18.5 – Selecting the best model

Introduction

This is a long entry in our textbook, many topics to cover. We discuss aspects of model fitting, from why model fitting is done to how to do it and what statistics are available to help us decide on the best model. Model selection very much depends on what the intent of the study is. For example, if the purpose of model building is to provide the best description of the data, then in general one should prefer the full (also called the saturated) model. On the other hand, if the purpose of model building is to make a predictive statistical model, then a reduced model may prove to be a better choice. The text here deals mostly with the later context of model selection, finding a justified reduced model.

From Full model to best Subset model

Model building is an essential part of being a scientist. As scientists, we seek models that explain as much of the variability about a phenomenon as possible, but yet remain simple enough to be of practical use.

Having just completed the introduction to multiple regression, we now move to the idea of how to pick best models.

We distinguish between a full model, which includes as many variables (predictors, factors) as the regression function can work with, returning interpretable, if not always statistically significant output, and a saturated model.

The saturated model is the one that includes all possible predictors, factors, interactions in your experiment. In well-behaved data sets, the full model and the saturated model will be the same model. However, they need not be the same model. For example, if two predictor variables are highly collinear, then you may return an error in regression fitting.

For those of you working with meta-analysis problems, you are unlikely to be able to run a saturated model because some level of a key factor are not available in all or at least most of the papers. Thus, in order to get the model to run, you start dropping factors, or you start nesting factors. If you were unable to get more things in the model, then this is your “full” model. Technically we wouldn’t call it saturated because there were other factors, they just didn’t have enough data to work with or they were essentially the same as something else in the model.

Identify the model that does run to completion as your full model and proceed to assess model fit criteria for that model, and all reduced models thereafter.

In R (Rcmdr) you know you have found the full model when the output lacks “NA” strings (missing values) in the output. Use the full model to report the values for each coefficient, i.e., conducting the inferential statistics.

Get the estimates directly from the output from running the regression function. You can tell if the effect is positive (look at the estimate for sample — it is positive) so you can say — more samples, greater likelihood to see more cases of cancer.

Remember, the experimental units are the papers themselves, so studies with larger numbers of subjects are going to find more cases of diabetes. We would worry big time with your project if we did not see statistically significant and positive for sample size.

Example: A multiple linear model

For illustration, here’s an example output following a run with the linear model function on a portion of the Kaggle Cardiovascular disease data set.

The variables (and data types) were

SBP = systolic blood pressure (from ap_hi)

age.years = Independent variable, ratio scale (calculated from age in days)

BMI = Dependent variable, ratio scale (calculated from height and weight variables)

active = Independent variable, physical activity, binomial

alcohol = Independent variable, alcohol use, binomial

cholesterol = Independent variable, cholesterol levels (scored 1 = “normal,” 2 = “above normal, and 3 = “well above normal”)

gender = Independent variable, binomial

glucose = Independent variable, glucose levels (scored 1 = “normal,” 2 = “above normal, and 3 = “well above normal”)

smoke = Independent variable, binomial

R output

Call:
lm(formula = SBP ~ BMI + gender + age.years + active + alcohol + 
cholesterol + gluc + smoke, data = kaggle.train)

Residuals:
Min 1Q Median 3Q Max 
-33.359 -9.701 -1.065 8.703 45.242

Coefficients:
             Estimate Std. Error t value   Pr(>|t|) 
(Intercept)   76.7664    14.9311   5.141  0.00000155 ***
BMI            0.8395     0.2933   2.862  0.00522 ** 
gender        -2.0744     1.6102  -1.288  0.20092 
age.years      0.2742     0.2403   1.141  0.25672 
active         7.1411     3.4934   2.044  0.04383 * 
alcohol       -7.9272     9.1854  -0.863  0.39039 
cholesterol    7.4539     2.3200   3.213  0.00182 ** 
gluc          -1.2701     2.9533  -0.430  0.66817 
smoke         11.4132     6.7015   1.703  0.09197 . 
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 14.96 on 91 degrees of freedom
Multiple R-squared: 0.2839, Adjusted R-squared: 0.221 
F-statistic: 4.51 on 8 and 91 DF, p-value: 0.0001214

Question. Write out the equation in symbol form.

We see from the modest R² (adjusted) that the model explains little variation in systolic blood pressure; only BMI, physical activity (yes, no), and serum cholesterol levels, were statistically significant at 5% level. Smoker (yes, no) approached significance (p-value = 0.092), but I wouldn’t go on and on about positive or negative coefficients and effects.

We report the value (I would round to 7.5) for cholesterol on SBP (Table 1), and note that those who had elevated serum cholesterol levels tended to have higher systolic blood pressure (e.g., the sign of the coefficient — and I related the coefficient back to the most important thing about your study — the biological interpretation). Repeat the same for the other significant coefficients.

Now’s a good time to be clear about HOW you report statistical results. DO NOT SIMPLY COPY AND PASTE EVERYTHING into your report. Now, for the estimates above, you would report everything, but not all of the figures. Here’s how the output should be reported in your Project paper.

Table 1. Coefficients from full model.

term     Coefficient Error  P-value* 
Intercept       76.8  14.93 < 0.00001
BMI              0.8   0.29   0.00522
gender          -2.1   1.61   0.20092 
age.years        0.3   0.24   0.25672 
active           7.1   3.49   0.04383
alcohol         -7.9   9.19   0.39039 
cholesterol      7.5   2.32   0.00182 
gluc            -1.3   2.95   0.66817 
smoke           11.4   6.70   0.09197

* t-test

Looks better, doesn’t it?

Once you have the full model, we should run simple diagnostics on the regression fit to the data (VIF, Cook’s Distance, basic diagnostic plots), then proceed to use this model for the inferential statistics.

> vif(model.1)
BMI gender age.years active alcohol cholesterol gluc smoke 
1.112549 1.091154 1.081641 1.048519 1.096382 1.178635 1.168442 1.131081

No VIF was greater than 2, suggesting little multicollinearity among the predictor variables.

> cooks.distance(model.1)[which.max(cooks.distance(model.1))]
60 
0.355763

This person had high SBP (180), was a smoker, but active and normal BMI.

Use the significance tests of each parameter in the model from the corresponding ANOVA table. Now, where is the ANOVA table? Remember, right after running the linear regression,

Rcmdr: Models → Hypothesis testing → ANOVA tables

Accept the default (partial marginality), and, Boom! … the ANOVA table you should be familiar with.

From the ANOVA table you will tell me whether a Factor is significant or not. You report the ANOVA table in your paper. You describe it.

Now, the next step is to decide what is the best model. It then guides you to the next step which is to decide whether a better model (fewer parameters, Occam’s razor) can be found. Identify the parameter from the ANOVA table with the highest P-value and remove it from the model when you run the regression again. Repeat the steps above, return the ANOVA table, checking the estimates and P-values, until you have a model with only statistically significant parameters.

Find the best model

We may be tempted to run the linear model again, but with non-significant predictors excluded. For example, re-run the model on SBP but with just BMI, active, cholesterol and excluding smoke.

We can calculate the AIC, Akaike Information Criterion, the lower value is considered the better model.

AIC(model.1)
[1] 835.4938

AIC(model.2)
[1] 834.0782

The full model had eight predictor variables, excluding the intercept, whereas the reduced model had just three predictor variables, again, not including the intercept. Unsurprisingly, the reduced model is better. But now, a stronger test — what about smoking, where the p-value approached 5%? We ran the linear model (model.3).

> AIC(model.3)
[1] 831.9495

We can use an anova test or the likelihood ratio test to compare the models. The anova test, the models must be nested.

stats::anova(model.2, model.3)
Analysis of Variance Table

Model 1: SBP ~ BMI + active + cholesterol
Model 2: SBP ~ BMI + active + cholesterol + smoke
  Res.Df   RSS Df Sum of Sq      F   Pr(>F) 
1     96 22205 
2     95 21307  1     898.1 4.0043 0.04824 *
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

and

lmtest::lrtest((model.2,model.3)
Likelihood ratio test

Model 1: SBP ~ BMI + active + cholesterol
Model 2: SBP ~ BMI + active + cholesterol + smoke
 #Df  LogLik Df  Chisq Pr(>Chisq) 
1  5 -412.04 
2  6 -409.97  1 4.1286 0.04216 *
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Both tests tell us the same thing — our model with four predictors, including smoking, is better.

2 December 2024 — Past this point, edits pending

But first, I want to take up an important point about your models that you may not have had a chance to think about. The order of entry of parameters in your model can effect the significance and value of the estimates themselves. The order of parameter model entry above can be read top to bottom. Age was first, followed in sequence by CalsPDay, CholPDay, and so on. By convention, enter the covariates first (the ratio-scale predictors), that’s what I did above.

Here’s the output from a model in which I used a different order of parameters.

Anova(LinearModel.2, type="II")
Anova Table (Type II tests)

Response: BMI
              Sum Sq    Df    F value     Pr(>F)
Sex             1.52     1     0.0478    0.82764
Smoke           8.84     1     0.2782    0.59964
Age             0.62     1     0.0196    0.88918
CalsPDay       10.79     1     0.3394    0.56215
CholPDay       92.37     1     2.9065    0.09286 .
Sex:Smoke       3.35     1     0.1055    0.74631
Residuals    2129.34    67

The output is the same!!! So why did I give you a warning about parameter order? Run the ANOVA table summary command again, but this time select Type III type of test, i.e., ignore marginality

> Anova(LinearModel.2, type="III")
Anova Table (Type III tests)

Response: BMI
               Sum Sq    Df    F value       Pr(>F)
(Intercept)   1544.34     1    48.5929   1.708e-09 ***
Sex              4.69     1     0.1475     0.70214
Smoke           11.82     1     0.3720     0.54400
Age              0.62     1     0.0196     0.88918
CalsPDay        10.79     1     0.3394     0.56215
CholPDay        92.37     1     2.9065     0.09286 .
Sex:Smoke        3.35     1     0.1055     0.74631
Residuals     2129.34    67

The output has changed — and in fact it now reports the significance test of the intercept. This output is the same as the output from the linear model. Try again, this time selecting Type I, sequential

> anova(LinearModel.2)
Analysis of Variance Table

Response: BMI
           Df   Sum Sq    Mean Sq  F value    Pr(>F)
Sex         1     0.68      0.681   0.0214   0.88402
Smoke       1     2.82      2.816   0.0886   0.76690
Age         1     3.44      3.436   0.1081   0.74333
CalsPDay    1     2.27      2.272   0.0715   0.78998
CholPDay    1    96.30     96.299   3.0301   0.08633
Sex:Smoke   1     3.35      3.354   0.1055   0.74631
Residuals  67  2129.34     31.781

Here, we see the effect of order. So, as we are working to learn all of the issues of statistics and in particular mode fitting, I have purposefully restricted you to Type II analyses — obeying marginality correctly handles most issues about order of entry.

Easy there…. Take a deep breath, and guess what? Your best model needs to have significant parameters in it, right? Your best fit model then is Model 5. And that model will be your candidate for best fit as we proceed to complete our model building.

Now we proceed to gain some support evidence for our candidate best model. We are going to use an information criterion approach.

Use a fit criterion for determining model fit

To help us evaluate evidence in favor of one model over another there are a number of statistics one may calculate to provide a single number for each model for comparison purposes. The criteria model evaluators available to us include Mallow’s Cp, adjusted R², Akaike Information Criterion (AIC) or the Bayesian Information Criterion (BIC) to select best model.

We already introduced the coefficient of determination R² as a measure of fit – in general we favor models with larger values of R². However, values of R² will always be larger for models with more parameters. Thus, the other evaluators attempt to adjust for the parameters in the model and how they contribute to increased model fit.  For illustrative purposes we will use Mallow’s Cp. The equation for Mallow’s Cp in linear regression is

where p is the number of parameters in the model. Mallow’s Cp is thus equal to the number of parameters in the model plus an additional amount due to lack of fit of the model (i.e., large residuals). All else being equal we favor the model in which the Cp is close to the number of parameters in the model.

In Rcmdr, select Models → Subset model selection … (Fig. 1)

Figure 1. Rcmdr popup menu, Subset model selection…

From the menu, select the criterion and how many models to return. The function returns a graph that can be used to interpret which model is best given the selection criterion used. Below is an example (although for a different data set!) for Mallow’s Cp (Fig. 2).

Figure 2. Mallow’s Cp plot

Lets break down the plot. First define the axes. The vertical axis is the range of values for the Cp calculated for each model. The horizontal axis is categorical and reads from left to right: Intercept, Inspection[T.Yes], etc., up to Suspended. Looking into the graph itself we see horizontal bars — the extent of shading indicates which model corresponds to the Cp value. For example, the lowest bar which is associated with the Cp value of 8 extends all the way to the right of the graph. This says that the model evaluated included all of the variables and therefore was the saturated or full model. The next bar from the bottom of the graph is missing only one block (lgCarIns), which tells us the Cp value 6 corresponds to a reduced model, and so forth.

Cross-validation

Once you have identified your Best Fit model, then, your proceed to run the diagnostics plots. For the rest of the discussion we return to our first example.

Rcmdr: Models → Graphs → Basic diagnostic plots.

We’ll just concern ourselves with the first row of plots (Fig. 3).

Figure 3. Diagnostic plots

The left one shows the residuals versus the predicted values — if you see a trend here, the assumption of linearity has been violated. The second plot is a test of the assumption of normality of the residuals. Interpret them (residuals OK, Residuals normally distributed? Yes/ No), and you’re done. Here, I would say I see no real trend in the residuals vs fitted plot, so assumption of linear fit is OK. For normality, there is a tailing off at the larger values of residuals, which might be of some concern (and I would start thinking about possible leverage problems), but nothing dramatic. I would conclude that our Model 6 is a good fitting model and one that could be used to make predictions.

Now, if you think a moment, you should identify a logical problem. We used the same data to “check” the model fit as we did to make the model in the first place. In particular if the model is intended to make predictions it would be advisable to check the performance of the model (e.g., does it make reasonable predictions?) by supplying new data, data not used to construct the model, into the model. If new data are not available, one acceptable practice is to divide the full data set into at least two subsets, one used to develop the model (sometimes called the calibration or training dataset) and the other used to test the model. The benefits of cross-validation include testing for influence points, over fitting of model parameters, and a reality check on the predictions generated from the model.

A note on best practices

In the 1980s, researchers often “removed” the effect of a covariate by first fitting a regression and then saving the residuals to use in separate analyses, such as t-tests, ANOVAs, or variance component estimates (eg, Dohm et al 2001). For example, in a study of plant growth, scientists might regress final plant height on initial size to remove the effect of size, then compare the residuals between fertilizer groups. While this approach adjusts for the covariate, it is problematic because residuals are not raw data: they are correlated, have constrained sums, and underestimate uncertainty. This can distort group differences, especially if the covariate distributions differ among groups, and it ignores the variability introduced by the first regression. This approach was common in ecology, evolutionary biology, physiology, and psychology before modern mixed modeling and generalized regression workflows became standard.

Modern best practices instead recommend fitting a single unified model that includes both the covariate and the grouping variable directly. Using the same plant growth example, a linear model such as estimates the effect of fertilizer while properly accounting for initial plant size. This approach provides valid standard errors, allows correct inference about group differences, and can accommodate interactions (eg, if the relationship between and differs by group) or hierarchical data structures using mixed-effects models. Additionally, modern workflows include diagnostics such as residual plots, Cook’s Distance, and variance inflation factors to check model assumptions and influential points. Use mixed models when data are hierarchical or repeated data. Random effects handle variation among groups, individuals, years, etc., in a single coherent model. When covariates need to be “controlled for,” we use partial regression. This happens inside the model, not by subtracting covariate effects first. By analyzing all relevant predictors in a single model, researchers obtain more accurate estimates, properly reflect uncertainty, and avoid the statistical pitfalls of the old residual-based two-stage approach.

Thus, there are several statistical reasons the older approach fails under modern scrutiny. First, Residuals are not independent observations. Residuals are model-dependent quantities, not raw data. Once you fit a regression, the residuals are correlated with each other, constrained to sum to zero, and, likely heteroscedastic if the model was imperfect. Using them in a new test violates standard assumptions of the second test.

Second,  you underestimate uncertainty. If you perform tests on residuals, the standard errors do not include uncertainty from the first regression, which may lead to inflated Type I error and an overly confident “adjusted” group comparisons.

Third, the old approach creates a “two-stage model” that breaks the likelihood framework. A two-stage analysis discards the correct joint likelihood and therefore produces biases p-values, and interferes with valid confidence intervals or predictions.

Under the modern one-stage models include ML/REML, generalized estimating equations, and fully integrated Bayesian models, which keeps the model coherent. Use of mixed models make residualization unnecessary. In studies with repeated measures, blocks, sites, subjects, or years, the old approach cannot partition sources of variation correctly.

In short, the advantages of the one-stage approach is that resulting models have valid uncertainty, correct group comparisons, better handling of interactions and hierarchy, full likelihood-based inference. They fit a single unified regression or mixed-effects model with all relevant predictors, examine diagnostics, and make inferences directly. This is one more benefit of the increased power and availability of more powerful computers and more complete development of mixed methods modeling.

Questions

1. Write up three learning outcomes for this page. Hint: Point your favorite generative AI to this page and ask for help.

Quiz Chapter 18.5

Selecting the best regression model

Start Quiz

Question

Confirm

✕

✓

Your answer:

Correct answer:

Next

Please wait..

↺

You got {{SCORE_CORRECT}} out of {{SCORE_TOTAL}}

Your Answers

Chapter 18 contents

• Introduction
• Multiple linear regression
• Nonlinear regression
• Logistic regression
• Generalized Linear Squares
• Selecting the best model
• Compare two linear models
• References and suggested readings (Ch17 & 18)

14.5 – Nested designs

Introduction

Crossed versus nested design

Factors are independent variables whose values we control and wish to study because we believe they have an effect on the dependent variable. While it is logical to think of factors and levels within factors as independent variables fully under our control, a moments reflection will come up with examples in which the groups (levels) depend on the factor.

Crossed – each level of a factor is in each level of the other factor. This was illustrated in Chapter 14.1 on Crossed, balanced, fully replicated two-way ANOVA.

Nested – levels of one factor are NOT the same in each of the levels of the other factor. Nested designs are an important experimental design in science, and they have some advantages over the 2-way ANOVA design (for one), but they also have limitations.

Classic examples of nesting: culturing and passage of cell lines in routine cell colony maintenance means that even repeated experiments are done on different experimental units. Cells derived from one vial are different from cells derived from a different vial. Similarly, although mice from an inbred strain are thought to be genetically identical, environments vary across time, so mice from the same strain but born or purchased at different times are necessarily different. These scenarios involving time create a natural block effect. Thus, cells are nested by block effect passage number and mice are nested by block effect colony time. We introduced randomized block design in the previous Chapter 14.4.

Statistical model

If Factor B is nested within Factor A, then a group or level within Factor B occurs only within a level of Factor A. Like the randomized block model, there will be no way to estimate the interaction in a nested two-way ANOVA. Our statistical model then is

Examples

Example 1. Three different drugs, 6 different sources of the drugs. The researcher obtains three different drugs from 6 different companies and wants to know if one of the drugs is better than another drug (Factor A) in lowering the blood cholesterol in women. There is always the possibility that different companies will be better or worse at making the drug. So the researchers also use the Factor Source (Factor B) to examine this possibility. Unfortunately they can not obtain all drugs from the same sources. This leads to a Nested ANOVA — notice that each drug is obtained from a different source.

We CANNOT perform the typical two-factor ANOVA because we cannot get a mean of the different drugs by combining the same levels of the Sources: the data is NOT crossed. The Sources of the drugs (Factor B) are NESTED within the type of Drug (Factor A): each source is only found in one of the Drug categories. So, we can’t calculate a mean for the Drug levels independent of the SOURCE from which the drug came.

Table 1. Example of a nested design

Scroll to end of this page to get the data set in stacked worksheet format, or click here.

Compare Table 1 to CROSSED data structure (Table 2) — a typical two-factor ANOVA — which would look like

Table 2. Contents of Table 1 presented as crossed design

We can take a mean of the different drugs by combining the same levels of the Sources. Here’s the nested design (Table 3).

Table 3. Group means, nested design

We can take a mean of the different drugs by combining the same levels of the Sources. Here’s the crossed design (Table 4).

Table 4. Groups means by crossed design.

Why the “?” in Table 2 and 4? Manufacturing source 1 & 2 do not sell Drug B and Drug C. So, there cannot be a crossed design.

Why can’t we just use a One-Way ANOVA? Can’t we just ANALYZE the three DRUGS separately, ignoring the source issue (after all, the drugs are not all made by the same manufacturer)? But it is not a one-way ANOVA problem… Here’s why.

The researcher suspects that the response of a particular drug might be dependent upon the particular source from which the drug was purchased. So, the type of source from which the drug was purchased is another FACTOR. Thus, drugs from one source might have more (less) affect compared to drugs from another source regardless of the type of drug. However, each drug is NOT available from each source. Thus the research design can NOT be crossed and Drug is NESTED within Source.

We can ask ONLY two questions (hypotheses) from this NESTED ANOVA research design:

H_O: There is no difference in the average effect of the drugs on (tumor size, cholesterol level, blood pressure, etc.)

H_A: There is a difference in the average effect of the drugs on (tumor size, cholesterol level, blood pressure, etc.)

H_O: There is no difference in the average effect of the drugs on (tumor size, cholesterol level, blood pressure, etc.) purchased from different manufacturers.

H_A: There is a difference in the average effect of the drugs on (tumor size, cholesterol level, blood pressure, etc.) purchased from different manufacturers.

Notice that we do NOT examine the effect of the interaction between Drug type and source of the drug. Why not?
Table 5. Sources of Variation in Nested ANOVA

Source of Variation	Sum of Squares	DF	Mean Squares
Total
Among all subgroups
Among Groups
Among Subgroups
Error	Subtract all of the subgroup Sums of Squares from the Total Sums of Squares

Testing nested ANOVA with one main factor

Perhaps surprisingly given the number of terms above, there are only two hypothesis tests, and, only one of REAL interest to us. There are exceptions (eg, quantitative genetics provides many examples), but we are generally most interested in the among group test — this is the test of the main factor. In our example, the main factor was DRUG and whether the drugs differed in their effects on cholesterol levels. The second test is important in the sense that we prefer that it contributes little or no variation to the differences in cholesterol levels. But it might.

Table 6. F statistics for nested ANOVA

F for the main effect is given as
F for the subgroup is given by
and of course, use the appropriate DF when testing the F values!! The Critical Value F0.05 (2), df numerator, df denominator

One way to look at this: it would not make sense to conclude that an effect of the main group was significant if the variation in the subgroups was much, much larger. That’s in part why we test the main effect with the subgroups MS and not the error MS. If variation due to the nested variable is not significant, then it is an estimate of the error variance, too.

The nested model we are describing is a two factor ANOVA, but it is incomplete (compared to the balanced, fully crossed 2-way design we’ve talked about before). We don’t have scores in every cell. Instead, each level of nested factor is paired with one and only one level of the other factor. In our example, Source is paired with only one other level of the other factor Drug (eg, Source 1 goes with Drug 1 only), but the main effect is paired with 2 levels of the nesting factor (eg, Drug 1 is manufactured at Source 1 and Source 2).

Some important points about testing the null hypotheses in a nested design. For one, the test of the effect of the nesting factor (Source) is confounded by the interaction between the main factor. We don’t actually know if the interaction is present, but we also get no way to test for it because of the incomplete design. We must therefore be cautious in our interpretation of the effect of the nested factor.

Consider our example. We want to interpret the effect of source as the contribution to the response based on variation among the different suppliers of the drugs. It might be good to know that some drug manufacturer is better (or worse) than others. However, differences among the sources for the different drugs are completely contained in the main effect factor (the test of effects of the different drugs themselves on the response). Therefore, the observed differences between sources COULD be entirely due to the effects of the different drugs and have nothing to do with variation among sources!!

Questions

1. Identify the response variable and whether the described factor (in all caps) is suitable for crossed design or nested design
   a. In a breeding colony of lab mice, BREEDERS are used to generate up to five LITTERS; effects on offspring REPRODUCTIVE SUCCESS.
   b. Effects of individual TEACHERS at different SCHOOLS on STUDENT LEARNING in biology.
   c. Lisinopril, an ACE-inhibitor drug prescribed for treatment of high blood pressure, is now a generic drug, meaning a number of COMPANIES can manufacture and distribute the medication. Millions of DOSES of lisinopril are made each year; drug companies are required by the FDA to record when a dose is made and to record these dates by LOT NUMBER.
2.  Work the example data set provide in this page. After loading the data set into Rcmdr (R), use linear model. The command to nest requires use of the forward slash, /. For example, if factor b is nested within factor a, then a/b. The linear model formula then,

Model <- lm(Obs ~ a/b, data=source)

1. Describe the problem, i.e., what is a? What is b? What are the hypotheses?
2. What is the statistical model?
3. Test the model.
4. Conclusions?

Quiz Chapter 14.5

Nested designs

Start Quiz

Question

Confirm

✕

✓

Your answer:

Correct answer:

Next

Please wait..

↺

You got {{SCORE_CORRECT}} out of {{SCORE_TOTAL}}

Your Answers

Data set used in this page

Chapter 14 contents

• Introduction
• Crossed, balanced, fully replicated
• Sources of variation
• Fixed effects, Random effects
• Randomized block design
• Nested designs
• Some other ANOVA designs
• Rcmdr Multiway ANOVA
• More on the linear model in Rcmdr
• References and suggested readings

14.2 – Sources of variation

Introduction

Sources of variation, or components of the two-way ANOVA include two factors, each with two or more levels (groups), and collectively, factors are often referred to as the main effects in these types of ANOVA. The other source of variation in a two-way ANOVA is the interaction between the two factors. Below, I have listed the important components, although I have not included how the sum of squares are calculated. You are expected to know the sources of variation for this most basic two-way ANOVA table (Fig 1). You should also be able to solve any missing elements in one of these tables by utilizing any included information.

Figure 1. ANOVA table for two-way, balanced, replicated design.

Taking each row from Figure 1 one at a time we have

Source	DF	Mean Squares	F-statistic
First Factor

where Source refers to the source of variation, DF refers to Degrees of Freedom, a is the number of levels (groups) of the first factor, SS refers to Sum of Squares, and MS refers to the Mean Squares.

Next is the second factor

Source	DF	Mean Squares	F-statistic
First Factor

where b is the number of levels (groups) of the second factor. Next is the interaction between the first and second factors.

Source	DF	Mean Squares	F-statistic
Interaction

and lastly the Within-cell Error or residual source of variation

Source	DF	Mean Squares	F-statistic
Error

where n is the number of experimental units for each group. Note that if the sample size differs for one or more groups (levels),then the design would be unbalanced and this formula does not work to determine the degrees of freedom. The total degrees of freedom for the two-way ANOVA is simply N – 1, where N is the sample size for the entire problem; a little algebra shows that N may be calculated as

Unbalanced designs

An unbalanced design implies that observations are missing value for one or more groups. What to do if data are missing? Decision depends on how the data are missing (see Chapter 5). For example, if data are missing at random with respect to treatment, then this should not affect inference. If data are missing not at random, then inference, logically, must be impacted. Calculating the ANOVA, moreover, becomes a different matter. In the one-way ANOVA, no real problem arises although setting up contrasts among the levels requires a weighting term to be factored into the calculations. For higher-level ANOVA involving two or more factors the sums of squares for treatment effects are no longer simple partitioning into the different sources of variation. The sources overlap and the order by which the Factors enter into the statistical model now affects the calculations. Thus, while setting up the calculations for the balanced design is straight-forward, perhaps surprisingly, if group sizes differ, this simple relationship for calculating the degrees of freedom, sums of squares, and Mean squares  become an unsolvable problem. This problem is largely solved by the general linear model.

Questions

1. Based on the results of a two-way ANOVA, the error sums of squares (SSE) was computed to be 160. If we ignore one of the factors and perform a one-way ANOVA using the same data, will the SSE be the same as in the two-way ANOVA, or will it increase?  Decrease? Explain your choice.
2. While conducting a two-way ANOVA, you conclude that a statistically significant interaction exists between factor 1 and factor 2. What should be your next step? Do you drop the interaction term from the model and redo the analysis or do you report the results of factor effects including the non-significant interaction?

Quiz Chapter 14.2

Sources of variation

Start Quiz

Question

Confirm

✕

✓

Your answer:

Correct answer:

Next

Please wait..

↺

You got {{SCORE_CORRECT}} out of {{SCORE_TOTAL}}

Your Answers

Chapter 14 contents

• Introduction
• Crossed, balanced, fully replicated
• Sources of variation
• Fixed effects, Random effects
• Randomized block design
• Nested designs
• Some other ANOVA designs
• Rcmdr Multiway ANOVA
• More on the linear model in Rcmdr
• References and suggested readings

12.1 – The need for ANOVA

Introduction

Moving from an experiment with two groups to multiple groups is deceptively simple: we move from one comparison to multiple comparisons. Consider an experiment in which we have randomly assigned patients to receive one of three doses of a statin drug (lower cholesterol), including a placebo (eg, Tobert and Newman 2015). Thus, we have three groups or levels of a single treatment factor and we’ll want to test the null hypothesis that the group (level) means are all equal as opposed to the alternative hypothesis in which one or more of the group means, eg, group A, Group B, group C, are different.

The correct procedure is to analyze multiple levels of a single treatment with a one-way analysis of variance followed by a suitable post-hoc (“after this”) test. Two common post-hoc tests are Tukey’s range test (aka Tukey’s HSD [honestly significant difference] test), which is for all pairwise comparisons, or the Dunnett’s test, which compares groups against the control group. Post-hoc tests are discussed in Chapter 12.6.

For our three group experiment, how many  pairwise comparisons can be tested? Therefore, , and we have

Thus, for our three groups, A, B, and C, there are three possible pairwise comparisons.

How many pairwise comparisons for a four group experiment? Check your work, you should get .

The multiple comparison problem

Let’s say that we’re stubborn. We could do many single two-sample t-tests — certainly, your statistical software won’t stop you — this is a situation that calls for statistical reasoning. Here’s why we should not: we will increase the probability of rejecting a null hypothesis when the null hypothesis is true (eg, discussion in Jafari and Ansari-Pour 2019). That is, the chance we will commit a Type I error increases if we do not account for the lack of independence in these sets of pairwise tests evaluated by t-tests. This is multiplicity or the multiple comparison problem.

Review: when we perform a two-sample t-test we are willing to reject a true null hypothesis 5% of the time. This is what is meant by setting the critical probability value (alpha) = 0.05. By “willing” we mean that we know that our conclusions could be wrong because we ar3 working with samples, not the entire population. (Of course at the time, we have no way of actually knowing WHEN we are wrong, but we do want to know how likely we could be wrong!) However, if we compare three population means we have three separate null hypotheses.

H_O: one or more of the means are different.

But if we conduct these as separate independent sample t-tests, then we are implicitly making the following null hypothesis statements:

Thus, we have a 5% chance of being wrong for the first hypothesis and/or a 5% chance of being wrong for the second hypothesis and/or a 5% chance of being wrong for the third hypotheses. The chance that we will be wrong for at least one of these hypotheses must now be greater than 5%.

For three separate hypotheses there is a 14% chance of being wrong when we have the probability value for each individual t-test set at α = 0.05. How did we get this result? The point is that these tests are not independent, they are done on the same data set, therefore you can’t simply apply the multiplication probability rule.

Here’s how to figure this: for the set of three hypotheses, the probability of incorrectly rejecting at least one of the null hypotheses is

So, for three t-tests on the same experiment, the Type I error overall tests (experiment-wise) is actually 14% not 5%. It gets worse as the number of combinations (groups and therefore hypotheses) increases. For four groups, Type I error is actually

That’s 26.5%, not 5%.

If we have just five populations means to compare the probability of rejecting a null hypothesis when it is true climbs to 60%! How did this happen? The probability or correctly rejecting all of them is now

So, the probability of incorrectly rejecting one test (Type I error) is now

instead of the 5% we think we are testing.

The key argument for why you must use ANOVA to analyze multiple samples instead of a combination of t-tests!! ANOVA guarantees that the overall error rate is the specified 5%.

Why is the Type I error not 5% for each test? Because we conducted ONE experiment, we can conduct only ONE test (we could be right, we could be wrong 5% of the time). If we conduct the experiment over again, on new subjects, each time resulting in new and therefore independent data sets, then Type I error = 5% for each of these independent experiments.

Now, I hope I have introduced you to the issue of Type I error at the level of a single comparison and the idea of an experiment, holding Type I error-rates at 5% across all hypotheses to be evaluated in an experiment. You may wonder why anyone would make this mistake now. Actually, people make this “mistake” all the time and in some fields like evaluating gene expression for microarray data, this error was the norm, not the exception (see for example discussion of this in Jeanmougin et al 2010).

To conclude, if one does multiple tests on the same experiment, whether it is t-tests or some other test for that matter, then our subsequent tests are related. This is what we mean by independence in statistics — and there are many ways that nonindependence may occur in experimental research. For example, we introduced the concept of pseudoreplication, when observations are treated as if they are independent, but they are not (see Chapter 5.2). The “multiple comparison problem” specifically refers to the lack of independence when all the data set from a single experiment is parsed into lots of separate tests. Philosophically, there must be a logical penalty — and that is reflected in the increase in Type I error.

Clearly something must be done about this!

Corrections to account for the multiple comparisons problem

The family-wise (aka experiment-wise) error rate for multiple comparisons is kept at 5%, and each individual-wise comparison is compared against a more strict (ie, smaller Type I error rate). Put another way, the family-wise error rate is the chance of a number of false positives: making a mistake when we consider many tests simultaneously. The simplest correction for the individual-wise error rate is the Bonferroni correction: test each individual comparison at Type I error equal to , where C is the number of comparisons. The Šidák correction, , and the Holm-Bonferroni method are alternatives with reportedly better statistical properties (eg, reduced Type II error rates; Ludbrook 1998, Sauder and DeMars 2019).

ANOVA is a solution

One possible solution for getting the correct experiment-wise error rate: adjust for differences in probability for multiple comparisons with the t-test. We used post-hoc tests presented above: you could evaluate the tests after accounting for the change in Type I error. This is what is done in many cases. For example, in genomics. In the early days of gene expression profiling by microarray, it was common to see researchers conduct t-tests for each gene. Since microarrays can have thousands of genes represented on the chip, then these researchers were conducting thousands of t-tests, arranging the t-tests by P-value and counting the number of P-values less than 5% and declaring that the differences were statistically significant.

This error didn’t stand long, and there are now many options available to researchers to handle the “multiple comparisons” problem (some probably better than others, research on this very much an ongoing endeavor in biostatistics). The Bonferroni correction was an available solution, largely replaced by the Holm (aka Holm-Bonferroni) method (Holm 1979). Recall that the Bonferroni correction judged individual p-values statistically significant only if they were less than , where, again, α is the family-wise error rate (eg, α = 5%) and C was the number of comparisons. The Holm method orders the C p-values from lowest to highest rank. The method then evaluates lowest p-value, if less than , then reject hypothesis for that comparison. Proceed to next p-value, if less than , then reject hypothesis for that comparison, and so on until no more comparisons are less than .

A MUCH better alternative is to perform a single analysis that takes the multiple-comparisons problem into account: Single Factor ANOVA, also called the one-way ANOVA, plus the post-hoc tests with error correction. We introduce one-way ANOVA in the next section. Post-hoc tests are discussed in Chapter 12.6.

Questions

1. You should be able to define and distinguish how Bonferoni correction, Dunnett’s test, and Tukey’s test methods protect against inflation of Type I error.
2. What will be the experiment-wise error rate for an experiment in which there are only two treatment groups?
3. Experiment-wise error rate may also be called _______________ error rate.
4. List and compare the three described posthoc approaches to correct for multiple comparison problem,
5. Glycophosphate-tolerant soybean is the number one GMO (genetically modified organism) crop plant worldwide. Glycophosphate is the chief active ingredient in Roundup, the most widely used herbicide. A recent paper examined “food quality” of the nutrient and elemental composition of plants drawn from fields which grow soy by organic methods (no herbicides or pesticides) and GMO plants subject to herbicides and pesticides. A total of 28 individual t-tests were used to compare the treatment groups for different levels of nutrients and elements (eg, vitamins, amino acids, etc.,); the authors concluded that 10 of these t-tests were statistically significant at Type I error rate of 5%. Discuss the approach to statistical inference by the authors of this report; include correct use of the terms experiment-wise and individual-wise in your response and suggest an alternative testing approach if it is appropriate in your view.
6. In Chapter 4.2 we introduced a data set from an experiment. The experiment looked at DNA damage quantified by measuring qualities in a Comet Assay including the Tail length, the percent of DNA in the tail, and olive moment. The data set is copied to end of this page. In the next chapter I’ll ask you to conduct the ANOVA on this experiment. For now, answer the following questions.
   a. What is the response variable? Defend your choice(s).
   b. Why is this an ANOVA problem? Include as part of your explanation a statement of the null hypothesis.

Quiz Chapter 12.1

The need for ANOVA

Start Quiz

Question

Confirm

✕

✓

Your answer:

Correct answer:

Next

Please wait..

↺

You got {{SCORE_CORRECT}} out of {{SCORE_TOTAL}}

Your Answers

Data used in this page

comet assay dataset

Chapter 12 contents

• Introduction
• The need for ANOVA
• One way ANOVA
• Fixed effects, random effects, and agreement
• ANOVA from “sufficient statistics”
• Effect size for ANOVA
• ANOVA posthoc tests
• Many tests one model
• References and suggested readings