1. INTRODUCTION:
In household surveys gathering information
on inflation expectations, Barnett et al (2010) found a wide increase in cross
sectional dispersion during the recession periods, a study of U.K. It is at
odds with standard rational expectations models that are at the essential of
most central banks forecasting models. Recent studies has focused on explaining
the deviations from full information rational expectations models in frameworks
in which some form of limited information or bounded rationality is assumed {see
Mankiw and Reis (2002)}.
One of the sources of bounded rationality
of consumers is biases associated with the mode of responding. Depending on the
literature one can find three to four distinctive response styles that affect
responding and thus results of a research.
Fischer, Zealand, Fontaine, Vijver, and
Hemert (2009) point to three
response styles namely acquiescence, extreme response style and social
desirability. Baumgartner and Steenkamp (2006) add to this list mid-point responding,
which is sometimes seen as a mirror image of the extreme response style (Moors, 2003). With respect to economic surveys, acquiescence,
which is associated with agreeing to the item content, seems not to be the
case, as questions are formulated in a way which does not allow for agreeing.
Extreme response style is mostly assessed as tendency to endorse either the
most positive or the most negative categories at both sides of the rating scale
(Weijters, Geuens, and Schillewaert, 2010), which is also hard to be found in
economic surveys as they usually focus on one area of interest (economic
environment) and respondents usually have in the area predetermined opinions
affecting their own answers or even a pattern of responding associated with
whole questionnaire. Midpoint responding for balanced items can be justified by
lack of peculiar opinions with respect to the developments of the economy and
thus might be present in economic surveys. Social desirability, which is often
referred as tendency to responding in a way that presents them favorably
according to current cultural norms (Mick, 1996).
Nevertheless, the social survey biases have
been mostly investigated with social surveys. Therefore, the developed
framework needs to be modified in order to detect peculiarities of economic
surveys. We hypothesize that in economic surveys the patterns of responding
might interact combining social desirability associated with current (or more
persistent moods) regarding the state of the economy ad extreme response style
associated with adopting a simple negative or positive way of thinking. Due to
difficulties associated with questions about the state of the economy and the
fact that respondents are very often asked to make forecasts, although they are
not specialized in this field, their answers might be based on very simple
positive or negative way. Following the results for other developing countries,
we might expect that the prevailing bias in economic surveys might be an
extreme but only negative response style.
An important fact is that it is not
possible to detect any response style with a single item. One of the most
promising techniques in analyzing response patterns is statistical modelling (Baumgartner and Steenkamp, 2006) among which the use of latent class models
seems to be one of the most promising (Kankaraš, Moors, and Vermunt, 2011; Moors,
2003; M. Morren, Gelissen, and Vermunt, 2012). In this paper we focus on one of the areas present
in economic questionnaires namely inflation expectations and without a priori
assumptions we check with latent class approach for possible mode of answering
in the Indian survey.
According to the RBI’s Inflation
expectation survey of household’s questionnaire which comprises a question on
future price expectations s i.e. Block 2: Expectations of respondents on prices
in next 3 months? The respondents covered in the survey may incorporate extreme
(negative) responses to complain against their economic situation or their
dissatisfaction with the government. Such behavior results in certain patterns
of responding which have already been detected by (Bovi, 2009).So, in this paper we particularly address the issue
of identifying negative response style using data from the Inflation
expectations survey of households. We select data from a single quarter - the
first quarter of 2012 (27th round data). The paper is arranged in
the following sections. In section 2, the data of Inflation expectations survey
is described. In section 3, we discuss how latent class models may be used to
handle the extreme response bias in the Inflation expectations survey. Section
4 discussed the latent class model with all possible indicators, applied to the
measurement of extreme response bias. Section 5 concludes with a discussion on
results.
2. Inflation Expectations Survey of
Households - description of the survey and basic statistics
Description of the data
Reserve Bank of India has been conducting Inflation
Expectations Survey of Households (IESH) on a quarterly basis, since September
2005. The survey elicits qualitative and quantitative responses for three-month
ahead and one-year ahead period on expected price changes and inflation.
Inflation expectations of households (HHs) are subjective assessments and are based
on their individual consumption baskets and therefore, may be different from
the official inflation numbers released periodically by the government. Again,
they are rather not assumed to be based on forecasts of any official measure of
inflation, though these inflation expectations provide useful inputs on
directional movements of future inflation. In the present study, data from the
Jan – March, 2012 (27th round), IESH surveys were used in the
analysis; a total of 3792 interviews were used, where94.8% is the response rate
of 27th round.
Sample Coverage and Data collection
The survey is conducted simultaneously in 12 cities
that cover adult respondents of 18 years and above. The major metropolitan
cities, viz., Delhi, Kolkata, Mumbai and Chennai are represented by 500
households each, while another eight cities, viz., Jaipur, Lucknow, Bhopal,
Ahmedabad, Patna, Guwahati, Bengaluru and Hyderabad are represented by 250
households each. The respondents having a view on perceived current inflation
are well spread across the cities to provide a good geographical coverage. The
survey schedule is organized into seven blocks covering the respondent profile
(block 1), general and product-wise price expectations (block 2 and 3),
feedback on RBI's action to control inflation (block 4), current and expected
inflation rate (block 5), amount paid for the purchase of major food items
during last one month (block 6) and the expectations on changes in income/wages
(block 7).
The response options for price changes are
(i) price increase more than current rate, (ii) price increase similar to
current rate, (iii) price increase less than current rate, (iv) no change in
prices, and (v) decline in prices. The inflation rates are collected in
intervals - the lowest being ‘less than 1 per cent’ and the highest being ‘16
per cent and above’ with 100 basis point size for all intermediate classes.
Findings of IESH for last three rounds.
3. Methodology of Latent Class Analysis
A characteristic feature of the latent
class analysis is that the latent variables and items are discrete. The LCA
uses the cross sectional data to measure the latent class variable.
It identifies the classes, their
probabilities, relates probabilities to covariates and classifies individuals
into classes. The latent class model assumes that there are a finite number of
preference classes with homogeneous preferences within each class. These
preferences are assumed to vary significantly across classes. Answers to
attitudinal questions are used to estimate the unconditional probability that a
respondent belongs to class, then the conditional probability that an
individual belongs to a certain class can be estimated. The model assumes
conditional independence – that is, within a latent class the observed
(indicator) variables are independently distributed – although that assumption
can be relaxed (eg., Sinclair and Gastwirth, 1996; Vacek, 1985). Almost all the
attempts to assess error in survey variables using LCA that we are aware of are
based on the assumption of conditional independence (eg. Biemer and Wiesen,
2002; Biemer and Witt, 1996, Sinclair and Gastwirth, 1998).
LCA produces unconditional probabilities gk=P(c=k)
that represent the probability that the units being classified (in our case,
the survey respondents) being to each class of the latent variable. These
unconditional probabilities estimate the unconditional probability of a
respondent belonging to certain class (or the relative size of each latent
class). In addition, LCA models also produce estimates of item (question)
response probabilities conditional on class membership.
In this article, we will use the set of
assumptions proposed by Sinclair and Gastwirth (1996) and by Biemer (2004). We
fit the LCA models based on the question on expectations of respondents on
prices in next 3 months and next one year in general (D), food products (E),
non-food products (F), household durables (G), housing (H) and services (I),
with polytomous options of responses (i.e. price increase more than current
rate; price increase similar to current rate; price increase less than current
rate; no change in prices; decline in prices), based on the questionnaire of
Inflation expectations survey of households, by Department of Statistics and
Information Management, Reserve Bank of India. Using different grouping variables
to achieve identifiable models i.e. gender (B), and the age (C) groups of the
respondents are considered here as grouping variables.
3.1 Choosing the number of classes
To find the number of latent classes of latent
variable, poLCA software package in R statistical computing environment has
been used. In IESH survey of RBI, the respondents were asked to provide answers
(on the Likert scale) to several expectations on prices in next three months.
The response variables were the manifest variables of a latent class model. The
basic model considered is with no covariates.
Table 3: Parameters on converged latent class models
without covariates
|
Number of
classes
|
p
|
LL
|
AIC
|
BIC
|
|
1
|
24
|
-21561.42
|
43170.83
|
43320.61
|
|
2
|
49
|
-16872.17
|
33842.34
|
34148.13
|
|
3
|
74
|
-16626.84
|
33401.68
|
33863.49
|
|
4
|
99
|
-15504.40
|
31206.8
|
31824.62
|
|
5
|
124
|
-15043.78
|
30335.56
|
31109.40
|
|
6
|
149
|
-14942.39
|
30183.79
|
31112.64
|
|
7
|
174
|
-14883.24
|
30114.48
|
31200.35
|
Table 3 summarizes the statistics which supports the
existence of heterogeneity in the responses of the respondents. It suggests the
existence of a number of latent classes. The best fitting model considered
based on information criteria i.e. Akaike Information Criteria (AIC) and
Bayesian Information Criteria (BIC). The BIC is considered to be more
conservative due to its more stringent penalty for the number of additional
parameters (Schwartz, 1978).
Table 3 depicts the heterogeneity in the data. It suggests
the existence of a number of latent classes. The minimum BIC statistics is
associated with 5 latent classes suggesting that model with 5 classes is
optimal.
Class 1
About 19% of respondents were probabilistically
assigned to this class. Around 99% of this class respondents stated that in
general the price may increase more than the current rate in coming 3 months.
About 86% of respondents are in favour that the price increase more than the
current rate for food products. About 72% of the respondents expectation that
the price increase more than the current rate for housing. Finally, around 73%
of respondents expect that for services, the price increase more than the current
rate.
Class 2
About 57% of respondents were probabilistically
assigned to this class. Respondents are in favour of price increase more than
the current rate for next 3 months in all the indicator variables, i.e. in
general (100%); for food products (Less than 90%); for non-food products (Less
than 89%); for household durables (Around 82%); for housing (Around 95%) and
for services (less than 91%).
Class 3
Only 1.8% of respondents are in favour of no change in
prices for the next three months. Around 90% of the respondents are in favour
that in general there will be no change in prices for the next quarter. Around
74% of the respondents feels that no change in prices for household durables in
the next quarter. But for class three, the heterogeneity level in the
respondents is very high for food products, household durables, housing and
services.
Class 4
Around 7% of the respondents expect that the price
increase less than current rate. All feels that in general the price increase
less than the current rate. Around 75% of them are in faour that the food
products price will increase but with less than current rate for the next
quarter. Around 79% respondents of this class expect that the price for non
food products increase but with less than the current rate. Finally, about 70%
of them are in favour of price increase but with less than current rate for
services.
Class 5
About 15% of respondents are in favor that the prices
will increase with the similar to current rate. All respondents in general are
in favor that the price increase similar to current rate. About 88%, 86%,
&&% and 81% of respondents of class 5 are in favor of food products,
non-food products, household durables, housing and services, respectively.
Table 4 illustrates the estimated class conditional
response probabilities for the indicators C, D, E, F, G and H with each column
corresponding to a latent class of consumer sentiment, and each row
corresponding to classes of the indicator variable response.
The classification error probabilities for the
indicator variable C are 
for 
. Examples are when a respondent feels that there will
no change in price in next quarter, but respond to ‘C’ by saying that the
prices decline in next quarter, i.e. 
. Some of the response probability of different
indicators is as under:
; 
; 
; 
; 
; 
; 
; 
.
From the above, one can interpret the following
points:
i) Some respondents are complaining against
the government by giving the responses in terms of response (error)
probabilities.
ii) Respondents have finding difficulties to
respond correctly due to certain reasons, i.e. not able to understand the
question properly, wording in the statement of the questions is not clear.
CONCLUSION:
LCA method have considerable more potential for providing
more valid estimates of IESH; however with this method one can clearly identify
the classes of the latent variable and also identify the problems in the
questionnaire and the wording of questions. In this paper, identified the
heterogeneity in the data by defining five latent classes of the latent
variable and the problems in the questionnaire are identified in terms of
response probabilities which would have been very difficult to discover by
using the other analysis. From the analysis, it is found that the respondents
have given the false response to the questions E and F i.e. non-food products
and household durables, respectively. The analytical technique described in
this paper could provide a means for evaluating the error in the new design.
REFERENCES:
1.
Barnett, A., Mumtaz, H.,
Paustian, M. and Pezzini, S. (2010). Household inflation expectations in the U.
K.: exploiting the cross sectional dimension. Mimeo, Bank of England, November.
2.
Baumgartner, H. and
Jan-Benedict E.M. Steenkamp (2006). Response Biases in Marketing Research,
[in:] Rajiv Grover and Marco Vriens (eds.), Handbook of Marketing Research,
Thousand Oaks, CA: Sage, pp. 95-109.
3.
Biemer, P. P. (2004). An
analysis of classification error for the revised current population survey
employment questions. Survey Methodology, 30, 2, 127-140.
4.
Biemer, P. and Witt, M.
(1996).Estimation of Measurement Bias in Self-Reports of Drug
5.
Use with Applications to
the National Household Survey on Drug Abuse. Journal of Official Statistics,
Vol.12, No.3, pp. 275–300.
6.
Biemer,
P.P., Wiesen, C. (2002). Measurement
error evaluation of self-reported drug use: A latent class analysis of the US
National Household Survey on Drug Abuse. Journal of the Royal Statistical Society, Series A:
Statistics in Society. 165, (Part 1), 97-119.
7. Bovi,
M. (2009).Economic versus Psychological Forecasting.Evidence from consumer
confidence surveys. Journal of Economic Psychology, 30, 4, 563-574.
8. Fischer,
R., Fontaine, J. R. J., van de Vijver, F. J. R., and van Hemert, D. A.
(2009). What is style and what is bias in cross-cultural comparisons? An
examination of acquiescent response styles in cross-cultural research. In A.
Gari and K. Mylonas (Eds.), Quod erat demonstrandum: From herodotus’
ethnographic journeys to cross-cultural research (pp. 137-148). Athens: Pedio.
Retrieved from
http://iaccp.org/drupal/sites/default/files/spetses_pdf/16_Fischer.pdf
9. Kankaras,
M., Moors, G., and Vermunt, J.K (2010). Testing for measurement invariance with
latent class analysis. In: E. Davidov, P. Schmidt, and J. Billiet (eds.),
Cross-Cultural Analysis: Methods and Applications, 359-384. New York:
Routledge.
10.
Morren, M., Gelissen, J.
P. T. M., Vermunt, J. K.(2011). Dealing with extreme response style in cross
cultural research: A restricted latent class factor analysis approach, 41, 1,
13-47.
11.
Morren, M., Gelissen, J.
P. T. M., Vermunt, J. K. (2012). The impact of controlling for extreme
responding on measurement equivalence in cross-cultural research. Methodology:
European Journal of Research Methods for the Behavioral and Social Sciences,
Vol 8(4), 2012, 159-170.http://dx.doi.org/10.1027/1614-2241/a000048
12.
Sinclair,
M. D. and Gastwirth, J. L. (1996).On Procedures for Evaluating the
Effectiveness of Reinterview Survey Methods: Application to Labor Force Data.
Journal of the American Statistical Association, 91, 435, 961-969.
13.
Sinclair,
M. D. and Gastwirth, J. L. (1998).Estimates of the errors in classification in
the labour force survey and their effects on the reported unemployment rate.
Survey Methodology, 24, 2, 157-169.
14.
Mick,
D. G. (1996).Are Studies of Dark Side Variables Confounded by Socially
Desirable Responding? The Case of Materialism. Journal of Consumer Research,
September, 23 (2), 106-121.
15.
Mankiw, N. G. and Reis,
R. (2002).Sticky Information Versus Sticky Prices: A Proposal To Replace The
New Keynesian Phillips Curve. Quarterly Journal of Economics, 117, 4,
1295-1328.
16.
Moors.G.
(2003).Diagnosing response style behavior by means of a latent class factor
approach. Socio-demographic correlates of gender role attitudes and perceptions
of Ethnic discrimination reexamined. Quality and Quantity, 37, 277-302.
17.
Schwartz G (1978).
Estimating the Dimension of a Model. The Annals of Statistics, 6, 461–464.
18.
Vacek, Pamela M.,
"The Effect of Conditional Dependence on the Evaluation of Diagnostic
Tests," Biometrics 41, (1985) 959-968.
18.Weijters, B., Geuens, M. and Schillewaert, N. (2010).
The individual consistency of acquiescence and extreme response style in self
report questionnaires. Applied Psychological Measurement, 34, 105-121.