“Must See TV”: Predicting the Nielsen Ratings
Statistics Data Analysis Problem
“Must See TV”: Predicting the Nielsen Ratings
Introduction:
The Nielsen ratings gauge the number of people who are watching television programs
and the characteristics of those audiences, which in turn is used by both advertisers and
television programmers. Nielsen ratings are used as currency in the market of advertiser-paid
television. When advertisers want to reach certain audiences, they place ads on television shows
whose viewers display the characteristics of their target market. The larger the audience of a
particular show, the more money the station can charge the advertisers – advertising rates are
based on per thousand viewers. Hence, the networks have a serious interest in increasing their
Nielsen share to generate a greater revenue stream from advertising. It is also helpful for
programmers to know which shows are being watched so that they may discontinue shows that
are not making money (by not drawing in enough advertising revenue). Concurrently, the
advertisers are interested in the ratings so they can gauge the size and characteristics of their
audiences. The most desirable audience for advertisers is the 18-49 demographic. This
demographic, in turn, becomes the most desirable target audience for television programmers.
There are two major periods, called sweeps, in May and November, during which Nielsen
conducts a complete diary measurement across the nation from all 210 markets. This assessment
is completed by sending out a diary to viewers in which they record what show they watched,
what channel they watched, as well as who was watching. Nielsen asks the participants to
complete this survey per quarter hour of viewing and the participants then mail in their
completed diaries. These sweeps periods are when the stations air their best programs and top
news stories in an effort to increase viewership.
Page 2
Statistics Data Analysis Problem
Nielsen selects its TV sample from a diverse group of people – from renters to
homeowners, low income to high income, families with children versus no children, etc. Over
5,000 households are used in the sample, containing over 13,000 people. After this diverse
sample is selected, the families must agree to participate in the research study. To ensure that
Nielsen’s sample audience is reflective of the nation as a whole, they compare their audience’s
characteristics to the US Census Bureau data.
The stations use Nielsen because it is not easy to determine how many viewers a show
has at any one time as they are aired from a satellite or cable system. Nielsen estimates the
audience by taking a sample and then counting the number of viewers in that sample. The
ratings refer to the percent tuned to a particular program during the average minute. Successful
programs are defined by having the larger audiences and the coveted 18-49 demographic.
Goals of the Analysis:
This paper will analyze the various components that could possibly predict the Nielsen
ratings. We will look at network, type of show, awards nominations, day of the week and
percent share within the program’s timeslot. Is there a particular station or network that is
represented in the highest rated programs? Does day of the week factor into the audience size?
Do awards nominations impact the amount of viewership? Does the type of show impact
viewership? How important is it to win in your program’s timeslot to get better overall ratings?
We anticipate that some of these variables will predict the likelihood of a program falling into
the top ratings, which in turn will dictate the amount of advertising revenue a particular station
can generate. As previously stated, these questions are of great relevance to both advertisers and
television programmers.
Page 3
Statistics Data Analysis Problem
Although the 18-49 demographic is not a predictor variable, since it is a simultaneous
output of the Nielsen report, we will attempt to prove the correlation between high overall ratings
with those of the 18-49 demographic. This will show that programmers schedule programs
whose audiences reflect the characteristics desired by advertisers.
It is important to note that both advertisers and television programmers are frustrated
with Nielsen’s data-gathering methods for a variety of reasons. However, they are stuck with
Nielsen for a lack of anything better. We will mention the sources of such frustration, just to
indicate that Nielsen ratings are perhaps not the best judge of whether a show deserves a top
rating. Taking that into consideration, we will still try to model Nielsen ratings based on the
predictor variables mentioned above.
Sample Bias:
Nielsen households must all agree to participate in the survey, which is a potential source
of self-selection bias. We must ask ourselves if those households that agree to participate in the
survey take on certain characteristics that other households do not. Although Nielsen compares
its audience to the national population, this issue may still present a self-selection bias.
A second issue, that Nielsen readily admits, concerns whether the audience is actually
watching the television program or if their televisions are merely turned on. There really is no
way to determine the accuracy of this information except for the reliance on the diary that the
audience fills out during the year. This not only frustrates the programmers, but the advertisers
as well. While Nielsen measures the viewership of programs, it does not measure the viewership
of commercials, which is what the advertisers are interested in.
Thirdly, the diaries that the selected participants fill out during the sweeps periods have
issues of their own. The diaries are administered to supplement the metered data by obtaining a
Page 4
Statistics Data Analysis Problem
more precise breakdown of who is watching what shows. However, according to some critics,
the diary favors programs that air at the beginning of the diary week – Thursdays. Viewers are
apt to be more diligent about recording their viewing habits towards the beginning of the diary
week versus the latter half. This might explain, which we will discuss in detail, the tendency for
Thursday shows to be among the top-rated programs.
Finally, because Nielsen measures in-home viewing, rather than out-of-home viewing,
the ratings of shows that are watched in groups at locations such as bars, dorms, health clubs, etc.
could be deflated. Such programs that are under-measured are sports events, such as “Monday
Night Football”, which are typically watched in a bar or large gathering area. Another problem
regards the representation of certain audiences that tend to watch shows in large groups. For
instance, Boston has the highest concentration of college students of any city in the country.
Consequently, these students, who are within the desired 18-49 age demographic, are not fully
represented in their viewing habits since Nielsen collects its data solely through in-home
viewing.
The Data:
The data presented in our analysis is based on the 1998-1999 television season, starting in
September of 1998 and ending in September of 1999. We have included the Nielsen top-rated
200 programs for that season. The ratings were obtained from the Nielsen Media Research
report for prime-time, network-aired television programs. The response of interest is the overall
Nielsen share (percent of the average audience) for each show. The share is the determinant of
the actual number rating (1-200) for each show, however share is a more accurate target variable
as it determines the relevant audience size – which is ultimately used for advertising rates. The
predicting variables considered in the analysis are as follows:
Page 5
Statistics Data Analysis Problem
1. Network: The networks represented are NBC, CBS, ABC, FOX, UPN and WB.
These were obtained from the Nielsen report that corresponded each show to its
network. Only prime-time shows on these networks were included in this Nielsen
report for the top 200 shows.
2. New Show versus Recurring: A list of all new shows for the 1998 season was
obtained from the Infoplease website (www.infoplease.com). This includes both new
shows that started in September of 1998 as well as new shows that come in as
replacements in January of 1999.
3. Golden Globe Nomination: All shows that had Golden Globe nominations for the
following categories have been identified: Television Series – Drama; Actress in a
TV Series – Drama; Actor in a TV Series – Drama; TV Series – Comedy; Actress in a
TV Series – Comedy; Actor in a TV Series – Comedy. This data was obtained from
the Entertainment Weekly Online website (www.ew.com).
4. Emmy Nomination: All shows that had Emmy nominations for the following
categories have been identified: Outstanding Comedy Series; Outstanding Drama
Series; Outstanding Lead Actor in a Comedy Series; Outstanding Lead Actress in a
Comedy Series; Outstanding Lead Actor in a Drama Series; Outstanding Lead
Actress in a Drama Series; Outstanding Supporting Actor in a Comedy Series;
Outstanding Supporting Actress in a Comedy Series; Outstanding Supporting Actor
in a Dramatic Series; Outstanding Supporting Actress in a Dramatic Series;
Outstanding Guest Actor in a Comedy Series; Outstanding Guest Actress in a
Comedy Series; Outstanding Guest Actor in a Drama Series; Outstanding Guest
Actress in a Drama Series; Outstanding Variety, Music or Comedy Series;
Page 6
Statistics Data Analysis Problem
Outstanding Variety, Music or Comedy Special; Outstanding Performance in a
Variety or Music Program; Outstanding Animated Program; Outstanding Writing for
a Comedy Series; Outstanding Writing for a Drama Series; Outstanding Writing for a
Variety or Music Program. This data was obtained from the Entertainment Weekly
Online website (www.ew.com).
5. Day of the Week: The program schedule for the 1998-1999 season was obtained
from the Infoplease website (www.infoplease.com).
6. Type of Show: The shows were categorized in the following categories: comedy;
drama; news; sports.
7. Share in Program’s Timeslot: The Nielsen report provided each program’s share
within its timeslot.
Descriptive Statistics:
Before modeling the regression, we looked at the general characteristics of the data. It is
important to note that three of our predictors are categorical (day of week, type of show and
network), so we were not able to produce descriptive statistics on these three variables.
However, for the other 4 predictors and the target variable, the descriptive data is below. For a
frame of reference, the highest share value achieved for the 1998-1999 season (which was the
NBC television drama, E.R.) was 14.6%, which corresponds to an average audience size of 14.5
million viewers. This show had a 25% share in its Thursday night timeslot. The minimum share
value was 1.2% (which was UPN’s Home Movies), which corresponds to an average audience
size of 1.2 million viewers. This show had a 2% share in its Monday night timeslot.
Variable N Mean Median TrMean StDev SE Mean
HH AA % 200 5.576 5.950 5.441 2.896 0.205
HH US Sh 200 9.520 10.000 9.333 4.938 0.349
Premiere 200 0.2700 0.0000 0.2444 0.4451 0.0315
Globes N 200 0.0800 0.0000 0.0333 0.2720 0.0192
Emmy Nom 200 0.1200 0.0000 0.0778 0.3258 0.0230
Page 7
Statistics Data Analysis Problem
Variable Minimum Maximum Q1 Q3
HH US AA 1.200 14.600 2.725 7.500
HH US Sh 2.000 25.000 5.000 13.000
Premiere 0.0000 1.0000 0.0000 1.0000
Globes N 0.0000 1.0000 0.0000 0.0000
Emmy Nom 0.0000 1.0000 0.0000 0.0000
Network:
We first compare overall share based on the network that the show is on. The following
graph shows side-by-side boxplots of average share separated by the six possible networks.
Relative frequencies can also be determined from these boxes. The three major networks, NBC,
ABC and CBS, all seem to have the same average total median share within the top 200
programs. FOX’s median share is below those of the three major networks. These four
networks have a moderate amount of variability from their median share, whereas the UPN and
WB networks have little to no variability. This is because the UPN and WB do not have many
top rated shows in terms of share. The “lion’s share” of ratings is held by the three major
networks.
15
10
HH US AA%
5
0
ABC CBS FOX NBC UPN WB
Network
Page 8
Statistics Data Analysis Problem
Type of Show:
Does the type of show have any impact on the total share? Sports and news programs
have similar median shares, which is reflected in side-by-side boxplots in the graph below, and
represent an extremely small portion of the shows in the top 200. This must mean that those
news and sports shows that are included got very high ratings when they aired. Comedy and
drama programs have the same median shares and represent the bulk of the shows represented.
The boxplots show high variability because their representation covers the entire list – from
highest-rated to least-rated. There is one outlier based on type of show in the drama category.
This happens to be the top-rated show on the Nielsen ratings at 14.6% – E.R.
15
10
HH US AA%
5
0
Comedy Drama New s Sports
Type
Day of Week:
There seems to be a fair representation of each day of the week’s programs throughout
the entire list, except for Saturday programming. Saturday has the least variability and its
programming seems to fall in the bottom half of the ratings list with little to no representation in
the upper half. Most of the medians are within the same range of about 5-7%. The outlier in this
graph is a highly rated (8.4%) Saturday show on CBS, which is Walker Texas Ranger.
Page 9
Statistics Data Analysis Problem
Although there is no clear winner among day of the week, it is important to note that of
the top nine rated programs, five of them fell back-to-back on the NBC Thursday night lineup –
otherwise known as “Must See TV”. Unfortunately, we do not have this specific information for
all of the programs included on the list, but the Thursday night lineup has been consistently
present in the top of the Nielsen ratings. This indicates that from 8pm until 11pm, most viewers
start watching Friends and do not change the channel until after E.R., which airs at 10pm.
15
10
HH US AA%
5
0
F M S SA T TH W
Day of Wk
Emmy and Golden Globe Nominations:
The median overall share is higher for those shows that have either Golden Globe or
Emmy nominations as seen below in the side-by-side boxplots. Those programs with
nominations also have less variability around their medians. This seems like it would be a fairly
accurate predictor given the differences in medians between shows with and without
nominations. The outlier in the Emmy nominated category is, again, E.R. This seemed
surprising at first sight, but there is only one show between the outlier and the upper limit of the
third quartile.
There is one caveat to mention with respect to the awards nominations. For programs
such as the CBS Sunday Movie and Dateline, awards are nominated based upon particular
Page 10
Statistics Data Analysis Problem
movies or news stories. These types of nominations were excluded from this analysis since
awards are not granted for the overall programs and the inclusion would have overemphasized
the award.
15 15
10 10
HH US AA%
HH US AA%
5 5
0 0
0 1 0 1
Emmy Nomination? Globes Nomination?
Premiere Status:
Whether or not the show is new that season seems to have no bearing on the Nielsen
ratings, which is evidenced in the boxplots below. This makes sense because old shows that
were not watched were most likely canceled, so they would not be in the pool to bring the ratings
down for the old shows.
15
10
HH US AA%
5
0
0 1
Premiere?
Page 11
Statistics Data Analysis Problem
Share of Timeslot:
The variable with the most predictive power seems to be the share of the timeslot. This
seems to indicate that if a network wins its timeslot, the higher its overall share will be. This
makes sense as the highest overall rated shows grab almost a quarter of the total audience at that
particular time. The most popularly viewed shows (according to Nielsen) attract a large percent
of viewers during that timeslot. This might also indicate that the largest audiences are drawn to
the same viewing hours. The correlation between overall share and share of timeslot is
extremely high at .986 and has a p-value of zero, which indicates rejecting the null hypothesis
that the two are not correlated. The scatter plot below shows the strong correlation as well.
15
10
HH US AA%
5
0
0 5 10 15 20 25
HH US Shr
Multiple Regression Models:
Multiple Regression Models Using 7 Predictor Variables
Our initial model uses all 7 predictor variables in the belief that some of the categorical
variables could have some predictive power of the relative success of a show. The results
follow:
Page 12
Statistics Data Analysis Problem
Analysis of Variance for HH US AA, using Adjusted SS for Tests
Source DF Seq SS Adj SS Adj MS F P
Network 5 1101.709 0.342 0.068 0.45 0.811
HH US Sh 1 520.899 373.588 373.588 2468.53 0.000
Type 3 1.986 0.788 0.263 1.74 0.161
Premiere 1 0.610 0.298 0.298 1.97 0.162
Globes N 1 1.239 0.077 0.077 0.51 0.477
Emmy Nom 1 1.393 0.192 0.192 1.27 0.262
Day of W 6 13.976 13.976 2.329 15.39 0.000
Error 181 27.393 27.393 0.151
Total 199 1669.205
Term Coef StDev T P
Constant -0.1813 0.1356 -1.34 0.183
HH US Sh 0.58334 0.01174 49.68 0.000
Premiere 0.09378 0.06678 1.40 0.162
Globes N 0.0978 0.1371 0.71 0.477
Emmy Nom 0.1369 0.1217 1.12 0.262
Also, from the above data, we derived a standard error of 0.39, an R2 of 98.36% and an
overall F-statistic of 604.05. All of these factors indicate a good fit in the regression, however,
the network, type of show, premiere status, Golden Globes nomination and Emmy nomination all
have high p-values which indicate that they are not strong predictors and should therefore be
removed from the analysis. We subsequently performed the following analysis removing these
variables.
Multiple Regression Model Using 2 Predictor Variables
Analysis of Variance for HH US AA, using Adjusted SS for Tests
Source DF Seq SS Adj SS Adj MS F P
HH US Sh 1 1622.32 1600.78 1600.78 1.0E+04 0.000
Day of W 6 16.92 16.92 2.82 18.07 0.000
Error 192 29.96 29.96 0.16
Total 199 1669.20
Term Coef StDev T P
Constant -0.09756 0.06457 -1.51 0.132
HH US Sh 0.585795 0.005784 101.28 0.000
Day of W
F -0.45486 0.07626 -5.96 0.000
M 0.42188 0.06163 6.85 0.000
S 0.02782 0.06431 0.43 0.666
SA -0.5933 0.1102 -5.38 0.000
T 0.14330 0.06402 2.24 0.026
TH 0.21767 0.07018 3.10 0.002
Page 13
Statistics Data Analysis Problem
Based upon the above data, we derived the standard error of the estimate to be 0.40, an R2
of 98.21% and an overall F-statistic of 1463.61. These statistics support an extremely strong
regression. Both variables show small p-values which indicates that tail probabilities are
sufficiently low to reject the null hypothesis that there is no relationship between the predictor
variables and the response variable. As such, each predictor variable in this model is considered
statistically significant.
In addition, we note that the general linear model does not provide us with VIF statistics
therefore it is difficult for us to determine whether there is any multi-collinearity exhibited here.
However, since all variables have low individual p-values and there is a high overall F-statistic,
there is no indication that there is any multi-collinearity. This is a limitation in using this model.
Assumptions
We then tested the assumptions made with regard to the above model using 2 predictors. Based
on the normal probability plot of the residuals, it appears that the residuals are normally
distributed. In addition, the residuals versus the fitted values graph indicates no pattern to the
data, therefore showing constant variance. When looking at the residuals versus the one
continuous variable in the model, share of timeslot, it also appears that there is no pattern to the
data. Therefore, our regression assumptions appear to be valid.
Normal Probability Plot of the Residuals Residuals Versus the Fitted Values
(response is HH US AA) (response is HH US AA)
3 1
2
Normal Score
1
Residual
0
0
-1
-2
-1
-3
-1 0 1 0 5 10 15
Residual Fitted Value
Page 14
Statistics Data Analysis Problem
Residuals Versus HH US Sh
(response is HH US AA)
1
Residual
0
-1
0 5 10 15 20 25
HH US Sh
Implications of the Model
Based upon the above model, it appears that the only two relevant factors are the
program’s share of its timeslot and the day of the week. However, considering the day of the
week to be a strong predictor seems counterintuitive since the side-by-side boxplots showed that
there wasn’t a significant difference in the medians of the different days of the week and that the
only real difference was in the variability. We have previously shown that there is a very strong
correlation between the average audience and the share of the timeslot, so it seems reasonable
that it is a very strong predictor in our model. This model therefore answers the question of how
important it is for a program to win its timeslot.
However, timeslot share and average audience are actually derived from the same data
and may be so closely correlated that the timeslot share obscures the importance of the other
variables and makes all but the day of the week seem irrelevant. This can be seen by the
apparent importance given to the day of the week by this model, which had seemed to be one of
the weakest predictors when we had simply observed the data.
Page 15
Statistics Data Analysis Problem
Multiple Regression Model Using 6 Predictor Variables (excluding Share of Timeslot)
We then performed a regression on all the predictor variables, except share of timeslot, to
determine whether other variables are significant in determining the Nielsen ratings. The results
are as follows:
Analysis of Variance for HH US AA, using Adjusted SS for Tests
Source DF Seq SS Adj SS Adj MS F P
Network 5 1101.709 712.070 142.414 64.64 0.000
Type 3 45.813 55.330 18.443 8.37 0.000
Premiere 1 11.684 3.662 3.662 1.66 0.199
Globes N 1 68.949 12.843 12.843 5.83 0.017
Emmy Nom 1 22.789 21.624 21.624 9.82 0.002
Day of W 6 17.280 17.280 2.880 1.31 0.256
Error 182 400.980 400.980 2.203
Total 199 1669.205
Term Coef StDev T P
Constant 5.9713 0.2107 28.34 0.000
Premiere -0.3259 0.2528 -1.29 0.199
Globes N 1.2452 0.5157 2.41 0.017
Emmy Nom 1.4215 0.4537 3.13 0.002
Based upon the above regression, we calculated a standard error of the estimate of 1.48,
an R2 of 75.98% and an overall F-statistic of 33.86. However, it appears that there are variables
that should be removed based upon their high tail probabilities and low t-statistics, premiere
status and day of the week. This seems more in line with what we would expect based upon the
boxplots we looked at earlier. Therefore, we reran the regression eliminating these variables.
Multiple Regression Model Using 4 Predictor Variables (excluding Share of Timeslot)
We obtained the following results when we reran the regression:
Analysis of Variance for HH US AA, using Adjusted SS for Tests
Source DF Seq SS Adj SS Adj MS F P
Network 5 1101.71 772.18 154.44 69.34 0.000
Type 3 45.81 71.82 23.94 10.75 0.000
Globes N 1 75.08 13.79 13.79 6.19 0.014
Emmy Nom 1 25.65 25.65 25.65 11.52 0.001
Error 189 420.95 420.95 2.23
Total 199 1669.20
Page 16
Statistics Data Analysis Problem
Term Coef StDev T P
Constant 5.9623 0.1910 31.21 0.000
Network
ABC 1.4305 0.2322 6.16 0.000
CBS 2.4050 0.2373 10.13 0.000
FOX -0.1917 0.2453 -0.78 0.436
NBC 1.9114 0.2312 8.27 0.000
UPN -3.1593 0.2509 -12.59 0.000
Type
Comedy -1.0941 0.2235 -4.90 0.000
Drama -1.1076 0.2294 -4.83 0.000
News 0.1340 0.3193 0.42 0.675
Globes N 1.2835 0.5159 2.49 0.014
Emmy Nom 1.4890 0.4387 3.39 0.001
As we can see above, all of the predictor variables seem to be statistically significant at a
98% level. In addition, we calculated the standard error of the estimate to be 1.49, the overall F-
statistic to be 55.98 and the R2 to be 74.78%. Again, these factors indicate that our model is
statistically significant. Since the overall F-statistic increased from the previous model and the
R2 decreased only minimally, this appears to be a more appropriate model. The R2 indicates that
the model explains about 75% of the variability of the Nielsen ratings. The relevant variables
seem to be the network, the type of show, and whether or not it was nominated for a Golden
Globes award or an Emmy award. For the categorical variables, all categories seem to add some
element of predictability to our model, with the exception of the FOX network and the News
type of program, indicated by their high p-values. Therefore, we can be comfortable in using
this model for all categorical variables except these two.
Again, we are not able to obtain VIF statistics for these variables since many are
categorical. Therefore, we cannot determine if multi-collinearity is inflating the overall F-
statistic. Since, however, individual F-statistics and the overall F-statistic are both sufficiently
high, there is no indication that this is a problem. This is yet another limitation to the general
linear regression model.
Page 17
Statistics Data Analysis Problem
Assumptions:
We must now check the assumptions for this model as done for the previous model. The normal
probability plot of the residuals and the histogram of the residuals indicate that the residuals are
normally distributed. The residuals versus the fitted values however, seems to violate the non-
constant variance and exhibits homoscedasticity. This assumption violation indicates that the
results may not be trusted. The reason for the non-constant variance is not necessarily based
upon more variability in the data when overall rating is higher, but indicates that these predictor
values will play less of a role in the higher rated programs. This will be another limitation to this
model.
Normal Probability Plot of the Residuals Histogram of the Residuals
(response is HH US AA) (response is HH US AA)
3
40
2
30
Normal Score
1
Frequency
0 20
-1
10
-2
-3 0
-5 0 5 -5 0 5
Residual Residual
Residuals Versus the Fitted Values
(response is HH US AA)
5
Residual
0
-5
1 2 3 4 5 6 7 8 9 10
Fitted Value
Page 18
Statistics Data Analysis Problem
Conclusion:
Removing the variable for share of timeslot seems to provide us with a less accurate but
more intuitive model. If included in the model, the share of timeslot would be the only variable
needed to predict the Nielsen ratings (since day of the week only adds 1% to our R2). The
following model accounts for 98% of the variability in the Nielsen audience percentage.
Average Audience Percentage = -0.09756 + (0.585795 * HH Share) +
Coefficient for Day of the Week
Therefore, winning your timeslot appears to be critical to achieving high Nielsen ratings.
Once the share of timeslot variable is removed from the model, other variables, which
were previously obscured, come to the forefront as significant. Without using the timeslot share
variable, our model can be represented as follows:
Average Audience Percentage = 5.9623 + (1.2835 * Globes Nomination) +
(1.4890 *Emmy Nomination) + Coefficient for Network + Coefficient for Type
An estimated 75% of the variability is explained by this model which concludes that awards
nominations, network and type of show are the most appropriate predictors (of those within our
analysis) of a program’s success (as defined by high ratings).
As stated earlier, this model has certain limitations. The first rests in the initial
accumulation of the data by Nielsen Media Research, which is inherently biased. The second is
based upon the fact that we do not know the true VIF’s for the categorical variables in our model
and cannot determine the presence of multi-collinearity. And finally, the third surrounds the
violation of the constant variance assumption in this model.
Other variables definitely would have an impact in determining the success of a particular
show. One that we believe would be particularly helpful involves the affect that one show has on
the success or failure of surrounding shows. We have previously seen this phenomenon through
Page 19
Statistics Data Analysis Problem
the Thursday night NBC lineup which contains both Friends and E.R. and which, no doubt, has
an impact on the success of other programs like Jesse and Veronica’s Closet – two shows which
consistently get terrible reviews.
Critics’ reviews could have been useful in the model as possible predictor variables,
however, in our humble opinion, we do not believe that this will hold much predictive power.
Furthermore, we believe that there would be a high degree of non-constant variance as some of
the top-rated shows (Veronica’s Closet and Jesse) were consistently panned by the critics and
have even been cancelled for the 1999-2000 season. However, some of the lower rated shows,
such as X-Files, 3rd Rock from the Sun, and The Practice, are critically acclaimed yet do not
receive a large audience.
Another important point to note is the high correlation between the overall household
ratings and the ratings for the 18-49 demographic. The correlation is almost 90%, which is not
surprising given that the networks gear their shows toward this audience that is the most desired
by advertisers.
Finally, the most important thing to keep in mind is that people are fickle, writers get
hired and fired, stars come and go and particularly long-running shows lose their creative edge.
While this model can use several variables to predict ratings, it cannot predict such subjective
factors. We, therefore, suggest that one use caution when predicting the ratings based on the
variables included in this model.
Page 20