Relationships Among Home Literacy Behaviors,Materials,Socioeconomic Status,and Early Literacy Outcomes Across 14 Low- and Middle-Income Countries

Introduction

Most first- and second-grade children in developing countries are not fluent readers. In many low- and middle-income countries (LMICs), more than half of children cannot read one word, even after 2 years of schooling (Gove & Cvelich, 2011). Many youth in the age of 15 to 24 years still cannot read; in Chad, Central African Republic, and Niger, for example, less than 40% of youth are literate (UNESCO, 2017). These findings led to an intense focus on early grade reading by international aid agencies, such as the United States Agency for International Development (USAID; 2011) and the United Kingdom’s Department for International Development (DFID; 2018). In its 2011-2015 Education Strategy, USAID set the ambitious goal of ensuring that 100 million children in developing countries would become readers by 2015, work that has continued to date (USAID, 2018).

As a major funder of education projects in developing countries, USAID’s prevailing focus on children learning to read in formal settings—to the near exclusion of home-based learning and early childhood literacy—has had great influence on the research conducted and the evidence available about what works to promote learning (Y.-S. G. Kim, Boyle, Zuilkowski, & Nakamura, 2017). The focus of most reading research and interventions in low-resource settings around the world has therefore been on classrooms: How do teachers teach reading, and what can be done to improve the instructional process? This close focus on schools ignores the fact that successful literacy learning does not occur in a vacuum, nor does it begin in a first-grade classroom, but far earlier in a child’s life and in the larger context of their home and community literacy environments (Farver, Yiyuan, Lonigan, & Eppe, 2013; Hood, Conlon, & Andrews, 2008; McCoy et al., 2018). In the U.S.-centered literature, it is well documented that the home literacy environment (HLE) is related to both language and literacy outcomes in young children (Burgess, Hecht, & Lonigan, 2002; Snow, Burns, & Griffin, 1998; Sylva et al., 2011; Weigel, Martin, & Bennett, 2006). The impact may be even larger in LMICs, where children spend a larger proportion of their time out of school (Dowd, Friedlander, & Guajardo, 2012). In a randomized control trial in Rwanda, a version of Save the Children’s (SC’s) Literacy Boost program that included community involvement activities resulted in better literacy outcomes for children than did a model focused only on teacher training (Friedlander & Goldenberg, 2016).

Although the body of literature examining HLEs in developing countries is growing (Chansa-Kabali, Serpell, & Lyytinen, 2014; Dowd & Pisani, 2013; Friedlander, 2013; Kalia & Reese, 2009), there remains a great deal of variation in how HLEs are conceptualized and measured and, in particular, how HLEs may vary across households in countries with low-literacy levels and poor access to reading materials. In the United States, household socioeconomic status (SES) has been shown to be related to HLEs (Hart & Risley, 1995; Rowe, 2008), yet to date, research in LMICs has not necessarily considered these factors independently. Many studies and international assessments use aspects of HLE—for example, books in the home as a component of Programme for International Student Assessment’s (PISA’s) Index of Economic, Social, and Cultural Status—alongside wealth indices to assess household SES, demonstrating the complexity of defining these constructs (Hatch, 2016). Other studies define HLEs as including a combination of both reading materials (e.g., books) and the behaviors in which caregivers engage to support children’s reading (e.g., reading to the child; Dowd & Pisani, 2013). Understanding the relationships between HLE characteristics, household poverty, and children’s literacy development is critical as the limited available evidence suggests that although SES may predict HLEs and child learning outcomes (Dowd et al., 2018; Hart & Risley, 1995; Rowe, 2008), there can also be a broad range of HLEs for children at similar socioeconomic levels (Tusiime, Friedlander, & Malik, 2014).

The present study therefore addresses questions that are central to promoting broader understandings of reading development in LMICs. In particular, we compare the empirical fit of four conceptual models of children’s HLEs across 14 LMIC sites to generate a more robust, evidence-based definition of HLEs for developing countries. First, we examine a model in which books and reading behaviors are merely additional indicators of SES alongside nonreading-related assets (one latent factor). Second, we examine a model where the HLE (inclusive of reading materials and behaviors) is separate from SES (two latent factors). Third, we test a model in which reading materials variables are combined with household asset indicators, whereas reading behavior indicators remain separate (two latent factors). Fourth, we examine a model that comprised of three separate latent factors—SES, reading behaviors, and reading materials. These four models are described in greater depth later. We then add an early literacy outcome variable to the best-fitting model to examine relationships between the latent contextual variables and children’s literacy abilities. In doing so, we aim to examine whether children’s HLE is a mediator of the well-documented SES–literacy skill relationship. Creating a clearer definition of HLEs and their relationships with SES and literacy outcomes could help inform donors, program managers, and policy makers focused on improving early literacy globally.

Background

Literacy instructional approaches and interventions that focus solely on classrooms miss opportunities to improve children’s skills. A “life-wide” approach that considers the multiple ecological contexts in which children are embedded may better leverage the supports that children have at home and in their communities (Dowd et al., 2017; Dowd, Pisani, Dusabe, & Howell, 2018). For example, children may have more one-on-one reading time with a parent, grandparent, or sibling than with a teacher. This is particularly true in LMIC contexts, where early grade classes are often significantly larger than in the United States and other wealthy countries (UNESCO, 2017). Indeed, research is clear that there are associations between aspects of children’s HLE and early literacy skills in a range of national contexts, including the United States (Burgess et al., 2002; Bus, van Ijzendoorn, & Pellegrini, 1995; S. Kim, Im, & Kwon, 2015; Sénéchal & LeFevre, 2014), Zambia (Lyytinen, Serpell, & Chansa-Kabali, 2014), Ghana (Wolf & McCoy, 2019), Ecuador (Rindermann & Carl, 2017), and China (Zhao, Zhang, Chen, Zhou, & Zuo, 2016).

Although the literature on early literacy in the United States and elsewhere is clear that HLEs play an important role in skill development, there is little agreement on what, exactly, HLE is (Niklas, Nguyen, Cloney, Tayler, & Adams, 2016). Both theoretical and operational definitions vary widely. Researchers have conceptualized HLEs to have as few as one and as many as ten dimensions (Gonzalez et al., 2011). On the simpler end of this spectrum, HLEs are sometimes defined using a single indicator. For example, “The HLE can be defined as shared reading activities in the home” (Schmitt, Simpson, & Friend, 2011, p. 410). Other scholars, in contrast, conceptualize HLEs more as an umbrella of various reading behaviors. As Sawyer and colleagues (2014) explain, “HLE is generally conceived as a multidimensional construct, the dimensions of which relate differentially to various child outcomes” (pp. 67-68). Similarly, Hess and Holloway use a five-dimensional model, including the value attributed to reading and achievement, material availability, joint reading, and verbal skill development (Hess & Holloway, 1984).

Much of the examination of various models of HLEs has been in the context of the U.S. research base (Burgess et al., 2002, for example). It is problematic to rely on this evidence for decision making in LMIC settings because Western-developed models of HLEs pay relatively little—or no—attention to the availability of materials. This exclusively behavioral approach assumes that books are freely available in schools, libraries, and other community settings for children and families to access. In LMICs, however, this is generally not the case. Even access to textbooks is often limited or restricted in schools (Lee & Zuilkowski, 2015). The limited international research available on this topic underlines the importance of including reading materials as a potentially distinct component of the HLE. In Park’s (2008) cross-national study, the number of books in the home was a statistically significant predictor of children’s reading outcomes in all 25 countries examined. Also, the number of books more strongly mediated the effect of parental education on literacy than did reading activities, casting doubt on the international applicability of approaches using shared reading as the sole indicator of HLEs. The few available studies in lower-income countries, including India (Kalia & Reese, 2009) and Zambia (Chansa-Kabali et al., 2014), have generally used materials as core indicators of HLEs. It is possible that the presence of these literacy materials in homes in low-income settings reflects positive household attitudes about the value of reading. Indeed, in places where children often have no books at all (beyond perhaps a school textbook), those materials and the attitudes they convey may be quite powerful.

Beyond materials, HLEs may look different in other ways in contexts of extreme poverty relative to high-income settings. It is clear that family background influences the processes through which children learn to read (Hoff, 2006). Are caregivers literate? Do they have sufficient free time in which to read to children and to model reading behaviors? Can they afford to purchase children’s reading materials? What literacy-supporting practices are common among other caregivers in their social class? Are the household’s living conditions healthy and supportive of children’s development generally? Where reading materials and literate family members are scarce, is the home or the broader community the more relevant unit of investigation for the learning environment? All of these questions are relevant to children’s literacy learning outcomes. Research from LMIC settings suggests that children who live in settings where the HLE is poor—those in homes without books, or those where no one can read—benefit most from reading interventions that address the lack of rich opportunities to learn at home (Dowd & Pisani, 2013). It is therefore possible that the HLE is even more important to literacy development in LMICs than in high-income countries, making it critical to understand what it looks like and how it can be improved in these settings.

The body of research drawn on above is largely based in the United States and other wealthy, Western countries. Although a growing number of literacy-related studies are emerging from LMICs, particularly in the context of donor-funded development projects, these studies are generally highly localized—reporting outcomes in a specific country. The existing global, cross-national assessments that measure literacy, including PISA and PIRLs, include few lower LMICs. Our study aims to fill this gap by using a unique cross-national data set comprised of LMICs in Central America, Asia, the Pacific, and Africa to examine the contributions of SES and various components of HLEs to early literacy outcomes.

Method

We constructed the data set for this cross-national analysis by compiling 18 separate baseline surveys conducted for SC’s Literacy Boost programs in 14 LMICs: Bangladesh, Burundi, Egypt, El Salvador, Ethiopia, Indonesia, Kenya, Malawi, Mozambique, the Philippines, Senegal, South Africa, Tanzania, and Vietnam. Literacy Boost was designed by SC in 2008 to improve children’s basic reading skills by both enhancing classroom reading pedagogy and engaging children, families, and communities in reading activities outside of school. Reflective in-service teacher training sessions delivered in partnership with Ministries of Education at or near school sites promote active application to lesson plans during the sessions. Community activities offer reading materials and practice opportunities and promote community members’ awareness and support of children’s literacy. In 2012, World Vision (WV) joined SC to reach more children with this intervention. Today, Literacy Boost reaches more than 4 million children in 33 countries, often in multiple sites within a country.

Sample

Each of the 18 country-level data sets has between 20 and 86 schools and between 309 and 1,610 children. In each case, the schools included in the data set either were participating in the Literacy Boost intervention or are comparison counterpart schools, often to be phased into the Literacy Boost treatment at the close of an evaluation period. Where possible, the assignment to treatment or comparison was random. Where this was not possible, matching was done on key contextual characteristics as well as common indicators such as SES, language, and urbanicity.

In each school, approximately 20 children—10 boys and 10 girls—were randomly selected to participate in the study. The classroom sample size varied across data sets, with some classroom samples as small as one child and others as large as 40. Across the 14 countries and 18 independent baseline samples, 14,178 children in the first and second grades participated in the baseline studies. Their mean age was 8.4 years at the time of baseline. Given the range of development levels and education systems represented in the data set, there is wide variation across country sites in many of the descriptive variables. Table 1 compares the samples by country site.

OPEN IN VIEWER

Table 1.

Sample Comparison Across Countries.

Country	n	Female child (%)	Age in years (M)	Electricity (%)	Story books in the home (%)	Number of family members who read to the child in the past week (M)	Percent of letters correctly identified (M)
Bangladesh	1,531	49	8.3 (1.1)	68	48	1.6 (1.1)	79 (23)
Burundi	550	51	9.2 (1.6)	7	4	1.6 (1.7)	67 (31)
Egypt	800	50	8.8 (0.7)	n/a	36	2.0 (1.6)	84 (23)
El Salvador	875	51	8.2 (1.0)	n/a	59	1.9 (1.6)	77 (21)
Ethiopia	1,029	48	9.8 (2.2)	12	2	1.6 (1.6)	60 (34)
Indonesia	2,332	49	7.3 (1.1)	73	53	1.5 (1.4)	72 (29)
Kenya	942	49	8.0 (4.8)	6	31	2.2 (1.7)	62 (21)
Malawi	595	49	8.9 (1.6)	6	10	0.9 (1.3)	16 (19)
Mozambique	702	52	10.9 (1.7)	25	<1	0.6 (0.9)	21 (27)
The Philippines	1,610	50	7.4 (1.3)	59	11	1.8 (1.6)	53 (35)
Senegal	817	49	8.1 (0.6)	8	15	1.4 (1.5)	20 (19)
South Africa	798	56	7.9 (0.9)	97	37	2.2 (1.5)	86 (21)
Tanzania	1,035	49	9.1 (1.3)	10	15	0.9 (1.3)	61 (38)
Vietnam	562	48	7.2 (0.6)	91	28	0.8 (1.0)	90 (12)

Note. Standard deviations in parentheses.

Data Analytic Strategy

Our analytic strategy involved two stages. The first stage involved fitting a series of measurement models to determine the ways in which indicators of reading behaviors, reading materials, and assets do or do not represent distinct constructs. We first examined the asset indicators in a separate measurement model, with the goal of generating a simplified model. Initially, we had more than 40 varying asset indicators across the 14 sites and 18 data sets. We dropped asset indicators with factor loadings less than .5 and then confirmed that all countries still had at least three asset indicators among the retained set. The final set of asset indicators is displayed in Appendix. We then replicated this process independently for both reading materials and for reading behaviors. We used natural logs of the reading behavior variables to address nonlinearity.

The second stage involved testing four hypothesized models explaining the relationships between SES, reading materials, and reading behaviors in the data set. These models are shown in Figure 1:

Model 1 was driven by the perspective, held by some academics and practitioners in the field of education in LMICs, that reading materials and behaviors are proxies, alongside assets, for household SES. In this framework, books in the home are simply another indicator of wealth, the same as a refrigerator or a motorbike. As the simplest model, this is our base model.

Model 2 has two latent factors, grouping together reading materials and behaviors but separating them from assets. This is how the HLE is conceptualized in much of the U.S.-centric literature on reading development (Hess & Holloway, 1984).

Model 3 also has two latent variables but combines the reading materials variables with the asset variables as indicators of household SES, leaving the reading behaviors variables separate. This model acknowledges that, in LMICs, reading materials are not freely available. Books—even school books—must generally be purchased. Although poor families may be just as able to engage in behaviors, the cost of reading materials may mean that the presence of books, and so on, is more indicative of SES than of HLEs.

Model 4 contains three latent variables—household assets, reading behaviors, and reading materials. This framework hypothesizes that reading materials are not just an indicator of wealth and that families of varying levels of SES will make different decisions about investing in reading materials. It also assumes that reading behaviors are a separate, though related, factor from materials. Although materials are needed for some behaviors to take place, some other behaviors that are supportive of language development, such as telling stories and singing songs, can be done without materials. We hypothesize that this model will best fit the cross-country data set, allowing assets, reading materials, and reading behaviors to be estimated separately.

OPEN IN VIEWER

Figure 1.

Alternative models of the relationships between SES, reading behaviors, and reading materials. (a) Model 1, (b) Model 2, (c) Model 3, and (d) Model 4.

We fit each model using the MLR (maximum likelihood using robust standard errors) estimator in Mplus. To address missing data, we used a full information maximum likelihood (FIML) approach. Because the samples varied in their overall development levels as well as on other factors, we used fixed effects to control for observable and unobservable effects related to country site. After fitting each model, we used the Akaike information criterion (AIC) and the Bayesian information criterion (BIC) to examine the relative fit of these models. Other fit statistics, such as chi-square and root mean square error approximation (RMSEA), are not available with confirmatory factor analysis models with categorical variables (using MLR). A difference of at least two units in either AIC or BIC has been shown to indicate a statistically significant difference in fit between two models (Burnham & Anderson, 2004).

After selecting the best-fitting model, we extended it by adding in the percentage of letters identified correctly by the child as an observed outcome variable. This model used clustered standard errors at the school level and added a variable controlling for the effect of gender on the outcome. We then examined the direct paths between the latent and observed variables in the model to assess the relations between children’s contextual variables and their literacy outcomes across countries.

Findings

Research Question 1: How Might SES and HLEs Be Operationalized in LMIC Contexts?

To address this research question, we compared four possible models, as discussed earlier. Based on both AIC and BIC, Model 4, with three latent variables, was the best fit for the data (see Table 2). This indicates that household SES, reading materials, and reading behaviors should be measured as three distinct factors in this set of LMICs.

OPEN IN VIEWER

Table 2.

Comparing the Four Hypothesized Structural Models.

Model	Latent variables	Description	Log likelihood	Degrees of freedom	AIC	BIC
1	1	All as one latent variable	–104,051.1	1,001	208,262.1	208,864.9
2	2	SES, reading materials & reading behaviors	–96,904.4	983	194,004.9	194,743.2
3	2	SES & reading materials, reading behaviors	–94,057.3	983	188,310.7	189,049.0
4	3	SES, reading materials, reading behaviors	–91,439.8	964	183,113.6	183,995.1

Note. SES = socioeconomic status; AIC = Akaike information criterion; BIC = Bayesian information criterion.

An examination of the factor loadings for household assets, reading materials, and reading behaviors indicates moderate to strong correlations between the indicators and the latent variables. The standardized loadings for the 18 indicators retained on the SES latent variable were all statistically significant at p < .01 and ranged from .503 to .936. The three highest leadings were for electricity (.936), stove (.932), and television (.921). The three lowest loadings were for car (.503), latrine (.606), and water (.626). Turning to the literacy materials latent variable, the six retained indicators were all statistically significant and had loadings ranging from .568 to .922. The loadings were as follows: posters (.922), magazines (.696), newspapers (.684), religious books (.597), coloring books (.568), and storybooks (.539). Seven indicators were retained for the literacy behaviors latent variable. The loadings for these indicators, all used in a natural log transformation, were statistically significant and ranged from .507 to .824. The number of household members who play games with the child and who can read had lower loadings, whereas the loadings for the number of people the child sees reading (.736) and the number who read to the child (.824) had stronger loadings.

Research Question 2: Do Children’s HLEs Partially Explain the Relation Between Their Household SES and Their Early Literacy Outcomes in LMIC Contexts?

To address our second research question, we used the latent variables defined by Model 4 (above) to consider whether reading materials and behaviors affected the relationships between children’s SES and literacy skills, operationalized using the percentages of letters that children could correctly identify. This allowed us to examine the relative relationships between SES, reading materials, reading behaviors, and early literacy skills. The results are shown in Figure 2 and Table 3.

OPEN IN VIEWER

Figure 2.

Final structural equation model results.

OPEN IN VIEWER

Table 3.

Results for Final Model, Including Loadings on Latent Variables and Relationships Between the Latent and Observed Variables.

Latent variable	Observed variable	Unstandardized coefficient		Standardized coefficient
		b	SE	β
SES	Improved floor	0.924	(0.071)	.696
	Electricity	2.639	(0.282)	.941
	Refrigerator	2.312	(0.155)	.925
	TV	2.317	(0.184)	.925
	Water	0.762	(0.168)	.624
	Bed	2.255	(0.587)	.921
	Stove	2.475	(0.357)	.933
	Latrine	0.729	(0.078)	.608
	Motorbike	1.024	(0.058)	.732
	Farmcart	0.836	(0.286)	.660
	Air conditioning	1.926	(0.280)	.896
	Computer	1.676	(0.182)	.869
	Car	0.560	(0.059)	.506
	Atari/game system	2.031	(0.437)	.905
	Charcoal	2.166	(0.893)	.915
	Pig	0.824	(0.130)	.654
	Sheep	1.233	(0.115)	.791
	Donkey	1.009	(0.174)	.727
Reading materials	Religious	0.442	(0.156)	.595
	Magazines	0.616	(0.755)	.718
	Newspapers	0.575	(0.124)	.693
	Story books	0.385	(0.061)	.541
	Coloring books	0.420	(0.085)	.575
	Posters	1.438	(0.423)	.924
Reading behaviors	n people who can read	0.191	(0.015)	.545
	n people the child sees reading	0.366	(0.007)	.736
	n people who read to child	0.428	(0.006)	.824
	n people who tell stories to child	0.365	(0.007)	.683
	n people who help child study at home	0.333	(0.007)	.680
	n people who play games with child	0.265	(0.010)	.507
	n people who sing to child	0.364	(0.013)	.685
Direct paths	Reading materials on SES	0.867	(0.280)	.543
	Reading behaviors on SES	0.245	(0.021)	.409
	Percent letters on SES	–0.002	(0.061)	–.014
	Percent letters on reading materials	0.066	(0.033)	.580
	Percent letters on reading behaviors	0.003	(0.005)	.009
	Percent letters on female	0.031	(0.005)	.045

Note. Adjusted for cross-data set mean differences on SES, reading materials, and reading behaviors. Standard errors clustered at the school level. SES = socioeconomic status.

SES has a moderately strong, positive correlation with reading materials (β = .543, SE = 0.170, p = .001) and a weaker, though still moderate, positive relationship with reading behaviors (β = .409, SE = 0.033, p < .001). Given that most reading materials need to be purchased, the stronger relation between family financial status and literacy-related materials was expected. The positive relation between SES and reading behaviors was also as expected, as SES and literacy are correlated, indicating that higher SES parents are more likely to be able to read themselves and thus able to read to their children (Blunch & Pörtner, 2004; de Baldini Rocha & Ponczek, 2011; Denny, Harmon, & O’Sullivan, 2003; Reder, 2010). Reading materials were also statistically significantly predictive of the percentage of letters identified by children (β = .580, SE = 0.285, p < .05), suggesting that exposure to print may be important to early literacy skill development across LMICs.

Surprisingly, the coefficient for the direct path between SES and percent of letters identified by children was not statistically significant. This suggests that, after controlling for reading materials and reading behaviors, family SES does not explain additional variation in children’s outcomes. Also contrary to our hypotheses, we did not observe a statistically significant relationship between reading behaviors and children’s literacy outcomes. Gender was also significantly associated with performance on the literacy outcome (β _female =.045, SE = .008, p < .001).

We conducted a series of robustness checks to confirm the pattern of results discussed above. First, we fit the same model within single-country data sets for four country sites—Bangladesh, Egypt, Indonesia, and Tanzania—purposively selected to ensure geographic and income-level diversity. Second, we split the country-level data into various sets using natural break points in the country-level statistics—adult literacy levels less than and greater than 80%, gross domestic production per capita below and above US$3,000, and Human Development Index scores below and greater than .6—and conducted multigroup analyses. Third, we repeated the full group modeling with three alternative outcomes measuring varying levels of literacy development—decoding, reading fluency (correct words per min), and whether the child was a “reader,” using STC’s threshold of 10 words correct per min. In all cases, the findings above—moderate correlations between SES and both reading materials and reading behaviors, but only reading materials being positively associated with the outcome—consistently held.

Discussion

Our analyses demonstrate that, across a diverse set of 18 project sites in 14 LMICs, household assets, reading materials, and reading behaviors can be conceptualized as distinct but interrelated constructs. In particular, these results indicate that books and other reading materials are not merely a sign of wealth in the way that a television or a refrigerator is in a LMIC. The presence of books, posters, newspapers, and so on additionally signals that a family has decided to spend some of their resources on reading materials, which likely has implications for literacy-related practices in the home. Studies investigating children’s literacy development should therefore collect information on reading materials in the home, as knowing about the household’s SES will not necessarily provide the full picture of the learning-related resources available to children.

We also found that the conceptualization of HLEs as one latent factor, including materials and behaviors, does not fit this international data set as well as the model that considers these constructs as distinct. This is particularly true in light of Park’s (2008) analyses, which showed a U-shaped relationship between countries’ economic development levels and the number of books in the home. There are several implications of this result. First, the provision of books does not necessarily indicate that they will be used with children in an interactive way. Second, the separation of reading materials from reading behaviors underlines that families can enhance their literacy practices, even in the absence of a wealth of materials. For example, two of our indicators for the reading behaviors latent variable are singing songs to the child and telling stories to the child, practices that are highly supportive of oral language and vocabulary development (Curenton, Craig, & Flanigan, 2008; Isbell, Sobol, Lindauer, & Lowrance, 2004; Speaker, Taylor, & Kamen, 2004) and can be done by a nonreader caregiver without materials.

We used our conceptualization of children’s experiences as shown in Model 4 to examine connections between household SES, reading materials, reading behaviors, and children’s early literacy outcomes, with the aim of providing useful evidence to guide programmatic decision making. Given highly limited resources in LMICs, how should funds be spent, if the goal is to improve literacy? Our findings indicate that for the development of early literacy skills in low-literacy contexts, investments in reading materials may be particularly beneficial. Initially, we found this result questionable, as it appears to conflict with research on shared reading and parent involvement in literacy skill development in the United States and other high-income countries. Book flood models (Elley, 2000), although popular as a development approach for a period in the 1980s and 1990s, have largely been eclipsed by approaches that focus more on building specific foundational skills of early reading, including phonological and morphological awareness (Y.-S. G. Kim et al., 2017). However, the patterns of associations are robust across project sites, categories of countries, and early literacy outcomes. We therefore are confident in the recommendation that early literacy interventions in low-resource contexts include a strong materials component, in alignment with the literature that has identified the importance of a print-rich environment for beginning readers (Davidse, de Jong, Bus, Huijbregts, & Swaab, 2011; Kalia & Reese, 2009; Kirby & Hogan, 2008; Park, 2008).

At the same time, we also acknowledge the possibility that the presence of reading materials in the home may also serve as a proxy for unobserved yet salient factors that influence children’s learning outcomes, leading to an artificial inflation of the association between reading materials and children’s literacy outcomes. For example, it is likely that parents who invest in purchasing reading materials may also be investing in their children’s learning in other ways, both in terms of financial inputs (e.g., paying for higher quality school experiences) and in the ways that they interact with or socialize their children toward reading (McCoy, Zuilkowski, & Fink, 2015; Yeung, Linver, & Brooks-Gunn, 2002). As such, although investments in reading materials may be an important starting point for supporting student learning, we encourage additional research and intervention to support other, unmeasured characteristics of HLEs and of the use of skills in daily life outside the home that may benefit children over time. Finally, we encourage the exploration of whether the home is the correct focus of research into learning environments in all contexts and advocate investigation of the use of skills in the broader community, especially understanding recent large-scale investments that aim to promote reading culture alongside school-based improvement packages.

We acknowledge that our analysis has several limitations. First, given the cross-sectional and nonrandomized nature of our data set, our analyses cannot support causal conclusions. Our data set includes few common demographic or household-level variables, and it is possible that our lack of comparable indicators of SES may have introduced bias into our models. Second, the sample of countries used in this analysis is not random, and the samples within each country are not nationally representative. However, the geographic range of countries as well as the representation of more than a dozen LMICs that receive international development aid suggests that the patterns observed here may be more widely applicable, at least in low-resource contexts. Third, although the use of different indicators at the national level likely allowed for better cultural specificity in measuring our study constructs, they may also have introduced biases into the analysis. Our use of structural equation modeling with latent variables is aimed at addressing some of the measurement issues, while also allowing for cross-national analysis. Nevertheless, future research should consider a broader set of both HLE processes and learning-related outcomes to more fully understand the phenomena at play.