A Statistical Analysis of The Problems Faced by The Customers in Online Shopping

Introduction

• It’s easy. It’s quick and it’s literally a click away. That’s online shopping for customers. But is it that simple? Not exactly. There are many issues with online shopping faced by the consumers. And these challenges are not limited to fake products or hidden costs.

• Customers like to shop at websites that offer them convenience, are easy to browse through, aesthetically appealing and provide relevant information. When the website is not optimized right, it leads to the case of abandoned carts, order cancellations or returns.

Objecives

There are two objectives of this project. These are:

• To find out the most and the least serious issues that may cause the dissatisfaction of a customer towards online shopping.

• To find out various factors that may cause the dissatisfaction of a customer towards online shopping.

Survey procedure:

• Due to limitations of time and scope, my survey is limited to people of age group 18 - 24 years.
• I prepared a google form where I listed 19 questions which are expected to throw light on the objectives of the project.

• I circulated the google form to my M. Stat. batchmates of Indian Statistical Institue, Delhi.

• I also circulated the form to the M. Stat. students of Indian Statistical Institue, Kolkata.

• In all 71 students have responded to this survey.

Coding of responses

If you look at each of the 19 questions in the google form, you will find that each question has 5 options. Those are

• Extremely important.
• Quite important.
• Moderately important.
• Slightly important.
• Not important at all.

Coding of responses (Contd.)

I code these options as

• Extremely important => 5.
• Quite important => 4.
• Moderately important => 3.
• Slightly important => 2.
• Not important at all => 1.

Data matrix

• Now I have a data matrix which has 71 rows and 19 columns.

• The entries of this data matrix are postive integers between 1 to 5.

• Let’s take a look at the data matrix in tabular form by clicking here.

Likert Analysis

• Main focus is to test the null hypothesis of no difference in the responses between the questions.

• The responses here are not on a simple linear scale. Rather these are on likert scale.

• Mean of a number of likert scores is very difficult to interpret.

• The nature of the likert scale prevents the calculation of a valid standard deviation.

• Classical parametric methods for each question based on assumption of normality are not valid.

• Responses to many individual questions can be pooled and the central limit theorem kicks in.

• Not a very good advice either, as the assumption of independence is being made.

• Respondents who are dissatisfied regarding one aspect will also tend to be dissatisfied about other aspects, violating the assumption of independence.

• Suggestions for null hypothesis testing by Kruskall Wallis tests can be given.

• Does not test a very meaningful hypothesis nor does it provide a basis for comparing effect sizes.

• So the conventional approach involves splitting the data and looking at the proportion of responses that fall above or below some cut off point.

• Leads to discussions of the proportion of students who are dissatisfied with the issue revealed by a question.

Distribution of responses among various questions

Implementing Chi-square test

• The data can be simplified into binary classes and we can count the number of responses falling into each class.

• The measure typically used is the proportion or percentage of students who are dissatisfied i.e. giving a score above 3.

• Then the data can be quite validly analyzed using a Chi-square test.

## 
##  Pearson's Chi-squared test
## 
## data:  tb1
## X-squared = 273.23, df = 18, p-value < 2.2e-16

• The p-value of the above test is quite less than 0.05.

• So at 5% level of significance, this test shows the presence of significant differences between the responses when scores over 3 are scored as ones and scores below are scored as zeros.

• Could be used to look at stronger feelings towards a question by changing the splitting rule.

• Suppose scores over 4 are scored as ones and scores below are scored as zeros.

• Rerunning the test we see that p-value is very less than 0.05.

An even simpler way

• The confidence intervals for the proportion of dissatisfaction for the questions have not yet been extracted.

• For this purpose the Binomial test can be applied and this will be very simple.

• For each question it tests the null hypothesis that the true proportion of dissatisfaction is equal to the overall proportion of dissatisfaction.

• The Binomial test also provides confidence intervals through the Clopper and Pearson procedure.

• I can now use the Binomial test to build a function that takes the original vector of likert scores and returns the percentage dissatisfied, with upper and lower bounds for a 95% confidence interval along with a p-value for the significance of differences from the baseline value.

• The results can be tabulated for each group using dplyr function.

• Except for the questions 7, 8, 9, 11 and 16, all the questions show significant p-values.

• It is more useful to look directly at the confidence intervals, as they show the range of possible scores that could be obtained by chance.

• A quick and intuitive way of looking at the data is to plot the confidence intervals after ranking the scores.

Some findings

• Question 12 is the most serious question. Question 12 assesses the customer dissatisfaction if the payment confirmation is missing.

• Question 17 is the least serious question. Question 17 assesses the customer dissatisfaction if there are too many options in the website to choose from.

• Looking at the proportion of dissatisfied individuals and the confidence intervals for the questions, we can group the questions.

Adding depth in the analysis

• The analysis can be re-run using any cut off point in order to add depth in the analysis.

• We set the cutoff point to 4 to analyze extreme dissatisfaction.

• Question 12 still remains the most serious question and question 17 still remains the least serious question.

• Note that question 14 accounts for a large percentage of highly dissatisfied individuals. It is the issue of uclear website policies for return and refund.

Confidence intervals for likert means

• It is possible to produce confidence intervals for the mean through bootstrapping.

• This involves resampling with replaecements from the data.

• Repeat the resampling thousand of times and exclude the extreme values which occur very infrequently and get a bootstrapped confidence interval for the mean by calculating it for all the random samples.

• Occasionally breaks down for small samples but in general quite robust and will never throw up values outside the bounds of the data.

Exploratory factor analysis

• Basic assumption is the existence of sufficient correlations among data in the data matrix.

• Overall KMO test value is 0.67.

• Test result is mediocre, but I will still proceed for factor analysis.

• By principal component analysis, I get that 91% of the total variation is explained by the first 7 PCs.

• So consider an orthogonal factor model with 7 factors.

• Firstly, factor analyze the sample covariance matrix by the iterative PC method with varimax rotation.

• Let’s have a look at the loading matrix.

Loading matrix

## 
## Loadings:
##             PA4    PA1    PA5    PA6    PA2    PA7    PA3   
## question_1                 0.355                            
## question_2  -0.108                       0.850  0.124       
## question_3                               0.603 -0.173       
## question_4   0.297  0.246         0.528         0.127       
## question_5   0.664  0.202         0.176 -0.179  0.111       
## question_6   0.624  0.148                              0.151
## question_7   0.462 -0.123  0.207  0.187 -0.147  0.164  0.257
## question_8   0.159         0.108                       0.703
## question_9   0.403                0.315  0.129  0.239       
## question_10  0.124  0.258         0.171 -0.117  0.812  0.273
## question_11  0.228                0.289         0.436 -0.363
## question_12 -0.119         0.482        -0.179  0.316 -0.157
## question_13         0.256  0.260  0.215 -0.170  0.181  0.107
## question_14         0.265  0.464  0.307         0.123  0.306
## question_15  0.226         0.662  0.116                     
## question_16  0.599  0.317  0.425         0.126        -0.104
## question_17  0.101  0.389  0.180  0.752  0.126  0.107       
## question_18  0.239  0.745  0.193  0.222                     
## question_19  0.116  0.774         0.179         0.153       
## 
##                  PA4   PA1   PA5   PA6   PA2   PA7   PA3
## SS loadings    1.911 1.764 1.397 1.370 1.281 1.205 0.962
## Proportion Var 0.101 0.093 0.074 0.072 0.067 0.063 0.051
## Cumulative Var 0.101 0.193 0.267 0.339 0.406 0.470 0.521

Factors identification

• Corresponding to a particular question look at the absolute values of the estimated loadings and place the issue revealed by that question under that factor under which it has the maximum estimated loading.

• In this way all the questions can be classified under 7 factors as follows

• Factor 1: Questions 18 and 19.
• Factor 2: Questions 2 and 3.
• Factor 3: Question 8.
• Factor 4: Questions 4, 5, 6, 7, 9 and 16.
• Factor 5: Questions 1, 12, 13, 14 and 15.
• Factor 6: Question 17.
• Factor 7: Questions 10 and 11.

Factor naming

Looking at the questions I name the factors as

• Factor 1: Website design factor.
• Factor 2: Package handling factor.
• Factor 3: Delivery time factor.
• Factor 4: Product tracking and return time factor.
• Factor 5: Product quality, payment confirmation and customer care services factor.
• Factor 6: Number of options factor.
• Factor 7: Payment method factor.

Confirmatory factor analysis

• To see whether the assumed data structure is correct or not.

• High CFI and TLI values (0.881 and 0.847 respectively) are indicators of good fit.

• SRMR value (0.085) between 0.05 and 0.10 indicates an acceptable fit.

• RMSEA value (0.058) between 0.05 and 0.08 indicates a fit close to good. So we can say that our model fitting is acceptable.

Final list of the factors

Factor 5 includes product quality factor, which should not go with payment confirmation and customer care services factors. This is happening just because of the data. Treating it as a separate factor I can list down the factors as

1. Product quality not up to the mark.
2. Bad way of handling the packages.
3. Delayed delivery.
4. Bad product tracking facilities and problems and delay in case of product return.
5. Missing payment confirmation and bad customer care services.
6. Missing of the desired payment methods.
7. Too many options to choose from.
8. Complicated website designs and boring interfaces.

Conclusion

Websites designed using the right tools can overcome most of these challenges while you upgrade your web shop to the latest version. With the help of the mentioned solutions, you can get your website optimized and help visitors overcome these online shopping challenges to boost your ecommerce sales. Hence, going that extra mile for your customers and addressing their pain points will surely pay off in future. Remember, putting products on the display is not enough, conversions happen when everything goes well till end.

R - CODE

THANK YOU