Poverty and Proximate Barriers to Learning: Vision Deficiencies, Vision Correction and Educational Outcomes in Rural Northwest China

Abstract

Uncorrected vision may present a significant barrier to educational mobility in poor communities in low and middle income countries. Focusing on the case of rural Northwest China, we analyze the Gansu Survey of Children and Families (2,000 children; 100 rural villages) and the Gansu Vision Intervention Project (a randomized trial; 19,185 students, 165 schools, two counties). Four main findings emerge: significant unmet need for vision correction; socioeconomic gradients in vision correction; somewhat greater vulnerability to vision problems among higher socioeconomic status and more academically engaged children; and significant favorable effects of vision correction on math and literacy performance and class failure.

Introduction

Social scientists have long taken an interest in the mechanisms by which socioeconomic disadvantages in households translate to educational disadvantages for children. Researchers working from various frameworks have developed theories that emphasize socialization within families, the social networks and patterns of interaction that parents use to communicate with the school system, the cultural experiences and tools that aid children in their self-presentations to and interactions with teachers, and the different kinds of schools and teachers to which impoverished children have access (Buchmann and Hannum 2001; Hannum and Buchmann 2005; Huisman and Smits 2009). In high-poverty communities around the world, particularly those in low- and middle income countries, more proximate barriers also impede the day-to-day process of learning for children. For example, poverty can mean that children’s studies are hindered by the inability to purchase supplies to take notes or do assignments. Children can go to school hungry or poorly nourished, and thus less able to focus.

One potentially important mechanism by which poverty may affect a child’s day-to-day learning experiences is uncorrected vision problems. Bundy, Joshi, Rowlands & Kung (2003) report that about 10 percent of school-age children in developing countries have refraction errors, almost all of which can be corrected with properly fitted eyeglasses. Most children with refraction problems in low income countries do not have glasses. Few studies have investigated the impact of eyeglasses on school achievement, and none have investigated poor eyesight and educational achievement from a stratification perspective, by considering the social location of vision deficiencies and vision correction along with the impact of vision correction on outcomes. To address this gap, we ask first whether there are differences by child characteristics and by educational aptitude in the risk of poor vision, and in access to vision correction. We then investigate whether vision correction matters for educational outcomes—performance on standardized achievement tests and class failure.

Framework and Hypotheses

Despite the self-evident problems imposed by poor vision on classroom functioning and the potential for a relatively cheap and easy ameliorative intervention, there has been very little research on the impact of poor vision on students’ academic performance.¹ One published study found large impacts of poor vision among primary school children in northeast Brazil: children with poor vision had a 10 percent higher probability of dropping out of school, an 18 percent higher probability of repeating a grade, and scored about 0.2 to 0.3 standard deviations lower on achievement tests (Gomes-Neto, Hanushek, Leite & Frota-Bezzera, 1997). Another study in the U.S. found that vision problems among school-aged urban minority youth have a negative impact on academic achievement through effects on sensory perceptions, cognition, and school connectedness (Basch, 2011). A straightforward set of hypotheses exists: the most economically disadvantaged children lack access to vision correction, and uncorrected vision is thus a mechanism by which economic deprivation translates to a poorer opportunity to learn.

Yet, while the logical relationship between economic deprivation and vision correction is straightforward, the relationship between poverty and risk of poor vision is more complex. In a number of studies, poor vision has been associated with higher levels of education and test scores—attributes in turn often associated with higher family socioeconomic status. For example, studies of youth and young adults in Singapore show a positive association between educational attainment and the prevalence and severity of myopia (Au Eong, Tay, & Lim, 1993; Tay, Au Eong, Ng & Lim, 1992). Similarly, a study of 18 year-old men in Denmark showed that those with myopia had higher levels of education and higher test scores than those without myopia (Teasdale, Fuchs, and Goldschmidt, 1988). This situation renders the potential impact of vision correction on educational inequality difficult to isolate, even if the expected impact of vision correction on achievement—the main question addressed in earlier studies by economists—is clear.

Vision problems afflict a significant minority of school-aged children in China. One study in Shunyi District, northeast of Beijing, found that 12.8 percent of children age 5 to 15 years had vision problems, of which 90 percent were due to refraction errors (Zhao, Pan, Sui, Munoz, Sperduto & Ellwein, 2000; Zhao, Mao, Luo, Li, Munoz & Ellwein, 2002). Only 21 percent of the children with vision problems had glasses (Zhao et al., 2000). Girls and older children had higher risk of myopia than boys and younger children: myopia was minimal among five year-olds, but rose to 37 percent among 15 year-old boys and 55 percent among 15 year-old girls (Zhao et al., 2000). The authors conclude that over 9 percent of children could benefit from glasses. A study of junior high school students in Yangxi County, a rural setting in western Guangdong, showed that myopia² affected 36.8 percent of 13-year-olds, with the rate increasing to 53.9 percent among 17-year-olds (He, Huang, Zheng, Huang & Ellwein, 2007, p. 374). Of children with impairment in both eyes, only 46.5 percent were wearing glasses (p. 376).

To our knowledge, the data collection projects reported on in the current study, namely a longitudinal survey of 2,000 rural children and a randomized trial involving 19,185 students in 165 schools in one of China’s poorest provinces, are the first in China to link vision problems to educational achievement. The achievement effects of glasses provision in the randomized trial data have been analyzed in Glewwe, Park and Zhao (2006). Using a variety of estimation strategies, Glewwe and his colleagues showed that, after one year, provision of eyeglasses increased student performance by 0.15 to 0.30 standard deviations of the distribution of grades. The current paper complements Glewwe et al.’s (2006) report by presenting survey-based estimates utilizing standardized curriculum and literacy tests; by investigating the determinants of vision problems and access to vision correction; and by considering the relevance of vision correction as a protective factor in class failure.

Data and Methods

Study Site and Data

The study site is Gansu Province, in northwestern China. In 2000, the year our first wave of data was collected, Gansu’s population was 25.6 million people, 76 percent of whom resided in rural areas (Gansu Bureau of Statistics, 2001). In 2004, almost one in five people ages 15 and above was not literate, compared to just over one in ten for China as a whole (National Bureau of Statistics, 2006a). Official estimates of rural per capita income for 2004 rank Gansu 30th out of 31 provinces—below Tibet and above Guizhou (National Bureau of Statistics, 2006b).

The GSCF is a longitudinal survey of 2000 children in 20 counties who were 9 to 12 years old when they were first interviewed in the year 2000 (GSCF-1), and who were re-interviewed at ages 13 to 16 in 2004 (GSCF-2). GSCF−1 sought to estimate the individual, household, school, and community determinants of educational outcomes in rural, underdeveloped areas. GSCF−2 maintained the education-related focus of GSCF−1, but added a significant health component. In all, 1,918 target children from GSCF-1 were followed up at GSCF-2. However, about 13 percent of the children were not in school in 2004. Since this study focuses on the impact of vision problem and correction on school achievement, these cases are excluded from the sample.

The 2004 data collection effort also included an add-on project, not part of the GSCF sample, called the Gansu Vision Intervention Project (GVIP). In this project, a randomized evaluation was conducted to measure the impact on education outcomes of providing eyeglasses to vision-impaired children. Two counties in Gansu Province were selected as study sites. All townships in each county were first ranked by rural income per capita. In each county, starting with the first two townships, one was randomly assigned to receive treatment, and the other, to serve as a control. Then, all primary schools in each township either all received treatment, or all served as controls. In the process of implementing the project, a few control townships mistakenly received glasses. These townships, as well as the treatment townships that were originally paired with them, were dropped from the current analysis. The final sample includes 19,185 students in grades 3 to 5 in the 2004–2005 academic year, in 165 schools.³

Measurement

Vision

The first part of our analysis focuses on vulnerability to poor vision. Table 1 shows descriptive statistics on vision in both datasets. In 2004, in both the GVIP and the GSCF data, eye examinations were administered by Center for Disease Control personnel in Gansu. The examination employed was a domestic one used for screening purposes in schools by the Center for Disease Control. Scores ranged from 2.4 to 5.3 in the GVIP data and from 3.4 to 5.9 in the GSCF data. The Center for Disease Control and Prevention defines a 4.8 score in either eye as a cutoff for requiring glasses, and that standard, inclusive, is used here to define the outcome variable poor vision in both datasets. Vision problems afflict a significant minority of children in rural Gansu. In the GVIP data, 11 percent of the children suffer from poor vision, and in the GSCF, 17 percent of the children do (see Table 1). To capture full information from the eye scores from both eyes, we calculated a continuous average vision measure (vision score: average) and an absolute deviation measure (the absolute value of the deviation in score between the two eyes, vision score: difference).

Table 1.

Vision Descriptives

	Mean or Proportion	SD	95% Conf. Interval Lower	Upper	N
GVIP:
Poor Vision, 2004*	0.108	0.002	0.103	0.112	19143
Vision Score: Average	5.030	0.002	5.027	5.033	19143
Vision Score: Difference	0.048	0.001	0.047	0.050	19143
Wearing Glasses before Project	0.011	0.102	0.009	0.012	19178
Received Glasses from Project	0.057	0.232	0.054	0.060	19020
GSCF:
Poor Vision, 2004	0.166	0.372	0.149	0.184	1780
Vision Score: Average	5.008	0.007	4.995	5.022	1780
Vision Score: Difference	0.076	0.004	0.068	0.085	1780
Wears Glasses, 2004	0.071	0.257	0.060	0.084	1814

^*Poor vision is coded as “1” if either eye has a vision score less than or equal to a 4.8 cutoff.

Vision correction

The second part of the analysis focuses on access to vision correction. The GVIP data and the GSCF data contain reports about whether children wear glasses. In the GVIP data, this information is reported by teachers involved in the project regarding children’s status before the project provided glasses, and in the GSCF data, by children⁴ themselves. In the GVIP data, 1 percent of all children wore glasses prior to the start of the project. In the GSCF data, 7 percent of all children wore glasses in 2004.⁵

The GVIP project also contains a variable received glasses, which refers to the children who accepted glasses as part of the GVIP project. Among all children in the GVIP data, 6 percent received glasses, a number that constitutes 45.6 percent of all children who had vision problems and 71.76 percent of children with vision problems in treatment townships.⁶

Educational Outcomes

Table 2 shows current and prior educational achievement measures in both datasets. In the GVIP data, we employ an outcome set to one if the child failed in math, Chinese, or science in spring 2005. Failure means receiving a grade of below 60 percent. Failure in these main subjects is significant, as it may lead to the student’s repeating of the grade. In the sample, about 11 percent of all children had failed one or more subjects (see Table 2).

Table 2.

Prior and Current Achievement

All Cases		Mean					(SD)		CI		N
GVIP:	Fail Rate (Proportion)	0.11					0.31		0.10	0.11	18860
	Prior Achievement Scale	82.40					8.08		82.29	82.52	18973
GSCF:	Literacy Test Score	0.00					1.00		−0.05	0.05	1754
	Achievement Test: Math	0.00					1.00		−0.05	0.05	1719
	Achievement Test: Chinese	0.00					1.00		0.05	0.05	1719
	Math Grade in 2000	73.89					14.71		73.21	74.56	1828
	Chinese Grade in 2000	72.45					13.28		71.84	73.06	1822
	Cognitive Test in 2000	49.97					19.81		49.07	50.87	1865
	Self-report: Good Language Ability in 2000	0.38					0.48		0.35	0.40	1851
	Self-report: Good Math Ability in 2000	0.43					0.50		0.41	0.45	1852
By Vision Problem Diagnosisa		With Vision Problems					Without Vision Problems
		Mean	(SD)	CI		N	Mean	(SD)	CI		N
GVIP:	Fail Rate (Proportion)	0.10	0.30	0.08	0.11	2025	0.11	0.31	0.10	0.11	16818
	Prior Achievement Scale	82.25	8.12	81.89	82.60	2039	82.42	8.07	82.30	82.55	16911
GSCF:	Literacy Test Score	0.28	0.93	0.17	0.39	260	−0.05	0.99	−0.10	0.00	1316
	Achievement Test: Math	0.16	0.93	0.05	0.27	267	−0.01	1.01	−0.07	0.04	1334
	Achievement Test: Chinese	0.16	0.97	0.04	0.27	267	−0.02	0.99	−0.07	0.04	1334
	Math Grade in 2000	76.23	13.53	74.63	77.83	277	73.56	14.83	72.78	74.35	1381
	Chinese Grade in 2000	74.38	11.70	72.99	75.77	276	72.04	13.55	71.32	72.75	1376
	Cognitive Test in 2000	51.91	19.76	49.59	54.23	282	49.47	20.01	48.43	50.52	1410
By Glass-Wearing Status (2004)		Wearing Glasses Before Project					Not Wearing Glasses Before Project
GVIP:		Mean	(SD)	CI		N	Mean	(SD)	CI		N
	Fail Rate	0.05	0.22	0.02	0.09	199	0.11	0.31	0.10	0.11	18660
	Prior Achievement Scale	85.15	7.46	84.11	86.20	200	82.37	8.08	82.26	82.49	18772
GSCF:		Wearing Glasses					Not Wearing Glasses
	Literacy Test Score	0.56	0.79	0.41	0.71	111	−0.03	0.99	−0.08	0.02	1536
	Achievement Test: Math	0.33	0.95	0.16	0.49	124	−0.02	1.00	−0.07	0.02	1593
	Achievement Test: Chinese	0.29	1.01	0.11	0.47	124	−0.02	0.99	−0.07	0.03	1593
	Math Grade in 2000	79.08	11.92	76.99	81.18	127	73.78	14.53	73.07	74.48	1656
	Chinese Grade in 2000	76.35	10.13	74.57	78.13	127	72.36	13.17	71.72	72.99	1651
	Cognitive Test in 2000	55.55	19.17	52.21	58.89	129	49.62	19.79	48.67	50.57	1685

^aVision problem is coded as “ 1” if either eye has a vision score less than or equal to a 4.8 cutoff (in 2004), else “0”.

In the GSCF data, we use results from three tests administered as part of the project: a literacy assessment, a curriculum-based math achievement test, and a curriculum-based language achievement test. The literacy test had a mean of 20.5; a standardized version is used here. Math and language achievement tests had means of about 17 and 21, respectively. The numbers presented here are standardized by grade, as tests were grade-specific.

Prior engagement and achievement

In the GVIP, prior achievement is measured as a scale (average) of reported scores for math, science, and language for each semester in grades one and two. The scale has high internal reliability, with a Cronbach’s Alpha score of .94. The average is about 82 for those with and without vision problems (see Table 2.)

In the GSCF, prior achievement is measured as math score (grade) and language score (grade) reported by teachers in the 2000 round of the survey. In the sample, the average math score is 74 and the average language score is about 73. To further control prior ability, we also add a cognitive test score. This standard test of cognitive ability, developed for the project at the Institute of Psychology at the Chinese Academy of Science, had a mean score of about 50.

Finally, we included measures of child’s self-assessed math ability and language ability in 2000. The ability variables are based on children’s answers to questions “Compared with your classmates, what is your math level?”, and “Compared with your classmates, what is your language level?” The original five-category responses, very poor, below average, average, above average, and excellent, are recoded into two categories for each question, with 1 for “above average” and “excellent” and 0 for the other categories. Using these definitions, about 37 percent of children viewed their language ability favorably, while 44 percent viewed their math ability in this way.

Table 3 shows other background characteristics of children in the GVIP and GSCF.

Table 3.

Sample Background Characteristics

	Among All Cases			Among Children with Vision Problems*			Among Children Who Wore Glasses before the Project
GVIP:	Mean (Proportion)	(SD)	N	Mean (Proportion)	(SD)	N	Mean (Proportion)	(SD)	N
Household Head’s Education (Years)	8.62	2.26	19172	8.43	2.23	2063	9.38	2.79	202
Non-farm Household Head (1=Yes)	0.14	0.35	19181	0.11	0.32	2063	0.38	0.49	202
Child Gender (1=Male)	0.53	0.50	18817	0.48	0.50	2062	0.50	0.50	202
Child Age at Survey	10.69	1.95	19146	10.97	1.71	2063	10.11	1.97	202
Ethnicity (1=Minority)	0.14	0.35	19181	0.12	0.32	2063	0.25	0.44	202
Height	137.77	8.95	19148	139.40	9.17	2063	138.66	9.93	202
Grade 3	0.33	0.47	18817	0.22	0.42	2063	0.18	0.38	202
Grade 4	0.33	0.47	18817	0.31	0.46	2063	0.35	0.48	202
Grade 5	0.34	0.47	18817	0.47	0.50	2063	0.47	0.50	202
	Among All Cases			Among Children with Vision Problems*			Among Children Who Wore Glasses in 2004
GSCF:	Mean	(SD)	N	Mean	(SD)	N	Mean	(SD)	N
	(Proportion)			(Proportion)			(Proportion)
Father’s Education (Year)	6.97	3.52	1865	7.19	3.53	282	8.09	3.30	129
Mother’s Education (Year)	4.12	3.49	1863	4.63	3.60	282	5.56	3.82	129
Logged Family Wealth in 2004 (RMB)	8.49	1.12	1866	8.56	1.19	282	8.85	1.15	129
Child Gender (1=Male)	0.53	0.50	1866	0.50	0.50	282	0.51	0.50	129
Child Age in 2000	11.02	1.09	1866	11.14	1.06	272	11.55	1.04	129
Height	156.00	9.14	1783	157.64	8.73	295	159.00	8.65	120
Skinfold	0.96	0.50	1782	1.04	0.54	294	1.02	0.44	120
Village Has Bus Station (1=Yes)	0.67	0.47	2000	0.73	0.45	295	0.74	0.44	129

^aVision problem is coded as “1” if either eye has a vision score less than or equal to a 4.9 cutoff (in 2004), else “0”.

Socio-economic characteristics

In the GVIP data, we have just two simple variables measuring socio-economic status—household head’s years of schooling and head’s non-farm occupational status. On average, the household heads have 8.6 years of schooling, and 14 percent of them are not farmers (see Table 3). In the GSCF data, we employ measures of mother’s education, father’s education, and logged household wealth. On average, mothers have 4.2 years of schooling, while fathers have a little less than 7 years of schooling.

Control variables

In the GVIP data, we also include information on the age, grade, sex, and ethnicity of the child. The mean age of the children in the analytic sample is 10.69. All children in the analytic sample are in grades 3, 4, and 5 in primary school; 47 percent of children in the sample are girls, and 14 percent are minorities (see Table 3). Nearly all of the minority children are Tibetans. In the GSCF data, we include age and sex of the child. These variables were measured in 2000, so the average age of children was about 11 (15 in 2004), and 47 percent are female.

We also include controls for height (in both the GVIP, average 138 centimeters, and the GSCF, average 156 centimeters) and triceps skinfold thickness (GSCF only, average .96 centimeters) as measures of early and recent nutritional adequacy. Height is also an indicator of general health in early childhood. Finally, in models of glasses-wearing in the GSCF, we control for whether there is a bus station in the village where the child resides, as an indicator of possible transportation barriers to glasses provision.

Methodological Approach

Our analysis employs logistic regressions of vision problems and access to eyeglasses with random effects for schools (in the GVIP data) or villages (in the GSCF data). We show in these analyses that there are considerable differences across social groups in the propensity to wear glasses. We employ two strategies to address this difference in propensity to wear glasses, in order to investigate the impact of glasses-wearing on achievement.

First, for analyses of the GSCF data, we focus on determinants of poor vision and glasses-wearing in 2004. Next, we employ propensity score matching to address selection bias associated with gaining access to glasses. With a propensity score matching approach, we assume that pertinent differences between those with and without glasses can be captured by observable variables, and select from the non-treated a “control group” in which the distribution of observed variables is as similar as possible to that in the treated group (glasses-wearers). We use the psmatch2 program in Stata to estimate propensity scores for glasses-wearing, with kernel matching. We use logit models for estimation of propensity scores. In the models, we included all predictor variables that were part of our analysis of glasses-wearing, and we imposed a common support structure (for a straightforward discussion of the implications of model choice, matching choice, and common support, see Caliendo and Kopeinig, 2008). Further investigation showed that all significant differences in predictors in the original sample were eliminated in the matched sample. We present bootstrapped estimates of average treatment effects on the treated, with standardized literacy, language, and math scores as outcomes.

Second, using the randomized intervention data—the GVIP data—we present logistic regressions of class failure to estimate the impact of receiving glasses on school progress, at the margins. We include random effects for schools and control for background factors that might be associated with acceptance of the randomized offer of treatment and with class failure.⁷

Results

Analysis of Poor Vision

Children who can’t see what teachers are writing on the board and can’t do homework due to vision problems face evident barriers to learning. In the GSCF, a significant fraction of children themselves report experiencing vision-related barriers to learning. Almost one in five children reported having problems reading the blackboard, and 12 percent reported having difficulty doing homework because of eye problems. Nearly a quarter of students (23 percent) complained that their eyes hurt while doing homework because of poor light conditions at home (Figure 1). These self-reports indicate that vision problems—including strain related to poor lighting as well as poor eyesight—are experienced by children in ways that are detrimental to school engagement and performance.

Figure 1.

Proportion of Children Reporting Various Vision Problems

Turning to measured vision problems, in the GVIP sample, as noted above, about 11 percent of Children were diagnosed with poor eyesight (see Table 1). Results from a logistic regression analysis show that children who performed better early on in school, who were in non-farming households, who were girls, and who were taller had significantly higher risk of poor vision, though the non-farming household effects only achieved marginal significance (see Table 4). Alone or controlling for all other displayed variables in model 4, a one standard deviation increase in the prior achievement measure is associated with about a 6 or 7 percent increase in the odds of poor vision (1.008^8.08=1.07). Compared to children in farming households, odds of poor vision for children of parents in non-farming households were about 20 percent higher (100*[1.20–1], as shown in model 4 in the most conservative multivariate specification). In the same specification, boys’ odds of poor vision were about a quarter lower than girls’ (100*[1–.758], in model 4). Height is significantly positively related to vision problems in model 3 and model 4, with about a 2 percent increase in odds associated with a one-centimeter increase in height (100*[1.023–1]) (controlling for age and other variables in the model). Thus, the GVIP findings suggest that there is an elevated chance of poor eyesight among children who perform well; among children who are older and who are girls; among higher socioeconomic status children, as indicated by non-farm family status; and to children who were healthier in early childhood, also probably a proxy, in part, for socioeconomic status. There is no significant association with ethnicity or head’s education, in bivariate models or net of other controls in the models shown in Table 4.

Table 4.

Random Effects Logistic Regression Analysis of Poor Eyesight Diagnosis, GVIP Data

	1 OR/(SE)	2 OR/(SE)	3 OR (SE)	4 OR (SE)
Prior Achievement Scale	1.008** (0.004)			1.008** (0.004)
Household Head’s Education (Years)		0.998 (0.013)		0.997 (0.013)
Non-farm Household Head		1.190* (0.111)		1.200* (0.112)
Sex (1=Male)			0.755*** (0.037)	0.758*** (0.037)
Ethnicity (1=Minority)			0.970 (0.100)	0.969 (0.100)
Age			1.019 (0.012)	1.019 (0.012)
Height			1.023*** (0.003)	1.023*** (0.003)
/lnsig2u	0.688** (0.102)	0.686** (0.102)	0.681*** (0.101)	0.693** (0.103)
Log-Likelihood	− 6,104.31	− 6,105.16	− 6,056.42	−6,052.06

Note:

^***p<0.01,

^**p<0.05,

^*p<0.1

The GSCF project offers more detailed variables measuring children’s background. In the GSCF data, simple specifications show that mother’s education, child’s higher self-reported math ability in 2000, age, and height were at least marginally significant predictors of subsequent vision problems (see Table 5, models 1 to 5). Wealth, prior performance in math and language, prior cognitive development score, and sex are not significant here, though the odds-ratio for sex, like in the case of the GVIP, suggests lower odds of poor vision for boys. In the full model, model 6, mother’s education, Chinese and math ability, height, and age were at least marginally significant predictors. For example, each additional year of maternal education is associated with an increase of 4.6 percent in odds of a vision problem diagnosis (100*[1.047-1]). Reporting a high level of math ability early on is associated with 45 percent higher odds of poor eyesight, relative to reporting a lower ability (100*[1.452–1]). For Chinese ability, the figure is 31 percent (100*[1.31–1]). Finally, each year of age is associated with a 19 percent increase in odds of a vision problem diagnosis (100*[1.194–1]), and each centimeter increase in height is associated with about a 2 percent increase in odds (100*[1.017-1]).

Table 5.

Random Effects Logistic Regression Analysis of Poor Eyesight Diagnosis, GSCF Data

	1 OR/(SE)	2 OR/(SE)	3 OR/(SE)	4 OR/(SE)	5 OR/(SE)	6 OR/(SE)
Math Grade in 2000	1.010 (0.009)					1.007 (0.009)
Chinese Grade in 2000	1.002 (0.010)					0.999 (0.010)
Cognitive Test in 2000	1.005 (0.004)					1.003 (0.004)
Mother’s Education (Years)		1.043* (0.024)				1.047* (0.025)
Father’s Education (Years)		0.998 (0.022)				0.979 (0.022)
Logged Wealth			1.088 (0.074)			1.040 (0.074)
Child Reported Good Chinese Ability in 2000				1.260 (0.197)		1.310* (0.213)
Child Reported Good Math Ability in 2000				1.394** (0.216)		1.452** (0.234)
Sex (1=Male)					0.870 (0.142)	0.855 (0.144)
Age					1.152* (0.084)	1.194** (0.090)
Height					1.017* (0.009)	1.017* (0.010)
Skin Fold					1.239 (0.188)	1.254 (0.199)
/lnsig2u	0.721 (0.194)	0.699 (0.187)	0.728 (0.193)	0.779 (0.204)	0.683 (0.186)	0.709 (0.196)
Log-Likelihood	−712.26	−730.34	−731.82	−720.61	−723.32	−694.35

Note:

^***p<0.01,

^**p<0.05,

^*p<0.1

Overall, although the GVIP and GSCF offer different measures, neither suggests that the most socioeconomically disadvantaged are at particularly high risk of poor eyesight. In fact, analyses of both datasets suggest that there is a tendency for vision problems to be greater among higher socio-economic status children and among children who are more educationally engaged. This finding is consistent with available research conducted elsewhere.

Access to Vision Correction

In the GVIP sample, just one percent of all children reported wearing glasses before the project.⁸ However, there is a big gap between farming and non-farming households. In the full sample, about 0.8 percent of children in households headed by farmers were reported as wearing glasses prior to the project, compared to 2.8 percent of children in households headed by non-farmers (our calculations, not shown). Among children with poor eyesight, comparable figures were 1.95 and 7.3 percent (our calculations, not shown).

Table 6 shows results from a series of logistic regression models of glasses-wearing in the GVIP sample. Models 1 to 4 show that prior performance, head’s non-farm status, and height are associated with glasses-wearing, though prior performance is only marginally significant. Finally, model 5 re-estimates model 4, with both eye score measures included. This specification allows full control for measured vision in both eyes, and shows, again that household non-farm status, height, and prior performance (at a marginally significant level) are associated with glasses-wearing.

Table 6.

Random Effects Logistic Regression Analysis of Wearing Glasses, GVIP Data

	1 OR/(SE)	2 OR/(SE)	3 OR/(SE)	4 OR/(SE)	5 OR/(SE)
Prior Achievement Scale	1.021* (0.012)			1.021* (0.012)	1.020* (0.012)
Household Head’s Education (Years)		1.035 (0.033)		033 (0.033)	1.028 (0.033)
Non-farm Household Head		2.334*** (0.510)		2.406*** (0.525)	2.344*** (0.519)
Sex (1=Male)			0.877 (0.128)	0.863 (0.127)	0.894 (0.134)
Ethnicity (1=Minority)			1.148 (0.271)	1.244 (0.291)	1.205 (0.284)
Age				0.924 (0.045)	0.917* (0.045)
Height				1.036*** (0.009)	1.035*** (0.009)
Vision Score: Average					0.111*** (0.032)
Vision Score: Difference					1.316 (0.707)
/lnsig2u	3.273*** (0.870)	3.191*** (0.854)	3.401*** (0.902)	3.198*** (0.857)	3.289*** (0.875)
Log-Likelihood	−954.91	−946.47	−956.06	−935.17	−905.95

Note:

^***p<0.01,

^**p<0.05,

^*p<0.1

Logistic regression analysis of glasses-wearing in the GSCF shows that without adjusting for other factors, there are significantly higher odds of wearing glasses among children with higher test scores (in math and in the cognitive test) (model 1), children who report better academic ability (in Chinese) (model 2), children of better educated parents (model 2), and wealthier children (model 3) . There is a marginally significant difference by sex, with boys less likely to wear glasses. Age and height are significantly positively related to wearing glasses (model 5). In Model 6, with all predictors from models 1 to 5 controlled, mother’s education, Chinese ability, family wealth, children’s age and height significantly predict glasses wearing, though mother’s education, Chinese ability, and height are only marginally significant in this specification. Model 7 adds an indicator for whether the village has a bus station, which has a marginally significant positive effect on glasses-wearing, but does not change other patterns of significance other than for the prior math performance measure, which becomes marginally significant and positive again in this specification.

Model 8 includes controls for average vision and vision deviation, both of which show highly significant effects on glasses-wearing. This specification shows that, net of measured need for glasses, among all youth in the sample, wealthier youth, youth with better-educated fathers, older youth, and youth in villages with a bus station were more likely to be wearing glasses, though the bus station effect is only marginally significant.

Wealth differences are striking. In the raw data, about 4 percent of children in the bottom wealth quintile (measured in the earlier survey wave) wore glasses, as did about 9 to 11 percent in the top two wealth quintiles (our calculations). Among children with poor eyesight, the corresponding range was 10 percent for children in the poorest quintile of household wealth to over one-third in the top two quintiles.

Impact of Glasses on Achievement

We know from earlier work that has investigated various estimates, including difference-indifference estimates and instrumental variable approaches, that providing glasses to children in the GVIP sample had an impact on learning, as measured by grades standardized at the school level (Glewwe et al., 2006). Here, we complement this work with GSCF estimates, which can be produced based on standardized achievement tests rather than grades. However, the impact of glasses is harder to convincingly isolate in the GSCF survey, because of selection issues described in the preceding sections.

To address selectivity, we use model 8 in Table 7 to estimate propensity scores of wearing glasses, then present estimates of the average treatment effect on the treated for the matched samples⁹ produced by this exercise. Results are shown in Table 8. For the literacy outcome, the average treatment effect on the treated is .34 standard deviations. For the language achievement outcome, the effect is not significant. For the math achievement outcome, the effect is .26 standard deviations.¹⁰ We can’t completely rule out the possibility that our strategy for matching the treatment and control samples has not fully accounted for pertinent differences in unmeasured variables. However, our finding of significant effects of glasses-wearing on literacy and math scores are consistent with significant positive effects for grades found by Glewwe, Park and Zhao (2006) using an experimental design.

Table 7.

Random Effects Logistic Regression Analysis of Wearing Glasses in 2004, GSCF Data

	1 OR/(SE)	2 OR/(SE)	3 OR/(SE)	4 OR/(SE)	5 OR/(SE)	6 OR/(SE)	7 OR/(SE)	8 OR/(SE)
Math Grade in 2000	1.026** (0.014)					1.022 (0.014)	1.024* (0.014)	1.021 (0.016)
Chinese Grade in 2000	0.997 (0.014)					0.993 (0.015)	0.992 (0.015)	0.988 (0.017)
Cognitive Test in 2000	1.011** (0.005)					1.006 (0.006)	1.006 (0.006)	1.005 (0.006)
Child Reported Good Chinese Ability in 2000		1.628** (0.349)				1.503* (0.351)	1.520* (0.353)	1.367 (0.359)
Child Reported Good Math Ability in 2000		1.140 (0.242)				1.211 (0.281)	1.219 (0.282)	1.250 (0.328)
Mother’s Education (Years)			1.081** (0.033)			1.055* (0.035)	1.058* (0.035)	1.036 (0.038)
Father’s Education (Years)			1.066** (0.034)			1.054 (0.036)	1.054 (0.036)	1.089** (0.042)
Logged Wealth				1.336*** (0.121)		1.257** (0.125)	1.243** (0.122)	1.269** (0.140)
Sex (1=Male)					0.659* (0.163)	0.702 (0.176)	0.710 (0.178)	0.923 (0.261)
Age					1.572*** (0.168)	1.608*** (0.176)	1.615*** (0.176)	1.580*** (0.197)
Height					1.032** (0.015)	1.027* (0.015)	1.028* (0.015)	1.015 (0.016)
Skin Fold					1.009 (0.225)	1.098 (0.250)	1.101 (0.252)	0.975 (0.264)
Village Bus Station (1=Yes)							1.762* (0.522)	1.766* (0.545)
Vision Score: Average								0.020*** (0.007)
Vision Score: Difference								0.185*** (0.095)
/lnsig2u	0.974 (0.341)	1.111 (0.366)	0.829 (0.302)	0.982 (0.333)	0.844 (0.316)	0.655 (0.281)	0.552 (0.257)	0.392 (0.267)
Log-Likelihood	−431.53	−443.17	−439.53	−442.50	−398.81	−376.32	−374.51	−294.10

Note:

^***p<0.01,

^**p<0.05,

^*p<0.1

Table 8.

Propensity Score Matching Results for Eyeglass Provision, GSCF

	Treatment	Control		Bootstrapped			95% Confidence Interval
	N	N	ATT	Std. Err.	z	Sig	Lower Bound	Upper Bound
Standardized Literacy Assessment	94	1398	0.343	0.109	3.15	0.002	0.130	0.557
Standardized Language Curriculum Test	106	1,450	0.134	0.119	1.12	0.262	−0.100	0.368
Standardized Mathematics Curriculum Test	106	1,450	0.257	0.105	2.45	0.014	0.052	0.462

Note: Propensity score equations are same as Model 8 in Table

The Impact of Glasses on Class Failure

Finally, we consider whether glasses’ effect on achievement matters at the margins, for failure. Table 9 shows results from a logistic regression analysis of failure in math, Chinese, or science, with a positive outcome indicating failure in at least one of these subjects. Column 1 shows an analysis using the full sample, and column 2 shows the same analysis estimated on a sample of children with poor eyesight. In the first case, the odds of failing a class are reduced by about 44 percent (100*[1–.561]), and in the second case, the odds of failing a class are reduced by 35 percent (100*[1–.644]) among children who received glasses from the project. In the second case, results are marginally significant.¹¹ Other results suggest that males, early high achievers, and those in a higher grade are less likely to fail. Among those with poor eyesight, children in non-farming households were less likely to fail.

Table 9.

Logistic Regression Analysis of Class Failure, GVIP Data

	1 OR/(SE)	2 OR/(SE)
Received Glasses(Among Not Wearing Glasses before)	0.561*** (0.074)	0.644* (0.153)
Previous Achievement Scale	0.892*** (0.004)	0.900*** (0.011)
Household Head’s Education (Years)	0.980 (0.014)	0.998 (0.048)
Non-farm Household Head	0.945 (0.113)	0.397** (0.168)
Sex (1=Male)	0.895** (0.049)	0.719* (0.132)
Ethnicity (1=Minority)	0.906 (0.106)	0.575 (0.266)
Grade	0.796*** (0.033)	0.717** (0.100)
Age	1.010 (0.014)	1.012 (0.060)
Height	1.006 (0.004)	1.011 (0.012)
/lnsig2u	1.410** (0.198)	1.930** (0.529)
Log-Likelihood	−4,949.72	−527.28
Estimation Sample	Full Sample	Poor Eyesight

note:

^***p<0.01,

^**p<0.05,

^*p<0.1

Summary

Results from these analyses show that a significant fraction of children in Gansu face vision problems, and few have access to glasses. Moreover, access to vision correction is strongly associated with a child’s socioeconomic background—farming versus non-farming status and wealth. While access to glasses is lowest among the poorest, vision problems themselves may be somewhat selective of better-off children and more academically engaged students. Our analyses suggest that vision correction matters for standardized literacy and math tests, and for the likelihood of failing classes.

Conclusions

In low- and middle- income countries, economic deprivation often translates to proximate barriers to day-to-day educational functioning for children within the school system. In Gansu, children themselves report that poor eyesight impedes their educational experience, and our findings are consistent with this. perception. About 11 percent of third to fifth graders in the GVIP and about 17 percent of 13 to 16 year-olds in the GSCF had measured vision problems. Yet, just 1 percent of the GVIP sample and 7 percent of the GSCF sample wore glasses in 2004, and access to vision correction shows a sharp socioeconomic gradient in both datasets.

Significantly, vision problems themselves are selective of better-off children and more academically engaged students, and this selectivity makes isolating the causal impact of glasses-wearing a difficult task. Our propensity score matching estimates based on the GSCF suggest a significant effect of glasses-wearing on standardized math and literacy tests, though not on language tests. Analysis of the GVIP intervention data shows that those who received glasses were less likely to fail a class. While we cannot firmly rule out all sources of selectivity in glasses-wearing in the GSCF or in accepting glasses in the GVIP, our findings are consistent with the commonsense notion that correcting vision supports learning.

Thus, results attest most clearly to a significant unmet need for vision correction. This finding is consistent with Bundy et al.’s (2003) characterization of the situation of children in developing countries more broadly. This need, together with evidence suggesting that wearing glasses supports learning, underscores the potential value of glasses provision as an aid to educational functioning for students in impoverished areas in developing country settings. At the same time, together with earlier findings, the academic and socioeconomic selectivity in vision problems documented here suggests that vision interventions will be unlikely to target the most impoverished, most educationally vulnerable children in these areas. Selectivity issues also indicate the need for further empirical studies that test the impact of vision correction on learning outcomes.