48, Linear modeling Thus the weighted observations are homoskedastic since Var(e7) = 07. Regressing y} on the xj, yields BY = (XY X*) IXY? = (X'WX) OX Wy, (4.17) where W is the diagonal matrix with diagonal entries w;. The estimator 3* is called the weighted least squares estimator. Since (3° is the least squares imator computed relative to the model y* = X* + ¢* it follows that BY ~N {B,0%(X'WX)"} Weighted least squares estimators have all the desirable properties of least squares estimators, provided of course correct weights are used. 4.8 Grouped and ungrouped data Insurance data are often represented in tabular or grouped format. As an exam- ple, consider an analysis of claim sizes in the vehicle insurance data, by age and sex of the driver. For policies which had a claim, mean claim sizes by age and sex are en in Table 4.1. For male drivers, mean claim size decreases with increasing age, For female drivers the pattern is not as clear. The ques- tion arises as to the appropriate analysis for data in this form. Although there are 12 cells of data, the table is based on the 4624 policies which had a claim. A regression analysis with claim size as the response, and driver's age and sex as explanatory variables, is appropriate. This may be carried out in two ways: (the “grouped data” approach. This is based on the 12 observations in ‘Table 4.1, with regression weights equal to the number of observations in each cell; ii) the “ungrouped data” approach. This is based on the 4624 raw observations. The two approaches yield the same parameter estimates. The grouped data approach has been, and still is, popular, as data representation and storage is dramatically simpler in cases such as the current example. Computation is also simplified. However, in realistic analyses many explanatory variables are typi- cally considered, meaning that the summary table is highly multidimensional. In addition, with modern computing the saving in compytation time is mini- mal, Thus while a summary table such as Table 4.1 is a useful preliminary data inspection tool, its use in analysis does not present any advantage over use of the raw observations: In any case, if there is at least one explanatory variable which is continuous, then the grouped data approach cannot be used. ‘The ungrouped data approach is used throughout this text, i.e, analyses are performed on the basis of raw observations such as individual policies. The4.9 Transformations to normality and linearity 49 Table 4.1. Vehicle insurance: mean claim size (dollars) by age and sex of driver, number of observations in brackets i ers age categor e 1 O03 es Male 2118 1945 1869 1800 1724 1546 (276) (537) (688) (643) (322) (182) Female 3285 2381 1991 2143 1734 2198 (220) 395) (425) (461)_(292)_(183) grouped and ungrouped approaches are contrasted in the context of logistic regression, in Section 7. 4.9 ‘Transformations to normality and linearity Transformations of the response and explanatory variables are frequently car- ried out in order to view and analyze data on an appropriate scale. This has been discussed in Section 1.3. When performing a linear regression, data is transformed in order to satisfy the model assumptions of linearity between y and the explanatory variables, homoskedasticity and normality of ¢. While normality of the explanatory variables is not necessary, when an variable is strongly skewed a linear relationship with y is unlikely. Transformations which normalize, or at least symmetrize, the distribution of an explanatory variable frequently also linearize its relationship with y or some transform of y. Third party claims. Consider a model for the number of claims in an area as a function of the number of accidents. A scatterplot of claims against acci- dents is shown in the left panel of Figure 4.1, This is strongly heteroskedastic. Diagnostic checks on a linear model based on these data indicate clear viola- tion of the homoskedasticity assumption, The concentration of points around the origin make it difficult to discern the relationship between y and x. The top two panels of Figure 4.2 show the histograms of the number of accidents and the number of claims, Both are strongly skewed to the right. The bottom two panels are histograms of the log-transformed data, which appear more normal than those of the raw data, A scatterplot of the log-transformed data is shown in the right panel of Figure 4.1. This displays linearity and approx- imate homoskedasticity, and therefore a linear model based on log claims as, the response and log accidents as the explanatory variable, is more amenable to analysis using the normal linear model than one based on the raw data. ‘A GLM analysis is more appropriate for these data and is discussed in Section 6.2.50 Linear modeling a ; ° ; 8 | a. Be a & a, & 2600000 6000 8000 7s 6 8 ‘cients Log accionts Fig. 4.1. Scatterplots of number of accidents and number of claims, raw and log scales ] 84) 8 at 28 3. Be . o (MMlana = a2 3 2000” ado0 " edo lains 8 2 Bo ce &= e 8 . oid etrrre stertree Log acients Log elie Fig. 4.2. Histograms of number of accidents and number of claims, raw and log scales ing transformations. The above example demonstrates the use of the logarithmic transformation to both stabilize the variance of y and linearize its relationship with «. If the variance of y is of the form Var(y) = $H?, where @ > 0 is a scale parameter, then the transformation g(y), which makes the variance approximately independent of the mean, is such that the derivative of g(y) with respect to y is y~?/?. This is established from oy) 9(H) + G(H(y-H) => ——Var{o(u)} © {5(u)}PVar(y) »4.10 Categorical explanatory variables st where g(y) is the derivative of g evaluated at y. Integrating (1) = w-?/? with respect to 1 yields go) { y 2 forp #2 In(y)— forp=2. This is the Box-Cox transformation described in Section 1.3. As an example, when y has variability proportional to the mean (such as in the case of Poisson. data), Var(y) = dy, p = 1 and hence the variance stabilizing transform is gy) = VJ. If the standard deviation is proportional to the mean and hence p = 2, then the log transform stabilizes the variance. Variance stabilizing transformations were important before the GLM methodology became available, When non-normal responses were modeled, the only option was to find a transformation which would render the data amenable to analysis using the normal linear model. For example, when ana- lyzing count data which had the Poisson distribution, a normal linear model analysis was performed using the square root of y as the response. With GLMs the data is modelled directly, using the appropriate response distribution. 4.10 Categorical explanatory variables When a potential explanatory variable is categorical, then it is “dummied up” for inclusion into a multiple linear regression, For example, suppose the area of residence of a policyholder is considered to be a potential explanatory variable in a model for claim size y. If there are three areas A, B and C, then two indicator (dummy) variables «7 and za are defined: ‘Suppose area is the only explanatory variable. Then the relationship is modeled as ¥ © Bo+ Bras + Bowe - This states that y = ip in area C, y © By + Br in area A, and y © By + Be in area B. Thus is the difference between areas C and A, while (3 is the difference between areas C and B. Note the following: (i) An indicator variable 273 for C is not used. The “left out” category is called the base level. In the above example, area C is the base level,52 Linear modeling meaning that differences are measured between C and cach of the other levels. The choice of base level is up to the analyst: this is discussed in more detail below. (ii) In general, when the explanatory variable has r levels, r — 1 indi- cator variables are introduced, modeling the difference between each category and the base level. (iii) IL is not sensible to define, for example, a variable « = 1,2 or 3 accord- ing to whether the level is A, B or C, since y © Bg + Ait implies equal spacing between A, B and C, i.e. a difference of Ay between arcas A and B; and 3; between areas B and C. Choice of base level. The base level should not be sparse. To explain this suppose one has a categorical explanatory variable with r levels, and level r has been chosen as the base level. In the extreme scenario, if there are no cases having level r then for each case one of «1,...,2,-1 is always equal to 1, and the others equal to zero. This in turn implies that ary +--+ + 2p—1 = 1 and hence the X matrix is singular: the sum of the last r — 1 columns equals the intercept. More realistically, if the base level has very few cases, then for most cases sry +++++ar1 = 1, implying near linear dependency between the columns of X. Although @ is computable, it would be numerically unsta- ble, analogous to the result obtained when dividing by a number close to zero. Any level which is not sparse is an appropriate base level. Since 8; is the difference in the effect of the explanatory variable at level j compared with the base level, it is convenient to choose the base level as the “normal” or “usual” level, against which other levels are to be compared. This is often the level having the most cases. For example, in the vehicle insurance data set, the most commonly occurring vehicle body type is “Sedan,” which comprises almost a third of the cases. Comparing other body types with Sedan makes good sense, and makes the latter a good choice as the base level. Note, however, that one is not limited to making comparisons relative to the base level. Differences between non-base levels are of the form 8; — Bx. SAS notes. The software chooses the base level as the highest level ~ numer- ically or alphabetically. When levels are coded numerically, the highest level is often the category “Other.” This is not a good choice of base level as it is usually sparse, and comparisons relative to “Other” are generally not helpful. In this case a more suitable base level needs to be specified. The terminology for base level in the SAS manual is “reference level.”4.11 Polynomial regres ion 53 4.11 Polynomial regression Given a single explanatory variable x consider y © Bo + Aix + Box This is a linear regression since the right hand side is linear in the 3 coeffi- cients. The relationship between y and x, however, is quadratic. In essence there is one explanatory variable, albeit used twice in different forms. A unit increase in a: has the effect of changing y by about ( ++ 22. and hence the slope of the relationship depends on the value of «x. ‘This idea can be extended by defining further variables «°, 2, and so on, Incorporating more polynomial terms permits an increasingly complicated response structure. When a polynomial term of degree m, ic. x”, is included in a model, all lower order terms x, x”, ... ,2”"—? are generally included. Note that fitting polynomials of an order which is unnecessarily high, results in models with fitted values close to the observations but low predictive ability. This is illustrated in Figure 4.3, in which a small data set (n = 20), simu- lated with a quadratic relationship between x and y, has had polynomials of degree m = 1,2, 10 and 19 fitted. Clearly the linear relationship (m = 1) is m= si is 2 mat we ob & 2 Fig. 4.3. Polynomial fits, simulated data set inappropriate; m = 2 approximates the general shape of the relationship, with- out following local variation too closely; mm = 10 produces fitted values much closer to the observations, but which tend to follow small local variations; and ‘m = 19 (which fits 20 parameters to 20 observations) produces a perfect but useless fit, which “joins the dots.” The As from the degree 10 fit will have lower precision, and the resulting fitted values will be less reliable, than those from the quadratic model, Judgement about what degree of polynomial to fit to54 Linear modeling a continuous explanatory variable is guided by significance testing of the coef ficients (Section 4.15), and model sclection criteria (Section 4.19). Typically ‘one starts with a linear term, and adds in increasingly higher order terms. This is illustrated, in the context of logistic regression, in Section 7.3. Numerical difficulties. These can arise when high order terms are included in a model, due to the finite length of computer storage. If « is large (negative or positive), then overflow occurs in the computation of x” when m is large Frequently a further result of this is that underflow occurs in the corresponding A. This problem is avoided if « is suitably scaled before entering it into the regression. Specifically, denoting max as the maximum value of x in the data, and min as the minimum, then (max + @min)/2 is the midrange of x, and (max nin) /2 is half the range of x. The linear transformation Bmin)/2 lies between —1 and 1, Overflow and underflow problems are overcome by including 2* rather than « in the polynomial regression. Further problems of collinearity (see Section 4.14) can arise when comput ing high order polynomial regressions. Successive powers of «x may be highly correlated, leading again to numerical instability and consequent unreliability of the estimates. This problem is avoided by the use of orthogonal polynomi- als. Interested readers are referred to Chambers and Hastie (1991). Orthogonal polynomials are used in a regression model for mortality in Section 6.2. 4.12. Banding continuous explanatory variables ‘When the relationship between explanatory variable and response is not linear an alternative, safer, approach to finding a parametric form of the relationship is to band or categorize the explanatory variable. Although banding results in loss of information, a model employing a banded version of a continuous variable is sometimes considered more practical than one which employs the (possibly misspecified) continuous variable. However, caution is required in the choice of cutoffs and there is no general rule as to the optimal choice of cutoffs. © Excessively wide bands hide important features, Each band must have adequate numbers to ensure stable estimates. © Ideally bands correspond to realistic groups. For example, external factors such as industry standards can guide appropriate levels of banding. ‘An example of banding a continuous explanatory variable in the context of logistic regression is given in Section 7.3.4.13 Interaction 355 4.13 Interaction Consider the model with two explanatory variables x, 22 and with Y Bo + Ares + Pasta + Byrne This is a linear regression with x3 = 229, called an interaction term. A unit increase in 21 increases y by about (iy + yaa, and hence the effect of 1 depends on 272. Similarly the effect of a unit increase in «2 is about (22 + 8321 which depends on 1. Thus speaking of the effect of a unit increase in, for example, 27, holding all other variables constant, does not make sense. The variables xy and 2° are called the main effects and are distinguished from the interaction effect 3 = ear. Interacting an ordinary quantitative variable with an indicator variable is a useful technique of linear modeling. Assume 2 is continuous, and 2 is an indicator taking on the values 0 and 1 (indicating one of two groups). Then v={ Bot Arey if ry (Bo + 82) + (i + Bs)a1 ife Thus the effect of zz, that is the intercept and slope ~ depends on the group indicated by «2. For the base group 2 = 0, the intercept and slope of the relation of y to zr are (iy and (41, respectively. The intercept and slope when ©2 = 1 are fp + fz and f, + 3. Hence these are an increase of 2 and Bs relative to the intercept and slope when «2 = 0. This modeling idea is easily extended to more than two groups and a number of quantitative variables, so that the intercept and slopes with respect to quantitative variables may differ across many groups. As an illustration of interaction, the age of a driver may have an effect on the probability of a claim, but this effect may differ according to whether the car is a sports car or otherwise. More generally the effect of the age of a driver may vary with the type of car. 4.14 Collinearity Consider the effect of smoking on mortality. ‘The following explanatory variables are available: 2; the age at which smoking commenced, x2 the number of years since smoking initiation and 23 the current age. Obviously tg = 0) +2) and Bo + Bre + Bote + Byars = Bo + (1 + is) + (Ba + Bs)ra - Hence the three variables 21, :r and ars explain the same as any two of the variables. Thus the individual effects of each of the three variables cannot be assessed. The variables are perfectly “collinear.” In terms of the multiple least squares model (4.6), the matrix X has an exact linear dependence between its columns and hence X’X is singular.56 Li near modeling Using perfectly collinear variables. Many programs, when encountering perfect linear dependence in variables, exclude one or more variables to ensure the perfect collinearity disappears. Alternatively the software may just “crash” or even worse, report nonsensical results. Near collinearity. The smoking example above is a case of exact collinearity. Itis more usual (0 encounter situations of “near” and less obvious collinearity. For example, in a data set where the ‘ases are policyholders, area of resi- dence, income level and education level will be correlated. The result of the inclusion of highly correlated explanatory variables is the near singularity of X'X and difficulties in the computing the inverse (X'X)~!. If the inverse is. computable, then it will typically have large diagonal entries, implying large estimation variances on one or more of the components of f. The existence of collinearity may also mislead the analyst in the choice of explanatory vari- ables. If 2; and, are highly correlated and both predictive of the response, then each one, in the presence of the other, will not contribute much additional explanation. Hypothesis tests for each variable, assuming that the other one is in the model, will show non-significance, The solution is to include only one of them in the model. 4.15. Hypothesis testing Isa given coefficient J; equal to zero? This is equivalent to asking whether 7; is relevant to the prediction of the response. More generally it may be asked if a group of coefficients are all zero i.e. if, for a given “hypothesis” matrix Cit is the case that C8 = 0. For example, if C picks out the trailing coefficients in: 0 o10 0 Oo--. oOo1 0 Cc , O- 000-1 then the hypothesis {8 = 0 is the hypothesis that the final coefficients in are all zero. More generally it may be asked whether a given coefficient is equal to some number different to zero, or whether for example the sum of coefficients equal a given number. All these hypotheses can be written as C3 = r, where C and r are given. Throughout the following discussion it is assumed that C has full row rank and hence there is no redundancy in the hypotheses contained inC. Assume the hypothesis is that a subset of the ( coefficients are equal to zero, and that the number of rows of C (i.e. the number of parameters being, restricted to zero) is q. Further, let the number of parameters in the unrestricted4.15. Hypothesis testing 37 model be p. Then the number of unknown parameters in the restricted model is p — q. To facilitate the discussion, write f as the unrestricted MLE of 3, and let denote the MLE of when is maximized subject to the restrictions CA = 0. For the restricted model the X’ matrix has the columns corresponding to the restricted parameters deleted, and is denoted as X* ‘The residual sums of squares for the unrestricted and restricted models are (y — XB)'(y — XA) and (y — X*A)'(y — XA), respectively. The difference (u— X*B)'(y — X°B) — (y— XB (y - XB) = A'X'y - BX"'y (4.18) is a measure of the decrease in model fit due to the imposition of the restric- tions. It is positive (as the residual sum of squares for the unrestricted model ‘must be less than that for the restricted model) and if itis large then one would conclude that the deterioration in model fit as a result of the restrictions is too great. In that case the unrestricted model is chosen, i.e. the conclusion is made that CB 4 0. The commonly used F test uses as test statistic the difference (4.18), scaled by the estimate of the variance of the unrestricted model: GX'y—BX"W)/a any where 6? = (y— X8)'(y—X8)/(n—p). This test statistic, under the assump- tions summarized in (4.12), has the Fy,n—p distribution, if the restricted model is correct. One rejects the restricted model if it is “too large,” i.e. in the upper tail of the Fy,n—p distribution. Testing a single coefficient. When the hypothesis is that a single coefficient is zero, ie. 3; = 0, then (4.19) reduces to B ob; where j is diagonal element j of (X’X)~1. From (4.14), the denominater of (4.20) is the estimated variance of §;. The signed square root of (4.20) is the well-known “t statistic”, which is J; divided by its estimated standard error: ~Fin-ps (4.20) Bee tics. 421) o Testing the effect of a categorical variable. A test for the significance of a categorical explanatory variable with r levels involves the hypothesis that the coefficients of all r — 1 indicator variables are zero. For example, consider a model with a continuous explanatory variable «ry and a categorical explana- tory variable with three levels, and corresponding indicator variables ar and58 Linear modeling 2. In this case the hypothesis concerning the significance of the categorical variable is Bo oo10 Br oe G 00 t) Ba Bs ie. Bz = fy = 0. Sometimes the F test for the significance of a categorical explanatory variable is rejected, but tests for the coefficients of the individual indicator variables are not all rejected. This may signal the need for judicious grouping, of levels, particularly levels which are sparse. Care needs to be taken in this exercise, as wholesale grouping of, for example, all non-significant levels, may not make any sense. It is important not to combine category levels which are dissimilar in nature and do not form a sensible group. Statistical versus practical significance. Classical hypothesis testing was developed in an era when data sets were small, typically of the order of tens or hundreds of observations. Insurance data sets are usually orders of magnitude larger than this. The effect of increasing sample size is to increase the preci- sion of parameter estimates. Consider, for simplicity, a test of the hypothesis A; = 0. Anestimate J; which is close to zero, based on a small data set, may lead to non-rejection of the hypothesis since its standard error is large, result- ing in a small value for the t statistic (4,21). However, the identical A; based on a very large data set will have a small standard error. This results in a large t statistic and rejection of the hypothesis. However, one has to think about the magnitude of the effect, and whether the statistical significance translates into a practically useful effect or relativity. When the effect of an explana- tory variable is statistically significant but small relative to the effects of other explanatory variables, it is acceptable to exclude it from a final model. In cases such as these, model selection criteria (Section 4.19) lead to the exclusion of cally significant explanatory variables that have little explanatory power. 4.16 Checks using the residuals A fitted model must be checked. These checks are aimed at confirming whether the assumptions underlying the fitted model are applicable. The distributional assumptions of model (4.12) are: (i) The mean of y varies linearly with the 2’s. (ii) The error terms € are normally distributed,4.16 Checks using the residuals 59 (iii) The variance of the y's — equivalently the variance of the e's does not vary with the 2's, i.c. is constant. (Equivalent terms are homoskedas- ticity, constant variance, homogeneity of variance.) (iv) The e's are independent. ‘These assumptions must be checked. Note there is no assumption of normality of the explanatory variables z:;. ‘The assumptions on ¢ are not directly verifiable, since ¢ is not observable. However the ¢; can be estimated: é Yi — (Bo + Brin +--+ + Byarip) = 4s — These are called the residuals. nee y = XB +e, B= (XIX) X'y = (XX) X(XB +6) = B+ (X'X) Xe, it follows that XB+e-XB = {I-X(X'X)1X'fe The matrix H = X(X'X)~1X is called the “hat” matrix because it puts a hat Hy. if M = 1-H then M and hence = Me. Itis easy to verify MM = M Var(é) = Var(Me) = 07M = 0%(I— H). Thus Var(é;) = 02(1 — hy) where hi is diagonal element i of H. Much of model checking is based on the residuals, é, or alternatively the studentized residuals, defined as & “WE: (4.22) Checking for normality. Normality is checked by examining the distribution of either the residuals é; or the studentized residuals r;. The best way is via the pp-plot discussed in Section 1.6, where the percentile ranks of the é or ry are compared to those of the normal, Checking for constant variance. If the variance of e, or y, does not vary with the explanatory variables, then a plot of é or ri against each a, is a random scatter. The residuals are also commonly plotted against the fitted values 9, again yielding a random scatter if the constant variance assumption is true. Non-random scatters thus suggest heteroskedasticity.60 Linear modeling Checking for independence. If Var(e) = 71, then Var(é) = 07M. Off- diagonal elements of M are ofien relatively small and hence the é& are generally approximately independent. While there exist formal tests for cor- relation, typically one checks for serial correlation (i.e. correlation between suecessive observations) by plotting the é;’s in an order wherein one suspects lack of independence, for example the order of occurrence in the sample. If there is no serial correlation then one would expect to see no pattern in this plot. 4.17 Checking explanatory variable specifications Model diagnostics also provide guidance for identifying the correct specifica- tion of explanatory variables in the model. Two plots are useful and extend to the GLM: added-variable plots and partial residual plots. Added-variable plots. Added-variable plots are useful for detecting whether a continuous explanatory variable ; should be in the model, and for identi- fying the possible cause of unduly large residuals. In this plot the effect of 3 on the regression equation is isolated, given that all the other explanatory variables are in the model. An added-variable plot for «j, is a plot of y against ; when both variables are adjusted for all the other explanatory variables. Adjusting for all other variables means performing a regression of y and «; on all the other variables, and computing the residuals. This is done for y and 2; as the responses in two regressions. The residuals from the two regressions are plotted against cach other. The least squares slope associated with this plot is the regression coefficient on 2; when regressing y on all the «variables including i. A random scatter, i.e. a point cloud with zero slope, indicates that 7; should not be in the model. An linear plot means that x; should be included in the model, i.e. the rela- tionship between y and ; is linear, given the other explanatory variables. © Curvature in the plot indicates that a transformation is needed for 2. Partial residual plots. A nonlinear relationship between y and 2; is more easily detected with a partial residual plot. The partial residuals for «, are defined as (ui — G8) + 258; 5 where jj and (; are based on the model including all explanatory variables (including 3). Thus it is the residual with the explanation afforded by 23 added back in,4.18 Outliers 61 The partial residuals are plotted against 3: the plot will have slope fj, and a curvature indicates a departure from linearity in the relationship between y and 2p/n indicates the case is unusual in the space of explanatory variables, or equivalently is contributing disproportionately to its own fitted value. Leverages provide a useful mechanism for diagnosing subtle data errors undetectable using simple checks on sirigle explanatory variables. 4.19 Model selection Every explanatory variable added to a model improves fit. However adding unwarranted variables decreases the precision of parameter estimates. While an explanatory variable may be statistically significant, adding it to the model may not be worthwhile, because the improvement in fit may be outweighed by loss of estimation precision. The central concept is the bias-variance tradeoff. Bias refers to the lack of fit. Variance in this context means the variance of the regression estimates. Increasing the number of explanatory variables increases the number of parameters p and decreases the errors 1; — §j of the fit. A large p leads to low bias. However, the more parameters in the model, the greater the variances of the 3;. The aim is to find a model compromising between: © a large number of parameters and consequently a close fit to the observa- tions. In this case parameter estimates 3; have low precision (high standard deviation); and ¢ a small number of parameters with a less good fit. In this case parameter estimates 3; have relatively high precision (low standard deviation). There are a number of criteria which balance the goodness of fit of a model with a penalty term for the number of parameters. The best known are Akaike’s Information Criterion (AIC) and the Bayesian Information Criterion (BIC): AIC = —2 + 2p, BIC = —-2£+plnn. (4.23) Here é is the log-likelihood of the model given in (4.15). A good fit means a high value for the likelihood, and therefore a low value for —2¢.4.19 Model selection 63 The first two terms in the expres yielding ‘ion for € in (4.15) are usually omitted, s AIC +2p, BIC= 4 +pinn, (4.24) a where $ is the sum of squared residuals. The S/a? term is the bias term quan- tifying the fit of the model to the observations. A good fit yields a low value for S. The second term in cach of the criteria is the variance term, penalizing for the number of parameters in the model. A model with a large number of parameters and a good fit has a low value for the bias term, but the criterion is pushed up by the variance term. Conversely a model with small p has a higher bias term but lower variance term, Other things equal, the model with the lowest value of AIC or BIC is selected, A decision has to be made whether to use AIC of BIC. The latter applies a greater penalty for the number of parameters, so tends to choose models with fewer explanatory variables compared to AIC. When nis large, as is the case in most insurance data sets, the BIC tends to selects model which most analysts consider too simple. In this case the AIC is preferable. Models compared using AIC or BIC must be based on the same set of obser- vations. ‘This impacts when deciding between models which include at least one explanatory variable with missing values. Standard software omits from analysis any cases which have missing values on any of the variables in a model. This means that if one or other model includes a variable with miss- ing values, then the models being compared will be based on different sets of observations, rendering the comparison invalid. The solution is to perform model selection on the subset of cases which have no missing values for any of the potential explanatory variables Stepwise regression. Stepwise regression is an automated form of model selection. Stepwise algorithms are available in most statistical software. The methodology has fallen out of favor in recent years, as it is ad hoc, and yields different results when implemented in different ways. It is often used to avoid careful thought about the model. It encourages the notion that model selection can be performed automatically. ‘Model selection is as much art as science. Understanding of the subject mat- ter, and the interplay between all of the variables, is critical. Judgements about the inclusion or exclusion of a variable is validly made using information other than statistical criteria, For example, the expense of measuring or recording a variable, or the political sensitivity of a variable, may override purely statistical considerations,