Posts Tagged ‘statistical significance’

Forecast Friday Topic: Multicollinearity – Correcting and Accepting it

July 22, 2010

(Fourteenth in a series)

In last week’s Forecast Friday post, we discussed how to detect multicollinearity in a regression model and how dropping a suspect variable or variables from the model can be one approach to reducing or eliminating multicollinearity. However, removing variables can cause other problems – particularly specification bias – if the suspect variable is indeed an important predictor. Today we will discuss two additional approaches to correcting multicollinearity – obtaining more data and transforming variables – and will discuss when it’s best to just accept the multicollinearity.

Obtaining More Data

Multicollinearity is really an issue with the sample, not the population. Sometimes, sampling produces a data set that might be too homogeneous. One way to remedy this would be to add more observations to the data set. Enlarging the sample will introduce more variation in the data series, which reduces the effect of sampling error and helps increase precision when estimating various properties of the data. Increased sample sizes can reduce either the presence or the impact of multicollinearity, or both. Obtaining more data is often the best way to remedy multicollinearity.

Obtaining more data does have problems, however. Sometimes, additional data just isn’t available. This is especially the case with time series data, which can be limited or otherwise finite. If you need to obtain that additional information through great effort, it can be costly and time consuming. Also, the additional data you add to your sample could be quite similar to your original data set, so there would be no benefit to enlarging your data set. The new data could even make problems worse!

Transforming Variables

Another way statisticians and modelers go about eliminating multicollinearity is through data transformation. This can be done in a number of ways.

Combine Some Variables

The most obvious way would be to find a way to combine some of the variables. After all, multicollinearity suggests that two or more independent variables are strongly correlated. Perhaps you can multiply two variables together and use the product of those two variables in place of them.

So, in our example of the donor history, we had the two variables “Average Contribution in Last 12 Months” and “Times Donated in Last 12 Months.” We can multiply them to create a composite variable, “Total Contributions in Last 12 Months,” and then use that new variable, along with the variable “Months Since Last Donation” to perform the regression. In fact, if we did that with our model, we end up with a model (not shown here) that has an R2=0.895, and this time the coefficient for “Months Since Last Donation” is significant, as is our “Total Contribution” variable. Our F statistic is a little over 72. Essentially, the R2 and F statistics are only slightly lower than in our original model, suggesting that the transformation was useful. However, looking at the correlation matrix, we still see a strong negative correlation between our two independent variables, suggesting that we still haven’t eliminated multicollinearity.

Centered Interaction Terms

Sometimes we can reduce multicollinearity by creating an interaction term between variables in question. In a model trying to predict performance on a test based on hours spent studying and hours of sleep, you might find that hours spent studying appears to be related with hours of sleep. So, you create a third independent variable, Sleep_Study_Interaction. You do this by computing the average value for both the hours of sleep and hours of studying variables. For each observation, you subtract each independent variable’s mean from its respective value for that observation. Once you’ve done that for each observation, multiply their differences together. This is your interaction term, Sleep_Study_Interaction. Run the regression now with the original two variables and the interaction term. When you subtract the means from the variables in question, you are in effect centering interaction term, which means you’re taking into account central tendency in your data.

Differencing Data

If you’re working with time series data, one way to reduce multicollinearity is to run your regression using differences. To do this, you take every variable – dependent and independent – and, beginning with the second observation – subtract the immediate prior observation’s values for those variables from the current observation. Now, instead of working with original data, you are working with the change in data from one period to the next. Differencing eliminates multicollinearity by removing the trend component of the time series. If all independent variables had followed more or less the same trend, they could end up highly correlated. Sometimes, however, trends can build on themselves for several periods, so multiple differencing may be required. In this case, subtracting the period before was taking a “first difference.” If we subtracted two periods before, it’s a “second difference,” and so on. Note also that with differencing, we lose the first observations in the data, depending on how many periods we have to difference, so if you have a small data set, differencing can reduce your degrees of freedom and increase your risk of making a Type I Error: concluding that an independent variable is not statistically significant when, in truth it is.

Other Transformations

Sometimes, it makes sense to take a look at a scatter plot of each independent variable’s values with that of the dependent variable to see if the relationship is fairly linear. If it is not, that’s a cue to transform an independent variable. If an independent variable appears to have a logarithmic relationship, you might substitute its natural log. Also, depending on the relationship, you can use other transformations: square root, square, negative reciprocal, etc.

Another consideration: if you’re predicting the impact of violent crime on a city’s median family income, instead of using the number of violent crimes committed in the city, you might instead divide it by the city’s population and come up with a per-capita figure. That will give more useful insights into the incidence of crime in the city.

Transforming data in these ways helps reduce multicollinearity by representing independent variables differently, so that they are less correlated with other independent variables.

Limits of Data Transformation

Transforming data has its own pitfalls. First, transforming data also transforms the model. A model that uses a per-capita crime figure for an independent variable has a very different interpretation than one using an aggregate crime figure. Also, interpretations of models and their results get more complicated as data is transformed. Ideally, models are supposed to be parsimonious – that is, they explain a great deal about the relationship as simply as possible. Typically, parsimony means as few independent variables as possible, but it also means as few transformations as possible. You also need to do more work. If you try to plug in new data to your resulting model for forecasting, you must remember to take the values for your data and transform them accordingly.

Living With Multicollinearity

Multicollinearity is par for the course when a model consists of two or more independent variables, so often the question isn’t whether multicollinearity exists, but rather how severe it is. Multicollinearity doesn’t bias your parameter estimates, but it inflates their variance, making them inefficient or untrustworthy. As you have seen from the remedies offered in this post, the cures can be worse than the disease. Correcting multicollinearity can also be an iterative process; the benefit of reducing multicollinearity may not justify the time and resources required to do so. Sometimes, any effort to reduce multicollinearity is futile. Generally, for the purposes of forecasting, it might be perfectly OK to disregard the multicollinearity. If, however, you’re using regression analysis to explain relationships, then you must try to reduce the multicollinearity.

A good approach is to run a couple of different models, some using variations of the remedies we’ve discussed here, and comparing their degree of multicollinearity with that of the original model. It is also important to compare the forecast accuracy of each. After all, if all you’re trying to do is forecast, then a model with slightly less multicollinearity but a higher degree of forecast error is probably not preferable to a more precise forecasting model with higher degrees of multicollinearity.

The Takeaways:

  1. Where you have multiple regression, you almost always have multicollinearity, especially in time series data.
  2. A correlation matrix is a good way to detect multicollinearity. Multicollinearity can be very serious if the correlation matrix shows that some of the independent variables are more highly correlated with each other than they are with the dependent variable.
  3. You should suspect multicollinearity if:
    1. You have a high R2 but low t-statistics;
    2. The sign for a coefficient is opposite of what is normally expected (a relationship that should be positive is negative, and vice-versa).
  4. Multicollinearity doesn’t bias parameter estimates, but makes them untrustworthy by enlarging their variance.
  5. There are several ways of remedying multicollinearity, with obtaining more data often being the best approach. Each remedy for multicollinearity contributes a new set of problems and limitations, so you must weigh the benefit of reduced multicollinearity on time and resources needed to do so, and the resulting impact on your forecast accuracy.

Next Forecast Friday Topic: Autocorrelation

These past two weeks, we discussed the problem of multicollinearity. Next week, we will discuss the problem of autocorrelation – the phenomenon that occurs when we violate the assumption that the error terms are not correlated with each other. We will discuss how to detect autocorrelation, discuss in greater depth the Durbin-Watson statistic’s use as a measure of the presence of autocorrelation, and how to correct for autocorrelation.

*************************

If you Like Our Posts, Then “Like” Us on Facebook and Twitter!

Analysights is now doing the social media thing! If you like Forecast Friday – or any of our other posts – then we want you to “Like” us on Facebook! By “Like-ing” us on Facebook, you’ll be informed every time a new blog post has been published, or when other information comes out. Check out our Facebook page! You can also follow us on Twitter.

Advertisements

Analyzing Subgroups of Data

July 21, 2010

The data available to us has never been more voluminous. Thanks to technology, data about us and our environment are collected almost continuously. When we use a cell phone to call someone else’s cell phone, several pieces of information are collected: the two phone numbers involved in the call; the time the call started and ended; the cell phone towers closest to the two parties; the cell phone carriers; the distance of the call; the date; and many more. Cell phone companies use this information to determine where to increase capacity; refine, price, and promote their plans more effectively; and identify regions with inadequate coverage.

Multiply these different pieces of data by the number of calls in a year, a month, a day – even an hour – and you can easily see that we are dealing with enormous amounts of records and observations. While it’s good for decision makers to see what sales, school enrollment, cell phone usage, or any other pattern looks like in total, quite often they are even more interested in breaking down data into groups to see if certain groups behave differently. Quite often we hear decision makers asking questions like these:

  • How do depositors under age 35 compare with those between 35-54 and 55 & over in their choice of banking products?
  • How will voter support for Candidate A differ by race or ethnicity?
  • How does cell phone usage differ between men and women?
  • Does the length or severity of a prison sentence differ by race?

When we break data down into subgroups, we are trying to see whether knowing about these groups adds any additional meaningful information. This helps us customize marketing messages, product packages, pricing structures, and sales channels for different segments of our customers. There are many different ways we can break data down: by region, age, race, gender, income, spending levels; the list is limitless.

To give you an example of how data can be analyzed by groups, let’s revisit Jenny Kaplan, owner of K-Jen, the New Orleans-style restaurant. If you recall from the May 25 post, Jenny tested two coupon offers for her $10 jambalaya entrée: one offering 10% off and another offering $1 off. Even though the savings was the same, Jenny thought customers would respond differently. As Jenny found, neither offer was better than the other at increasing the average size of the table check. Now, Jenny wants to see if there is a preference for one offer over the other, based on customer age.

Jenny knows that of her 1,000-patron database, about 50% are the ages of 18 to 35; the rest are older than 35. So Jenny decides to send out 1,000 coupons via email as follows:

  

$1 off

10% off

Total Coupons

18-35

250

250

500

Over 35

250

250

500

Total Coupons

500

500

1,000

Half of Jenny’s customers received one coupon offer and half received the other. Looking carefully at the table above, half the people in each age group got one offer and the other half got the other offer. At the end of the promotion period, Jenny received back 200 coupons. She tracks the coupon codes back to her database and finds the following pattern:

Coupons Redeemed (Actual)

  

$1 off

10% off

Coupons Redeemed

18-35

35

65

100

Over 35

55

45

100

Coupons Redeemed

90

110

200

 

Exactly 200 coupons were redeemed, 100 from each age group. But notice something else: of the 200 people redeeming the coupon, 110 redeemed the coupon offering 10% off; just 90 redeemed the $1 off coupon. Does this mean the 10% off coupon was the better offer? Not so fast!

What Else is the Table Telling Us?

Look at each age group. Of the 100 customers aged 18-35, 65 redeemed the 10% off coupon; but of the 100 customers age 35 and up, just 45 did. Is that a meaningful difference or just a fluke? Do persons over 35 prefer an offer of $1 off to one of 10% off? There’s one way to tell: a chi-squared test for statistical significance.

The Chi-Squared Test

Generally, a chi-squared test is useful in determining associations between categories and observed results. The chi-squared – χ2 – statistic is value needed to determine statistical significance. In order to compute χ2, Jenny needs to know two things: the actual frequency distribution of the coupons redeemed (which is shown in the last table above), and the expected frequencies.

Expected frequencies are the types of frequencies you would expect the distribution of data to fall, based on probability. In this case, we have two equal sized groups: customers age 18-35 and customers over 35. Knowing nothing else besides the fact that the same number of people in these groups redeemed coupons, and that 110 of them redeemed the 10% off coupon, and 90 redeemed the $1 off coupon, we would expect that 55 customers in each group would redeem the 10% off coupon and 45 in each group would redeem the $1 off coupon. Hence, in our expected frequencies, we still expect 55% of the total customers to redeem the 10% off offer. Jenny’s expected frequencies are:

Coupons Redeemed (Expected)

  

$1 off

10% off

Coupons Redeemed

18-35 45 55 100
Over 35 45 55 100
Coupons Redeemed 90 110 200

 

As you can see, the totals for each row and column match those in the actual frequency table above. The mathematical way to compute the expected frequencies for each cell would be to multiply its corresponding column total by its corresponding row total and then divide it by the total number of observations. So, we would compute as follows:

Frequency of:

Formula:

Result

18-35 redeeming $1 off: =(100*90)/200

=45

18-35 redeeming 10% off: =(100*110)/200

=55

Over 35 redeeming $1 off: =(100*90)/200

=45

Over 35 redeeming 10% off: =(100*110)/200

=55

 

Now that Jenny knows the expected frequencies, she must determine the critical χ2 statistic to determine significance, then she must compute the χ2 statistic for her data. If the latter χ2 is greater than the critical χ2 statistic, then Jenny knows that the customer’s age group is associated the coupon offer redeemed.

Determining the Critical χ2 Statistic

To find out what her critical χ2 statistic is, Jenny must first determine the degrees of freedom in her data. For cross-tabulation tables, the number of degrees of freedom is a straightforward calculation:

Degrees of freedom = (# of rows – 1) * (# of columns -1)

So, Jenny has two rows of data and two columns, so she has (2-1)*(2-1) = 1 degree of freedom. With this information, Jenny grabs her old college statistics book and looks at the χ2 distribution table in the appendix. For a 95% confidence interval with one degree of freedom, her critical χ2 statistic is 3.84. When Jenny calculates the χ2 statistic from her frequencies, she will compare it with the critical χ2 statistic. If Jenny’s χ2 statistic is greater than the critical, she will conclude that the difference is statistically significant and that age does relate to which coupon offer is redeemed.

Calculating the χ2 Value From Observed Frequencies

Now, Jenny needs to compare the actual number of coupons redeemed for each group to their expected number. Essentially, to compute her χ2 value, Jenny follows a particular formula. For each cell, she subtracts the expected frequency of that cell from the actual frequency, squares the difference, and then divides it by the expected frequency. She does this for each cell. Then she sums up her results to get her χ2 value:

  

$1 off

10% off

18-35 =(35-45)^2/45 = 2.22 =(65-55)^2/55=1.82
Over 35 =(55-45)^2/45 = 2.22 =(45-55)^2/55=1.82
     

χ2=

2.22+1.82+2.22+1.82  

=

8.08  

 

Jenny’s χ2 value is 8.08, much higher than the critical 3.84, indicating that there is indeed an association between age and coupon redemption.

Interpreting the Results

Jenny concludes that patrons over the age of 35 are more inclined than patrons age 18-35 to take advantage of a coupon stating $1 off; patrons age 18-35 are more inclined to prefer the 10% off coupon. The way Jenny uses this information depends on the objectives of her business. If Jenny feels that K-Jen needs to attract more middle-aged and senior citizens, she should use the $1 off coupon when targeting them. If Jenny feels K-Jen isn’t selling enough Jambalaya, then she might try to stimulate demand by couponing, sending the $1 off coupon to patrons over the age of 35 and the 10% off coupon to those 18-35.

Jenny might even have a counterintuitive use for the information. If most of K-Jen’s regular patrons are over age 35, they may already be loyal customers. Jenny might still send them coupons, but give the 10% off coupon instead. Why? These customers are likely to buy the jambalaya anyway, so why not give them the coupon they are not as likely to redeem? After all, why give someone a discount if they’re going to buy anyway! Giving the 10% off coupon to these customers does two things: first, it shows them that K-Jen still cares about their business and keeps them aware of K-Jen as a dining option. Second, by using the lower redeeming coupon, Jenny can reduce her exposure to subsidizing loyal customers. In this instance, Jenny uses the coupons for advertising and promoting awareness, rather than moving orders of jambalaya.

There are several more ways to analyze data by subgroup, some of which will be discussed in future posts. It is important to remember that your research objectives dictate the information you collect, which dictate the appropriate analysis to conduct.

*************************

If you Like Our Posts, Then “Like” Us on Facebook and Twitter!

Analysights is now doing the social media thing! If you like Forecast Friday – or any of our other posts – then we want you to “Like” us on Facebook! By “Like-ing” us on Facebook, you’ll be informed every time a new blog post has been published, or when other information comes out. Check out our Facebook page! You can also follow us on Twitter.

Forecast Friday Topic: Multicollinearity – How to Detect it; How to Correct it

July 15, 2010

(Thirteenth in a series)

In last week’s Forecast Friday post, we explored how to perform regression analysis using Excel. We looked at the giving history of 20 contributors to a nonprofit organization, and developed a model based on the recency, frequency, and monetary value (RFM) of their past donations. We derived the following regression equation:

We were pleased to see that our model had a coefficient of determination – or R2=0.933, indicating that our model explained 93.3% of the change in the donor’s current contribution (our Ŷ). But we were a little disheartened when we looked at the t-statistics of each of our regression coefficients. Recall that we found our recency coefficient was not significant:

Parameter

Coefficient

T-statistic

Significant?

Intercept

87.27

4.32

Yes

Months since Last

(1.80)

(1.44)

No

Times Donated

2.45

2.87

Yes

Average Contribution

0.35

3.26

Yes

Yet, most direct marketing professionals know clearly that RFM theory postulates that all three variables are significant indicators of whether and how much a donor will give (or a customer will buy). When our model doesn’t replicate what a tried and true theory has long maintained, there could possibly be something wrong.

Multicollinearity

Most times, when something doesn’t look right in the results of a regression model, it is safe to assume that one of the regression assumptions has been violated. The problem is trying to determine which assumption – or assumptions – was violated. Since the coefficient for “Months Since Last Contribution” has a t-statistic that indicates it isn’t statistically significant, we might suspect that the specification assumption is violated: that is, we may believe that “Months Since Last Contribution” is an extraneous, irrelevant variable that should not have been included in the model and, thus, be removed.

But is that really the case? There can be other reasons why a parameter estimate does not come up significant. If two or more independent variables are highly correlated, the resulting multicollinearity can cause the regression model to assign a statistically insignificant parameter estimate to an important independent variable. So, how can we detect multicollinearity?

Detecting Multicollinearity: Correlation Matrix

The first step in detecting multicollinearity is to examine the correlation among the independent variables. We do this by looking at a correlation matrix. You can run a correlation matrix in Excel by using its Data Analysis ToolPak. Looking at the correlation matrix for our variables, we find:

Correlation Matrix – Original Variables

Variable

Contribution Y

Months Since Last Donation X1

Times Donated in last 12 months

X2

Average Contribution in last 12 months

X3

Contribution (Y)

1.00

  

  

  

Months Since Last Donation – X1

-0.93

1.00

  

  

Times Donated in last 12 months – X2

0.89

-0.88

1.00

  

Average Contribution Last 12 mo. – X3

0.88

-0.84

0.69

1.00

 

A correlation of 1.00 means two variables are perfectly correlated; a correlation of 0.00 means there is absolutely no correlation. The cells in the matrix above, where the correlation is 1.00, shows the correlation of an independent variable with itself – we would expect a perfectly correlated relationship. What is most important to us are the numbers below the 1.00 correlations. The first column shows our dependent variable, “Contribution”. As you go down the column, row by row, you see that each of our independent variables is strongly correlated with the dependent variable, indicating that they are all strong predictors.

The correlation between “Months Since Last Donation” (X1) and the donor’s Contribution (Y) shows a correlation that is almost perfectly negative (-0.93), while those correlations of the dependent variable with each of the other two independent variables is almost perfectly positive with the contribution (0.89 and 0.88). When writing these in shorthand, we use the Greek letter rho, ρ, to denote correlation. Hence, to show the correlation between each independent variable with the dependent variable, we would express them as follows:

ρX1Y = -0.93

ρX2Y = 0.89

ρX3Y = 0.88

But now, let’s look at the correlations among our independent variables:

ρX1X2= -0.88

ρX1X3= -0.84

ρX2X3= 0.69

 

Notice that all of our independent variables are highly correlated with one another. The relationship between “Times Donated in Last 12 Months” and “Average Contribution in Last 12 Months” is not as strong as the correlation between those individual variables with “Months Since Last Donation,” but the correlation is still very strong.

Hence, we can conclude that multicollinearity is present in this model.

Correcting Multicollinearity: Dropping Variables

In today’s post, we will discuss one of the remedies for multicollinearity – dropping a highly correlated independent variable. Next week, we’ll discuss the other approaches to correcting multicollinearity. Sometimes, when a variable is “iffy,” we can save ourselves some trouble and just kick it out. If we were to ignore “Months Since Last Donation,” and run our regression with the remaining two variables, we end up with the following regression equation:

Ŷ= 60.68 + 3.37X2 + 0.45X3

We get R2 =0.924, suggesting that we didn’t lose much explanatory power by excluding “Months Since Last Donation.” We also get an F statistic of 103.36, much higher than the 73.90 we had in our original model. A higher F-statistic indicates a model that is more statistically valid. It also reflects the exclusion of one or more extraneous variables. Also, the t-statistics for both independent variables are significant, and they’re even higher than they were in the original model, further indicating increased validity:

Parameter

Coefficient

T-statistic

Significant?

Intercept

60.68

7.24

Yes

Times Donated

3.37

5.83

Yes

Average Contribution

0.45

5.49

Yes

Dropping “Months Since Last Donation” from our analysis worked here. However, dropping variables without a rational decision process can cause new problems. In some cases, dropping a variable can result in specification bias, as we saw in our previous example of predicting profit margin for savings and loan associations a few weeks ago. So, consider dropping variables cautiously.

Next Forecast Friday Topic: More Multicollinearity Remedies

Today, we described one of the ways to remedy multicollinearity – dropping variables. Next week, we will explore two other ways of correcting multicollinearity: obtaining more data and transforming variables. We will also discuss the pitfalls of all three of these remedies, and we will discuss when it’s not worth it to reduce the impact of multicollinearity.

*************************************

Let Analysights Take the Pain out of Forecasting!

Multicollinearity is but one of the many problems you can encounter when forecasting. Let Analysights walk you through the forecasting process so that you can spend more time making strategic decisions and less time trying to guess first where business is going. We will make your forecasting efforts seamless, so you can concentrate on running your business. Check out our Web site or call (847) 895-2565.

Multiple Regression: Specification Bias

July 1, 2010

(Eleventh in a series)

In last week’s Forecast Friday post, we discussed several of the important checks you must do to ensure that your model is valid. You always want to be sure that your model does not violate the assumptions we discussed earlier. Today we are going to see what happens when we violate the specification assumption, which says that we do not omit relevant independent variables from our regression model. You will see that when we leave out an important independent variable from a regression model, quite misleading results can emerge. You will also see that violating one assumption can trigger violations of other assumptions.

Revisiting our Multiple Regression Example

Recall our data set of 25 annual observations of U.S. Savings and Loan profit margin data, shown in the table below:

Year

Percentage Profit Margin (Yt)

Net Revenues Per Deposit Dollar (X1t)

Number of Offices (X2t)

1

0.75

3.92

7,298

2

0.71

3.61

6,855

3

0.66

3.32

6,636

4

0.61

3.07

6,506

5

0.70

3.06

6,450

6

0.72

3.11

6,402

7

0.77

3.21

6,368

8

0.74

3.26

6,340

9

0.90

3.42

6,349

10

0.82

3.42

6,352

11

0.75

3.45

6,361

12

0.77

3.58

6,369

13

0.78

3.66

6,546

14

0.84

3.78

6,672

15

0.79

3.82

6,890

16

0.70

3.97

7,115

17

0.68

4.07

7,327

18

0.72

4.25

7,546

19

0.55

4.41

7,931

20

0.63

4.49

8,097

21

0.56

4.70

8,468

22

0.41

4.58

8,717

23

0.51

4.69

8,991

24

0.47

4.71

9,179

25

0.32

4.78

9,318

Data taken from Spellman, L.J., “Entry and profitability in a rate-free savings and loan market.” Quarterly Review of Economics and Business, 18, no. 2 (1978): 87-95, Reprinted in Newbold, P. and Bos, T., Introductory Business & Economic Forecasting, 2nd Edition, Cincinnati (1994): 136-137

Also, recall that we built a model that hypothesized that S&L percentage profit margin (our dependent variable, Yt) was positively related to net revenues per deposit dollar (one of our independent variables, X1t), and negatively related to the number of S&L offices (our other independent variable, X2t). When we ran our regression, we got the following model:

Yt = 1.56450 + 0.23720X1t – 0.000249X2t

We also checked to see if the model parameters were significant, and obtained the following information:

Parameter

Value

T-Statistic

Significant?

Intercept

1.5645000

19.70

Yes

B1t

0.2372000

4.27

Yes

B2t

(0.0002490)

(7.77)

Yes

We also had a coefficient of determination – R2 – of 0.865, indicating that the model explains about 86.5% of the variation in S&L percentage profit margin.

Welcome to the World of Specification Bias…

Let’s deliberately leave out the number of S&L offices (X2t) from our model, and do just a simple regression with the net revenues per deposit dollar. This is the model we get:

Yt = 1.32616 – 0.16913X1t

We also get an R2 of 0.495. The t-statistics for our intercept and parameter B1t are as follows:

Parameter

Value

T-Statistic

Significant?

Intercept

1.32616

9.57

Yes

B1t

(0.16913)

(4.75)

Yes

 

Compare these new results with our previous results and what do you notice? The results of our second regression are in sharp contrast to those of our first regression. Our new model has far less explanatory power – R2 dropped from 0.865 to 0.495 – and the sign of the parameter estimate for net revenue per deposit dollar has changed: The coefficient of X1t was significant and positive in the first model, and now it is significant and negative! As a result, we end up with a biased regression model.

… and to the Land of Autocorrelation…

Recall another of the regression assumptions: that error terms should not be correlated with one another. When error terms are correlated with one another, we end up with autocorrelation, which renders our parameter estimates inefficient. Recall that last week, we computed the Durbin-Watson test statistic, d, which is an indicator of autocorrelation. It is bad to have either positive autocorrelation (d close to zero), or negative autocorrelation (d close to 4). Generally, we want d to be approximately 2. In our first model, d was 1.95, so autocorrelation was pretty much nonexistent. In our second model, d=0.85, suggesting the presence of significant positive autocorrelation!

How did this happen? Basically, when an important variable is omitted from regression, its impact on the dependent variable gets incorporated into the error term. If the omitted independent variable is correlated with any of the included independent variables, the error terms will also be correlated.

…Which Leads to Yet Another Violation!

The presence of autocorrelation in our second regression reveals the presence of another violation, not in the incomplete regression, but in the full regression. As the sentence above read: “if the independent variable is correlated with any of the included independent variables…” Remember the other assumption: “no linear relationship between two or more independent variables?” Basically, the detection of autocorrelation in the incomplete regression revealed that the full regression violated this very assumption – and thus exhibits multicollinearity! Generally, a coefficient changing between positive and negative (either direction) when one or more variables is omitted is an indicator of multicollinearity.

So was the full regression wrong too? Not terribly. As you will find in upcoming posts, avoiding multicollinearity is nearly impossible, especially with time series data. That’s because multicollinearity is typically a data problem. The severity of multicollinearity can often be reduced by increasing the number of observations in the data set. This is often not a problem with cross-sectional data, where data sets can have thousands, if not millions of observations. However, with time series data, the number of observations available is limited to how many periods of data have been recorded.

Moreover, the longer your time series, the more you risk structural changes in your data over the course of your time series. For instance, if you were examining annual patterns in bank lending within a particular census tract between 1990 and 2010, you might have a reliable model to work with. But let’s say you widen your time series to go back as far as 1970. You will see dramatic shifts in patterns in your data set. That’s because prior to 1977, when Congress passed the Community Reinvestment Act, many banks engaged in a practice called “redlining,” where they literally drew red lines around some neighborhoods, usually where minorities and low-income households were, and did not lend there. In this case, increasing the size of the data set might reduce multicollinearity, but actually cause other modeling problems.

And as you’ve probably guessed, one way of reducing multicollinearity can be dropping variables from the regression. But look what happened when we dropped the number of S&L offices from our regression: we might have eliminated multicollinearity, but we gained autocorrelation and specification bias!

Bottom Line:

The lesson, for us as forecasters and analysts, therefore is that we must accept that models are far from perfect and we must weigh the impact of various regression model specifications. Is the multicollinearity that is present in our model tolerable? Can we add more observations without causing new problems? Can we drop a variable from a regression without causing either specification bias or material differences in explanatory power, parameter estimates, model validity, or even forecast accuracy? Building the model is easy – but it’s these normative considerations that’s challenging.

Next Forecast Friday Topic: Building Regression Models Using Excel

In next week’s Forecast Friday post, we will take a break from discussing the theory of regression analysis and look at a demonstration of how to use the “Regression Analysis” tool in Microsoft Excel. This demonstration is intended to show you how easy running a regression is, so that you can start applying the concepts and building forecasts for your business. Until then, thanks again for reading Forecast Friday, and I wish you and your family a great 4th of July weekend!

*************************

Analysights is now on Facebook!

Analysights is now doing the social media thing! If you like Forecast Friday – or any of our other posts – then we want you to “Like” us on Facebook! By “Like-ing” us on Facebook, you’ll be informed every time a new blog post has been published, or when other information comes out. Check out our Facebook page.

Forecast Friday Topic: Multiple Regression Analysis

June 17, 2010

(Ninth in a series)

Quite often, when we try to forecast sales, more than one variable is often involved. Sales depends on how much advertising we do, the price of our products, the price of competitors’ products, the time of the year (if our product is seasonal), and also demographics of the buyers. And there can be many more factors. Hence, we need to measure the impact of all relevant variables that we know drive our sales or other dependent variable. That brings us to the need for multiple regression analysis. Because of its complexity, we will be spending the next several weeks discussing multiple regression analysis in easily digestible parts. Multiple regression is a highly useful technique, but is quite easy to forget if not used often.

Another thing to note, regression analysis is often used for both time series and cross-sectional analysis. Time series is what we have focused on all along. Cross-sectional analysis involves using regression to analyze variables on static data (such as predicting how much money a person will spend on a car based on income, race, age, etc.). We will use examples of both in our discussions of multiple regression.

Determining Parameter Estimates for Multiple Regression

When it comes to deriving the parameter estimates in a multiple regression, the process gets both complicated and tedious, even if you have just two independent variables. We strongly advise you to use the regression features of MS-Excel, or some statistical analysis tool like SAS, SPSS, or MINITAB. In fact, we will not work out the derivation of the parameters with the data sets, but will provide you the results. You are free to run the data we provide on your own to replicate the results we display. I do, however, want to show you the equations for computing the parameter estimates for a three-variable (two independent variables and one dependent variable), and point out something very important.

Let’s assume that sales is your dependent variable, Y, and advertising expenditures and price are your independent variables, X1 and X2, respectively. Also, the coefficients – your parameter estimates will have similar subscripts to correspond to their respective independent variable. Hence, your model will take on the form:

 

Now, how do you go about computing α, β1 and β2? The process is similar to that of a two-variable model, but a little more involved. Take a look:

The subscript “i” represents the individual oberservation.  In time series, the subscript can also be represented with a “t“.

What do you notice about the formulas for computing β1 and β2? First, you notice that the independent variables, X1 and X2, are included in the calculation for each coefficient. Why is this? Because when two or more independent variables are used to estimate the dependent variable, the independent variables themselves are likely to be related linearly as well. In fact, they need to be in order to perform multiple regression analysis. If either β1 or β2 turned out to be zero, then simple regression would be appropriate. However, if we omit one or more independent variables from the model that are related to those variables in the model, we run into serious problems, namely:

Specification Bias (Regression Assumptions Revisited)

Recall from last week’s Forecast Friday discussion on regression assumptions that 1) our equation must correctly specify the true regression model, namely that all relevant variables and no irrelevant variables are included in the model and 2) the independent variables must not be correlated with the error term. If either of these assumptions is violated, the parameter estimates you get will be biased. Looking at the above equations for β1 and β2, we can see that if we excluded one of the independent variables, say X2, from the model, the value derived for β1 will be incorrect because X1 has some relationship with X2. Moreover, X2‘s values are likely to be accounted for in the error terms, and because of its relationship with X1, X1 will be correlated with the error term, violating the second assumption above. Hence, you will end up with incorrect, biased estimators for your regression coefficient, β1.

Omitted Variables are Bad, but Excessive Variables Aren’t Much Better

Since omitting relevant variables can lead to biased parameter estimates, many analysts have a tendency to include any variable that might have any chance of affecting the dependent variable, Y. This is also bad. Additional variables means that you need to estimate more parameters, and that reduces your model’s degrees of freedom and the efficiency (trustworthiness) of your parameter estimates. Generally, for each variable – both dependent and independent – you are considering, you should have at least five data points. So, for a model with three independent variables, your data set should have 20 observations.

Another Important Regression Assumption

One last thing about multiple regression analysis – another assumption, which I deliberately left out of last week’s discussion, since it applies exclusively to multiple regression:

No combination of independent variables should have an exact linear relationship with one another.

OK, so what does this mean? Let’s assume you’re doing a model to forecast the effect of temperature on the speed at which ice melts. You use two independent variables: Celsius temperature and Fahrenheit temperature. What’s the problem here? There is a perfect linear relationship between these two variables. Every time you use a particular value of Fahrenheit temperature, you will get the same value of Celsius temperature. In this case, you will end up with multicollinearity, an assumption violation that results in inefficient parameter estimates. A relationship between independent variables need not be perfectly linear for multicollinearity to exist. Highly correlated variables can do the same thing. For example, independent variables such as “Husband Age” and “Wife Age,” or “Home Value” and “Home Square Footage” are examples of independent variables that are highly correlated.

You want to be sure that you do not put variables in the model that need not be there, because doing so could lead to multicollinearity.

Now Can We Get Into Multiple Regression????

Wasn’t that an ordeal? Well, now the fun can begin! I’m going to use an example from one of my old graduate school textbooks, because it’s good for several lessons in multiple regression. This data set is 25 annual observations to predict the percentage profit margin (Y) for U.S. savings and loan associations, based on changes in net revenues per deposit dollar (X1) and number of offices (X2). The data are as follows:

Year

Percentage Profit Margin (Yt)

Net Revenues Per Deposit Dollar (X1t)

Number of Offices (X2t)

1

0.75

3.92

7,298

2

0.71

3.61

6,855

3

0.66

3.32

6,636

4

0.61

3.07

6,506

5

0.70

3.06

6,450

6

0.72

3.11

6,402

7

0.77

3.21

6,368

8

0.74

3.26

6,340

9

0.90

3.42

6,349

10

0.82

3.42

6,352

11

0.75

3.45

6,361

12

0.77

3.58

6,369

13

0.78

3.66

6,546

14

0.84

3.78

6,672

15

0.79

3.82

6,890

16

0.70

3.97

7,115

17

0.68

4.07

7,327

18

0.72

4.25

7,546

19

0.55

4.41

7,931

20

0.63

4.49

8,097

21

0.56

4.70

8,468

22

0.41

4.58

8,717

23

0.51

4.69

8,991

24

0.47

4.71

9,179

25

0.32

4.78

9,318

Data taken from Spellman, L.J., “Entry and profitability in a rate-free savings and loan market.” Quarterly Review of Economics and Business, 18, no. 2 (1978): 87-95, Reprinted in Newbold, P. and Bos, T., Introductory Business & Economic Forecasting, 2nd Edition, Cincinnati (1994): 136-137

What is the relationship between the S&Ls’ profit margin percentage and the number of S&L offices? How about between the margin percentage and the net revenues per deposit dollar? Is the relationship positive (that is, profit margin percentage moves in the same direction as its independent variable(s))? Or negative (the dependent and independent variables move in opposite directions)? Let’s look at each independent variable’s individual relationship with the dependent variable.

Net Revenue Per Deposit Dollar (X1) and Percentage Profit Margin (Y)

Generally, if revenue per deposit dollar goes up, would we not expect the percentage profit margin to also go up? After all, if the S & L is making more revenue on the same dollar, it suggests more efficiency. Hence, we expect a positive relationship. So, in the resulting regression equation, we would expect the coefficient, β1, for net revenue per deposit dollar to have a “+” sign.

Number of S&L Offices (X2) and Percentage Profit Margin (Y)

Generally, if there are more S&L offices, would that not suggest either higher overhead, increased competition, or some combination of the two? Those would cut into profit margins. Hence, we expect a negative relationship. So, in the resulting regression equation, we would expect the coefficient, β2, for number of S&L offices to have a “-” sign.

Are our Expectations Correct?

Do our relationship expectations hold up?  They certainly do. The estimated multiple regression model is:

Yt = 1.56450 + 0.23720X1t – 0.000249X2t

What do the Parameter Estimates Mean?

Essentially, the model says that if net revenues per deposit dollar (X1t) increase by one unit, then percentage profit margin (Yt) will – on average – increase by 0.23720 percentage points, when the number of S&L offices is fixed. If the number of offices (X2t) increases by one, then percentage profit margin (Yt) will decrease by an average of 0.000249 percentage points, when net revenues are fixed.

Do Changes in the Independent Variables Explain Changes in The Dependent Variable?

We compute the coefficient of determination, R2, and get 0.865, indicating that changes in the number of S&L offices and in the net revenue per deposit dollar explain 86.5% of the variation in S&L percentage profit margin.

Are the Parameter Estimates Statistically Significant?

We have 25 observations, and three parameters – two coefficients for the independent variables, and one intercept – hence we have 22 degrees of freedom (25-3). If we choose a 95% confidence interval, we are saying that if we resampled and replicated this analysis 100 times, the average of our parameter estimates will be contain the true parameter approximately 95 times. To do this, we need to look at the t-values for each parameter estimate. For a two-tailed 95% significance test with 22 degrees of freedom, our critical t-value is 2.074. That means that if the t-statistic for a parameter estimate is greater than 2.074, then there is a strong positive relationship between the independent variable and the dependent variable; if the t-statistic for the parameter estimate is less than -2.074, then there is a strong negative relationship. This is what we get:

Parameter

Value

T-Statistic

Significant?

Intercept

1.5645000

19.70

Yes

B1t

0.2372000

4.27

Yes

B2t

(0.0002490)

(7.77)

Yes

So, yes, all our parameter estimates are significant.

Next Forecast Friday: Building on What You Learned

I think you’ve had enough for this week! But we are still not finished. We’re going to stop here and continue with further analysis of this example next week. Next week, we will discuss computing the 95% confidence interval for the parameter estimates; determining whether the model is valid; and checking for autocorrelation. The following Forecast Friday (July 1) blog post will discuss specification bias in greater detail, demonstrating the impact of omitting a key independent variable from the model.