What factors drive heterogeneity of preferences for micro-health insurance in rural Malawi?

Abstract

Investigating the factors that drive differences in preferences for health insurance products among rural populations is a relevant policy issue that has so far received little attention. This study used a discrete choice experiment to explore heterogeneity of preferences for a prospective micro-health insurance (MHI) product in Malawi. Through an extensive qualitative study, six attributes, each associated with three levels, were derived and used to construct a D-efficient design. The attributes included unit of enrollment, management structure, health service benefit package, copayment levels, transportation coverage and monthly premium. The experiment was interviewer administered to a stratified random sample of household heads and their spouse(s). Using mixed logit and generalized multinomial logit models, respondent characteristics were interacted with MHI attributes to explore heterogeneity of preferences. The results showed that those in the higher age group (≥55 years) and those from households with higher household expenditure had significantly higher preferences for comprehensive and medium benefit packages than for a basic package. Those from households that incurred any healthcare expenditure within the past 4 weeks had lower preferences for the core family as a unit of enrollment than the individual, and higher preferences for coverage of transport costs. Women and non-micro-finance members had higher preferences for 25% copayment than for 50% copayment. There was evidence of scale heterogeneity signifying that the observed preference variations could have resulted from scale and variance differences, rather than real variations in the taste of respondents. To attract the relatively older and wealthier, prospective MHI should offer comprehensive health services benefit packages. Premium exemptions or subsidies should also be offered to the poor. Lower copayments can provide an incentive for women and non-micro-finance members, whilst coverage of transport costs can also attract those with recent history of incurring out-of-pocket healthcare expenditure to accept MHI.

Discrete choice experiment, health insurance, heterogeneity of preferences, micro-health insurance, rural Malawi

Key Messages

Preferences for micro-health insurance (MHI) among rural residents in Malawi significantly differ if the MHI scheme defines the core nuclear family or extended family as the unit of enrollment, is managed by an external NGO, charges no or 25% copayments, offers a medium or a comprehensive service benefit package, covers all healthcare-related transport costs and based on the level of the insurance premium.

A comprehensive health service benefit package could significantly increase the preferences of the relatively old and rich community residents for MHI; lower copayments can provide a significantly higher incentive for women and non-micro-finance members to accept MHI. Coverage of transport costs by an MHI can attract those who have ever incurred any out-of-pocket healthcare costs, and lower premiums can also positively influence preferences for MHI among the poor and the vulnerable.

In the absence of a real market for MHI in rural Malawi, a significant level of observed heterogeneity in preferences for hypothetical design features emanated from differences in the scale and variance among respondents, rather than real underlying variations in taste across observed respondents’ characteristics.

Introduction

In recent times, health insurance, and specifically micro-health insurance (MHI), has been promoted as one among a pool of existing tools to extend financial protection to low-income households in low- and middle-income countries (LMICs). However, enrollment into and subsequent retention of membership within such health insurance systems have remained very low within sub-Saharan Africa (SSA) ( Atinga et al. , 2015 ; Basaza et al. , 2007 ; De Allegri et al. , 2009 ; Dong et al. , 2009 ). In relation to very poor penetration rates across SSA countries, a number of social and economic theories have been used to explain consumer decisions to enroll or not to enroll in health insurance or to choose a particular insurance plan above another ( Schneider, 2004 ). A synthesis of the theoretical literature in addition to empirical evidence on determinants of enrollment reveals that three broad factors can explain the decision to enroll or not enroll in health insurance or to choose among health insurance plans. These include the characteristics of the decision unit such as the individual, the household or the community; the characteristics or design features of the insurance plan and the service delivery system characteristics (provider-related factors) ( Berki and Ashcraft, 1980 ; Carrin, 2003 ; Schneider, 2004 ).

An earlier conceptual model provided by Berki and Ashcraft (1980) comprehensively explains how these three sets of factors interact to influence the decision to purchase an insurance plan. According to Berki and Ashcraft (1980) , an insurance product is made up of the combination of the features of the insurance plan itself and of the health service delivery system (provider related) within which the insurance operates. Berki and Ashcraft (1980) identified the insurance benefit package, premium and cost-sharing arrangements such as copayment, deductibles, coinsurance and benefit ceilings as the main features of the insurance plan. The features of the delivery system include the quality of health services, the spatial location of providers, patient–provider interaction, provider’s office hours and waiting times ( Berki and Ashcraft, 1980 ).

The fundamental conceptual reasoning behind the Berki and Ashcraft (1980) model is that the decision unit normally evaluates the relative attractiveness of the product with respect to its specific features in order to reach a decision on whether to buy or not to buy the product. In turn, perceptions about the product features are shaped by the beliefs and experiences of the decision unit. Hence, the characteristics of the decision unit do significantly influence preferences and choice of insurance products ( Berki and Ashcraft, 1980 ).

In line with the financial loss (vulnerability) hypothesis, a household’s socioeconomic characteristics directly affect its risk-bearing ability and hence also preferences for health insurance products ( Berki and Ashcraft, 1980 ). Furthermore, in line with the risk perception hypothesis, predisposing factors such as family size, age, chronic and acute conditions, perceived health status and previous health service utilization experience all influence the risk of falling ill and the subsequent utilization of health services and hence preferences for health insurance cover ( Berki and Ashcraft, 1980 ).

The empirical literature on revealed preferences for MHI in SSA, as measured in relation to actual enrollment rates and its determinants, revealed that in Burkina Faso, age was negatively associated with the probability of enrolling in MHI ( Parmar et al. , 2014 ), but was positively associated with dropout ( Dong et al. , 2009 ). In Kenya, females ( Kimani et al. , 2012 ), and in Ghana, female-headed households ( Jehu-Appiah et al. , 2011 ) were found more likely to insure although Dong et al. (2009) found that in Burkina Faso, female-headed households were more likely to drop out of MHI. A household’s socioeconomic status was found to be positively associated with MHI enrollment status in a number of SSA settings ( De Allegri et al. , 2006a ; Evans et al. , 2005 ; Jehu-Appiah et al. , 2012 ; Osei-Akoto, 2003 ; Parmar et al. , 2014 ; Sarpong et al. , 2010 ). In Burkina Faso, having a high number of children was significantly positively associated with MHI enrollment status ( De Allegri et al. , 2006a ), whereas fewer children also increased the probability of dropping out from MHI ( Dong et al. , 2009 ). Those working in the formal sector in Kenya ( Kimani et al. , 2012 ) and members of micro-finance institutions (MFI) such as savings and credit cooperative (SACCO) organizations in Uganda ( Dekker and Wilms, 2009 ) were observed to be more likely to enroll in MHI. Additional characteristics identified as positively associated with enrolment in MHI were as follows: higher education of an individual ( Jütting, 2003 ; Kirigia et al. , 2005 ) or the household head ( Schneider and Diop, 2004 ); ethnicity ( De Allegri et al. , 2006a ; Jütting, 2003 ); residential locational (village) characteristics ( Jütting, 2003 ; Schneider and Diop, 2004 ) such as residing in an affluent residential area ( Kirigia et al. , 2005 ). Dong et al. (2009) also reported that the fewer the illness episodes within the past 3 months, the higher the probability of dropping out from MHI in Burkina Faso. In addition, qualitative research has identified sociocultural factors, such as people’s beliefs, practices and (mis-) understanding of the concept of illness, health risk and solidarity, as important elements influencing perceptions about and demand for health insurance in SSA ( Akazili et al. , 2002 ; Arhinful, 2003 ; De Allegri et al. , 2006b ; Goudge et al. , 2012 ; Sommerfeld et al. , 2002 ; Wiesmann and Jutting, 2000 ).

The evidence described above implies that people with different characteristics are likely to differ in their preferences for an MHI product and its specific insurance attributes. This means that for MHI to reflect the varied preferences of the target population, policy makers ought to offer differentiated products. However, across SSA, MHI schemes mostly offer largely homogenous products ( Rösner et al. , 2012 ; Wang et al. , 2012 ). This may be due to either the limited managerial capacity and higher administrative costs associated with offering complex differentiated products ( De Allegri et al. , 2009 ) or the limited empirical evidence capable of informing policy makers in the design of differentiated products. As presented above, empirical studies in SSA have mainly examined the association between individual and household characteristics and health insurance enrolment status. There remains limited empirical evidence as to how the characteristics of the decision unit and the insurance product features (insurance design and provider related) ( Bonan et al. , 2014 ; Donfouet and Makaudze, 2011 ; Onwujekwe et al. , 2010 ) interact to influence an individual or household health insurance choice and/or enrollment decision. An assessment of this interaction can explain heterogeneity in preferences for MHI design features and hence can provide the much-needed evidence to guide policy towards differentiated packages to meet heterogeneous preferences.

In line with the theoretical framework outlined above, our study postulates that community preferences for MHI are likely to be influenced simultaneously by the interaction of the insurance product attributes and by personal characteristics of the decision unit (respondent). This article therefore aims to examine these interactions in order to explore and explain heterogeneity in preferences for MHI product attributes, using data collected from a discrete choice experiment (DCE).

Methods

Study context

The study was conducted in the Thyolo and Chiradzulu districts of the Southern Region of Malawi, where the Malawian Union of SACCOs (MUSCCO) planned to pilot an MHI scheme.

Malawi currently implements an essential health package (EHP), funded from tax revenue and external donations, which offers health care free of charge at public health facilities ( Bowie and Mwase, 2011 ; Ministry of Health Malawi, 2010 ; Gwatkin et al. , 2007 ). The EHP officially covers the prevention, treatment and/or management of vaccine-preventable diseases, acute respiratory infections (including pneumonia), malaria, tuberculosis, acute diarrhoeal diseases (including cholera); HIV/AIDS and other sexually transmitted infections, schistosomiasis; malnutrition and nutritional deficiencies; common eye, ear and skin conditions; common injuries and emergencies, and cancer and other non-communicable diseases and reproductive health interventions to address adverse maternal/neonatal outcomes ( Bowie and Mwase, 2011 ; Ministry of Health Malawi, 2010 ). The cost of medications, laboratory examinations and other services (including surgical operations and transportation during referral) for the treatment and/or prevention of the above conditions at the local health facilities, public district hospitals and the public wards of teaching hospitals is supposed to be fully covered under the EHP. However, the cost of all services received from private wards of teaching hospitals, private and some mission health facilities, and general transport to health facilities, is not included in the EHP. The EHP is also supposed to be offered free of charge at selected mission facilities under service level agreements with the government ( Chirwa et al. , 2013 ; Ministry of Health Malawi, 2010 ). However, essential gaps in both financial protection and access to health care have been reported consistently within both the public and the private health sectors by a number of studies within Malawi. Empirical evidence suggests that wealthier and urban residents have relatively better access to public healthcare services than poor and rural dwellers ( Gwatkin et al. , 2007 ; Zere et al. , 2007 ). In addition to its limited availability, the quality of public health services in Malawi is reported to be very poor ( Mueller et al. , 2011 ). Shortages of medicines and health personnel, poor infrastructure and equipment, and poor access to transport and emergency services in the public sector as well as unaffordable out-of-pocket charges for medical treatment at private facilities have been repeatedly identified as the main barriers to universal health coverage in previous studies ( Bowie and Mwase, 2011 ; Chibwana et al. , 2009 ; Chirwa et al. , 2013 ; Makaula et al. , 2012 ; Mann, 2006 ; Mueller et al. , 2011 ; Muula and Maseko, 2006 ; Palmer, 2006 ; van den Akker and Lommers, 2011 ; Zere et al. , 2007 ; Abiiro et al. , 2014b ).

No social health insurance or MHI schemes exist in Malawi ( Phiri and Masanjala, 2012 ). Only the Medical Aid Society of Malawi in collaboration with a few employers provide private health insurance to formal sector employees and their households ( Makoka et al. , 2007 ). The premium rates charged by these schemes are very high and not affordable by the poor and by most low-income households employed within the informal sector ( Makoka et al. , 2007 ). Due to this restricted coverage of existing health insurance and the shortfalls in the public health system, private not for-profit institutions, especially MFIs, have initiated plans to provide MHI for their members ( Enarsson and Wirén, 2005 ). At the time of the study, none of these schemes were operational.

The MUSCCO is one of the largest MFIs in Malawi, which planned to launch an MHI scheme for the members of one of its largest community-based (Bvumbwe) SACCOs ( Enarsson and Wirén, 2005 ). The members of the Bvumbwe SACCO are spread mainly across the Thyolo and Chiradzulu districts. The benefit package and all other attributes that will make up a MUSCCO MHI product are yet to be clearly specified.

In 2008, the Chiradzulu and Thyolo districts had a total population of 878 401 (290 946 and 587 455, respectively), representing about 6.7% of the total national population of approximately 15 million ( Malawi National Statistical Office, 2008 ). Apart from the district capitals, the entire Thyolo and Chiradzulu districts are rural. At the time of the study, the districts counted a total of 37 public, 13 faith-based (mainly Christian Health Association of Malawi), and 4 private-for-profit facilities. Each district had two main district hospitals: one owned by the government and one by the Christian Health Association of Malawi ( Chiradzulu District Health Office, 2012 ; Ministry of Health Malawi, 2010 ).

Study design

This study was designed as a DCE. In a DCE, products or interventions (usually hypothetical) are described by their specific attributes (main characteristics), and each attribute is illustrated by a defined number of dimensions called attribute levels ( Louviere et al. , 2011 ). A set of hypothetical choice options are experimentally constructed using the defined attributes and attribute levels ( Johnson et al. , 2013 ; Hensher et al. , 2005 ). Respondents are presented with a set of two or more of these choice options and are asked to state their preferred alternative ( de Bekker-Grob et al. , 2012 ). The respondent’s choice indicates the preference or utility attached to an intervention and its attributes ( Lancsar and Louviere, 2008 ). We followed a widely recommended systematic process to construct and implement our DCE ( Bridges et al. , 2011 ; Coast et al. , 2012 ; Hensher et al. , 2005 ; Louviere et al. , 2010 ; Mangham et al. , 2009 ).

First, we derived broad attributes and attribute levels from a literature review and further shaped them into context-specific attributes using a rigorous qualitative study. A detailed description of the process of deriving attributes and attribute levels for this DCE is documented in Abiiro et al. (2014a ). The qualitative data were collected from 12 focus group discussions with community members and 8 key informant interviews with health workers, selected according to stratified purposive sampling. All data were tape-recorded, transcribed and thematically analyzed using NVivo 9 software.

Ten attributes were identified: premium level, premium collection modalities, premium structure, unit of enrolment, geographical level of pooling, management structure, health services benefit package, transportation coverage, copayment levels and provider network. We directly identified levels for all the attributes from the qualitative transcripts. Only the three most pertinent attribute levels from the perspective of respondents were specified for each attribute. The 10 attributes and their corresponding levels were subsequently scaled down to six, based on the outcome of further discussions with two experts of the DCE methodology and a group of five researchers familiar with Malawi and MHI ( Abiiro et al. , 2014a ). Attributes and/or levels that overarched/overlapped with other attributes were dropped to reduce cognitive inter-attribute correlation; attribute and relevant levels for which a clear preference was directly established in the qualitative study were also dropped to avoid dominance; attributes that were identified to be of secondary importance to respondents were also discarded ( Abiiro et al. , 2014a ). The final list of attributes (and their corresponding levels) comprised unit of enrollment (individual, core nuclear family and entire extended family), management structure (Bvumbwe SACCO, an external NGO and community committee), health service benefit package (basic: drugs only; medium: drugs, lab tests/x-rays and comprehensive: drugs, lab test/x-rays and surgical care), copayment [50% (half), 25% (quarter) and no copayment], transport (no transport, only during referral and emergencies, and all transport: always from home to health facility any time sick), and premium level per person per month irrespective of the unit of enrollment [100 Malawian kwacha (MWK), 300 MWK and 500 MWK] ( Abiiro et al. , 2014a ,c). The term “community committee” indicates a committee made of community members directly selected by the community to manage the scheme, a form of management known to the local communities. Surgical care was consistently reported in our prior qualitative work as one main secondary level service that was poorly accessed under the EHP and hence should be included in the definition of a comprehensive MHI benefit package ( Abiiro et al. , 2014a,b ).

Using the attributes and attribute levels developed from the qualitative study, we constructed a D-efficient design based on prior parameters from a pilot study. A detailed description of the pilot study, the construction of the D-efficient design using the Ngene software and its design properties are documented elsewhere ( Abiiro et al. , 2014c ). In brief, an unlabeled experiment of three choice alternatives, comprising two alternative MHI packages and an opt-out option, was constructed. The final design contained 18 choice sets, experimentally grouped into 3 blocks. Each block comprised six choice sets, differently ordered across blocks. All attributes except premium were effects coded. The opt-out represented no insurance and was associated with zero premium, no transportation costs covered and no attribute levels for all other attributes ( Abiiro et al. , 2014c ).

Sampling and field data collection

The DCE was nested within the second round of an existing panel household survey, comprising both microfinance (SACCO) and non-micro-finance member households. The overall panel survey covered a stratified sample of 1200 households comprising 55.62% micro-finance (MUSCCO-affiliated SACCO) members and 44.37% non-SACCO member households. The SACCO members were taken from a full sample of all loan group members from the Bvumbwe SACCO, and the non-SACCO member households were systematically sampled from the communities where the SACCO members were located using a simple random route systematic sampling procedure, implemented by means of the spinning of a bottle.

The experiment was only administered to a random sub-sample of the panel’s household heads and their spouses ( Abiiro et al. , 2014c ). We determined a theoretical minimum sample size for the DCE based on an S-error estimate of 215.5 from the D-efficient design. This S-error implied that a minimum of 216 respondents per block leading to a total 648 respondents was sufficient for parameter estimations ( ChoiceMetrics, 2012 ). We randomly sampled 504 out of the total panel sample of 1200 households, comprising 816 individuals.

A detailed description of the structure of the DCE questionnaire that was administered in March–May, 2013 has already been documented ( Abiiro et al. , 2014c ). We pre-tested the questionnaire prior to data collection. Trained research assistants manually administered the DCE questionnaire in the local language (Chichewa) using the pictorial images described in Figure 1 , but directly entered the data for all respondents onto the tablet computer. The other sections of the overall household survey were fully administered using tablets.

Figure 1

An example of a choice set set as was administered in the field (English version).

Drawing inferences from the existing theoretical and empirical literature on MHI enrollment in SSA as reported in the Introduction section, we expected a number of respondent characteristics to influence preferences for the specific attributes of MHI, potentially driving heterogeneity of preferences. The survey therefore also collected data on such characteristics as: the individual’s age, sex, education, microfinance (SACCO) membership and health insurance coverage statuses, number of children, form of employment, household’s location, household’s monthly total and health expenditures and chronic illness in the household, among others.

Data analysis

DCEs are grounded in consumer theory of economic rationality, utility maximization and random utility theory ( Hall et al. , 2004 ). The DCE approach is also consistent with neoclassical economics, and Lancaster’s economic theory of value ( Lancaster, 1966 ). The DCE responses were therefore analyzed based on the specification of a random utility model ( Hall et al. , 2004 ). This model assumed that the utilities that respondents would attach to any of the three choice alternatives were determined by the attributes and attribute levels of the alternatives ( Louviere et al. , 2010 ). This implied that if alternative j was chosen, j yielded the maximum utility among the three choice options that were presented to the respondents within a choice situation. The individual i’ s utility associated with option j in choice situation s is therefore illustrated as

$$U_{i}js={\beta }^{\prime }{X}_{i}js+{\epsilon }_{ijs},j=1,2,3;s=1,2,3,4,5,\mathrm{6\dots \dots }$$

(1)

where $X_{ijs}$ is a vector of the relevant attributes of MHI option j , β is a vector of the associated preference (utility) parameters and ${\epsilon }_{ijs}$ is an unobservable random component or error term. Assuming a specific parametric distribution of the random component allows a probabilistic analysis of individual choice behavior.

The probability for individual i choosing alternative j in a choice set s can be written as standard logit formula:

$$L_{is}= \frac{e^{{\beta }^{\text{'}}{X}_{js}}}{{\sum }_{j=1}^2{e}^{{\beta }^{\text{'}}{X}_{js}}}$$

The likelihood function is easily computed within the standard logit framework. Our prior estimation ( Abiiro et al. , 2014c ) using the conditional logit (CL) model revealed a violation of the assumption of independence of irrelevant alternatives (IIA) ( Louviere et al. , 2010 ). Besides, the CL model is only suitable for estimating average preferences and due to the assumption of an asymmetric heterogeneity structure ( Fiebig et al. , 2010 ; Keane, 1997 ), it is less capable of detecting preference heterogeneity, resulting from both observable and unobservable factors ( Hole, 2008 ). The CL model also assumes that the idiosyncratic error term is independently and identically distributed. This is not only practically unrealistic but also unsuitable for investigating scale heterogeneity ( Fiebig et al. , 2010 )

The key feature of a mixed logit model is that it allows for variation in the coefficient values across individual respondents (while remaining constant over choice situations for each respondent). In this way, the model is capable of detecting preference heterogeneity and is estimated based on maximum simulated likelihood or Bayesian procedure ( Hole, 2007 ). The output of a mixed logit model includes mean and SDs of random coefficients along with their respective confidence intervals. Mean coefficients represent the relative utility of each attribute conditional on other attributes while SDs reflect the degree of heterogeneity among respondents.

To explore heterogeneity in preferences for each of the attribute-levels, a basic mixed logit model (MXL 1) was estimated in STATA 12, using 500 random draws ( Hole, 2007 ). A fixed constant (opt-out) parameter and random parameters for all attribute levels were estimated. A normal distribution was assumed for all the random attribute level parameters.

MXL 1:

$$\begin{array}{l}U=\beta 0+\beta 1(\text{enroll as nuclear family})+\beta 2(\text{enroll as extended family})+\beta 3(\text{community management})+\beta 4(\text{NGO management})\\ +\beta 5(\text{comprehensive health\hspace{0.17em}services})+\beta 6(\text{medium health services})+\beta 7(\text{no copayment})+\beta 8(\text{25\% copayment})+\beta 9(\text{all transport})\\ +\beta 10(\text{transport during emergency and referral})+\beta 11\text{\hspace{0.17em}(premium)}+\mathrm{\epsilon \dots \dots \dots \dots }\end{array}$$

(2)

where β0 is the parameter for the opt-out, β1-11 are the parameters for each of the attribute levels and ɛ is the error term.

Those variables that produced significant SDs were considered as the attribute levels associated with significant heterogeneity in preferences ( Hole, 2008 ). In order to investigate the extent to which heterogeneity of preferences was driven by respondent observable characteristics, we extended MXL1 with a series of interaction terms between respondents’ characteristics and attribute levels (MXL2). We included interaction terms for only those respondents’ characteristics ( Table 1 ) that were established from the theoretical and empirical literature to influence MHI enrollment in SSA and were also associated with enough variation in our sample ( Table 3 ). We expected these variables to also influence preferences for specific MHI attribute levels. Based on what has been described in the theoretical and empirical literature on demand for MHI in SSA, Table 2 presents the hypothesized signs for the interaction terms included in the analysis.

Table 1.

Socio economic interaction variables and their coding.

Variable	Dummies	Coding
Sex	Male	0
	Female	1
Age	Younger age (<55 years)	0
	Older age (≥55)	1
Education level attained	Below secondary level	0
	Secondary and above	1
SACCO membership	No	0
	Yes	1
Number of children	Lower children (below median)	0
	Higher children (median and above)	1
Household monthly total non-health expenditure in MWK	Lower (below median)	0
	Higher (median and above)	1
Household health expenditure	No health expenditure	0
	Positive (incurred expenditure)	1
Chronic illness in household	No	0
	Yes	1

Table 2.

Prior expected signs of interaction terms (refer to Table 1 for the reference points of each respondent characteristic).

Interaction between dummy respondent characteristic and attribute/levels	Female	Older (≥55 years)	Education: at least secondary level	SACCO member	Higher no. of children	Higher household expenditure	Positive health expenditure	Chronic illness
Unit of enrollment (reference: Individual)
Core nuclear family	+	+/−	+/−	+/−	+	+	+/−	+/−
Entire extended family	+	+	−	+/−	+	+	+/−	+/−
Management (reference: Bvumbwe SACCO)
An external NGO	+	+	+	−	+	+/−	+/−	+
Community committee	−	+	−	−	+	+/−	+/−	+
Service benefit package (reference: basic)
Medium service package	+	+	+	+/−	−	+	+	+
Comprehensive service package	+	+	+	+/−	−	+	+	+
Copayment (reference: 50% copayment)
Quarter (25%)	+	+	−	−	+	−	+	+
No copayment	+	+	−	−	+	−	+	+
Transport (reference: no transport)
Emergency transport	+	+	+	+/−	+	−	+	+
All transport	+	+	+	+/−	+	−	+	+
Premium level per person per month	−	−	+	+	−	+	+/−	+/−

Table 3.

Respondents’ basic characteristics

Variable	Number of respondents	Number of observations	Percentage
Sex
Male	347	6243	43.08
Female	459	8247	56.92
Age (years)
Younger age (<55)	678	12 189	84.12
Older age (≥55)	128	2301	15.88
Education level attained
Below secondary level	677	12 186	84.10
At least secondary level	128	2304	15.90
SACCO membership
No	632	11 367	78.45
Yes	174	3123	21.55
Number of children
Lower: below the median (3)	272	4893	33.77
Higher: ≥3	534	9597	66.23
Household 4-week total non-health expenditure in MWK
Lower (below median)	403	7245	50
Higher (median and above)	403	7245	50
Household 4-week health expenditure
No health expenditure	468	8415	58.07
Positive expenditure	338	6075	41.93
Reported episodes of chronic illness in household (at least one)
No	441	7929	54.72
Yes	365	6561	45.28

MXL 2:

$$\begin{array}{l}U=\beta 0+\beta 1(\text{enroll as nuclear family})+\beta 2(\text{enroll as extended family})\cdots +\beta 11(\text{premium})\\ +\beta 12\text{(female}*\text{core family)}+\beta 13(\text{female}*\text{NGO management})+\cdots +\beta 59(\text{positive health expenditure}*\text{premium})+\mathrm{\epsilon \dots \dots }\end{array}$$

(3)

where β12-59 are the parameters for the interaction terms that were introduced for the socio-economic variables (dummies) illustrated in Table 1 .

The mixed logit model maintains the CL model assumption that the idiosyncratic error term should be independent and identically distributed ( McFadden, 1973 ). This makes the mixed logit model unable to account for scale heterogeneity ( Fiebig et al. , 2010 ). To account for the potential effect of scale and taste (preference) heterogeneity, a generalized multinomial logit model (G-MNL) that relaxes the above CL assumption ( Fiebig et al. , 2010 ) was estimated in STATA 12, using 500 random draws with gamma restricted to zero. Similar to the procedure and equations used for estimating the mixed logit models, two G-MNL models were estimated: a basic model (G-MNL1 without socioeconomic interactions) and an extended model (G-MNL 2 with socioeconomic interaction terms). Fixed parameters were estimated for the constant (opt-out) in both models and the interaction terms in G-MNL 2, whereas random parameters were estimated for the attribute levels in both models. A normal distribution was assumed for all random parameters. All parameters except that for the opt-out were scaled. The G-MNL 2 model, however, contained more interactions terms than the MXL 2 because there were more significant SDs in G-MNL 1 than in MXL 1, as shown in Table 4 . The G-MNL 2 therefore produced more parameters than MXL 2 as shown in Table 5 .

Table 4

. Results of the basic mixed logit and multinomial logit models

	MXL 1				G-MNL 1
	Mean coefficient values		SD		Mean coefficient values		SD
Attribute and attribute levels	β	P	β	P	β	P	β	P
Opt-out	−1.2246	0.000	–	–	−1.6519	0.000
Unit of enrollment (reference: individual)
Core nuclear family	0.17413	0.000	0.3547	0.000	0.0859	0.031	−1.3876	0.000
Entire extended family	0.1243	0.001	−0.1842	0.052	−0.1377	0.149	0.0669	0.497
Management (reference: Bvumbwe SACCO)
An external NGO	0.0339	0.295	0.1841	0.078	0.2777	0.012	−0.0620	0.555
Community committee	−0.0735	0.039	−0.1489	0.245	−0.1797	0.088	0.0526	0.420
Service benefit package (reference: basic)
Medium service package	0.3521	0.000	0.3379	0.000	0.9827	0.000	0.8046	0.000
Comprehensive service package	0.5278	0.000	0.4424	0.000	1.6023	0.000	1.7257	0.000
Copayment (reference: 50% copayment)
Quarter (25%)	0.1126	0.004	0.1551	0.086	0.1260	0.264	−0.6278	0.000
No copayment	0.1537	0.001	0.5148	0.000	0.1015	0.360	0.1173	0.172
Transport (reference: no transport)
Emergency transport	0.2047	0.000	−0.1457	0.188	0.4558	0.002	−0.7040	0.000
All transport	0.5967	0.000	0.4537	0.000	2.0877	0.000	1.4436	0.000
Premium level per person per month	−0.00081	0.000	−0.0032	0.000	−0.0021	0.000	0.0098	0.000
Number of observations	14 490				14 490
Number of respondents	806				806
Number of parameters	12				12
Wald χ 2	609.1				424.71
P > χ 2	0.0000				0.000
Pseudo- R2 (McFadden)	0.3704				0.3680
Akaike Information Criterion (AIC)	6928.61				6920.67
Log pseudo-likelihood	−3452.3071				−3448.3329

Table 5

. Exploring heterogeneity using expanded mixed logit and Generalized Multinomial logit models

	MXL 2				G-MNL 2
	Mean coefficient values		Standard deviations		Mean coefficient values		Standard deviations
Attribute and attribute levels	β	P	β	P	β	P	β	P
Opt-out	−1.2315	0.000	–	–	−1.6161	0.000	–	–
Unit of enrollment (reference: individual)
Core nuclear family	0.2553	0.004	0.3483	0.000	0.2733	0.016	0.9094	0.000
Entire extended family	0.1238	0.001	−0.1637	0.109	0.0393	0.577	−0.3579	0.000
Management (reference: Bvumbwe SACCO)
An external NGO	0.0348	0.277	0.1561	0.235	0.2136	0.034	−0.1466	0.039
Community committee	−0.0738	0.038	−0.15053	0.226	−0.1977	0.051	−0.1435	0.094
Service benefit package (reference: basic)
Medium service package	0.3553	0.001	0.2958	0.000	0.4243	0.004	0.5192	0.000
Comprehensive service package	0.3139	0.002	0.4261	0.000	0.4260	0.015	0.8639	0.000
Copayment (reference: 50% copayment)
Quarter (25%)	0.1203	0.002	0.1669	0.072	−0.0093	0.921	0.4394	0.001
No copayment	0.0306	0.747	0.4930	0.000	0.3006	0.003	0.8140	0.000
Transport (reference: no transport)
Emergency transport	0.2037	0.000	−0.1553	0.153	−0.2918	0.001	−0.1204	0.145
All transport	0.5844	0.000	0.0684	0.000	0.8539	0.000	1.0596	0.000
Premium level per person per month	−0.00002	0.963	−0.0032	0.000	0.00228	0.011	−0.0073	0.000
Significant interaction terms
Female × quarter co-payment					0.1379	0.009
Older age × comprehensive service package					0.2388	0.008
Older age × medium service package					0.2061	0.042
Higher household monthly expenditure × comprehensive service package	0.2461	0.002			0.3426	0.000
Higher household monthly expenditure × medium service package	0.2666	0.001			0.3584	0.000
Higher household expenditure × premium					−0.0007	0.024
SACCO × quarter co-payment					−0.1396	0.041
Higher children × medium service package	−0.1891	0.029
Positive health expenditure × core family	−0.1619	0.011			−0.1184	0.038
Positive health expenditure × all transport					0.2429	0.003
Positive health expenditure × Emergency transport					0.2016	0.013
Positive health expenditure × no copayment	0.1805	0.018
Chronic disease in the household × premium					−0.0008	0.011
Tau	–	–	–	–	2.3871	0.000
Model statistics
Number of observations	14 490				14 490
Number of respondents	806				806
Number of parameters	60 (59)				68
Wald χ 2	648.49				750.47
P > χ 2	0.0000				0.0000
Pseudo- R2 (McFadden)	0.3728				0.3782
Akaike Information Criterion (AIC)	6950.50				6874.17
Log pseudo-likelihood	−3416.2509				−3369.0861

Log likelihood ratio tests were used to compare goodness of fit and to test the extended models against the restricted models with no interactions. The theoretical validity of each of the extended models (MXL 2 and G-MNL 2) was assessed by examining the signs and significance of coefficient estimates of the interaction terms in relation to our priori hypotheses ( Table 2 ).

Results

Respondents’ characteristics

The DCE was successfully administered to 473 households comprising 814 household heads and/or their spouse(s). This figure represents a response rate of 94% of the sampled households. The non-respondents do not reflect individuals who purposely refused to be interviewed. Non-responses were recorded when sampled respondents were not found at home during three repeated visits by the study team or when the entire household had moved outside the study area. Due to their relatively greater familiarity with health insurance and its features, the eight respondents, representing (0.98%), who were already covered by private health insurance, were dropped from the analysis. The 806 respondents who were not covered by any health insurance were retained for the analysis. The mean age of the respondents was ≈40 (SD of ≈13). The majority of the respondents, lived in rural communities −738 (93%) compared with the national average of 85%, and worked in the informal sector −747 (93%) compared with the national average of 89% ( Danish Trade Union Council for International Development Cooperation, 2014 ). The median number of children per respondent was 3 (range, 0–14). The mean household size was 5 (SD 2) people, consistent with national estimates ( Malawi National Statistical Office, 2008 ). The median monthly household expenditure was MWK 15201.7 (range 650–358033.3 MWK) equivalent to about US$39 (1US$ = 386 MWK). The average daily per capita mean household expenditure was 171 MWK, which was quite higher than the 2011 national average of 150 MWK and the 118 MWK average for rural Malawi ( Malawi National Statistical Office, 2012 ) although it was lower than the current minimum wage of 511 MWK ( Danish Trade Union Council for International Development Cooperation, 2014 ). Table 3 provides further details on the respondents.

DCE choice responses and models estimation

Each valid DCE response, a chosen choice set, produced three observations. Because each respondent was presented with six choice sets, 18 observations were associated with each respondent who successfully completed all six choice questions. The 806 respondents produced 4830 valid responses, providing 14 490 observations for analysis (with six missing responses representing 18 observations). A total of 7122 (49.15%) observations were counted for MHI A, 7026 (48.49%) for MHI B and only 342 (2.36%) for the opt-out scenario.

Table 4 presents the results of the basic mixed logit and generalized multinomial logit models (MXL 1 and G-MNL 1) that were used to explore heterogeneity in preferences for the attribute levels. Both models were statistically significant ( P  < 0.001). By looking at the significance and magnitude of SDs in comparison to the mean coefficients, it can be seen that there exists some significant heterogeneity that varies across attribute levels. At the 5% significance level, six attribute levels including core family as a unit of enrollment, comprehensive service package, medium service package, no copayments, all transport, and premium, were associated with significant SDs in the MXL 1 model, reflecting significant heterogeneity within the sample. In the G-MNL1 model, seven attribute levels were associated with significant SDs, reflecting significant heterogeneity. These included five of those that were significant in MXL 1; core family as a unit of enrollment, comprehensive service package, medium service package, all transport, and premium and two additional attribute levels as quarter copayment and emergency transport.

In order to explore this heterogeneity further, we expanded the basic models with interaction terms. Dummy terms for each of the eight respondents’ characteristics presented in Table 1 were interacted with each of the six attribute levels in MXL 1 and the seven in G-MNL 1 that produced significant SDs, to estimate MXL 2 and G-MNL 2, respectively. The results of the log likelihood ratio test ( P  = 0.0107) confirmed that MXL 2 (log likelihood of −3416.2509) had significantly improved explanatory power in comparison to MXL1 (log likelihood of −3452.3071). A log likelihood ratio test ( P  < 0.0001) also established that the G-MNL 2 model (log likelihood of −3369.0861) had a significantly better explanatory power than the G-MNL 1 (log likelihood of −3473.2346). In addition, the co-efficient of the tau for G-MNL1 (2.0726) and G-MNL 2 (2.3871) were persistently statistically significant ( P  = 0.000), revealing strong evidence of scale heterogeneity in both models. The results of the expanded models are illustrated in Table 5 .

The MXL 2 and G-MNL 2 models were both statistically significant ( P  < 0.001). In each of the models, three mean coefficient estimates were statistically insignificant ( P  > 0.05). The insignificant mean parameters were different in the two models. Although management by an external NGO, no copayment and premium, were statistically insignificant in the MXL 2 model, the extended family as a unit of enrollment, management by community committee and quarter payment, were statistically insignificant in the G-MNL 2 model. The opt-out and all other attribute levels had mean coefficient values that were significantly different from zero, indicating that they had a significant influence on respondents’ choice of MHI package. All mean coefficient parameters of MXL 2 were associated with signs in line with our theoretical expectations ( Abiiro et al 2014c ). However, quarter copayment, emergency transport and premium did not have expected signs in the G-MNL 2.

For the sake of space, only those interaction terms that were statistically significant at the 5% significance level are reported in Table 5 . At the 5% significance level, five out of the total of 48 interaction terms in MXL2 and 11 out of 56 in G-MNL 2 were statistically significant. Almost all the significant interaction terms had theoretically valid signs in line with our prior expectations ( Table 2 ). Only the coefficient for the interaction term between higher household expenditure and premium did not have the hypothesized sign. All other interaction terms included in both models were statistically insignificant and hence did not contribute to explain heterogeneity of preferences for attribute levels.

Both models showed evidence of taste heterogeneity as seen by significant SDs for most attributes levels. As G-MNL 2 model showed the presence of significant scale and taste heterogeneity, we exclusively focused the rest of the Results and the Discussion sections on this model. All attribute levels in the G-MNL 2 model except management by community committee and emergency transport were associated with significant heterogeneity. The significant interactions terms are reported in Table 5 and described in the following paragraph.

The results revealed that women had significantly higher preferences for an MHI scheme that charged 25% copayment than 50%. Those in the higher age group (≥55 years) attached more importance to an MHI scheme covering comprehensive and medium service benefits. Also, those from households with higher household expenditure had significantly higher preferences for comprehensive and medium benefit packages, and their preferences for MHI significantly reduced with increasing premiums. Compared with their non-SACCO members, SACCO members had significantly lower preferences for an MHI scheme that charged 25% copayment. Those households that had a positive healthcare expenditure within the past 4 weeks had significantly lower preferences for the core family as the unit of enrollment and higher preferences for coverage of all transport costs including that of emergency transport. The preferences of respondents from households with reported chronic disease for MHI significantly reduced with increased premiums.

Discussion

This study explored and found evidence of significant heterogeneity associated with preferences for some design features of a prospective MHI package in rural Malawi. These included enrolling as a nuclear family or the entire extended family, management by an NGO, medium and comprehensive benefit packages, quarter (25%) copayments, no copayment, coverage of all transport costs and premiums. Management by a community committee and coverage of emergency transport were not statistically associated with significant heterogeneity. This implied that the taste of respondents for the two MHI attributes did not differ with respect to the observed respondent characteristics.

An understanding of the factors that could explain the observed significant heterogeneity is of importance to MHI policy makers. Our study explored and found that few observable respondents’ characteristics could significantly explain some of the heterogeneity in preferences for the attribute levels. These included age, sex, microfinance (SACCO) membership, household non-health (proxy for socioeconomic status) and health expenditures and the presence of chronic disease in the household.

First, in line with theoretical expectations, the findings showed that community preferences for coverage of a comprehensive and medium service packages differed by age group. In comparison to younger respondents, older respondents aged ≥55 years were more likely to choose insurance packages that covered a comprehensive and medium, rather than a basic service package. Our finding is in line with the risk perception hypothesis, confirming that older people perceived themselves to have a relatively higher probability of falling ill with complicated diseases that will require not just basic treatment but very comprehensive treatment including surgical operations and laboratory examinations ( Berki and Ashcraft, 1980 ).

Also, in relation to those from poor households, respondents from less poor households expressed relatively higher preferences for comprehensive and medium service benefit packages, compared with a basic service package. In line with the empirical literature ( De Allegri et al. , 2006a ), our finding suggests that higher socioeconomic status reflects an improved ability to pay when in need of basic, non-expensive services, but is not sufficient to guarantee continued ability to pay when more expensive care is required. In such rural communities, out-of-pocket payments for more comprehensive services could prove catastrophic even for relatively wealthy households. The results, however, suggest that compared with those from lower socioeconomic status households, those from relatively higher socioeconomic status households had reduced preferences for MHI as the premium level increased. Because this group usually has a higher ability to pay, our prior expectation was that they would rather have higher preferences for MHI despite increased premiums. The contrary finding can however be explained in relation to the concept of risk aversion which is central to the expected utility theory ( Schneider, 2004 ). As the premiums increase the probability of enrolling into health insurance among the rich falls because members of this group are often less risk averse due to their relatively higher ability to bear risk ( Schneider, 2004 ).

The evidence that those households that incurred any positive health expenditure within the prior 4 weeks expressed significantly lower preferences for the core family as a unit of enrollment than the individual is a possible indication of potential adverse selection ( Browne, 1992 ; Rajkotia and Frick, 2012 ). It is possible that the respondents from these households would like to have the opportunity to enroll only those individuals at higher risk of incurring healthcare costs. Also, because of the experience, these households had in relation to paying out-of-pocket for health care including cost of transport, within the past month, our study revealed that they were significantly more likely to choose insurance packages that cover all transport costs and emergency transport compared with those that cover no transport cost.

The finding that in relation to men, women were more likely to choose an MHI scheme that charged only 25% copayment compared with the one that charged 50% copayment, can be explained in relation to relative ability to pay. Within such rural communities, women are a vulnerable group that has little ownership over properties including animals and farm lands. Given the unpredictability of illness, they may therefore be unable to raise money to cover higher copayments at the time of seeking care. On the other hand, SACCO members were less likely to reject an MHI that charged higher copayments. It is possible that their participation in this MFI had improved their ability to pay than their non-SACCO counterparts or had put them in a better position to obtain credit to finance higher copayment. Another vulnerable group within rural settings, who often find it difficult to raise money, is those suffering from chronic diseases. It is therefore not surprising that their preferences for MHI reduced with increasing premiums.

This is the first study to examine how socioeconomic factors interact with health insurance policy attributes to influence preferences. Previous studies have only examined how individual socioeconomic status influenced health insurance enrollment in general ( Jehu-Appiah et al. , 2012 ; Osei-Akoto, 2003 ; Sarpong et al. , 2010 ; Parmar et al. , 2014 ), or willingness to pay the insurance premium in particular ( Bonan et al. , 2014 ; Donfouet and Makaudze, 2011 ; Dong et al. , 2003 , 2004 ; Onwujekwe et al. , 2010 ). Existing studies have focused on revealed preferences and were conducted in the presence of homogeneous insurance products. Within SSA, no study, neither experimental nor non-experimental, has examined how socioeconomic factors interact with health insurance policy attributes to influence preferences. More research is therefore required to adequately understand the socioeconomic determinants of the preferences for specific attribute levels.

Despite the significant interactions described above, most interactions tested in both models were not significant. This suggests that there may be other characteristics not included in these models that could better explain the preference heterogeneity found here. An example is proximity to private or public health facilities. Further work is required to identify such characteristics if valid targeted MHI products are to be offered in rural Malawi. Based on the evidence from this study, these variables cannot therefore be recommended as a basis for offering targeted differentiated MHI products within the rural context of Malawi.

Furthermore, it has already been argued that much of even taste heterogeneity within certain choice contexts can be considered as scale heterogeneity ( Fiebig et al. , 2010 ; Louviere et al. , 2008 ). A general alternative explanation for the findings as confirmed by the results of the generalized multinomial logit models is that the heterogeneity associated with certain attribute levels detected in the first mixed logit model and generalized multinomial logit model originated from differences in scale, rather than from actual underlying preference (taste) differences across the socioeconomic characteristics of the study population ( Swait and Louviere, 1993 ). This implies that the scale of the idiosyncratic error term was greater for some consumers than for others ( Fiebig et al. , 2010 ). Some respondents therefore probably paid less attention to the choice task and hence made more random choices than others ( Swait and Louviere, 1993 ).

A detailed discussion of the methodological considerations that potentially threaten the external validity of the study and hence the generalizability of the study findings, including potential hypothetical bias and preference stability, has been extensively documented elsewhere ( Abiiro et al. , 2014a , 2014c ). In addition, specific to the findings discussed in this article, we need to acknowledge that the D-efficient design of this study was based on the specification of a fixed effects CL model. The exploration of heterogeneity of preferences based on random parameters from mixed logit and generalized multinomial logit models although necessitated by the weakness of the CL model in detecting both taste and scale heterogeneity, could, therefore, have affected the optimality of the D-efficient design.

Conclusion

This study found evidence of significant preference heterogeneity for certain attribute levels of MHI. However, the respondent characteristics examined in our study could only explain a portion of the observed heterogeneity. The unique contribution made by our findings rests in their capacity to draw attention to the fact that preferences for MHI products are not homogeneous and that, unlike what has been practiced until now across settings that have implemented MHI schemes, there may be value in offering differential products even among apparently homogenous rural communities. Our work differentiates itself from prior evidence exploring demand for health insurance in LMICs because it focused specifically on a policy concern, i.e. differential preferences for MHI attributes, which cannot be observed in reality given the absence of market offers for differential MHI products. For instance, our findings indicate that while offering a comprehensive health service benefit package could be needed to attract the relatively old and least-poor community residents (even if premium should be higher) into MHI, lower copayments would be needed to reach the same goal for women and non-microfinance (SACCO) members. Coverage of transport costs appeared to be an attractive feature of a potential scheme especially for those who had previously incurred health out-of-pocket expenditure, probably indicating a higher awareness of how direct medical costs do not represent the single obstacle to access in Malawi. In line with the empirical literature indicating lower willingness and ability to pay by the very poor, we also found low premiums to be of fundamental importance to attract the poor and vulnerable into MHI. This last set of findings inevitably calls for the implementation of subsidized premiums for the poorest and most vulnerable to encourage equitable enrollment in the light of striving for universal health coverage. Likewise, encouraging household enrollment through the application of reduced household premiums (i.e. households enrolling all members get a discount) could prove to be an effective measure to avert adverse selection.

Our policy recommendations bring into question both a certain capacity to operate relatively complex MHI schemes offering differential products and the need to rely on external funds to subsidize schemes that would otherwise not be sustainable only through community contributions. Although little has so far been done to test whether MHI schemes would have capacity to offer differential products, several experiences, from Ghana to Rwanda to India, stand to prove the benefit of nationally subsidized schemes in advancing progress towards universal health coverage.

Ethical approval

Ethical approval for the study was obtained from the Ethical Commission of the Faculty of Medicine of the University of Heidelberg in Germany and from the National Health Science Research Committee in Malawi. Permission was also obtained from the District Commissioners, the District Health Officers, and the local authorities of the concerned communities before commencement of data collection. Written informed consent was obtained from all study participants.

Acknowledgements

We would like to thank the Health Financing Group of the Institute of Public Health, University of Heidelberg, and the staff and the field research assistants of Reach Trust, Malawi, for their support in the design and implementation of the study.

Funding

The entire study was funded by the German Research Society (DFG) (Grant no. AL 1361/2-1).

Conflict of interest statement . None declared.