

2.1. Predicting Spanish‐language preference

Because Hispanic ethnicity is the primary predictor of Spanish preference but is not well measured in the original administrative CMS race/ethnicity variable,27 we began by estimating the probability that a beneficiary is Hispanic using the MBISG approach. First, we estimated the population‐level distribution of beneficiary race/ethnicity (Hispanic, non‐Hispanic White, Black, Asian or Pacific Islander [API], American Indian/Alaska Native [AI/AN], or Multiracial) for each combination of Medicare administrative classification and age‐group (18‐34, 35‐54, 55‐64, 65‐74, 75‐84, 85+), as the average of self‐reported race/ethnicity in weighted CAHPS data, collapsing age‐groups where needed to avoid imprecise estimation. These population estimates were matched to beneficiaries by Medicare administrative race/ethnicity and age and used as a vector of six initial racial/ethnic probabilities. Second, an independent set of probabilities of each race/ethnicity for each beneficiary were calculated using only surname and address information.26 Third, these two estimates (from three information sources) were combined using Bayes’ theorem to develop a final, “updated” vector of six racial/ethnic probabilities for each beneficiary.

As noted above, virtually all Spanish responses outside of Puerto Rico (>99 percent) were by telephone because obtaining a paper Spanish survey required the beneficiary to call and request it. Very few people initiate a call, and when they do, efforts are made to convert the inbound request to telephone (Spanish) interviews to increase response rates. Hence, we do not have a direct measure of Spanish‐language preference for all sampled beneficiaries, but only a proxy based on completing the survey in Spanish by telephone when the offer is made to the few who call in or to the many who receive a call during phone follow‐up. We expect that this proxy is positively related to Spanish‐language preference, but underestimates it.

We used weighted logistic regression to predict Spanish completion of the 2010 MCAHPS survey among all U.S. mainland completions (n = 397 639) in the prior year's (2009) MCAHPS survey. The independent variables were as follows: the predicted probabilities of being Black, Hispanic, and API, the SSA‐based race/ethnicity variable, 2007‐2009 county‐level proportions of persons who both speak Spanish at home and have limited English proficiency (from the U.S. Census Bureau's ongoing American Community Survey, geocoded to a beneficiary's county of residence), receipt of a low‐income subsidy (limited to those with income <150 percent of the federal poverty line), beneficiary age and gender, and coverage type. The concordance of this logistic regression was 97 percent (P < 0.001). The most important predictors of Spanish completion (data not shown) were (a) the MBISG predicted probability of being Hispanic (odds ratio [OR] = 1.62 per 10 percent change), county‐level percent who speak Spanish at home (OR = 1.38 per 10 percent change), and eligibility for a low‐income subsidy (OR = 2.91).

Finally, we used the coefficients from the logistic regression model predicting Spanish response to the 2009 MCAHPS survey to calculate individual (proxy) probabilities of Spanish‐language preference among all 2010 MCAHPS beneficiaries for use in the bilingual mailing experiment. For brevity, in what follows we will refer to these predicted probabilities of Spanish telephone completion under 2009 survey protocols as the “predicted probability of Spanish preference.”

2.2. Bilingual mailing experiment sample and survey administration

A randomized bilingual mailing experiment was conducted as part of the 2010 MCAHPS survey. We restricted this experiment to those having fee‐for‐service with a free‐standing prescription drug plan (FFS + PDP), representing approximately 38 percent of Medicare beneficiaries. There were 177 139 FFS + PDP beneficiaries randomly sampled in 2010 for MCAHPS surveys. Among those beneficiaries, the distribution of the predicted probability of Spanish preference is skewed, with a mean of 2.2 percent and median of 0.1 percent. We randomized the 10 000 (5.6 percent) sample members with a predicted probability of Spanish response of at least 10 percent to one of two conditions: a standard mailing arm or a bilingual mailing arm. The two experimental arms had very similar distributions for the probability of Spanish response: The means (SDs) for both the standard mailing arm and bilingual mailing arm were 33.9 percent (13.0 percent) with a median of 33.2 percent.

As noted above, the standard MCAHPS approach (control) has a bilingual (English/Spanish) presurvey notification letter. The prenotification letter provides a telephone number to call in to complete the interview by phone in English or Spanish. Two English‐language mailings of the survey were sent to all beneficiaries in the U.S. mainland. Nonrespondents were then followed up by telephone, with options to complete the survey in English or Spanish.

The bilingual mailing (intervention) involved a double‐stuffed mailing with English and Spanish surveys, but was otherwise the same as the standard mailing with identical bilingual prenotification and follow‐up. Development of the Spanish version of the MCAHPS survey included initial professional translation followed by review by a bilingual committee from the CAHPS Cultural Comparability Team and a professional translator experienced in CAHPS instrument design issues. Studies have found empirical evidence of measurement equivalence for Spanish and English versions of CAHPS surveys.36

2.3. Overall effects of bilingual mailing on response rates

Because the detection of ineligibles might differ by language and mode, and our focus was on increasing total response to the survey, we used a conservative variant of AAPOR RR1 that treated the few ineligibles as nonrespondents. As shown in Table 1, those with a high predicted probability of Spanish preference had markedly lower response rates than other Medicare beneficiaries under the standard mail protocol (35.8 percent vs 57.1 percent, P < 0.0001), especially by mail (23.9 percent vs 46.6 percent, P < 0.0001). The bilingual mailing improved the mail response rate for the targeted beneficiaries by 20.1 percent or 4.8 percentage points from 23.9 percent (22.7 percent‐25.1 percent) to 28.7 percent (27.5 percent‐30.0 percent, P < 0.0001). Response rates by telephone follow‐up for high‐Spanish‐probability beneficiaries were not significantly different for the two mail protocols (11.3 percent bilingual mailing vs 12.0 percent standard, P = 0.30, which translates to 15.8 percent and 15.7 percent of mail nonrespondents, respectively). By calculating the ratio of the 4.2 percent increase (40.0 percent‐35.8 percent) in the overall response rate to the proportion who did not respond to the standard mailing (100−35.8 = 64.2 percent), we estimated that 6.5 percent of persons with high (>10 percent) predicted probability of Spanish response who would not have otherwise responded, responded by mail.

Table 1.

Response mode and language by experiment status, 2010 fee‐for‐service beneficiaries with prescription drug coverage (n = 177 139)

	≥10% Predicted Spanish‐preferring probability Participated in experiment					<10% Predicted Spanish‐preferring probability Did not participate in experiment
	Received bilingual mailing (n = 5000)		Did not receive bilingual mailing (n = 5000)			Did not receive bilingual mailing (N = 167 139)
	N	%	N	%	P‐valueb	N	%	P‐valuec
Resolution of cases
Spanish mail	763	15.3	0	0.0	<0.0001	0	0.0	—
English mail	674	13.5	1194	23.9	<0.0001	77 895	46.6	<0.0001
Spanish phone	436	8.7	455	9.1	0.53	208	0.1	<0.0001
English phone	128	2.6	143	2.9	0.39	17 284	10.3	<0.0001
Nonresponse/Incompletea	2999	60.0	3208	64.2	<0.0001	71 752	42.9	<0.0001
Response ratesa
Overall		40.0		35.8	<0.0001		57.1	<0.0001
Total mail		28.7		23.9	<0.0001		46.6	<0.0001
Total phone		11.3		12.0	0.30		10.5	<0.0001
Phone among mail nonrespondents		15.8		15.7	0.90		19.6	<0.0001
Percentage of responses in Spanish
All responses		59.9		25.4	<0.0001		0.2	<0.0001
Mail responses		53.1		0.0	<0.0001		0.0	—
Phone responses		77.3		76.1	0.63		1.2	<0.0001

^aA small number of ineligibles were pooled with nonrespondents to produce conservative response rates, since it is possible that the detection of ineligibles would differ by language and mode.

^bFisher's exact test P‐value comparing the two experimental groups.

Fisher's exact test P‐value comparing the two groups that did not receive bilingual mailing (participated vs did not participate in experiment).

2.4. Overall effects of bilingual mailing on language of response

With respect to telephone responses, 77.3 percent of bilingual mailing and 76.1 percent of other high‐probability Spanish phone respondents chose to respond in Spanish compared to only 1.2 percent of all other telephone respondents (see Table 1), providing support for the validity of the predictions of Spanish‐language preference. With respect to mail responses, 53.1 percent of responses to the bilingual mailing were in Spanish, compared to none (0.0 percent) in the absence of bilingual mailing (even among those with high predicted probabilities of Spanish preference). By calculating the percentage of beneficiaries who responded in Spanish (40.0 × 59.9 = 24.0 percent in bilingual arm, 9.1 percent in standard arm) and attributing 4.2 percent of the 24.0 percent bilingual arm Spanish response rate to the difference in overall response rates (24.0−4.2 = 19.8 percent), we estimate that 54.0 percent [(19.8–9.1)/19.8] of predicted Spanish respondents who would have responded by English under the standard protocol responded in Spanish instead.

2.5. Stratifying effects on response rate and language of response by level of predicted probabilities of Spanish preference

The sample of beneficiaries targeted for bilingual mailings is likely to consist of a mixture of beneficiaries with a true Spanish‐language preference and false positives‐beneficiaries who truly prefer to respond in English or another language. If our proxy predicted probabilities provide useful information about which potential respondents are most likely to be Spanish‐preferring, we would expect the proportion of those truly preferring Spanish and the percentage of responses that are in Spanish to be highest for respondents with the highest predicted probabilities and lowest for respondents with the lowest predicted probabilities. We might also expect that the response rates in the absence of bilingual mailing would fall as this probability increased (due to more frequent language mismatch) and perhaps that the increase in mail response rate with bilingual mailing would increase as this probability increased due to greater relevance.

We present more granular data on response rates and propensity to respond in Spanish to examine if patterns differ by level of predicted probability group. Table 2 distinguishes three levels: predicted probabilities of 40 percent or more (2939 cases), 25‐39 percent (4637 cases), and 10‐24 percent (2424 cases). Table 2 generally shows the same patterns as in Table 1, but graded by predicted probability of Spanish. As expected, standard protocol response rates were lowest in the group with the highest predicted probability of Spanish preference (32.8 percent overall, P < 0.05 vs all others 20.1 percent by mail, P < 0.005 vs all others). Under bilingual mailing, the percentages of all targeted beneficiaries responding by mail and in Spanish were 18.0 percent, 15.7 percent, and 11.1 percent in the three levels of predicted probability of preferring Spanish. Gains in mail response rates due to bilingual mailing were largest for respondents with >40 percent and 25 percent‐40 percent predicted probability (+6.1 percentage points, P < 0.001, and +5.2 percentage points, P < 0.001, respectively), but nonsignificant for respondents with <25 percent predicted probability (+2.7 percentage points, P = 0.15).

Table 2.

Response mode and language by experimental condition and Spanish‐preferring predicted probabilities, 2010 fee‐for‐service beneficiaries with prescription drug coverage (n = 177 139)

	Received bilingual mailing						Did not receive bilingual mailing
	Predicted probability of Spanish preference						Predicted probability of Spanish preference						P‐value (Received vs Did Not Receive)b
	>40% (n = 1470)		25‐40% (n = 2318)		10‐25% (n = 1212)		>40% (n = 1469)		25‐40% (n = 2319)		10‐25% (n = 1212)		>40%	25‐40%	10‐25%
	N	%	N	%	N	%	N	%	N	%	N	%
Resolution of cases
Spanish mail	265	18.0	363	15.7	135	11.1	0	0.0	0	0.0	0	0.0	<0.0001	<0.0001	<0.0001
English mail	120	8.2	319	13.8	235	19.4	296	20.1	561	24.2	337	27.8	<0.0001	<0.0001	<0.0001
Spanish phone	162	11.0	191	8.2	83	6.8	174	11.8	214	9.2	67	5.5	0.49	0.25	0.21
English phone	14	1.0	50	2.2	64	5.3	12	0.8	72	3.1	59	4.9	0.84	0.05	0.71
Nonresponse/Incompletea	909	61.8	1395	60.2	695	57.3	987	67.2	1472	63.5	749	61.8	0.003	0.02	0.03
Response ratesa
Overall		38.2		39.8		42.7		32.8		36.5		38.2	0.003	0.02	0.03
Total mail		26.2		29.4		30.5		20.1		24.2		27.8	0.0001	<0.0001	0.15
Total phone		12.0		10.4		12.1		12.7		12.3		10.4	0.58	0.04	0.20
Phone among mail nonrespondents		16.2		14.7		17.5		15.9		16.3		14.4	0.82	0.22	0.09
Percentage of responses in Spanish
All responses		76.1		60.0		42.2		36.1		25.3		14.5	<0.0001	<0.0001	<0.0001
Mail responses		68.8		53.2		36.5		0.0		0.0		0.0	<0.0001	<0.0001	<0.0001
Phone responses		92.0		79.3		56.5		93.5		74.8		53.2	0.68	0.25	0.63

Note: P‐values < 0.05 are bolded.

^aA small number of ineligibles were pooled with nonrespondents to produce conservative response rates, since it is possible that the detection of ineligibles would differ by language and mode.

^bFisher's exact test P‐value comparing the two experimental groups within each predicted probability stratum.

As further validation of the predicted probability approach, the percentage of mail responses that were in Spanish increased with the predicted probability of Spanish: 68.8 percent (>40 percent probability of Spanish preference), 53.2 percent (25 percent‐40 percent probability), and 36.5 percent (<25 percent probability), respectively. A similar pattern was seen in the proportion of phone responses that were in Spanish: 92.0 percent (>40 percent probability), 79.3 percent (25 percent‐40 percent probability), and 56.5 percent (<25 percent probability) for the bilingual mailing arm, and 93.5 percent (>40 percent probability), 74.8 percent (25 percent‐40 percent probability), and 53.2 percent (<25 percent probability) for the standard mailing arm (P > 0.05 within all three pairs). There is little evidence that bilingual mailings alter total phone response rates: 12.0 percent (>40 percent probability), 10.4 percent (25 percent‐40 percent probability), and 12.1 percent (<25 percent probability) for bilingual mailing, vs 12.7 percent (>40 percent probability), 12.3 percent (25 percent‐40 percent probability), and 10.4 percent (<25 percent probability) for standard mailing (P < 0.05 only for the 25 percent‐40 percent comparison in favor of standard mailing).

3. CONCLUSIONS

As is the case for many health surveys, the Spanish‐preferring subgroup of MCAHPS surveys has been greatly underrepresented in mail responses and overall. In the absence of a bilingual mailing intervention, the response rate for a group with a high predicted probability of preferring Spanish was 35.8 percent as compared to 57.1 percent for all other 2010 respondents; the response rate for just those truly preferring Spanish may be lower. Thus, this study shows the value of survey procedures to increase the representation of Spanish‐preferring beneficiaries in health surveys. In some cases, budgets may require targeted Spanish mailings when bilingual mailings for all are not affordable.

We showed that our predicted probabilities based on respondent's demographic characteristics, address, and surname drawing on a model predicting Spanish response in the prior year's survey were effective in identifying members of this formerly unidentifiable group. The targeted sample is indeed primarily Spanish‐preferring with a majority responding in Spanish by mail or by telephone when given the option.

The randomized intervention of mailing both Spanish‐ and English‐language surveys vs the standard English‐only mailing protocol increased mail response rates for this traditionally difficult‐to‐reach subgroup. Its results imply that 6.5 percent of persons with high (>10 percent) predicted probability of Spanish response, who would not have otherwise responded, responded by mail. Among those with the highest predicted probability (>40 percent), 8.0 percent responded by mail who would not otherwise have done so. These increases represent a 20 percent relative improvement in mail response rates for targeted beneficiaries compared to response rates using the standard English‐only mailing protocol. In addition, study results suggest that 54.0 percent of predicted Spanish respondents who would have responded in English under the standard protocol responded in Spanish instead (50.2 percent of the most likely [>40 percent] Spanish‐preferring respondents). This shift from English to preferred language may improve data quality if respondents are better able to understand the question and response choices in their preferred language. Additionally, the greater representation of Spanish‐preferring beneficiaries who may have had worse experiences improves the overall and plan‐specific measurement of patient experiences as well as our ability to measure and address disparities involving these beneficiaries. Although not measurable in this experiment, it should be noted that bilingual mailings to those with low predicted probabilities of Spanish preference (<10 percent) may be more inclusive, but less cost‐effective as there is a trade‐off between greater cost and greater total effects at lower cutoffs of predicted probabilities and greater efficiency at higher cutoffs.

The higher proportion of Spanish responses by telephone (>90 percent) than by mail (<70 percent) for the group with the highest predicted probabilities of Spanish preference may reflect limited literacy in both English and Spanish for some of the targeted beneficiaries,6 which may itself place some limits on the potential gains from bilingual mailing and underscore the need for telephone follow‐up of mail nonresponse to ensure broad representation.12, 20, 37

A strength of our approach is that MBISG estimates of Spanish preference for all Medicare beneficiaries are freely available from CMS via a Data Use Agreement. Moreover, with additional effort, our approach could be potentially modified to target beneficiaries who prefer other non‐English languages, such as Chinese. Doing so would require a modification of the MBISG algorithm to identify national origin groups with the Asian Pacific Islander group, which is technically feasible given the distinctiveness of Chinese surnames but it would require additional information from the Census Bureau. A limitation of our analysis is that it was restricted to beneficiaries with fee‐for‐service coverage who were members of prescription drug plans, a group that is less than half of all Medicare beneficiaries, and differs from other Medicare beneficiaries to some extent;38 the pattern of associations between response rate and propensity to respond in Spanish may not hold for other populations or survey versions.

Our approach to predicting the probability of race/ethnicity and language can be used in realms beyond survey design such as targeting public health, educational, or other interventions with language‐specific materials. Similarly, it may be useful in other aspects of survey protocols such as allocating bilingual telephone interviewers to cases for follow‐up. Our approach can also be adapted to making predictions of language‐minority populations even if all the information that we used in the analyses presented here was not available. Future research should test the effects of bilingual mailings on other populations including those younger than the Medicare population examined here.

In summary, we demonstrate that administrative data can be used to accurately predict Spanish‐language preference in an older population and that bilingual mailing to this population based on those predictions produces some increases in response rates. In addition, the bilingual mailing substantially shifts responses to the language preferred by the respondent. These results are notable and have important implications for a vulnerable subgroup that is significantly underrepresented in national health surveys. Furthermore, the usefulness of these methods need not be limited to linguistic subgroups. A similar approach might be taken to improving representation of other subpopulations that can be identified with probability models using administrative and/or geographically‐linked data, and for which some targeted modification of survey methods might improve response.

CONFLICT OF INTEREST

All authors declare no conflict of interest. However, please note that Sarah Gaillot is an employee of the sponsoring agency.

Supporting information

 

Elliott MN, Klein DJ, Kallaur P, et al. Using predicted Spanish preference to target bilingual mailings in a mail survey with telephone follow-up. Health Serv Res. 2019;54:5–12. 10.1111/1475-6773.13088