Constructing Practical and Realistic Asset-Based Socioeconomic Status Assessment Scale Using Principal Component Analysis for Urban Population of Puducherry, India.

BACKGROUND: Socioeconomic status (SES) is a key determinant of health. However, ascertaining the SES in developing countries is really challenging. Hence, we decided to develop an asset-based simple and rational SES tool for urban population of Puducherry and compare it with Modified Kuppuswamy's (MK) scale.
MATERIALS AND METHODS: Sequential mixed methods design was used. The list of local household assets to determine SES was created based on group interviews with stakeholders and review of literature. Then, survey was carried out among 500 urban households by trained medical interns after obtaining informed consent. EpiCollect-5, mobile-based software, was used to capture data. Principal component analysis (PCA) was carried out to construct a wealth index using SPSS version 24. The assets included in the final PCA were ranked based on their contribution to the index by linear regression.
RESULTS: The eigenvalue for the first principal component was 6.7 accounting for 33.6% of the variance in the original data. Finally, reduced 10-item-based SES scale was created and scoring system was formulated based on regression coefficient. The weighted kappa statistics and correlation coefficient measure of reliability between household quintiles on 20-item and 10-item reduced SES tool were 0.77 and 0.95, respectively. There was a moderate correlation between SES obtained from MK scale and newly constructed scale.
CONCLUSIONS: The newly devised SES scale is context specific, reliable, easy to administer, and quick to ascertain the SES and thus can be used for a similar context in future health research. Copyright:

INTRODUCTION

Socioeconomic status (SES) is a comprehensive term relating to the social and economic factors that determine the position of individuals or groups within a hierarchical society.[12] The SES influences the accessibility, affordability, acceptability, and utility of available health facilities and hence an important determinant of health status of people.[3] SES is measured to ascertain the household development and to decide for the provision of social security and welfare services and to avail various SES-based government incentives. Consumption expenditure, education, income, occupation, participatory wealth ranking, and subjective measures were commonly used in epidemiological investigations to ascertain the SES, however, each has its own limitations.[45678]

Nationwide representative surveys conducted all over the world use wealth index which is a household asset-based index as a measure to classify SES of people using principal component analysis (PCA). This has a higher predictive value than other methods of measuring SES.[79] However, data collection is time consuming and analysis and interpretation requires expertise. Notably, some of the assets used in the nationwide survey might not be context specific and outdated. Therefore, we decided to construct a household wealth quintile to ascertain the SES of urban population of Puducherry using the assets that are available locally and specific to urban context by PCA method. We also aimed to construct an SES tool with reduced assets and to check its correlation with the wealth quintile derived from the comprehensive list of assets and to compare the finding with commonly used modified Kuppuswamy's (MK) scale.

MATERIALS AND METHODS

Study setting

The study was planned and carried out by the epidemiology unit of the department of community medicine in a tertiary care teaching hospital in Puducherry. The data collection was done in Villianur, a town that comprised approximately of 2000 households, which is the headquarters of the Villianur taluk in the Union Territory of Puducherry.

Study design

It is a sequential exploratory mixed method design, having two phases which are sequential in nature.[10] In the first phase, the qualitative research method was carried out and it was followed by the quantitative phase [Figure 1]. This study was a part of a larger study on health insurance and its determinants among middle-class family members and the Institute Ethical Committee clearance (Code no.: 51/2017) was obtained along with the main study.

Figure 1

Flow diagram explaining the mixed-methods design and the steps followed in principal component analysis and validation of the study

Sample size and sampling

As a rule of thumb, a minimum of 10 participants per variable is necessary to avoid the statistical difficulties in PCA analysis. We included a sample size of 500 which was considered an adequate sample that could yield meaningful interpretation statistically.[11] Systematic random sampling was adopted to identify the households. The sampling interval was calculated to be four (total household/desired sample size).

Data collection

Phase-I (qualitative data collection and analysis)

Group interviews and free listing of assets were carried out among 10 study participants who were the head of households, mothers of under-five children, Anganwadi teachers, field workers, and staff nurses available in the study area. They were selected purposively on the basis of their perceived experience with the subject of interest (ability to discriminate SES based on the availability of household assets), willingness to participate, and the ability to communicate their opinions in an expressive and reflective manner.[12] The participants were informed to list the household assets that can discriminate the SES of the family. The assets used in the recent National Family Health Survey (NFHS-4) was used to assist participants to arrive at a comprehensive list.[13] The listed items were manually summarized and the commonly reported assets were included in the questionnaire. The same was then face validated with the field experts and was then pilot tested and finalized. Eventually, a 37-item questionnaire was developed to be used in the next phase of the study. The items are mentioned in Table 1.

Table 1

Description of final 37 assets included in the survey (n=500)

Assets	n (%)
Water usage
Disinfected drinking water	323 (64.6)
Wall-mounted water heater	207 (41.4)
Motor to pump groundwater	303 (60.6)
Cooking related
LPG as cooking fuel	497 (99.4)
Induction stove	239 (47.8)
Rice cooker	217 (43.4)
Microwave oven	58 (11.6)
Wet grinder (self-owned)	402 (80.4)
Wet grinder (government supply)	259 (51.8)
Mixer grinder (self-owned)	403 (80.6)
Mixer grinder (government supply)	265 (53)
Electronic goods
Refrigerator	473 (94.6)
Washing machine	326 (65.2)
Dish wash machine	8 (1.6)
Air conditioner	259 (51.8)
Inverter	150 (30)
Iron box	437 (87.4)
Ceiling fan	494 (98.8)
Recreational items
Black-and-white TV	1 (0.2)
Color TV (old model)	217 (43.4)
Color TV (LCD/LED)	241 (48.2)
Cable connection (set top box)	49 (9.8)
Telecommunication
Landline	133 (26.6)
Mobile phone (basic model)	257 (51.4)
Mobile phone (smart model)	383 (76.6)
Desktop/laptop (self-owned)	207 (41.4)
Laptop (government supply)	15 (3)
Internet connection	213 (42.6)
Housing
Own living house	330 (66)
Pucca type of living house	463 (92.6)
Availability of own land	141 (28.2)
Transportation
Bicycle	177 (35.4)
Scooter	230 (46)
Bike	371 (74.2)
Car	141 (28.2)
Others
Sewing machine	121 (24.2)
Livestock in home	29 (5.8)

LPG: Liquefied petroleum gas, LCD: Liquid crystal display, LED: Light emitting diode

Phase-II (quantitative study)

The questionnaire containing 37 assets identified in phase-I and MK scale was used to conduct a community-based survey. In the randomly selected household after obtaining written informed consent, the homemakers (women) were interviewed as they are better aware of the assets available. The survey was carried out by postgraduate and medical interns in the department of community medicine. Information was captured using an efficient mobile application, EpiCollect version 5.0.

Statistical analysis

Descriptive analysis of assets was done in frequency and percentage. PCA technique adopted by the World Bank was used to develop asset-based socioeconomic index or wealth index.[14] Correlation between the assets is used to generate a set of uncorrelated principal components.[15] Since the aim is to construct a single measure of SES, the first principal component was used to define the asset index.[16]

Households were divided into quintiles based on the composite wealth index score. Linear regression was then carried out to find the contribution of each asset in determining the wealth index. Again the PCA was carried out to find out the various composite wealth indices and wealth quintiles using the top 12, 10, 8, and 6 assets that were ranked based on regression coefficient. Pearson's correlation and weighted kappa agreement were carried out for statistical comparison of the quintiles generated using a different number of assets. Finally, a reliable and simple SES tool with reduced assets that can be quickly used in the field was prepared. The scoring system was developed based on the regression coefficient obtained from linear regression. The SES obtained based on a new reduced asset-based scale was compared with SES of the MK scale. Data analysis was performed in software SPSS version 24 (SPSS Inc, Chicago, IL, USA).

Selection of assets to run principal component analysis

The process of the final selection of assets to include in the PCA analysis is depicted in Figure 1. Assets that are correlated and unequally distributed between households will only be able to discriminate the SES better using PCA analysis.[16] Hence, five assets that were present in more than 95% or <5% of the households were excluded, following which preliminary PCA with 32 items was done. The correlation matrix of the PCA analysis which showed eight variables had either high (r > 0.80) or poor correlation (r < 0.10) and as a rule, these assets can also be not able to differentiate SES, so they were also removed.[15] Again, PCA with 24-items was carried out. Then, the wealth quintile generated was cross-tabulated with the assets included in the PCA. Four more variables that had meaningless distribution with wealth quintiles were also finally excluded [Figure 2]. After multiple PCAs and reiterations, 20 assets were ultimately identified to be able to discriminate the SES better.

Figure 2

Distribution of top ten household assets across the wealth quintiles of the study participants

RESULTS

Liquid petroleum gas (LPG) as cooking fuel was present in 99.4% of the households. Similarly, a ceiling fan for ventilation was seen in 98.8% of the households. On the contrary, dish wash machine, black-and-white television, and government-supplied laptop were found in 1.6%, 0.5%, and 3% of the households [Table 1].

PCA with the finally selected 20 assets showed the eigenvalue for the first principal component as 6.7 accounting for 33.6% of the variance in the original data. The Kaiser–Meyer–Olkin measure of sampling adequacy was 0.856 (P < 0.001). Wealth quintile was prepared ultimately based on the PCA analysis with 20 assets. The linear regression analysis with the composite wealth index score as dependent variable showed car ranked first and presence of motor to pump groundwater in the household ranked the last in the list based on regression co-efficient value. The list of assets with corresponding weights is shown in Table 2.

Table 2

List of assets ordered based on their priority in entrance to the linear regression model

Rank	Asset	Regression coefficient
1	Car	0.238
2	Refrigerator	0.234
3	Microwave oven	0.223
4	Wall-mounted water heater	0.221
5	Desktop/laptop (self-owned)	0.219
6	Inverter	0.218
7	Air conditioner	0.216
8	Mixer grinder (self-owned)	0.207
9	Internet connection	0.206
10	Wet grinder (self-owned)	0.205
11	Color TV (LCD/LED)	0.204
12	Washing machine	0.195
13	Mobile phone (smart model)	0.186
14	Color TV (old model)	0.175
15	Induction stove	0.174
16	Landline	0.172
17	Disinfected drinking water	0.137
18	Mobile phone (basic model)	0.136
19	Availability of own land	0.132
20	Motor to pump groundwater	0.089

LCD: Liquid crystal display, LED: Light emitting diode

The Spearman correlation coefficient for the 20-item and 10-item wealth quintile was 0.95 and was statistically significant (P < 0.001) and the value of weighted kappa agreement between them was 0.77 (95% confidence interval: 0.74–0.80). Their agreement was better than the agreement between 8 and 6 items with 20 item-based SES quintile and the same as that of 12-asset and 20-asset-based SES quintiles. Finally, a simple and reliable 10-asset-based SES scale was developed and the scoring system was generated based on the regression coefficient of the assets, as displayed in Table 2. The scoring system of the simplified 10-item household asset-based socioeconomic scale is shown in Table 3.

Table 3

The scoring system of the simplified 10-item household asset-based socioeconomic scale

Asset	Score
Car	4
Refrigerator	3
Microwave oven	2
Wall-mounted water heater	2
Desktop/laptop (self-owned)	2
Inverter	2
Air conditioner	2
Mixer grinder (self-owned)	1
Internet connection	1
Wet grinder (self-owned)	1
Total	20
SES	Range of score
I (upper)	15-20
II (upper middle)	12-14
III (lower middle)	9-11
IV (upper lower)	7-8
V (lower)	0-6

SES: Socioeconomic status

The SES was calculated for all participants using the MK scale. The SES was calculated with the newly constructed 10-item asset-based SES scale (based on PCA analysis). Wealth quintile calculated by PCA with 20-asste scale and with newly constructed 10-asset scale was correlated and the co-efficient obtained was 0.90 (P < 0.001). Correlation co-efficient between the SES ascertained using the MK scale and the newly constructed 10-asset scale was 0.52 (P < 0.001).

DISCUSSION

Household asset-based simple SES scale with a scoring system for urban population of Puducherry was constructed. There was no under representation of context-specific household assets used to classify SES. The agreement between the wealth quintile generated using PCA based on 20 assets and reduced 10 assets was excellent. However, the correlation between SES obtained between MK scale and the newly constructed 10-asset-based scale was moderate.

In the derivation of asset-based wealth index, different assets were used worldwide by various authors.[14] The assets included in the NFHS-4 survey were used as a reference; however, many modifications were made to it. Assets specific for rural areas such as animal-drawn cart, tractor, and thresher were removed. Commonly found assets such as cot, chair, mattress, and table that are unlikely to differentiate SES were excluded. Furthermore, to keep it context specific, mixer grinder, wet grinder, and laptop that were supplied by the local government free of cost in the study area were included. We also included many new items as per the result of qualitative interview with homemakers. Hence, the tool is unique to the urban context, so its usage suits better for any urban geographic area.

The eigenvalue indicates the percentage of variation in the total data explained for each principal component. If it is lesser, then it could be due to less assets included or the complexity of correlations between them.[14] The higher value in the present study gives assurance that assets were comprehensively included.

Other statistical issues in PCA related to a number of assets included are clumping and truncation.[71416] Clumping or clustering is described as households being grouped together in a small number of distinct clusters. Truncation indicates a more even distribution of SES but spreads over a narrow range resulting in difficulty to differentiate between socioeconomic groups. The histogram constructed in our study showed the absence of clumping and truncation that ensured the range of assets included was broad enough to avoid these issues in PCA.

Health researches carried out in India commonly used MK (1976) scale for urban setting; B G Prasad (1961) for urban and rural; and Uday Pareek scale (1964) for rural population.[1718] B G Prasad scale is purely income based and Uday Pareek is for rural setting; hence, we used the MK scale to compare the finding of our newly constructed tool. In our present study, we got moderate correlation and poor agreement between the SES classes obtained between the newly constructed 10-item asset-based scale and the MK scale. This was similar to a previous study that compared the wealth indices estimated between various standard tools, with the agreement ranged between kappa of 0.01 and 0.15.[19]

The reasons could be firstly our scale is purely asset based and MK scale takes into consideration the education, occupation, and per capita family income. Second, inquiring about income is considered as a sensitive issue and so it suffers validity. Third, the theories governing their construction were different.[20] Besides these shortfalls, many variables involved in the MK scale are subjective as clear definitions of the same are not provided.[3]

In the current study, sampling bias was reduced by including a broad range of urban region-specific assets that were able to discriminate SES and the same was confirmed by a higher and significant statistical measure of sampling adequacy. There was excellent and significant agreement between the wealth quintiles calculated based on the total of 20 assets and the final reduced 10 assets. As the study was based on a representative sample of 500 urban households, it has adequate discriminative power, and hence, the results are statistically generalizable and can be used in different surveys and studies among the urban population.

CONCLUSIONS

A simple and reliable household asset-based SES scale using the PCA technique was created for the urban population of Puducherry. This is a better way of estimating SES at field studies as it has fewer assets and avoids those problems associated with income and consumption-based methods such as recall bias, social desirability bias, seasonality, and data collection time.

Financial support and sponsorship

Nil.

Conflicts of interest

There are no conflicts of interest.