COVID-19 interactive data dashboard

This dataset includes observations covering six states (AP, Bihar, Jharkhand, MP, Rajasthan, UP) and three survey rounds. We did not have a single, unified frame from which to sample phone numbers. The final sample was assembled from several different sample frames, and the choice of frame sample frames varied across states and survey rounds. These frames comprise four prior IDinsight projects, and from an impact evaluation of the National Rural Livelihoods project conducted by the Ministry of Rural Development Each of these surveys sought to represent distinct populations, and employed idiosyncratic sample designs and weighting schemes.

Key features of each sample frame are summarised below:

Poshan Abhiyan Monitoring (PAM): Two-stage stratified cluster random sample.
Poshan Abhiyan SBCC Monitoring (SBCC): Two-stage stratified cluster random sample. Sample size (with phones/total): Representativeness: representative at the state level of all households with pregnant or lactating mother listed by an ASHA Anganwadi Workers (AWW)
State of Aadhaar Report (SOAR): Three-stage stratified cluster random sample.
National Rural Livelihoods Programme (NRLP): Sample was collected by the Deendayal Antyodaya Yojana-National Rural Livelihoods Mission (DAY-NRLM) and covers 9 states that were part of the National Rural Livelihoods Programme.
Soil Health Card (SHC) (Used for round 1 only): Within UP, we selected 4 districts that represent literacy and mobile phone penetration heterogeneity, as well as agro climatic variation. We then selected 24 villages from each district, and approximately 20 farmers per village. The farmers were selected by a random walk using WHO Extended Programme on Immunization (EPI) protocol.

For some state/rounds we randomly selected a subset of households from the frame of phone-owning households. Based on prior experiences we anticipated a 60% response rate. We took a subsample in those instances where we expected applying the 60% response rate to the sample would yield an adequate sample size. For instance, in round 2 we restricted the sample to 1500 households in Jharkhand and Rajasthan, which should have yielded approximately 900 completed surveys. In other cases we did not sample from phone owning households because the size of the frame of phone owners was small. For instance, in all rounds for both frames in AP we included every available phone number.

In order to create comparable state-level estimates from the successfully interviewed households – as well as to create correctly pooled estimates across the six states – we have applied weights to the information provided by the sampled households. The weights were calculated in several steps. Due to the variation in sampling frames and sampling procedures across states and across rounds, the precise weight procedures tend to be idiosyncratic to a given state/frame/round combination. The procedure detailed below gives a generalized set of steps, and notes significant state/frame deviations from the process.

1. Rescale base weights: Base weights reflect a probability of selection into the original sample, and can be interpreted as an expansion factor to some population. In cases where the sample frame does not include a base weight or a means of calculating a probability of selection equal weights are assigned to each unit in the frame. State-wise details follow:
   • Bihar: The sample frame used in round 1 provides a base weight. The sample frame used for rounds 2 and 3 do not provide a base weight, nor does it provide a means of estimating probability of selection.
   • Madhya Pradesh: The sample frame used in round 1 provides a base weight. The sample frame used for rounds 2 and 3 do not provide a base weight, nor does it provide a means of estimating probability of selection.
   • Uttar Pradesh: The sample frame used in rounds 1, 2, and 3 do not provide a base weight, nor does it provide a means of estimating probability of selection.
   • Andhra Pradesh: The source datasets provide base weights for all three rounds, which expand to the state population (SOAR) or the population of households listed on ASHA/AWW rosters (SBCC).
   • Jharkhand, Rajasthan: The source dataset provides base weights which expand to full district populations. These districts were chosen with PPS from geographic strata to represent the rural population of the two states.
2. Apply an adjustment for noncoverage (i.e. selection into the frame): Next, create weighting classes within districts based on socioeconomic covariates known from respective sample frames. Within these weighting classes sum the base weights, and divide this into the sum of the weights for “covered” households (i.e. those with a mobile phone). Next, multiply the base weights from by the inverse of these proportions. State-wise details follow:
   • Bihar, Madhya Pradesh, Uttar Pradesh: We define the weighting classes by two categorical variables from the NRLP data: caste (SCST/OBC/General) and household income (five quintiles).
   • Andhra Pradesh: We define weight classes by caste (SCST/OBC/General), and PPI quintile (if the household is from the SBCC sample) or ration card status (if the household is from the SOAR sample).
   • Jharkhand, Rajasthan: Unfortunately weighting class adjustments are not possible in these states due to insufficient covariate data.
3. Apply an adjustment for selection into the sample: In certain states (Bihar, Jharkhand, Rajasthan) we randomly selected a subset of the covered households. In these cases we apply an adjustment factor to the weight to reflect each unit’s probability of probability of selection.
4. Rescale the weights from step 3 to account for nonresponse: The correction for nonresponse also employs weighting classes, and follows the steps from 2 exactly. Nonresponse adjustments are applied to the output of step 3.
5. Calibrate the weights from step 4
   • Process: Poststratification attempts to correct for known differences between a sample and a target population. The process entails adjusting the sample weights such that their sum within each poststratum equals the known population total for that poststratum. Specifically, we use the raking to margins method, which recursively recalibrates weights to marginal totals of the chosen poststratification covariates until the weights converge.
   • Data: We define poststrata for SCST status and religion using population totals (household level) for caste and religion from the 2011 population census.

Instrument: The survey questionnaires covered the following subjects.

• Agriculture: COVID-19-related changes in price realisation, acreage decisions, input expenditure, access to credit, access to fertilisers, etc.
• Income and consumption: Changes in wage rates, employment duration, consumption expenditure, prices of essential commodities, status of food security etc.
• Migration: Rates of in-migration, migrant income and employment status, return migration plans etc.
• Access to relief: Access to in-kind, cash and workfare relief, quantities of relief received, and constraints on the access to relief.
• Health: Access to health facilities and rates of foregone healthcare, knowledge of COVID-19 related symptoms and protective behaviours.
  
  While a number of indicators were consistent across all three rounds, questions were added and removed as and when necessary to account for seasonal changes (i.e: in the agricultural cycle).

The India COVID-19 surveys were conducted using Computer Assisted Telephone Interview (CATI) techniques. The household questionnaire was implemented using the CATI software, SurveyCTO. The software was deployed through surveyors’ smartphones, who called respondents via mobile, and recorded their responses over the phone. If unreached, surveyors would attempt to call back respondents up to 7 times, often seeking explicit appointments for suitable times to avoid non-responses. Validation and consistency checks were incorporated into the SurveyCTO software to avoid human error. Extreme values and outliers were scrutinised through a real time dashboard set up by IDinsight. Surveys were also audio audited by monitors to check for consistency and accuracy of question phrasing and answer recording. Finally, supervisors also randomly back-checked a subset of interviews to further ensure data accuracy. IDinsight cleaned and labelled the data for further processing and analysis. Johns Hopkins University examined the data for discrepancies and errors and merged the dataset with their proprietary spatial data. All personally identifiable information has been removed from the datasets.