Sparkling Data

Top 5 Diagnoses account for 20% of CMS Inpatient Medicare Spending for 2014

 

The Centers for Medicare and Medicaid (CMS) recently released Inpatient Charge Data for FY 2014 . CMS now provides annual data starting from 2011, although 2014 is the first year for which data is available for all diagnosis groups. In this post, we analyze this data from both the CMS and Provider perspectives.

 

Data Characteristics

 

The Inpatient data describes the following data elements:

- DRG: A Diagnosis Related Group is a statistical system of classifying inpatient stays into groups for the purposes of payment. The CMS data provides average payments for providers organized by individual DRGs.

- Provider: Individual providers are described by their id, name and address.

- HRR: The Hospital Referral Region is a key concept from the Dartmouth Atlas of Healthcare. HRRs represent regional healthcare markets for tertiary medical care, and were defined by determining where patients were referred to for major cardiovascular surgical procedures and for neurosurgery. Each provider belongs to a single HRR, although of course an HRR can contain many providers.

- Total Discharges: This is simply the count of the number of discharges for that DRG and Provider.

- Charges:

Average Covered Charges represent hospital billing costs, which are used as negotiating points and not particularly meaningful in terms of the actual amounts paid by Medicare. These are the charges that would be billed to a patient without insurance.

Average Total Payments include the co-payments and deductibles that the patient is responsible for, as well as additional payments by third parties, in addition to Medicare payments.

Average Medicare Payments represent the average amount that Medicare pays to the provider for Medicare’s share of the MS-DRG.

 

A detailed explanation of the data elements can be found here (PDF) on the CMS web site. A simple graphical representation of the key data elements can be found in an earlier post describing the CMS 2011 data set.

The rest of this post primarily uses the Average Medicare Payments to analyze medicare spending; the average covered charges and total payments from the dataset are ignored.

Also this post makes the following assumption for calculating total Medicare payments for a given DRG and Provider.

For a given DRG for a given Provider:

Total Medicare Inpatient Spending =

(Average Medicare Payments * Total Discharges)

 

Questions

 

We look at addressing the following questions, which may provide actionable insights to CMS planners for reducing costs and to individual providers for increasing revenues.

For CMS

Where is the money going? Which DRGs have seen an increase in payments over the past four years? Which DRGs and locations account for the largest amounts of spending in 2014? Those would be good targets to focus on for cost reduction in the future.

For Providers

What are the opportunities to earn increased payments from Medicare?

For a provider their capacity is limited, so assuming utilization is high the number of discharges cannot be easily increased. Also, payment is mandated by CMS, so that’s not really under their control. One potential axis of freedom is to enter or exit certain DRGs (services) based on need or lack of it. Accordingly, it may be useful to look for underserved regions: HRRs that lack providers for high-payment DRGs. At the same time there are a small number of providers that are so dominant nationally within certain DRGs that it may not make sense to challenge them in their area of expertise.

 

Analysis: 2014 Medicare Payments

 

The top 5 Diagnosis Related Groups (DRGs) in terms of Medicare spending are consistent over the past four years. These DRGs are the following:

⇒  470 - MAJOR JOINT REPLACEMENT OR REATTACHMENT OF LOWER EXTREMITY W/O MCC

⇒  871 - SEPTICEMIA OR SEVERE SEPSIS W/O MV 96+ HOURS W MCC

⇒  853 - INFECTIOUS & PARASITIC DISEASES W O.R. PROCEDURE W MCC

⇒  291 - HEART FAILURE & SHOCK W MCC

⇒  460 - SPINAL FUSION EXCEPT CERVICAL W/O MCC

 

From the 2014 data, we can look at how the payment for the top 5 most expensive DRGs is distributed by the average medical payment per discharge vs the total number of discharges, as shown in the chart below. The size of the bubble in this chart represents the total Medicare payment for those DRGs.

It is interesting to note that there are two distinct groups:

1. Three of the top 5 expensive DRGs (470, 871, 291) have a high number of discharges but a low average cost per discharge
2. The remaining two (460 and 853) do not have as many cases, but they have a high average cost per discharge

 

 

 

In more detail, here is the percentage of total Medicare spend focused on the top 5 DRGs.

 

 

We can also split it by region for these high-cost DRGs, payments for the top regions are given below.

 

 

Analysis: Long-Term Trends

 

From 2011 to 2014, Medicare payments have continued to rise for DRGs 470 (Major Joint Replacement) and 871 (Septicemia/Sepsis), while the other three have remained relatively flat. These two DRGs, 470 and 871, have significantly higher costs than the others over the entire time period of the data.

 

 

For comparison, we can look at spending over these four years for the top 20 DRGs. While some DRGs have risen significantly in the past couple of years [e.g. 208 - Respiratory System and 219 - Cardiac Valve], the Major Joint Replacement and Septicemia DRGs far outpace spending on other diseases and conditions, as shown in the chart below.

(Note: some DRGs show a value of 0 prior to 2014 because CMS Inpatient data only provided information for the top 100 DRGs in earlier years.)

 

 

Analysis: Providers

 

Given that a large percentage of Medicare spending is concentrated on the top 25 DRGs, it is interesting that there are a number of Hospital Referral Regions (HRRs) that have no Medicare billing providers at all for at least 5 DRGs out of those top 25. The geographical map below highlights those regions.

(Click on the map image to navigate to an interactive web page that shows the list of underserved DRGs for the marked regions.)

 

At the same time, there are some dominant providers that handle a very large number of discharges for specific diseases and conditions. The geographical map below highlights individual providers that handle greater than 1% of all discharges nationally, in terms of Medicare billing, for some of the top 25 DRGs.

(Click on the map image to navigate to an interactive web page that shows the list of high-volume DRGs for the marked providers.) 

 

Future Directions

 

CMS has recently released hospital quality data, so it would be interesting to correlate that data with average Medicare payments and total number of discharges. For example, do hospitals that generate a lot of revenue in specific DRGs have higher quality ratings? Also, how does the Medicare payment data relate to general health data within a hospital referral region?

Now that we have specific data over the past several years for DRGs and Providers, it may be possible to make some predictions about future expenses and thus highlight the potential for cost increases. This would point to areas of focus for containing increasing healthcare costs.

 

Analysis of HIPAA Breach data

According to HIPAA regulations under the HITECH Act, a covered entity must notify the Secretary of the Department of Health and Human Services of the federal government if it discovers a breach of unsecured protected health information. This notification must be made relatively quickly (usually within 60 days) if the breach affects 500 or more individuals; for fewer affected individuals this notification can be carried out at the end of the calendar year.

The US Department of Health and Human Services posts a list of breaches on their web site, which covers all reported breaches since the data started getting collected back in late 2009. In this post, we cover an initial analysis of this dataset.

Note: Some cleanup was carried out on the raw data. Reports that did not indicate the state where the breach took place were dropped; if other types of information were not specified it was set to 0 or ‘Unknown’ as appropriate. However this only affected a small number of data points.

The summary dashboard is shown below.

Next let's look at each chart one by one. (You can click on an image to see a larger view.)

 

Breaches by Entity Type

This chart shows the number of incidents reported per year by each type of entity. Healthcare providers are consistently the biggest source of incorrect disclosure of protected health information, followed recently by health plans. Business associates on the other hand have been making progress in improving their HIPAA compliance in order to prevent breaches.

[Note that the reporting is usually done by the covered entity, so it is possible that some breaches are incorrectly attributed to the healthcare provider even after originating with the business associate.]

Surprisingly the total number of breaches has not increased significantly over the past few years even as healthcare organizations delve deeper into electronic health records and data exchange. This is good news for the industry.

 

Types of Breaches

The top four types of breach account for 85% of all breaches:

1. Theft:  45%
2. Unauthorized Access/Disclosure:  22%
3. Hacking/IT Incident:  11%
4. Loss:  7%

Although healthcare entities pay a great deal of attention to network and data center security, it turns out that hacking is responsible for only a small percentage of improper disclosures; the most common scenarios involve theft of some type of computer (laptop, desktop or network server) or increasingly, a mobile device.

Of course, it’s possible that the percentage of breaches due to hacking is so small precisely because these entities put so much of their energy into back-end security, so that the thefts stand out due to a form of survivorship bias.

In any case, it’s clear that healthcare entities in general, and healthcare providers in particular, should put more effort into reducing the theft of laptops and other mobile devices.

 

Location

The heat map below shows the number of breaches by state. California is the clear winner (loser?) here, followed by Texas, Florida and New York:

1. California:  178
2. Texas:  131
3. Florida:  110
4. New York:  101

It’s possible that these states have more breaches because they lead in the adoption of electronic records, data exchanges and interoperability, new health technology, and so on; if so, the numbers should even out in the next few years.

 

Number of Users Affected

The chart below shows the distribution of the breach counts for size of the breach, in terms of the number of users affected. As expected, the number of massive breaches affecting greater than 100,000 individuals is small although still alarming; surprisingly the number of tiny breaches affecting less than 500 individuals is also very small, perhaps those do not get reported as consistently. The majority of breaches affect between 1,000-10,000 individuals.

 

Conclusions

There are a few conclusions we can draw from the above analysis:

• Organizations that manage protected health data should continue their emphasis on data security.
• The data points to the next area of focus: to try to reduce computer and mobile device theft, and to reduce or eliminate the amount of unencrypted data stored on these devices at the same time.
• It is possible to manage high levels of data security even when health data is stored electronically and exchanged between disparate healthcare entities.

 

Next Steps

This was an initial analysis, but there are several interesting areas yet to be explored:

1. What do the individual trends for each breach type look like over time? For example, have thefts always been a major cause or have they increased more recently?
2. How do different breach types affect the number of individuals affected? Here we’ve looked mainly at different dimensions relative to the number of breaches rather than the number of individuals affected.
3. Do different entity types suffer from one type of breach more than from the others?
4. How quickly is the “Other Portable Electronic Device” group growing? Given the explosion of smartphones and other mobile devices, this is likely to be a growth area for health data breaches in the future.
5. Finally, it would be interesting to be able to study the number of breaches and individuals affected relative to the size of the reporting organization, e.g. are there more breaches at large hospitals or small clinics? Unfortunately HHS does not collect this information when breaches are reported.

 

2016 Campaign contributions: Is Hillary Clinton in bed with the Oil & Gas industry?

 

There has been some controversy recently about the supposedly cozy relationship between Democratic presidential candidate Hillary Clinton and the Oil & Gas industry - with the implication that the industry is a significant contributor to her 2016 presidential campaign in the Democratic primaries. Let’s see what the numbers tell us.

In the previous post, we looked at the change in momentum over time between the two leading candidates - Hillary Clinton and Bernie Sanders - in the Democratic presidential primaries. In this post, we take a look at their campaign contributions by industry, using data from the always amazing followthemoney.org , a free non-partisan data archive of contributions to political campaigns.

Since the total number of industries in the data set is quite large, we’ve chosen to depict the top seven industries that donated to each campaign, along with two additional ones: Oil & Gas and Mining. That should be sufficient to compare the relative importance of the different industries to each candidate’s fund-raising success.

 

        [You can view the live chart here: http://ftm.sparklingdata.com/industrycompare.htm ]

 

As the chart above shows, the Oil & Gas and Mining industries together contributed less than 1% of total campaign contributions to each of these two candidates. This can hardly be described as a “dependence” by either of these candidates; based on this data, the controversy has no element of truth.

 

Feeling the “Bern”: Visualizing the increased momentum of the Bernie Sanders campaign

 

The Bernie Sanders campaign has been an interesting sideshow in the Democratic presidential primary for 2016. Given the huge delegate lead accumulated by Hillary Clinton in the early days of polling, Senator Sanders’ campaign was always a long shot. As his campaign has been saying recently, it has always been about the message, not about winning.

And there is no question about it: that message has resonated. Let’s look at the numbers - specifically, campaign contributions to the Clinton and Sanders camps over time. Data for the chart below comes from the amazing followthemoney.org , a free non-partisan data archive of contributions to political campaigns.

 
      [You can view the live chart here: http://ftm.sparklingdata.com/demfundstrend.htm ]

 

The image above tells the story. In February 2016, contributions to the Sanders campaign were less than one-fifth (19.85%) of the contributions to the Clinton campaign; just a month later by March 2016, it was a different story, that ratio was roughly two-thirds (66.3%). It will be interesting to view the numbers for April to see if the trend continues, assuming of course that Senator Sanders stays in the race.

 

Analyzing CMS 2011 Data: DRG 292 Heart Failure

 

This is the first chart in the series for analyzing diagnosis and payment medicare data for 2011 published by the Centers for Medicare and Medicaid (CMS).

 

Analytics Charts

In the first chart, we look at the total discharges for the top 11 DRGs in 2011. DRGs are a classification system that groups similar clinical conditions (diagnoses) and the procedures furnished by the hospital during the stay, so this chart represents the top conditions or diagnoses - as measured by the total number of discharges - paid for by medicare in 2011.

In the remaining three charts shown below, we examine the data related to DRG-292: Heart Failure and Shock with CC. Specifically:

1. The geographical distribution of DRG 292 claims
2. The top providers for these claims
3. The relationship of payments to discharges for these claims. 

[Note: The Average Total Payments values have been scaled for display on this graph in order to show the Total Discharges bars. The scaling is applied uniformly to all payments values, so the relative bar heights are still correct.]

 

Observations

Some observations from the charts shown above:

1. DRG 292: Heart Failure and Shock with CC - is the fourth highest DRG with 222K claims in 2011
2. Geographically the claims are concentrated in the eastern half of the US and along the coastline in the west, although this is largely explained by regional population density.
3. Payments and Total Discharges are not as proportional as one would expect; presumably there are additional factors beyond simply the number of discharges. 

 

Analyzing CMS 2011 Data

 

The Centers for Medicare and Medicaid (CMS) publish historical medicare diagnosis and payment data. This series of posts analyzes this data for 2011.

Data

Source

Inpatient Prospective Payment System (IPPS) Provider Level Charges and Medicare Payments for the Top 100 Diagnosis-Related Groups (DRG)

http://www.cms.gov/Research-Statistics-Data-and-Systems/Statistics-Trends-and-Reports/Medicare-Provider-Charge-Data/Inpatient2011.html

Size:

163072 rows

Description

1. Medicare provider hospital
2. Hospital address and zip
3. DRGs: DRGs are a classification system that groups similar clinical conditions (diagnoses) and the procedures furnished by the hospital during the stay
4. Total Discharges: The number of discharges billed by the provider for inpatient hospital services
5. Average Covered Charges: The provider's average charge for services covered by Medicare for all discharges in the DRG
6. Average Total Payments: The average of Medicare payments to the provider for the DRG. This is essentially the entire payment received by the hospital, including:

• DRG amount, teaching,  disproportionate share, capital, and outlier payments for all cases
• Co-payment and deductible amounts that the patient is responsible for
• Payments by third parties for coordination of benefits

 

The data can be represented diagrammatically as follows:

 

 

Open Questions

1. Correlations between hospitals, diseases and locations: which hospitals have the most incidence of specific DRGs and is there a correlation with zip code?
2. For the top 10 DRGs by number of discharge, which hospitals are in the top (or bottom) 10 percentile in billing for those DRGs?
3. What are the top 10 DRGs with the biggest percentage difference in billing between the high billers and the low billers? High billing can be calculated as the average billing of the top 10 billers for that DRG, and likewise for low billing. Is there a location component to this difference in billing?
4. What are the top DRGs with the biggest percentage average difference between charges and payments? Similarly for top hospitals?
5. What are the top DRGs by billing and by payment costs? Is there a correlation with location?
6. Which hospitals have at least an average 10 percent higher billing rate than other hospitals for the same DRGs? Is there a correlation with location?

Are there additional interesting and meaningful questions to be answered? What would you like to know?

 

 

Finding the Pricing Sweet Spot from Sales Transaction Data

 

Pricing is one of the most important decisions you can make for your small or medium-sized business. At the same time, it is often the least-understood part of the business; many businesses simply continue to stick with the product prices they have already used in the past, regardless of changes in external conditions.

One aspect of the pricing decision is: what price-point are your current customers comfortable with? What price points are most popular for different product categories or customer segments?

The charts below map out the relationship between total sales, profits, number of customers and price point for our sample data set, with further narrowing down for two product categories.

Once you know your current average price point by category, you can consider options such as introducing additional high-end products or changing your marketing to try to raise the average price at which you sell your products.

And by identifying customer segments that respond positively to a higher price point, you can create special promotions targeting those customers.

 

POS Analytics Basics: Products, Sales and Customer Data

 

Point of Sale Data is one of the most valuable sources of business intelligence data that is often overlooked. This is data that every retail business collects as a matter of course, simply by doing what they do routinely. It can provide deep actionable insights that can be used to boost the profitability and success of your business.

The value of POS data lies in the fact that:

1. It is real - this data represents actual purchase decisions by individual consumers acting independently, not the output of a focus group
2. It is big - not in the sense of the Big Data hype, but because this is data the business collects every day, over time you necessarily have a lot of it
3. It is continuous - since you collect this data every day when you’re open for business, you can get valuable information about trends, cycles, causal effects and product chains

On a side note, if you collect additional data external to the POS system - such as customer demographic data or promotional data - and can connect it with your POS data, that makes your data analytics exponentially more valuable. For example, even without additional external data you can still observe buying patterns from point of sale data by identifying customers across individual sales transactions; but with additional customer information such as age, gender or location, you can build a much better Customer Engagement model.

In this post, we’ll look at how to get some key insights into your business using simple, easy-to-understand charts generated from POS data. Future posts will cover more complex calculations, advanced metrics and machine learning, but if you are a beginner to business data analytics then creating a simple dashboard using the graphs in this blog post is a good way to get started.

We are going to look at three key areas covered by transaction data coming from point of sale systems: Products, Sales and Customers. The charts in this blog post have been generated from a two-week sales transaction log of sample data.

 

Products

The charts below show the profit and revenue associated with individual products and with each product category. The third chart plots the product count of each product sold within that time interval.

From these charts you can find the most popular products, and those that are the most profitable. As a business owner, these are the products that you should keep an eye on - if supplier or inventory problems cause these products to become unavailable then the business will suffer.

In addition, using market basket analysis you can identify products that are often associated together in your sales transactions. This gives you an indication of which products you can try to upsell along with your popular products.

 

Sales

Sales transactions provide a rich treasure trove of data that can be analyzed along many dimensions. The charts below, which are part of a sales dashboard, show an analysis of sales distribution by retail location, salesperson and time for the given time period.

From these charts, you can easily figure out:

• Which retail locations are doing well and which are not,
• Which salespersons are the most effective, whether it’s because of their own ability or due to external factors, and
• Which times are busiest for sales, so you can staff up accordingly.

 

Customers

By analyzing customers for different transactions, you can find your most loyal star customers who provide the largest sales and profit for your business. This enables you to get to know them and reward them!

The chart below plots the number of customers by location. Along with sales by location, it provides insight into which geographical locations are the most popular with customers.

The image below shows a couple of real-life examples of the types of charts we've seen above and a resulting upsell offer (courtesy of Business Insider, Atlassian and Amazon):

All of the analysis shown above just scratches the surface of POS transaction data analytics. In future posts, we will examine more complex calculations with deeper insights.

 

Motorcycle Crash Data Analysis: Do helmets make riding safer?

 

In this post, let's analyze motorcycle safety data from the US Department of Transportation to get an objective, data-driven answer to the question: Are motorcyclists statistically less likely to get injured or worse if they're wearing a helmet?

 

Problem Definition

Given motorcycle safety data about fatalities, injuries, miles ridden and percentage helmet usage over the years 1994 - 2012, can we show an inverse correlation between percentage motorcycle helmet usage and motorcycle injuries and deaths?

 

Data Description

The data set was obtained from the Bureau of Transportation Statistics of the US Department of Transportation. It contains aggregate data about motorcycle fatalities and injuries, miles ridden, and motorcycle helmet and car safety belt use, for the years 1994-2012. This overall data is also broken out by the type of vehicle for cars and by rider vs passenger for motorcycles, but we don’t use that data in this analysis.

There is also data earlier than 1994, but it’s less comprehensive and less frequent so it was excluded from this analysis.

 

Data Preparation

There is only minimal cleanup needed for this data, for which I carried out the following actions:

• Remove the description non-data lines from the data files and clean up the “no data available” cells for Tableau
• Correlate the two types of data by year
• Transpose the data and add column names for running the visualization program

 

Data Analysis

The data set has the following dimensions by year for analysis:

1. Total fatalities, injuries and rider crashes
2. Number of fatalities, injuries and rider crashes per 100M vehicle-miles
3. Total vehicle-miles ridden
4. Overall motorcycle helmet use expressed as a percentage
5. Overall car safety belt use expressed as a percentage

 

Step I - General Trends

In this step, we take a look at the general trends for overall motorcycle ridership, helmet use and safety belt use.

According to the data, there was an almost 77% increase in motorcycle ridership in 2006-2007, which does not seem reasonable. [Perhaps there was a change in the DOT's data collection methodology? Within the data files the DOT notes that: “Since Vehicle-miles data for 2000 and later years are estimated using enhanced methodology, data for these years are not comparable with prior years or with numbers published in the previous NTS reports.”]. In any case, the overall ridership was pretty steady in the 10B vehicle-miles range in the earlier period, and fluctuates around 20B vehicle-miles in the later period.

Helmet usage tends to fluctuate between 50% and 65% of all riders, with a low in 2005 [48%] and a high in 2000 [71%] and another near-high in 2009 [67%].

It is interesting to see the relationship between Helmet Usage for motorcycle riders and Safety Belt Usage for cars. While helmet usage fluctuates up and down every year around the 50-65% range, car safety belt usage has climbed steadily over the years, starting from less than 60% in 1994 to 86% in 2012. Clearly public awareness campaigns for car safety belt use have been more widespread or more effective (or both) than for motorcycle helmet usage.

 

Step II - Motorcycle Safety Trends

Now let us look at motorcycle safety information: fatalities and injuries.

From the chart above, we can see that motorcycle fatalities have been steadily increasing from 2000 to 2008. The tremendous jump in ridership in 2007 does not seem to have caused a corresponding jump in fatalities - giving even more credence to the theory that the sudden change was due to a change in the data collection method, rather than in actual ridership.

The overall trends in fatalities and injuries seem to move in lock step, which makes sense. In terms of riders affected, injuries are an order of magnitude higher than fatalities, suggesting that a majority of motorcycle crashes are survivable.

For motorcycle safety, the chart above shows clearly that an increase in overall helmet use is correlated with a corresponding decrease in both fatalities and injuries - e.g. percentage helmet use was high in 2009 and 2011 and the number of fatalities per 100M vehicle-miles was correspondingly low. The converse also applies: in the years when helmet usage dropped, there was a rise in fatalities and injuries - e.g. in 2005.

 

Notes

Increasingly, many motorcycle riders today wear not just helmets, but body protection like padded jackets, kevlar pants, boots and gloves. (This may also help explain the consistent lower values for motorcycle injuries from 2007 to 2012.) Perhaps the USDOT should start collecting statistics about body protection in addition to helmet statistics?

 

Analyzing Crowdfunding: Solving the Localytics Data Visualization Challenge

 

Marketing Analytics Vendor Localytics has posted a data visualization challenge. Essentially it is a predictive analytics problem.

 

The Problem

 The challenge gives us a data set of 50K previous funding or viewing events at fictional crowdfunding company BootLoader. A variety of user data, such as age range, gender, location, marital status and device, are provided for each event. Each crowdfunding project belongs to one or more categories. The goal is to predict the user segments that would be interested in a specific new bicycle project that belongs to two categories: Sports and Environment.

Details of the project are given here:    http://data-viz-challenge.localytics.com/

Let us see what visualization of this data tells us about the user segments most likely to fund projects in general, and this project in particular.

 

Approach

In this blog post, we are going to use Tableau visualization charts to come up with our prediction.

 

Data Preparation

The data of 50000 events is available from github in the form of a JSON file. The first step is to convert this JSON to a CSV file for easy consumption by Tableau. I used an open source Python program called JSON2CSV [available here:  https://github.com/evidens/json2csv] for the conversion.

 

Analysis

The dimensions around which we have to analyze events are:

1. Event Type
2. Age Range
3. Gender
4. Marital Status
5. Geographic Location
6. Device Used
7. Project Category

 

Step I: All Categories

Let us do some basic data analysis around all the events in our data set. 

First, we can chart event counts independently by age range, marital status and city. We can also generate combined charts for event counts by device, age range, gender and marital status. The dashboard below shows the different charts for "all-events" analysis.

 

From the charts above, we can draw some initial observations for event counts across all categories:

• View Project > Fund Project, exactly as you would expect
• Gender does not seem to have any impact; M and F counts are almost even. (Luckily the counts for 'undefined' are low.)
• For age range, the 18-24 age range is golden; it accounts for half of all viewing and funding for projects
• Married users view and fund almost twice as many projects as single users
• Usage of Bootloader is heavily concentrated in a few cities: Atlanta, Lake Oswego, Portland, West Linn, ... - in that order
• iOS users view and fund more projects than Android users, especially married iOS users

 

Step II: Categories - Sports and Environment

Now let us analyze specifically the categories for the 3-speed bicycle project: Sports and Environment. Do the same patterns shown above apply, or is there anything different about user behavior for these categories?

 

Let's re-draw the same charts for event counts by device, age range, gender, marital status and city. The dashboard below shows these charts. In addition, it includes a Treemap that shows the counts for combined dimensions.

From the charts above, we can make the following observations for funding events across these two categories:

• Gender still does not have an impact; M and F counts are almost even.
• The age range of 18-24 outweighs all the other age ranges combined.
• Married users, especially married iOS users, are critical for funding projects.
• Distribution of project funding by city follows the same pattern as for all categories.

 

Step III: Unused Dimensions

For completeness, there are two other dimensions we have not considered, because they did not impact the specific objective of this visualization.

1. Client_time:  By charting viewing and funding events against the client's time of day, we can figure out what is the best time to send out our project emails
2. Amount:  By charting the amounts for different funding events, we can figure out the best price point for this project or the best amount to solicit in our project email

 

Conclusions

The objective of this visualization is to identify the types of users who would be interested in the bicycle project. 

 

Let's say that the specific goal is to send out email notifications in waves until a target number of users have contributed to funding the project, with the first wave going to the user group most likely to be interested, then the group next likely to be interested, and so on. 

Based on the visualizations above, these are the target user selection criteria for each successive wave of project emails:

1. First wave:  [ Age range: 18-24; Married; Device: iOS; City: Atlanta, Lake Oswego, Portland, West Linn ]
2. Second wave:  [ Remaining users in age range: 18-24 ]
3. Third wave:  [ Users not covered above, having device: iOS; Married; City location: top 10 cities ]