Skip to main content
Table of Contents

Methodology

The CART analysis is different in two key ways:

First, an algorithm selects the circumstances that shape the furthest behind and furthest ahead group, so the results are free of human bias and preconceptions. For each country and each indicator, a unique LNOB tree is produced and therefore the furthest behind in one development area (e.g. stunting in children) are not the same as those in another area (e.g. overweight in children).

Second, the CART analysis captures situations when advantages or disadvantages intersect. Frequently, circumstances overlap to create pockets of advantage or disadvantage. Without the CART analysis, revealing these intersections would require multiple calculations and group comparisons to identify who the furthest behind group is in each indicator.

The D-Index for barrier indicators (eg. stunting) does not make any sense and we never use it. Instead, we produce D-Index for the equivalent "non-barrier" indicators (eg. non-stunting), and we show that D-Index instead (D-Index for non-stunting). This is sometimes depicted with a (*) next to the indicator (under "Other indicators" in the Analysis tab).

It is possible to add a new indicator in the analysis.

The first step is to check if the indicator of interest is available in the DHS or MICS in your country. If yes, then you can proceed to add it yourself to the analysis (if you are interested in learning how, email us to join a training on using the R code at: escap-sdd@un.org) or you can contact us to request the analysis with the new indicator.

When a new indicator is explored, it is important to define the reference population of interest (is the indicator referring to women, households, children, men?) and the circumstances that most likely affect access (mother's education, sex, wealth of the household etc.). Once these parameters are defined and relevant data are cleaned, the code can be run to produce the CART and D-Index for this new indicator.

For more information, please see our reference guide: https://www.unescap.org/resources/leave-no-one-behind-decision-tree-user-reference-guide

The LNOB analysis (CART & D-Index) is recalculated when additional relevant circumstances are available. To date, ethnicity and religion are used as default and when available. When ethnicity is not available, we use language as a substitute. If both ethnicity & language are available, we prefer to use ethnicity.

For the “including Religion, Ethnicity and/or Language” analysis, the minimum size of the end nodes has been lowered from 9% to 5% (of the population), to be able to capture subgroups that are smaller in size. Therefore, variation in tree formation may be seen, even if the additional variables "ethnicity", "religion" and "language" or "caste" do not actually show up in the tree. These trees which allows smaller subgroups often are more complex and therefore perhaps more difficult to interpret, but they do provide more granular information.

There are several splitting methods for trees. For our data and desired output, we use analysis of variance (ANOVA), because we are not interested in prediction (at the end nodes), but rather a proportion of "1" (share of the end node group that has access to our indicator of interest). ANOVA calculates the average in each node (in our case, the proportion of "1"). To determine how many splits are made, we use the Complexity Parameter (cp), a threshold that tells our the algorithm when to stop splitting. You can read more about the statistics here: https://cran.r-project.org/web/packages/rpart/vignettes/longintro.pdf.

Platform

If you are interested in learning more about the ESCAP LNOB methodology, please navigate to our Training tab and follow the materials that are available. If you are a policymaker or researcher form one of our featured countries, please write to us (escap-sdd@un.org) so we can alert you about any upcoming live trainings (live trainings are open to all, but are primarily geared towards policymakers from the countries in question).

If you want to download the results of the ESCAP LNOB analysis, first select your country, year and indicators of interest, either in the "Analysis" or in the "Overview Results" pages. Once the results have loaded, you will notice a small download symbol at the top right of every graph and figure. Click on the download symbol. Two options for download will appear, for png or csv formats.

If you want to skip through all the analysis on the furthest behind (insightful LNOB trees, box plots, maps and graphs, which are all in the "Analysis" page), you can navigate directly to the "Overview Results" tab. There, you will find a drop down menu with three preselected filters (Indicator, Country and Year - 2019). Remove these preselections and put yours instead, then click "Go". The resulting table will produce all of the furthest behind groups for all the indicators you selected, in the years you selected. You will see the composition of the furthest behind groups, the size and access rate of this furthest behind group by indicator, and other interesting summary information. You can sort or download these results. Done!

General

The LNOB analysis found on this platform is designed to help policymakers and development practitioners access a quick, data-driven, reliable overview of who are the groups left furthest behind in a range of development areas in their country/ies of interest.

The ESCAP LNOB analysis uses microdata from recent (after 2010) Demographic and Health Surveys (DHS) and Multiple Indicator Cluster Surveys (MICS). If microdata from these surveys are not publicly available, we are not able to undertake the analysis. If you have access to other relevant microdata from your country, please contact us at escap-sdd@un.org

The ESCAP LNOB analysis focuses on 16 indicators, mostly aligned with the SDG indicators, drawing from available DHS and MICS microdata. Depending on whether the survey in question used for your country is DHS or MICS, some indicators may be missing. For example, indicators on perpetrated violence against women (VAW) are usually available only in DHS. Indicators on early childhood education have been mostly available in MICS. However, new DHS and MICS rounds tend to include increasingly more SDG-relevant questions, so there is hope for the future!

The platform only shows trees where average access to an opportunity is <99% or average barrier is >1%. It would be difficult to make meaningful splits for an indicator that the vast majority of people have access to.

This question is at the core of the LNOB analysis! Indicators are the "response variables" we are called to observe and monitor, such as access to electricity (an opportunity) or stunting in children (a barrier). Circumstances are the determinant factors or "independent variables" that we assume should NOT matter in shaping access to indicators. In reality, we see that they DO matter, and therefore the LNOB trees create subgroups, based on various interactions of these circumstances, that shape the groups that are furthest behind, or ahead.

Our indicators are split into two types: opportunities, that everyone should have access to (such as access to electricity), and barriers, that no one should experience (such as stunting in children). Our analysis, colour schemes and data insights are customized to reflect this fundamental difference in these two types of indicators.