READ CORE and READ CORE PLUS INTERVENTION BANGLADESH BASELINE September 2015 Silvia Diazgranados Harvard University Akter Hossain Liana Gertsch Jarret Guajardo Save the Children Acknowledgements to Sayed Mohammed, Akidul Islam, Parvin Shahana and the team of enumerators for their support in the data collection. Background and Context Bangladesh has made dramatic strides in improving access to basic education, driven by successful partnerships among government, donors and NGOs. Despite unprecedented gains in education access, there is a crisis in early grade learning. The 2008 National Assessment Survey (NAS) found only 12% of 3rd grade students and 14% of 5th grade students scored 80% of higher in Bangla competencies, including vocabulary, writing and reading comprehension. Research in Bangladesh shows a host of school related challenges, including lack of teacher preparation, lack of well-designed reading materials for children in schools and homes, limited contact hours in schools, limited opportunities to practice reading, limited community engagement and low literacy among parents, no culture of reading for pleasure, lack of interventions for children struggling to learn, etc. To address these problems and with the aim of improving reading competencies in children, Save the Children created the Reading Enhancement for Advancing Development (READ) program in collaboration with the Government of Bangladesh and with support from the US Agency for International Development (USAID). READ Core and READ Core Plus READ incorporates evidence-based practices that emphasize capacity and sustainability, focusing on four areas of intervention: 1) teacher education and continuous professional development; 2) reading assessment; 3) increased availability of reading material, and 4) increased opportunities in the community to read and to provide support to beginning readers by those outside the school walls. Specifically, the READ model provides participating schools with state-of-the-art strategies to teach and assess key reading skills, such as letter knowledge, phonemic awareness, vocabulary, fluency and comprehension. This year, 1465 participating schools (805 Government Public Schools (GPS) and 660 Newly Nationalized Public Schools (NNPS) received a READ Core package that includes: Teacher training on Early Grade Reading Instruction Teacher training on Early Grade Classroom Diagnostic and Formative Assessment Administrators (Head teacher, Upazila Education Officers (UEO), Upazila Resource Centre (URC) and Assistant Upazila Educatior Officers (AUEO)) training on academic supervision. Teachers´ preparation for Education Quality Improvement: In order to improve the quality of teachers’ instruction, good practices inside the class room are captured through a film and shared with teachers. Every partner organization films two teachers three times in a year. In total 14 teacher of 7 division will be filmed three times in a year and the videos will be used in trainings and forums. School-based book bank: READ supplied 70 titles and one book stand for all schools. The book stands are designed to display books and bring them within easy reach of children. The books are visible through transparent pockets so that children will be attracted to them, according to their diverse interests. Books are color-coded according to reading level so that children can choose books according to their reading skill. Each classes have a specific time to receive and read book from the book corner. Book captains are managing the book corner in each school. SMC Read Orientation Reading Sessions for Parents: Mothers’ gather at school to receive IAT report card, giving emphasis on parents support at home Additionally, out of 805 GPS, 80 participating schools were randomly selected to receive the READ Core Plus Package, so that children in those schools will benefit from all of the above, plus the following: Community-based book bank: To ensure children´s access to ample and diverse reading materials appropriate to their level of reading outside of school, each community is awarded with a book packet of 70 books which are specifically designed to help them develop their literacy. Specifically, the book bank has 70 illustrated books, 2 alphabet primers, 2 language games, and guidelines on how to read for children and how to read with children. Parents and students have access to the book bank and children are able to take one book for a week. CLV oversee community book banks and use a log to monitor book use and return. Reading Camp: Reading camps provide children with the chance to participate in fun, engaging literacy activities outside of the school environment. Two reading camps are established within each catchment area. Two groups comprising 1st and 2nd grade students will participate in the reading camps. Each reading camp provides 30 first and second grade children within the 80 participating sites with the opportunity to attend once per week, for 8 months of the year to (3032 sessions). Each camp is a 90 minute session where two CLV engage children in activities where they can practice their literacy skills through games, lecture, activities and story-telling. Specifically, during each session, children: 1) Sign their name, sing a song and discuss the agenda, 2) participate in story-telling (15-20 minutes), 3) listen to scripted lectures (30-35 minutes) that have been designed to promote phonemic awareness (7 sessions), letter knowledge (2 sessions) vocabulary (3 sessions), fluency (4 sessions) and comprehension (4 sessions), 4) practice what they learned (15 minutes), and 5) respond to questions to assess what they learned (10 minutes). Before departing, they can choose to borrow a book from the book bank to read with a Reading Buddy at the end of the session. Reading buddies: In order to provide support and individualized attention to struggling students who are falling behind and to increase the habit of reading for pleasure, Community Literacy Volunteers (CLVs) match pairs of children who live close by but have different levels of literacy skills to have competent readers supporting beginning readers. Older buddies are trained on how to read to the younger children, and have high levels of fluency and reading comprehension. The younger reader is encouraged to borrow books from the book bank and read them together with his or her ‘Buddy’ who is more advanced. Books are distributed at the reading camp, but the actual reading takes places at the community or at home. Parent workshops: Parents are invited to awareness sessions in support of reading where they learn concrete ideas about how to improve the language development and literacy of their children in their daily activities and how to establish a reading culture at home. Workshops are conducted once a month for seven months. Each 60-90 minute workshop includes 20-30 parents of students targeted at the reading camp. All READ Core Plus activities are implemented by Community Literacy Volunteers (CLV), who receive a small monthly compensation for their work. CLVs attend a two-day long training before initiation of activities, which are conducted by field officers from PNGO, and then a refresher training after the first year of implementation. READ Core Plus uses a cascade model to conduct trainings, where 14 Technical officers and field officers from the PNGO's attend a Train the Trainers Training conducted by Senior Technical Officers (STO), after which they proceed to conduct their own trainings for groups of 25-30 CLVs. All trainings follow the same training manual, which introduces READ Core Plus to the community, and provides information about the goals, routines, materials and activities of the program, with a focus on the promotion of literacy skills. Trainings include demonstrations, role plays and discussions about their roles and responsibilities. Trainings are followed by on-site visits once a month, conducted by Technical Officers (TOs) and Field Officers, who monitor the program and give support to CLV in their work. TOs and Field Officers work for PNGO’s, have professional degrees and themselves attend a number of trainings that prepare them in their roles as coaches of the CLVs. TOs are responsible to supervise three Field Officers, and Field Officers provide support to CLVs in two READ Camps. During their visits, Field Officers monitor sessions, observe lessons to ensure proper implementation of the program. TOs meet with field officers once a month to provide them guidance with their work. Finally, each READ community program has a Management Committee comprised of parents and community members, who provide a regular venue for the program and assistance to the CLVs, including maintenance of materials, etc. Table 1 provides a comparison of the supports provided by READ and READ Core Plus. Table 1. Resources and Activities of the READ Core and READ Core Plus models Supports READ CORE READ CORE PLUS Teacher Trainings X X School Management Committee X X Academic supervision for Bangla X X Monthly Monitoring and Evaluation X X Establishment of print reach X X X X environment IAT Conduction Community Volunteers Trainings X Reading Camp and storytelling X Reading Buddies X Parent reading-awareness workshops X Community Book Bank X CLV Monitoring and Evaluation X The Present Study: A Multi-Site Randomized Controlled Clustered Evaluation The present study uses a multi-site randomized controlled clustered evaluation to measure the value-added impact of the READ Core Plus model on children, by comparing it to the READ Core model. Specifically, our study aims to answer the following research questions: At baseline Are there statistically significant differences in the observable characteristics of children in the treatment (READ Core Plus) and control (READ Core) groups before initiation of activities? What are children's reading skills by treatment and District, before initiation of activities? At post-test Does receiving an offer to attend READ Core Plus improve students’ literacy skills as compared to children who only received an offer to attend READ Core?1 Does participating in READ Core Plus improve the literacy skills of students as compared to those who participate in READ Core?2 Method Participants and Sample Program participants are children in first, second and third grade who are at high risk of negative educational outcomes in Bangladesh. All children are benefiting from the READ Core intervention. Within the 805 schools that are receiving READ Core, 80 schools were randomly selected to benefit from the READ Core Plus model. The communities within the catchment area of a READ Core Plus will obtain additional supports such as weekly reading circles, a community book bank, and supplemental community-based reading activities. Children attending schools that were 1 2 Intent to Treat Analysis Treatment on the Treated Analysis randomly selected to receive the READ Core Plus model are in the treatment group, and children attending schools that only receive the READ Core model are in the control group. To select the sample of our study, we first randomly chose 70 READ Core Plus schools from among the 80 that are receiving the intervention. For each treatment school, we then located a GPS receiving READ Core within the same Upazila, to be used as a control group. Within each selected site, we then proceeded to randomly select 20 first grade students to participate in the study. Therefore, our study sample is comprised by 2800 first grade children in 140 GPS schools: 1400 grade students in 70 schools receiving the READ Core Package, and 1400 students in 70 schools receiving the READ Core Plus Package. Teachers and Head Teachers were also contacted in each participant school, for an approximate total of 140 adult educators. Figure 1 illustrates how the READ study sample was drawn from the population of READ. First, second and third grade students in GPS and NNPS Bangladesh Population Program Participants Point of randomization Study Participants 805 GPS Schools 660 NNPS READ CORE READ CORE+ 725 GPS 80 GPS 1283 grade-1 Children within 70 READ CORE schools (Control) 1221 grade-1 Children within 70 READ CORE+ Schools (Intervention) Figure 1. Population, READ and READ Core Plus program participants and study sample. Table 2 contains a description of students within different participating districts by treatment status and sex. Table 2. Study sample by Treatment Status, Sex and District Baseline Sample Treatment: READ Control: READ Core Treatment Status Core Plus Total District / Sex Male Female Male Female Barisal 108 107 93 91 399 Cox's Bazar 60 64 44 51 219 Dhaka 110 131 120 115 476 Khulna 136 124 120 138 518 Rangpur 124 127 120 122 493 Sylhet 97 95 106 101 399 Total 635 648 603 618 2504 Procedure Research assistants provided participants with written and verbal information about the study and the nature of the tests and questionnaires that they would be answering. Parents and children were informed that participation was voluntary and that their willingness of unwillingness to participate would not affect their relationship with the school. In May and June of 2015, before initiation of activities, research assistants visited selected sites and met individually with students who agreed to participate to administer the background information questionnaire and a literacy test for up to one hour of their time. Head Teachers filled out a questionnaire with information about the school and the community in which the school is located, for up to 15 minutes of their time. Teachers filled out a questionnaire with information about students' attendance, social skills and school engagement, for about one hour of their time. Participants provided their names as they normally do during tests. Their responses were transferred to a digital format, which contains their ID number. A key code connects subject’s data to their identity. Participants did not receive any payment for their participation in the study. Measures Predictors Treatment is the key exogenous predictor of the question of interest, a dummy variable that indicates whether children are in a school that was randomly assigned to be part of the treatment (READ Core Plus =1) or control (READ Core=0) groups. Outcomes Alphabet_Knowledge is a continuous individual level variable that reflects the sum of the number of letters/sounds that children are able to identify from the Bangla alphabet. (See Figure 2 in Appendix) Most_Used_Words is a continuous individual level variable that reflects the percentage of words read aloud correctly from a list of 20 words that are familiar to children (See Figure 3 in Appendix). Phonemic_Awarenessis an individual continuous level variable that reflects the percentage of similar beginning sounds and rhyming words that children are able to read correctly from a list of 20 (See Figure 4 in Appendix). Vocabulary is an individual level variable that reflects the percentage of words that children are able to retrieve correctly when prompted by a question or an image (See Figure 5 in Appendix). Reader is an individual-level dummy variable that reflects whether a child is an independent reader, defined as the ability to read 5 words correctly within the first 30 seconds of a subtest (See Figure 6 in Appendix). Fluency is an individual level continuous variable that reflects the number of words that children who are independent readers are able to read correctly in a minute using a 59-word connected text (See Figure 7 in Appendix). Accuracy is an individual level continuous variable that reflects the percentage of words that children who are independent readers are able read correctly in a 59-word connected text (See Figure 8 in Appendix). Reading comprehension is an individual level variable that reflects the number of comprehension questions (out of 10) that children who are independent readers are able to answer correctly after reading a passage (See Figure 10 in Appendix). Reading_with_Comprehension is a categorical variable that reflects the percentage of children who are 1) readers, 2) readers able to answer correctly more than 50% of reading comprehension questions, 3) readers who are able to answer more than 80% of reading comprehension questions correctly. Social_Strengths is an individual continuous level variable that describes teachers' perceptions of children’s prosocial skills, based on their answers to some items taken from the Strengths and Difficulties Questionnaire (Goodman, 1997). School_Engagement is an individual continuous level variable that reflects students’ emotional and behavioral signs of engagement and disaffection towards school, as reported by teachers based on the School Engagement vs. Disaffection scale (Furrer & Skinner, 2003. Table 2 contains a summary of the measures of our outcomes of interest, specifying their dimensions of interest, the instruments used to collect the data, the number of items and range of each measure, the reliability estimates in the pre-test sample, the respondents and the original instrument from which the items used were taken. Table 3 shows the means, standard deviations, number of items and Cronbach's alpha of key outcomes of interest Table 3. Description of variable, number of subjects, means, standard deviations and Cronbach's alpha of Key Outcomes of Interest Literacy Outcome Description Mean SD Alpha Alphabet knowledge Number of letters/sounds known of 50 31.62 18.06 0.98 Most Used Words Percentage of words read aloud correctly from the most frequently used words (out of 20) 0.32 0.38 0.97 Phonemic Awareness Percentage of similar Beginning words read correctly (out of 20) 0.36 0.24 0.87 Vocabulary Percentage of words retrieved correctly when shown images (out of 15) 0.7 0.23 0.83 Reader A child able to read the oral reading passage independently, defined as reading at least five words correctly in the first 30 seconds of the sub-test. 0.16 0.36 n/a Fluency Percentage of number of words read correctly per minute in a 59-word connected text. (readers only) 17.04 16.32 n/a 0.87 0.1 n/a 0.57 0.25 0.94 Accuracy Reading Comprehension Percentage of words in an 59-word connected text read correctly (readers only) Percentage of comprehension questions (of 5) answered correctly after reading a text read aloud (readers only) Covariates Age is a continuous variable that reflects that students' age in years (mean=6.74, sd=.97). Female is a dummy variable which indicates whether the participant is a girl (1) or a boy (0) (mean=.5, sd=.5). Home_Materials reflects the sum of 8 dummy variables that indicate whether the family has textbooks, magazines, newspapers, coloring books, children books, religious books, comics or internet at the house (mean=.15, sd=.49). SES is an exogenous predictor of socioeconomic status which reflects the sum of eleven dummy variables that indicate whether families have electric appliances such as TV, radio, refrigerator, and other possessions such as land, books and bicycles (mean=4.87, sd=1.88). ECD is a dummy variable that indicates whether children attended an early childhood education program (1) or not (0) (mean=.51, sd=.49). DistrictID is a vector of six dummy variables that indicate whether the school is located in Dhaka (reference category), Barisal, Khulna, Sylhet, Cox's Bazar or Rangpur. School_ID is a categorical variable that identifies all 140 participating schools. Analytic methods In order to answer our first research question and estimate whether there are statistically significant differences in a set of characteristics that were observed after randomization took place, but before initiation of activities, we used OLS regression estimates that adjusted for clustering at the school level to account for nesting of children within schools. The purpose was to determine if groups are balanced in a series of observable characteristics, but not if treatment has an effect on these variables. In fact, the data was collected before initiation of activities and therefore, the treatment could not have any effect on any outcome. Equation 1 describes the formula used for these analyses: (1) Outcomeis=Bois + B1(TREATMENTis) + Ɛi+ Us Outcome represents a set of variables that measure background characteristics such as age, gender, SES, ECD, afterschool location, supervision. Treatment reflects whether students are in the READ Core Plus (Treatment) or READ Core (control group). Level-1 and Level-2 population residuals are represented by 𝜀𝑖𝑗 and 𝑢𝑗 (with the latter being a random school intercept), each included in the model under the usual normal-theory assumptions. The parameter of interest is β1. If the estimated value of this parameter is not statistically significant, we will be able to conclude that there are not statistically significant differences between children in the treatment and control groups. In order to answer our second research question and identify the average baseline literacy skills of children by treatment status and district, after controlling for demographic information, we used multi-level models with school random effects and district fixed effects. Given the presence of floor and ceiling effects on some outcomes (e.g.: Alphabet knowledge, Most used words, Phonemic awareness, Vocabulary, Accuracy, Fluency), we used Tobit regressions (Tobin, 1958) to censor the data and obtain a more accurate average children performance for those outcomes. A typical equation is given: (2) Outcomeis=Bois + B1(TREATMENTis)+ αYis + γXis+ Ɛi + Us Outcome is a continuous variable that measures an individual score in a given skill that was observed before initiation of program activities:1) Alphabet knowledge; 2) Most Used Words; 3) Phonemic Awareness, 4) Vocabulary, 5) Reader, 6) Accuracy, 7) Fluency, 8) Reading Comprehension. Treatment reflects whether students are in the READ Core Plus (Treatment) or READ Core (control group). YS is a vector of covariates that include gender, grade, SES, home literacy and ECD attendance. Level-1 and Level-2 population residuals are represented by 𝜀𝑖𝑗 and 𝑢𝑗 . The parameter of interest is β1. If the estimated value of this parameter is not statistically significant, we will be able to conclude that at baseline, there are not statistically significant differences between children in the treatment and control groups in our key literacy outcomes of interest. Results Research Question 1: We conducted 50 regressions for a set of observable variables that were measured before initiation of activities. All models adjust for clustering so the coefficients, t-tests and p-values are all conditional on school effects. Table 4 shows regression results for the differences in means between READ Core Plus and READ Core for 22 exogenous variables. Exogenous variables are factors that cannot be changed by the actions of the participants, such as age, gender, SES, teacher’s level of education, school size, etc. Our analyses show that an alpha level of .05, there aren’t statistically significant differences between the treatment and control groups in exogenous characteristics that were measured before the implementation of the program. Table 4. Regressions showing differences in means (READ Core Plus minus READ Core), t-tests and constants for a set of exogenous variables that were measured at baseline Dimension Variable Male Age Demographic info SES Home literacy Teacher's age Teachers' gender Teacher Teacher's experience Characteristics Teacher's professional education Head teacher's Gender HeadTeacher Head teachers' experience School Resources School Size School Characteristics Number of teachers Std-Teacher Ratio Student Behavior Attended ECD Repeat Preschool Student Academic Trajectory Repeat First No Repetitions At school Reading materials At the Community *p<0.05 **p<0.01 *** <.001 Treatment (Difference in means) T-test Constant 0 0.01 0.01 -0.03 0.01 0.04 0.11 0.07 0.19 0.18 0.07 0.34 1.92 0.26 0.5 6.69 4.9 1.07 2.97 1.31 10.22 -0.09 0.07 0.06 -0.3 -4 -0.53 0.54 -0.03 -0.01 -0.16 0.02 0 0.007 -0.08 -1.18 0.92 0.05 0.2 0.35 -1.57 0.7 -0.48 -0.47 -0.88 1.07 0.01 0.27 -0.6 5.66 1.67 9.45 3.6 51.22 5.81 8.96 3.22 0.52 0.14 0.12 0.73 1.09 3.57 Table 5 shows regression results from a set of 28 endogenous characteristics that were measured before the implementation of the program. Endogenous variables are characteristics that are determined by the actions of participants and therefore, are subject to change during the implementation. For example, test scores, self-reports, school attendance, afterschool activities, etc. Given that children were randomly assigned to treatment and control groups, changes in endogenous variables observed after the implementation of the program can be interpreted as program outcomes. The results show that at an alpha level of .05 there aren’t statistically significant differences between the treatment and control groups in endogenous characteristics before initiation of activities. Table 5. Regressions showing differences in means (READ Core Plus minus READ Core), t-tests and constants for a set of endogenous variables that were measured at baseline. Dimension Variable Treatment (Difference in means) T-test Constant Student Academic Trajectory Miss school last week Study time -0.03 -0.94 -0.9 -0.33 0.44 120.22 Reading materials Home 0.04 1.17 0.13 -0.006 -0.01 0.009 -0.005 -0.002 -0.002 0.001 -0.38 -0.61 0.59 -0.35 -0.77 -0.26 0.22 0.09 0.82 0.12 0.12 0.005 0.95 0.32 0.003 1.18 0.003 -0.01 0.002 0.05 0.01 -0.06 0.08 -0.02 -0.08 -0.77 0.04 -1.78 0.93 0.42 0.19 -0.27 0.52 -0.26 -0.63 -0.52 -0.21 0.02 0.003 4.92 0.47 3.27 5.63 0.34 1.52 2.15 3.44 0.18 0.81 1.78 -0.19 -0.07 0.01 -0.01 -0.89 -0.43 0.22 -0.32 3.13 3.48 0.21 0.36 Afterschool Supervision Afterschool location Afterschool Activities Self-Care Parent Care Non-Parent Adult Care Sibling Care Mixed-Care Own Home Someone else's home School or other place for structured activities Somewhere to "hang out" Mixed location Doing homework Attending Afterschool program Receiving Private Tutoring Reading books at home Reading at community program Music, dancing, art Organized sports Playing games that are not sports Activities at mosque, temple, church Watching TV Hanging out with friends Volunteering at the community Working at a job Doing house chores Taking care of a sibling *p<0.05 **p<0.01 *** <.001 0.57 -0.003 1.01 -0.02 4.78 3.04 Figure 2 shows the percentage of children who report having different reading materials at home, school and the community, by treatment status. The regressions indicated that there are no differences between children in the treatment and control groups in those categories. Within those categories, we also do not observe differences, but can appreciate that a great percentage of children have texts available in school, but not at home or the community. Most other reading resources are available to them in the community, but not at home or school. Home School Control Internet Comics Coloring Books Children's books Newspapers Magazines Religious Texts Internet Comics Coloring Books Children's books Newspapers Magazines Religious Texts Internet Comics Coloring Books Children's books Newspapers Magazines Religious Texts 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% Community Treatment Figure 2. Percentage of children who report having different types of reading materials at home, school and the community, by treatment status. Research Question 2: At baseline, what are children's average literacy skills by treatment and District? Are there significant differences in literacy outcomes between children who received an offer to participate in READ CORE Plus and children who did not receive the offer? We did not find any statistically significant differences in the baseline literacy outcomes of children attending schools who received the offer to participate in READ Core Plus and those attending schools who received the offer to participate in READ Core. As can be seen in Table 6, there aren´t any differences in the alphabet knowledge, recognition of most used words, phonemic awareness, vocabulary, percentage of readers, fluency, accuracy, and reading comprehension skills of children in schools assigned to different treatment status (p>.001). However, we do observe a statistically significant differences in reading with comprehension which favors children in the treatment group (p<.01) (To see full regression models including covariates see Appendix. Table 7 includes models for low order literacy skills. Table 8 includes models for high order literacy skills). Table 6. Coefficients, standard errors and p-values for regressions of treatment on the key outcomes of interest, adjusting for clustering at the school level and controlling for demographic characteristics and District Fixed effects Literacy Outcome Coefficient Standard error p-value N Alphabet Knowledge -0.53 0.99 0.59 2177 Most Used Words -0.02 0.05 0.60 2177 Phonemic Awareness -0.01 0.02 0.76 2177 Vocabulary -0.01 0.01 0.68 2177 Reader -0.01 0.02 0.50 2177 Fluency 0.87 1.28 0.50 423 Accuracy 0.01 0.01 0.60 423 Reading Comprehension 0.10 0.06 0.07 423 Reading with Comprehension .118 0.04 0.01 423 *p<0.05 **p<0.01 *** <.001 Figure 3 shows average alphabet knowledge of children by treatment status and district. Specifically, we did not observe any statistically significant differences between children in the treatment and control groups in their alphabet awareness. On average, children in Barisal show the highest scores of all groups, but differences are only statistically significant as compared to those in Sylhet. Children in Sylhet exhibit scores that are statistically significantly lower (p<.05) than the scores in children from all other districts, with the exception of Cox’s Bazar, where children have similarly low scores. Alphabet Knowledge Number of letters in the Bangla alphabet identified correctly 50 45 40 35 30 25 20 15 10 5 0 Barisal Cox's Bazar Khulna Rangpur Sylhet Dhaka AVERAGE READ Core+ 36.31 31.36 32.98 32.04 27.85 32.29 32.14 READ Core 37.21 32.26 33.89 32.95 28.76 33.19 33.04 READ Core+ READ Core Figure 3. Average Alphabet Knowledge by Treatment Status and District (n=2177) Figure 4 shows the distribution of scores in the letter knowledge of children according to their district. In addition to a significant amount of zero scores, which reflect children who were unable to identify any letter in the alphabet, we observe a distribution that indicates that many children are able to identify approximately 45 letters in the alphabet and many of them are able to identify them all. As can be observed, children in Barisal exhibit the highest probability of being able to identify many letters in the alphabet, while children in Sylhet have the highest probability of being unable to identify any letter in the alphabet as well as the lowest one being able to identify 40 .03 .02 0 .01 kdensity ltrstot .04 .05 letters or more. 0 10 20 30 40 50 x Barisal Dhaka Rangpur Cox's Bazar Khulna Sylhet Figure 4. Histogram of Alphabet Knowledge, by District Figure 5 shows average percentage of most used words that children are able to recognize from a list of 20 familiar words, by treatment status and district. We observe that there aren't any statistically significant differences between children in the treatment and control groups in their ability to read familiar words. Children in Barisal show the highest performance and differences are statistically significant when compared to all other districts (p<.01). Children in Sylhet exhibit the lowest scores of all districts, but differences are only statistically significantly when compared to the performance of children in Barisal (p<.001). Most Used Words Percentage of most used words read correctly 60% 50% 40% 30% 20% 10% 0% Barisal Cox's Bazar Khulna Rangpur Sylhet Dhaka AVERAGE READ Core+ 54.46% 33.78% 32.67% 26.08% 26.19% 27.40% 32.63% READ Core 57.01% 36.32% 35.21% 28.62% 28.74% 29.95% 35.18% READ Core+ READ Core Figure 5. Percentage of Familiar Words Recognized by Treatment Status and District (n=2177) Figure 6 shows the distribution of scores for most used words, according to district. Once again, the distribution shows a great amount of children –especially in Sylhet- have a high probability of being unable to identify familiar words. We also observe that children in Barisal have the lowest probability of being unable to recognize familiar words, as well as the highest probability of being able to identify them. 4 3 kdensity muwpct 2 1 0 0 .2 .4 .6 .8 1 x Barisal Dhaka Rangpur Cox's Bazar Khulna Sylhet Figure 6. Histogram of Percentage of Most Used Words, by District Figure 7 shows average percentage of vocabulary words that children are able to retrieve after prompted by questions and images, by district and treatment status. We observe that there aren't any statistically significant differences between children in the treatment and control groups in their ability to retrieve vocabulary words. Children in Barisal exhibit the largest vocabulary words and differences are statistically significant when compared to all other districts (p<.001). Children in Dhaka also show a vocabulary that is statistically superior to the vocabulary of children in other districts (p<.001), but inferior to the vocabulary of children in Barisal (p<.001). Children in Rangpur exhibit the lowest performance of all districts, but differences are only statistically significant when compared to the higher performance of their peers in Dhaka (p<.001) and Barisal (p<.001). Vocabulary 100% Percentage of vocabulary words identified 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% Barisal Cox's Bazar Khulna Rangpur Sylhet Dhaka AVERAGE READ Core+ 83.54% 67.00% 64.64% 64.49% 68.59% 75.75% 70.43% READ Core 84.51% 67.97% 65.61% 65.46% 69.55% 76.72% 71.39% Axis Title READ Core+ READ Core Figure 7. Average Percentage of Vocabulary by Treatment Status and District (n=2177) Figure 8 shows the distribution of children who are able to read different percentages of vocabulary words, according to district. We observe that the distribution of children in Barisal and Dhaka have ceiling effects, reflecting the fact that many children are able to identify most vocabulary words as prompted in the test by facilitators. 3 2 kdensity Vocabpct 1 0 0 .2 .4 .6 .8 1 x Barisal Dhaka Rangpur Cox's Bazar Khulna Sylhet Figure 8. Histogram of Percentage of Vocabulary, by District Figure 9 shows average phonemic awareness, defined as the percentage of similar beginning sounds and rhyming words that children are able to read correctly from a list of 20, by district and treatment status. We observe that there aren't any statistically significant differences between children in the treatment and control groups in their phonemic awareness. Children in Cox's Bazar exhibit the highest phonemic awareness of all districts with differences that are statistically significant when compared to all other districts (p<.001). Children in Dhaka exhibit the lowest phonemic awareness, and differences are also statistically significant when compared to the performance of children in all other districts (p<.001). Phonemic Awareness 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% Barisal Cox's Bazar Khulna Rangpur Sylhet Dhaka AVERAGE READ Core+ 44.50% 51.73% 34.43% 34.90% 40.42% 24.00% 38.00% READ Core 45.49% 52.72% 35.43% 35.89% 41.41% 24.99% 39.00% READ Core+ READ Core Figure 9. Average Percentage of Phonemic Awareness by Treatment Status and District (n=2177) Figure 10 shows the distribution of percentage of phonemic awareness, according to district. We observe that the greatest proportion of children in Cox´s Bazar exhibit high phonemic awareness and only a very small proportion exhibit low phonemic awareness. In contract, we observe that the greatest proportion of children in Dhaka are unable to read correctly similar beginning sounds and rhyming words as observed in their high zero scores, as well as the lowest proportion of children who are able to do the same, as compared to all other districts. 2.5 2 1.5 1 .5 0 0 .2 .4 .6 .8 1 x Barisal Dhaka Rangpur Cox's Bazar Khulna Sylhet Figure 10. Histogram of Percentage of Phonetic Awareness, by District Readers Figure 11 shows the average percentage of children who are independent readers, defined as the ability to read 5 words within the first 30 seconds of a reading passage test, by treatment status and district. We observe that there aren't any statistically significant differences between the percentage of readers in the treatment and control groups. On average, readers constitute 18.11% of the READ Core+ group and 19.79% of READ Core group. Barisal has the highest percentage of readers with differences that are statistically significant when compared to all other groups (p<.001) and Rangpur the lowest one percentage of readers, but differences are only significant when compared to outcomes from Barisal. Readers Percentage of children able to read 5 words correctly within 30 seconds 40% 35% 30% 25% 20% 15% 10% 5% 0% Barisal Cox's Bazar Khulna Rangpur Sylhet Dhaka AVERAGE READ Core+ 32.64% 17.45% 18.22% 11.87% 14.36% 15.97% 18.11% READ Core 34.31% 19.12% 19.89% 13.54% 16.04% 17.65% 19.79% Figure 11. Average Percentage of Readers by Treatment Status and District (n=2177) High order literacy skills In the final sub-tests students who qualified as readers were assessed on their fluency, reading accuracy and reading comprehension. This section presents results for the 19% of students who qualified as readers. (See Table 8 in the Appendix) Reading Accuracy Figure 12 shows average reading accuracy, defined as the percentage of words read correctly in a reading passage of 59 words (only for readers), by treatment status and district. We observe that there aren't any statistically significant differences between children in the treatment (READ CORE Plus) and control (READ Core) groups in their reading accuracy. Children in Barisal exhibit the highest reading accuracy of all groups and their performance is statistically when compared with children in Dhaka (p<.001) and Khulna (p<.05), who exhibit the lowest performance of all groups. Accuracy 100% Percentage of passage read correctly 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% Barisal Cox's Bazar Khulna Rangpur Sylhet Dhaka READ Core+ 90.17% 87.48% 84.64% 86.93% 86.06% 83.78% READ Core 88.20% 85.50% 82.67% 84.95% 84.09% 81.81% Figure 12. Average Reading Accuracy of Independent Readers by Treatment Status and District (n=423) Figure 13 shows the distribution of reading accuracy scores by district. We observe that in most districts the distribution that is skewed right, reflecting the fact that, overall, readers are able to read with good accuracy. The distributions shows that Barisal has the highest proportion of children with high reading accuracy as well as the smallest proportion of children with low reading accuracy. A high proportion of children in Rangpur exhibit high reading accuracy, but more children in this district exhibit low reading accuracy scores than in other districts. Both Dhaka and Khulna exhibit a lower proportion of children reading with high accuracy. 5 4 0 1 2 3 kdensity Readpct 0 .2 .4 .6 .8 1 x Barisal Dhaka Rangpur Cox's Bazar Khulna Sylhet Figure 13. Histogram of Percentage of Reading Accuracy, by District Figure 14 shows average fluency of independent readers, defined as the number of words per minute read correctly, by treatment status and district. We observe that there aren't any statistically significant differences between children in the treatment and control groups in their reading fluency, but that there are some differences between districts. The highest fluency average is observed in Sylhet3 but differences are only statistically significant when compared to Khulna (p<.001) and Dhaka (p<.05). 3 To some degree the high average score observed in Sylhet is driven by an outlier which I examined carefully but decided to leave as a legitimate datapoint based on other equally high literacy scores. However, even after taking out the outlier, Sylhet contines having the highest fluency average. Fluency 25 Words per minute 20 15 10 5 0 Barisal Cox's Bazar Khulna Rangpur Sylhet Dhaka AVERAGE READ Core+ 19 21 11 17 21 16 10 READ Core 18 20 10 16 19 14 8 Figure 15. Average Reading Accuracy of Independent Readers by Treatment Status and District (n=423) Figure 16 shows the distribution of reading fluency by district. The distributions show that Sylhet and Cox’s Bazar have the largest proportion of children with high reading fluency. Sylhet exhibits a case of great reading fluency, an outlier that was confirmed to be a legitimate datapoint. We also observe that children in Khulna and Dhaka exhibit large proportions of children reading with low fluency. .08 .06 0 kdensity wcpm .04 .02 0 50 100 x Barisal Dhaka Rangpur Cox's Bazar Khulna Sylhet Figure 16. Histogram of Fluency Scores, by District Reading Comprehension Readers were also assessed on ten reading comprehension questions related to the reading passage. Figure 17 shows students' average reading comprehension, defined as the percentage of reading comprehension questions (out of 10) that readers are able to answer correctly after reading a passage, by treatment status and district. We observe that there aren't any statistically significant differences between children in the treatment (READ Core Plus) and control (READ Core) groups in their reading comprehension. Children in Cox’s Bazar exhibit the highest reading comprehension of all groups, with differences that are statistically significant when compared to Ranpur (p<.01) and all other districts (p<.001). Children in Dhaka exhibit the lowest reading comprehension, and significantly different from the average performance of children in Cox’s Bazar (p<.001), Rangpur (p<.001) and Barisal (p<.01). Reading Comprehension 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% Barisal Cox's Bazar Khulna Rangpur Sylhet Dhaka AVERAGE READ Core+ 59% 80% 53% 65% 51% 47% 59% READ Core 54% 75% 48% 60% 46% 42% 54% READ Core+ READ Core Figure 17. Average Percentage of Reading Comprehension by Treatment Status and District (n=386) Figure 18 shows the distribution of reading comprehension by district. The distributions show that Cox’s Bazar and Rangpur have the largest proportion of children with high reading comprehension and Dhaka the largest proportion of children with low reading comprehension. 2 1.5 1 kdensity comprpct .5 0 0 .2 .4 .6 .8 1 x Barisal Dhaka Rangpur Cox's Bazar Khulna Sylhet Figure 18. Histogram of Percentage of Reading Comprehension Questions Answered Correctly (readers only) by District (n=432) Table 9 in the Appendix shows the regression results for different types of reading comprehension questions. Figure 19 shows the percentage of different types of reading comprehension questions that students were able to answer correctly. Reading comprehension questions included summary, factual, inferential, and evaluative questions. For the summary question students were asked what the story was about and their responses were marked correct if they mentioned at least three of four main points of the story (characters, problem, action, resolution). Six literal questions asked children about information that was directly available in the text, such as "What was the name of the main character?" and "Where did the main character go?". Two inferential questions asked children about information that was indirectly available in the text. Finally, one evaluative question asked children for their opinion of the text, and children’s responses were scored correctly if they justified their opinion with information from the text. We did not observe statistically significant differences between children in the treatment and control groups in their summary, literal and evaluative comprehension, but we did observe differences in their inferential comprehension, which favor the treatment group (p<.05) (See Table 9). Overall, summary questions, literal questions and inferential questions were equally challenging for students, with correct response rates within 55% and 66%. The hardest reading comprehension questions for students in the sample were the evaluative questions, with correct response rate between 35% and 45%. Reading Comprehension by Type of Question 80% 70% 60% 50% 40% 30% 20% 10% 0% Reading Comprehension Summary* Reading Comprehension Literal Reading Comprehension Inferential* Reading Comprehension Evaluative READ Core+ 66% 62% 65% 46% READ Core 55% 61% 56% 38% Figure 19. Reading comprehension questions by (readers only) by Type (n=432) Figures 20-23 shows students’ reading comprehension by types of question (summary, literal, inferential and evaluative) in each district. We can observe that children in Cox’s Bazar exhibit the highest performance of all groups in all types of reading comprehension questions. In all figures, we observe that children in Cox’s Bazar (p<.001) and Rangpur (p<.001) have the highest performance of all groups, with differences that are statistically significant when compared to all other districts, but not among themselves. 2 1.5 1 .5 0 0 .2 .4 .6 .8 1 .8 1 x Barisal Dhaka Rangpur Cox's Bazar Khulna Sylhet 0 .5 1 1.5 2 2.5 Figure 20. Summary Reading Comprehension, by District. 0 .2 .4 .6 x Barisal Dhaka Rangpur Cox's Bazar Khulna Sylhet 0 .5 1 1.5 2 Figure 21. Literal Reading Comprehension, by District 0 .2 .4 .6 .8 x Barisal Dhaka Rangpur Figure 22. Inferential Reading Comprehension, by District Cox's Bazar Khulna Sylhet 1 2 1.5 1 .5 0 0 .2 .4 .6 .8 1 x Barisal Dhaka Rangpur Cox's Bazar Khulna Sylhet Figure 23. Evaluative Reading Comprehension, by District Readers with Comprehension Figure 24 shows the percentage of 1) non-readers, 2) readers, defined as those who are able to read five words correctly within the first 30 seconds, 3) readers with comprehension, defined as those who are able to answer correctly 50% of reading comprehension questions, and 4) readers with advanced comprehension, defined as those who are able to answer correctly 80% or more of reading comprehension questions. We do not observe statistically significant differences between nonreaders. We observe that children in the treatment group have less readers than children in the control group (p<.05). We do not observe any statistically significant differences between readers with comprehension and readers with advanced comprehension in the treatment and control groups. Treatment 85.01% Control 4.18% 6.63% 4.18% 83.01% 0% 10% Non-readers 20% Readers 30% 40% 6.16% 6.08% 4.75% 50% Readers with Comprehension 60% 70% 80% 90% 100% Readers with Advanced Comprehension Figure 24. Percentage of Non-Readers, Readers, Readers with Comprehension and Readers with Advanced Comprehension by Treatment Status Who is falling behind? Using the coefficients of different covariates from the multi-level regression models, we identified associations between the literacy subscales and students’ background information, in terms of the sex, socio-economic status and reading materials available at home. Table 7 presents our findings. We observe that boys and children in lower SES are falling behind girls and children in higher SES in their knowledge of the alphabet, ability to identify most used words, phonemic awareness, vocabulary, ability to read and fluency, but not in terms of their reading accuracy and reading comprehension. Children in house environments with few reading materials are falling behind children living in homes with more access to reading materials in terms of their phonemic awareness, vocabulary and reading fluency. Children who did not attend and ECD are falling behind children who had the opportunity to attend preschool in their alphabet knowledge, ability to identify most used words, and their vocabulary. Table 7: At Baseline, Who is Falling Behind? Sex SocioEconomic Status Reading Materials at home Previous ECD Attendance Alphabet Knowledge Boys*** Low SES*** -- No ECD*** Most Used Words Boys*** Low SES*** -- No ECD*** Phonemic Awareness Boys*** Low SES** Few reading materials** -- Vocabulary Boys** Low SES*** Few reading materials* NO ECD* % Readers Boys*** Low SES** -- -- Boys* Low SES*** Few reading materials*** -- Accuracy -- -- -- -- Reading Comprehension -- -- -- -- Fluency Discussion The present study uses a randomized controlled clustered design to estimate the value added of a community-based literacy intervention (READ Core Plus) over a core school-based literacy packet (READ Core), on a set of literacy outcomes of interest. The present report presents the analyses of the baseline data that was collected in May and June 2016. Analysis indicates that the randomization worked, as we found that children in treatment and control groups are perfectly balanced on a set of observable characteristics. Internal Validity Concerns about internal validity address the question of whether an intervention makes significant differences in the outcomes of participants (Shadish et al, 2002). Given that the present study uses an experimental design where schools were randomly assigned to treatment and control groups, we will be able to make causal inferences about the effects of READ Core Plus on our outcomes of interest. In fact, random assignment effectively eliminates the probability that children in treatment and control groups would differ in observable and unobservable characteristics that may have an impact on the outcomes of interest. A potential threat to the internal validity of an experimental study can be that, despite the random assignment, children in the treatment and control groups would still be different in observable and unobservable characteristics before the initiation of activities. One of the strengths of the present study is that we collected pretest scores, and were able to determine that groups are actually equal in expectation. In fact, after analyzing the baseline data we confirmed that before initiation of activities children in the control and treatment groups did not exhibit statistically significant differences in any observable characteristics. Therefore, any post-test changes in endogenous characteristics can be attributed to the causal effect of the offer to participate in READ Core Plus. Another potential threat to the internal validity of an experimental study can be diffusion effects. Specifically, if children in the treatment group are able to interact with children in the control group, the effects of the intervention may have an effect on them as well. In our study, diffusion effects are of low concern because schools were randomly assigned to treatment and control groups and therefore, children are not in contact with one another. Additionally, schools assigned to treatment and control groups are located within the same Upazila, but with distances that would make it unlikely for children to influence each other. A future threat to the internal validity of our experimental study is that students’ treatment status would change after their initial random assignment. We will address this problem by conducting intent-to-treat (ITT) and treatment on the treated analysis (TOT). ITT analysis examine the effect of being randomly assigned to a treatment group on several outcomes of interest, without regard for whether or not the chosen participants received the treatment or not, and independently of the dosage or amount of hours they actually attended the program. ITT estimates are based on the original intention to provide treatment to a given group of participants, and not to the treatment that they actually received. For this reason, ITT analysis will resemble the effect that READ Core Plus will have in the “real” world, where some children who received the offer to participate in the program will take full advantage of the opportunity, some will take inconsistent advantage and some will not take advantage. TOT estimates will determine the effect of READ Core Plus on children who did not change their treatment status and actually participated in the program. External Validity External validity concerns the inference about the extent to which the causal relationship that we will identify between READ and the outcomes of interest holds over variations in persons, settings, treatments and outcomes that were in the experiment (Shadish et al, 2002). Given the point of randomization used in the experiment-the criteria used to select students to participate in the lotteriesthe findings of the present study will only apply to first grade children in GPS schools. Given that Save the Children, USAID and donors have determined that READ should target schools from poor and marginalized communities in Bangladesh, the findings of this experimental study will only generalize to students at high risk of negative educational outcomes, who live in rural areas of Bangladesh. Specifically, the findings of this study will not generalize to students or settings that are not represented in this sample, such as:1) older students, 2) students from medium or high SES, 3) in urban areas, and 4) who attend Newly Nationalized Schools, Madrasas or private schools, 5) who live in other countries. In this regard, it is very likely that our findings would overestimate the effects of READ on children from more privileged backgrounds. References Murnane, R & Willet, J (2011) Methods Matter. Improving Causal Inference in Educational and Social Science Research. Oxford. Shadish, W. R., Cook, T. D., & Campbell, D. T. (2002). Experimental and Quasi-Experimental Designs for Generalized Causal Inference. Boston, MA: Houghton Mifflin Appendix Table 7. Multi-level regressions for baseline differences between children in the treatment (READ CORE Plus ) and control groups (READ CORE) in a key literacy outcomes, after accounting for demographic characteristics and district fixed effects (n=2177) Alphabet Knowledge Most Used Words Phonemic Awareness Vocabulary Reader Intercept 28.09*** (3.350) 0.227** (0.0741) 0.0803 (0.0444) 0.492*** (0.0416) 0.179* (0.0745) Treatment -0.907 (1.327) -0.0247 (0.0284) -0.00987 (0.0142) -0.00960 (0.0134) -0.0168 (0.0235) Age -0.193 (0.376) -0.0112 (0.00840) 0.0162** (0.00525) 0.0284*** (0.00492) -0.0185* (0.00882) Sex 2.485*** (0.685) 0.0567*** (0.0153) 0.0465*** (0.00970) 0.0227* (0.00908) 0.0522** (0.0163) SES 0.788*** (0.201) 0.0183*** (0.00452) 0.00790** (0.00283) 0.0124*** (0.00265) 0.0168*** (0.00476) 0.181 (0.717) 0.0166 (0.0160) 0.0279** (0.0100) 0.0221* (0.00938) 0.0287 (0.0168) 2.535*** (0.747) 0.0595*** (0.0167) -0.00175 (0.0105) 0.0254** (0.00980) 0.0184 (0.0176) Barisal 4.024 (2.186) 0.269*** (0.0469) 0.205*** (0.0238) 0.0780*** (0.0224) 0.167*** (0.0395) Cox’s Bazar -0.916 (2.773) 0.0652 (0.0593) 0.277*** (0.0294) -0.0874** (0.0277) 0.0147 (0.0487) Khulna 0.721 (2.129) 0.0548 (0.0456) 0.104*** (0.0228) -0.111*** (0.0215) 0.0225 (0.0378) Rangpur -0.223 (2.160) -0.0141 (0.0462) 0.109*** (0.0232) -0.113*** (0.0218) -0.0411 (0.0384) Sylhet -4.427 (2.308) -0.0131 (0.0496) 0.164*** (0.0254) -0.0718** (0.0239) -0.0161 (0.0421) Home Materials ECD Sigma u Sigma e Rho N 0.07 0.24 0.155 2177 Standard errors in parentheses ="* p<0.05 ** p<0.01 0.05 0.24 0.141 2177 0.06 0.22 0.0692 2177 0.05 0.2 0.0709 2177 0.09 0.37 0.0656 2177 *** p<0.001" Table 8. Multi-level regressions for baseline differences between children in the treatment (READ CORE Plus ) and control groups (READ CORE) in key advanced literacy outcomes (only for readers), after accounting for demographic characteristics and district fixed effects. Accuracy Fluency Reading Comprehension Reading with Comprehension Intercept 0.782*** (0.0602) -6.170 (7.325) 0.210 (0.107) 0.0648 (0.197) Treatment 0.0202 (0.0153) 2.703 (2.475) 0.0498* (0.0251) 0.118* (0.0459) Age 0.00567 (0.00766) 1.221 (0.888) 0.0293* (0.0138) 0.0407 (0.0252) Sex 0.0136 (0.0136) 2.885 (1.485) -0.00301 (0.0249) 0.00841 (0.0456) SES -0.000976 (0.00397) 1.485*** (0.444) 0.00384 (0.00723) 0.0112 (0.0132) Home Materials -0.0157 (0.0117) 5.968*** (1.409) -0.0435* (0.0206) -0.0861* (0.0378) ECD -0.00278 (0.0143) 2.872 (1.614) 0.00754 (0.0257) 0.00292 (0.0470) Barisal 0.0636** (0.0242) 4.008 (3.924) 0.123** (0.0398) 0.217** (0.0728) Cox’s Bazar 0.0371 (0.0331) 5.676 (5.591) 0.324*** (0.0529) 0.486*** (0.0970) Khulna 0.00757 (0.0249) -5.584 (4.009) 0.0596 (0.0409) 0.169* (0.0750) Rangpur 0.0312 (0.0270) 1.815 (4.212) 0.181*** (0.0452) 0.343*** (0.0828) Sylhet 0.0224 (0.0315) 10.29* (4.679) 0.0362 (0.0539) 0.106 (0.0987) 0.03 0.13 0.0598 423 10.24 13.92 0.351 423 0 0.25 0 423 0 0.45 0 423 Sigma_U Sigma_e Rho N Table 9. Multi-level regressions for baseline differences between children in the treatment (READ CORE Plus) and control groups (READ CORE) in different types of reading comprehension outcomes (only for readers), after accounting for demographic characteristics and district fixed effects. (n=423) Summary Comprehension Literal Comprehension Inferential Comprehension Evaluative Comprehension Intercept -0.0810 (0.201) 0.283** (0.105) 0.198 (0.167) 0.0879 (0.192) Treatment 0.119* (0.0468) 0.0169 (0.0244) 0.0957* (0.0390) 0.0862 (0.0448) Age 0.0739** (0.0257) 0.0274* (0.0134) 0.0261 (0.0214) 0.00254 (0.0246) Sex -0.00642 (0.0465) 0.0164 (0.0242) -0.0728 (0.0387) 0.0234 (0.0445) SES 0.000616 (0.0135) -0.000285 (0.00704) 0.00944 (0.0113) 0.0206 (0.0129) Home Materials -0.0457 (0.0385) -0.0478* (0.0201) -0.0385 (0.0321) -0.0255 (0.0369) ECD -0.0313 (0.0480) 0.00569 (0.0250) 0.0235 (0.0400) 0.0256 (0.0459) Barisal 0.0882 (0.0743) 0.111** (0.0387) 0.146* (0.0619) 0.182* (0.0711) Cox’s Bazar 0.387*** (0.0989) 0.311*** (0.0515) 0.319*** (0.0824) 0.352*** (0.0947) Khulna -0.0141 (0.0764) 0.0936* (0.0398) 0.115 (0.0637) -0.181* (0.0732) Rangpur 0.378*** (0.0844) 0.108* (0.0440) 0.268*** (0.0703) 0.248** (0.0808) -0.0521 (0.101) 0.0496 (0.0524) -0.00624 (0.0838) 0.129 (0.0963) 0 0.47 0 423 0 0.24 0 423 0 0.39 0 423 0 0.45 0 423 Sylhet Sigma_U Sigma_e rho N